From 029b84e2b7c4856473e2ace796f915d00695ca91 Mon Sep 17 00:00:00 2001 From: Samuel Audet Date: Sun, 26 Jul 2020 21:59:27 +0900 Subject: [PATCH] Development updates (#9053) * RL4J: Add generic update rule (#502) Signed-off-by: Alexandre Boulanger * Shyrma reduce (#481) * - start working on improving of cpu legacy code for reduce ops Signed-off-by: Yurii * - further work on improving legacy loops Signed-off-by: Yurii * - still working on improving reduce ops Signed-off-by: Yurii * - further work on improving reduce ops Signed-off-by: Yurii * - testing speed run of new reduce op Signed-off-by: Yurii * - working on improvement of default loop for reduce op Signed-off-by: Yurii * - update signatures of stuff which calls reduce ops Signed-off-by: Yurii * - make corrections in cuda reduce kernels Signed-off-by: Yurii * - change loop for default case in broadcast legacy ops Signed-off-by: Yurii * - comment some shape stuff Signed-off-by: Yurii * - comment unnecessary prints in RNGtests Signed-off-by: Yurii * - finish to resolve conflicts after master has been merged Signed-off-by: Yurii * - get rid of some compilation mistakes of cuda stuff Signed-off-by: Yurii * - minor changes Signed-off-by: Yurii * - further search for bug causing crash on java test Signed-off-by: Yurii * - add scalar case in reduce_ ... exec stuff Signed-off-by: Yurii * - minor corrections in NAtiveOps.cu Signed-off-by: Yurii * - add switch to scalar case execReduceXD functions Signed-off-by: Yurii * - add support for vectors old shape in ConstantShapeHelper::createShapeInfoWithNoUnitiesForReduce Signed-off-by: Yurii * - correct cuda mirrorPad Signed-off-by: Yurii * - add support for vectors old shape in cuda createShapeInfoWithNoUnitiesForReduce Signed-off-by: Yurii Co-authored-by: raver119 * Add support for CUDA 11.0 (#492) * Add support for CUDA 11.0 * libnd4j tweaks for CUDA 11 Signed-off-by: raver119@gmail.com * bindings update, again? Signed-off-by: raver119@gmail.com * * Update versions of JavaCPP Presets for FFmpeg, OpenBLAS, and NumPy * update API to match CUDA 8 Signed-off-by: raver119@gmail.com * * Update version of JavaCPP Presets for CPython * C++ updated for cuDNN 8.0 Signed-off-by: raver119@gmail.com * one more test Signed-off-by: raver119@gmail.com * one more test Signed-off-by: raver119@gmail.com * one more test Signed-off-by: raver119@gmail.com * 128-bit alignment for workspaces Signed-off-by: raver119@gmail.com * change seed in 1 test Signed-off-by: raver119@gmail.com * Fix dependecy duplication in python4j-parent pom * Fix group id for in python4j-numpy * few tests tweaked Signed-off-by: raver119@gmail.com * Remove macosx-x86_64-gpu from nd4j-tests-tensorflow * few minor tweaks for IndexReduce Signed-off-by: raver119@gmail.com * one test removed Signed-off-by: raver119@gmail.com Co-authored-by: raver119@gmail.com Co-authored-by: Serhii Shepel <9946053+sshepel@users.noreply.github.com> * RL4J: Add SyncTrainer and AgentLearnerBuilder for a few algorithms (#504) Signed-off-by: Alexandre Boulanger Co-authored-by: Alexandre Boulanger <44292157+aboulang2002@users.noreply.github.com> Co-authored-by: Yurii Shyrma Co-authored-by: raver119 Co-authored-by: Serhii Shepel <9946053+sshepel@users.noreply.github.com> --- arbiter/arbiter-core/pom.xml | 2 +- arbiter/arbiter-deeplearning4j/pom.xml | 2 +- arbiter/arbiter-server/pom.xml | 2 +- arbiter/arbiter-ui/pom.xml | 2 +- arbiter/pom.xml | 6 +- change-cuda-versions.sh | 6 +- datavec/datavec-api/pom.xml | 2 +- datavec/datavec-arrow/pom.xml | 2 +- .../datavec-data/datavec-data-audio/pom.xml | 2 +- .../datavec-data/datavec-data-codec/pom.xml | 2 +- .../datavec-data/datavec-data-image/pom.xml | 2 +- datavec/datavec-data/datavec-data-nlp/pom.xml | 2 +- datavec/datavec-data/datavec-geo/pom.xml | 2 +- datavec/datavec-data/datavec-hadoop/pom.xml | 2 +- datavec/datavec-data/pom.xml | 2 +- datavec/datavec-excel/pom.xml | 2 +- datavec/datavec-jdbc/pom.xml | 2 +- datavec/datavec-local/pom.xml | 2 +- datavec/datavec-python/pom.xml | 2 +- .../datavec-spark-inference-client/pom.xml | 2 +- .../datavec-spark-inference-model/pom.xml | 2 +- .../datavec-spark-inference-server/pom.xml | 2 +- .../datavec-spark-inference-parent/pom.xml | 2 +- datavec/datavec-spark/pom.xml | 2 +- datavec/pom.xml | 6 +- .../deeplearning4j-common-tests/pom.xml | 4 +- deeplearning4j/deeplearning4j-common/pom.xml | 2 +- deeplearning4j/deeplearning4j-core/pom.xml | 4 +- deeplearning4j/deeplearning4j-cuda/pom.xml | 10 +- .../convolution/CudnnConvolutionHelper.java | 29 +- .../cuda/lstm/ValidateCudnnLSTM.java | 4 +- .../deeplearning4j-datasets/pom.xml | 2 +- .../deeplearning4j-datavec-iterators/pom.xml | 2 +- .../deeplearning4j-utility-iterators/pom.xml | 2 +- deeplearning4j/deeplearning4j-data/pom.xml | 2 +- .../deeplearning4j-dataimport-solrj/pom.xml | 4 +- deeplearning4j/deeplearning4j-graph/pom.xml | 2 +- .../deeplearning4j-tsne/pom.xml | 2 +- .../deeplearning4j-manifold/pom.xml | 2 +- .../deeplearning4j-modelexport-solr/pom.xml | 4 +- .../deeplearning4j-modelimport/pom.xml | 4 +- .../pom.xml | 4 +- .../pom.xml | 2 +- .../pom.xml | 2 +- .../nearestneighbor-core/pom.xml | 4 +- .../pom.xml | 2 +- .../deeplearning4j-nlp-chinese/pom.xml | 2 +- .../deeplearning4j-nlp-japanese/pom.xml | 2 +- .../deeplearning4j-nlp-korean/pom.xml | 2 +- .../deeplearning4j-nlp-uima/pom.xml | 2 +- .../deeplearning4j-nlp/pom.xml | 2 +- .../deeplearning4j-nlp-parent/pom.xml | 2 +- deeplearning4j/deeplearning4j-nn/pom.xml | 2 +- .../deeplearning4j-json-server/pom.xml | 4 +- deeplearning4j/deeplearning4j-remote/pom.xml | 2 +- .../pom.xml | 4 +- .../pom.xml | 2 +- .../deeplearning4j-scaleout/pom.xml | 2 +- .../spark/dl4j-spark-nlp-java8/pom.xml | 2 +- .../spark/dl4j-spark-nlp/pom.xml | 2 +- .../spark/dl4j-spark-parameterserver/pom.xml | 2 +- .../spark/dl4j-spark/pom.xml | 2 +- .../deeplearning4j-scaleout/spark/pom.xml | 2 +- .../deeplearning4j-ui-components/pom.xml | 2 +- .../deeplearning4j-ui-model/pom.xml | 2 +- .../deeplearning4j-ui-standalone/pom.xml | 2 +- .../deeplearning4j-ui/pom.xml | 2 +- .../deeplearning4j-vertx/pom.xml | 2 +- .../deeplearning4j-ui-parent/pom.xml | 2 +- deeplearning4j/deeplearning4j-zoo/pom.xml | 2 +- deeplearning4j/dl4j-integration-tests/pom.xml | 2 +- deeplearning4j/pom.xml | 6 +- libnd4j/include/array/NDArray.h | 24 +- libnd4j/include/array/NDArray.hXX | 115 +- libnd4j/include/helpers/ConstantShapeHelper.h | 12 +- libnd4j/include/helpers/Loops.h | 923 ++++++++----- libnd4j/include/helpers/ShapeBuilders.h | 6 +- libnd4j/include/helpers/ShapeUtils.h | 6 + .../helpers/benchmark/ReductionBenchmark.h | 4 +- .../helpers/cpu/ConstantShapeHelper.cpp | 39 +- libnd4j/include/helpers/cpu/MmulHelper.cpp | 6 +- .../helpers/cpu/loops/ReductionLoops_bool.cpp | 11 +- .../cpu/loops/ReductionLoops_float.hpp | 13 +- .../helpers/cpu/loops/ReductionLoops_long.cpp | 13 +- .../helpers/cpu/loops/ReductionLoops_same.cpp | 14 +- .../helpers/cuda/ConstantShapeHelper.cu | 35 + .../include/helpers/cuda_off/MmulHelper.cu | 6 +- .../include/helpers/impl/ShapeBuilders.cpp | 22 +- libnd4j/include/helpers/impl/ShapeUtils.cpp | 11 + libnd4j/include/helpers/shape.h | 364 +++-- libnd4j/include/legacy/NativeOpExecutioner.h | 12 +- .../legacy/cpu/NativeOpExecutioner.cpp | 45 +- libnd4j/include/legacy/cpu/NativeOps.cpp | 128 +- .../legacy/cuda/NativeOpExecutioner.cu | 142 +- libnd4j/include/legacy/cuda/NativeOps.cu | 79 +- libnd4j/include/loops/cpu/broadcasting.hpp | 16 +- .../include/loops/cpu/broadcasting_bool.hpp | 16 +- .../include/loops/cpu/broadcasting_int.hpp | 16 +- .../include/loops/cpu/reduce/reduce_bool.cpp | 113 +- .../include/loops/cpu/reduce/reduce_float.hpp | 133 +- .../include/loops/cpu/reduce/reduce_long.cpp | 124 +- .../include/loops/cpu/reduce/reduce_same.cpp | 133 +- libnd4j/include/loops/cuda/broadcasting.chpp | 15 +- .../include/loops/cuda/broadcasting_bool.cu | 15 +- .../include/loops/cuda/broadcasting_int.cu | 15 +- libnd4j/include/loops/cuda/indexreduce.cu | 18 +- .../include/loops/cuda/reduce/reduce_bool.cu | 100 +- .../loops/cuda/reduce/reduce_float.chpp | 108 +- .../include/loops/cuda/reduce/reduce_long.cu | 110 +- .../include/loops/cuda/reduce/reduce_same.cu | 101 +- libnd4j/include/loops/cuda/reduce3.chpp | 19 +- .../include/loops/cuda/summarystatsreduce.cu | 19 +- libnd4j/include/loops/indexreduce.h | 4 +- libnd4j/include/loops/reduce_bool.h | 31 +- libnd4j/include/loops/reduce_float.h | 34 +- libnd4j/include/loops/reduce_long.h | 31 +- libnd4j/include/loops/reduce_same.h | 31 +- libnd4j/include/loops/summarystatsreduce.h | 2 +- .../generic/loss/absoluteDifference.cpp | 6 +- .../generic/loss/cosineDistance.cpp | 6 +- .../ops/declarable/generic/loss/hingeLoss.cpp | 6 +- .../ops/declarable/generic/loss/huberLoss.cpp | 6 +- .../ops/declarable/generic/loss/logLoss.cpp | 6 +- .../generic/loss/log_poisson_loss.cpp | 6 +- .../generic/loss/meanPairWsSqErr.cpp | 6 +- .../ops/declarable/generic/loss/meanSqErr.cpp | 6 +- .../generic/loss/sigmCrossEntropy.cpp | 6 +- .../generic/loss/softmaxCrossEntropy.cpp | 6 +- .../declarable/generic/nn/recurrent/sru.cpp | 4 +- .../declarable/helpers/cuda/activations.cu | 22 +- .../ops/declarable/helpers/cuda/hamming.cu | 10 +- .../ops/declarable/helpers/cuda/pad.cu | 7 +- .../ops/declarable/helpers/cuda/top_k.cu | 9 +- .../ops/declarable/helpers/cuda/transforms.cu | 29 +- .../declarable/impl/LegacyReduceBoolOp.cpp | 49 +- .../declarable/impl/LegacyReduceFloatOp.cpp | 44 +- .../declarable/impl/LegacyReduceLongOp.cpp | 49 +- .../declarable/impl/LegacyReduceSameOp.cpp | 50 +- .../ops/declarable/platform/cudnn/conv2d.cu | 23 +- .../ops/declarable/platform/cudnn/conv3d.cu | 24 +- .../platform/cudnn/depthwiseConv2d.cu | 23 +- libnd4j/include/system/pointercast.h | 2 + libnd4j/pom.xml | 4 +- .../layers_tests/ConvolutionTests1.cpp | 61 +- .../layers_tests/CudaBasicsTests1.cu | 1230 +++++++---------- .../layers_tests/DeclarableOpsTests3.cpp | 12 +- .../tests_cpu/layers_tests/LegacyOpsTests.cpp | 32 +- .../tests_cpu/layers_tests/NDArrayTests.cpp | 20 +- .../layers_tests/PlaygroundTests.cpp | 68 +- libnd4j/tests_cpu/layers_tests/RNGTests.cpp | 220 +-- .../api/memory/abstracts/Nd4jWorkspace.java | 7 +- .../nd4j-cuda-platform/pom.xml | 8 +- .../nd4j-backend-impls/nd4j-cuda/pom.xml | 8 +- .../nd4j/jita/workspace/CudaWorkspace.java | 15 +- .../java/org/nd4j/nativeblas/Nd4jCuda.java | 2 +- .../org/nd4j/nativeblas/Nd4jCudaPresets.java | 22 +- .../java/org/nd4j/nativeblas/Nd4jCpu.java | 2 +- .../nd4j-tests-tensorflow/pom.xml | 14 +- nd4j/nd4j-backends/nd4j-tests/pom.xml | 2 +- .../linalg/workspace/BasicWorkspaceTests.java | 44 +- .../workspace/WorkspaceProviderTests.java | 39 +- nd4j/nd4j-jdbc/nd4j-jdbc-hsql/pom.xml | 2 +- .../nd4j-parameter-server-client/pom.xml | 2 +- .../nd4j-parameter-server-node/pom.xml | 2 +- .../nd4j-parameter-server/pom.xml | 2 +- nd4j/nd4j-remote/nd4j-grpc-client/pom.xml | 2 +- nd4j/nd4j-remote/nd4j-json-client/pom.xml | 2 +- nd4j/nd4j-remote/nd4j-json-server/pom.xml | 2 +- nd4j/nd4j-serde/nd4j-aeron/pom.xml | 2 +- nd4j/nd4j-serde/nd4j-arrow/pom.xml | 2 +- nd4j/nd4j-serde/nd4j-kryo/pom.xml | 2 +- nd4j/nd4j-uberjar/pom.xml | 6 +- nd4s/pom.xml | 4 +- pom.xml | 18 +- python4j/pom.xml | 8 +- python4j/python4j-numpy/pom.xml | 4 +- rl4j/pom.xml | 6 +- rl4j/rl4j-ale/pom.xml | 2 +- rl4j/rl4j-api/pom.xml | 2 +- rl4j/rl4j-core/pom.xml | 2 +- .../org/deeplearning4j/rl4j/agent/Agent.java | 63 +- .../rl4j/agent/AgentLearner.java | 43 +- .../learning/algorithm/IUpdateAlgorithm.java} | 21 +- .../learning/algorithm/NStepQLearning.java | 104 ++ .../algorithm/dqn}/BaseDQNAlgorithm.java | 20 +- .../dqn/BaseTransitionTDAlgorithm.java} | 58 +- .../learning/algorithm/dqn}/DoubleDQN.java | 18 +- .../learning/algorithm/dqn}/StandardDQN.java | 17 +- .../{ => behavior}/ILearningBehavior.java | 2 +- .../{ => behavior}/LearningBehavior.java | 11 +- .../agent/learning/update/FeaturesLabels.java | 64 + .../rl4j/agent/learning/update/Gradients.java | 58 + .../{ => learning}/update/IUpdateRule.java | 2 +- .../agent/learning/update/UpdateRule.java | 54 + .../updater/GradientsNeuralNetUpdater.java} | 37 +- .../update/updater}/INeuralNetUpdater.java | 15 +- .../updater/LabelsNeuralNetUpdater.java | 81 ++ .../rl4j/agent/listener/AgentListener.java | 8 + .../agent/listener/AgentListenerList.java | 12 + .../agent/update/DQNNeuralNetUpdateRule.java | 56 - .../rl4j/agent/update/Gradients.java | 26 - .../rl4j/builder/BaseAgentLearnerBuilder.java | 168 +++ .../builder/BaseDQNAgentLearnerBuilder.java | 93 ++ .../rl4j/builder/DoubleDQNBuilder.java | 56 + .../rl4j/builder/INetworksHandler.java | 43 + .../rl4j/builder/NStepQLearningBuilder.java | 96 ++ .../rl4j/builder/StandardDQNBuilder.java | 57 + .../rl4j/builder/SyncNetworkHandler.java | 50 + .../rl4j/environment/IActionSchema.java | 2 + .../rl4j/environment/IntegerActionSchema.java | 9 +- .../ReplayMemoryExperienceHandler.java | 40 +- .../StateActionExperienceHandler.java | 17 +- .../learning/async/AsyncThreadDiscrete.java | 5 +- .../qlearning/discrete/QLearningDiscrete.java | 42 +- .../rl4j/network/CommonGradientNames.java | 5 + .../rl4j/network/CommonLabelNames.java | 7 + .../rl4j/network/IOutputNeuralNet.java | 5 + .../rl4j/network/ITrainableNeuralNet.java | 20 +- .../rl4j/network/NeuralNet.java | 2 +- .../rl4j/network/ac/ActorCriticCompGraph.java | 20 +- .../rl4j/network/ac/ActorCriticSeparate.java | 20 +- .../deeplearning4j/rl4j/network/dqn/DQN.java | 44 +- .../deeplearning4j/rl4j/policy/DQNPolicy.java | 12 +- .../deeplearning4j/rl4j/policy/EpsGreedy.java | 36 +- .../rl4j/policy/INeuralNetPolicy.java | 4 +- .../deeplearning4j/rl4j/policy/Policy.java | 8 +- .../deeplearning4j/rl4j/trainer/ITrainer.java | 56 +- .../rl4j/trainer/SyncTrainer.java | 64 + .../rl4j/agent/AgentLearnerTest.java | 14 +- .../deeplearning4j/rl4j/agent/AgentTest.java | 83 +- .../algorithm/NStepQLearningTest.java | 129 ++ .../algorithm/dqn}/DoubleDQNTest.java | 27 +- .../algorithm/dqn}/StandardDQNTest.java | 27 +- .../{ => behavior}/LearningBehaviorTest.java | 6 +- .../learning/update/FeaturesLabelsTest.java | 38 + .../agent/learning/update/GradientsTest.java | 40 + .../agent/learning/update/UpdateRuleTest.java | 67 + .../GradientsNeuralNetUpdaterTest.java | 56 + .../updater/LabelsNeuralNetUpdaterTest.java | 77 ++ .../NeuralNetUpdaterTest.java | 51 - .../builder/BaseAgentLearnerBuilderTest.java | 92 ++ .../ReplayMemoryExperienceHandlerTest.java | 13 +- .../StateActionExperienceHandlerTest.java | 22 +- .../discrete/QLearningDiscreteTest.java | 5 +- .../rl4j/learning/sync/support/MockDQN.java | 16 +- .../rl4j/policy/PolicyTest.java | 19 +- .../deeplearning4j/rl4j/support/MockDQN.java | 18 +- .../rl4j/support/MockNeuralNet.java | 15 +- .../rl4j/trainer/SyncTrainerTest.java | 59 + rl4j/rl4j-doom/pom.xml | 2 +- rl4j/rl4j-gym/pom.xml | 2 +- rl4j/rl4j-malmo/pom.xml | 2 +- scalnet/pom.xml | 2 +- 253 files changed, 5029 insertions(+), 3340 deletions(-) rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/{learning/sync/qlearning/discrete/TDTargetAlgorithm/ITDTargetAlgorithm.java => agent/learning/algorithm/IUpdateAlgorithm.java} (56%) create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearning.java rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/{learning/sync/qlearning/discrete/TDTargetAlgorithm => agent/learning/algorithm/dqn}/BaseDQNAlgorithm.java (70%) rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/{learning/sync/qlearning/discrete/TDTargetAlgorithm/BaseTDTargetAlgorithm.java => agent/learning/algorithm/dqn/BaseTransitionTDAlgorithm.java} (66%) rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/{learning/sync/qlearning/discrete/TDTargetAlgorithm => agent/learning/algorithm/dqn}/DoubleDQN.java (77%) rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/{learning/sync/qlearning/discrete/TDTargetAlgorithm => agent/learning/algorithm/dqn}/StandardDQN.java (79%) rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/{ => behavior}/ILearningBehavior.java (94%) rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/{ => behavior}/LearningBehavior.java (90%) create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabels.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/Gradients.java rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/{ => learning}/update/IUpdateRule.java (93%) create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRule.java rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/{update/neuralnetupdater/NeuralNetUpdater.java => learning/update/updater/GradientsNeuralNetUpdater.java} (55%) rename rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/{update/neuralnetupdater => learning/update/updater}/INeuralNetUpdater.java (63%) create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdater.java delete mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/DQNNeuralNetUpdateRule.java delete mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/Gradients.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilder.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseDQNAgentLearnerBuilder.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/DoubleDQNBuilder.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/INetworksHandler.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/NStepQLearningBuilder.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/StandardDQNBuilder.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/SyncNetworkHandler.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonGradientNames.java create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonLabelNames.java rename libnd4j/include/loops/cuda/reduce3.cu => rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/ITrainer.java (71%) create mode 100644 rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/SyncTrainer.java create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearningTest.java rename rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/{learning/sync/qlearning/discrete/TDTargetAlgorithm => agent/learning/algorithm/dqn}/DoubleDQNTest.java (74%) rename rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/{learning/sync/qlearning/discrete/TDTargetAlgorithm => agent/learning/algorithm/dqn}/StandardDQNTest.java (73%) rename rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/{ => behavior}/LearningBehaviorTest.java (93%) create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabelsTest.java create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/GradientsTest.java create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRuleTest.java create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/GradientsNeuralNetUpdaterTest.java create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdaterTest.java delete mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/NeuralNetUpdaterTest.java create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilderTest.java create mode 100644 rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/trainer/SyncTrainerTest.java diff --git a/arbiter/arbiter-core/pom.xml b/arbiter/arbiter-core/pom.xml index 76251f4cd..7c65b2efb 100644 --- a/arbiter/arbiter-core/pom.xml +++ b/arbiter/arbiter-core/pom.xml @@ -99,7 +99,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/arbiter/arbiter-deeplearning4j/pom.xml b/arbiter/arbiter-deeplearning4j/pom.xml index ec7e22d3c..92e3fb7aa 100644 --- a/arbiter/arbiter-deeplearning4j/pom.xml +++ b/arbiter/arbiter-deeplearning4j/pom.xml @@ -77,7 +77,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/arbiter/arbiter-server/pom.xml b/arbiter/arbiter-server/pom.xml index c4306b967..26aa80b07 100644 --- a/arbiter/arbiter-server/pom.xml +++ b/arbiter/arbiter-server/pom.xml @@ -63,7 +63,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/arbiter/arbiter-ui/pom.xml b/arbiter/arbiter-ui/pom.xml index 7392392db..88f39a310 100644 --- a/arbiter/arbiter-ui/pom.xml +++ b/arbiter/arbiter-ui/pom.xml @@ -37,7 +37,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/arbiter/pom.xml b/arbiter/pom.xml index 93f877968..ab8d72365 100644 --- a/arbiter/pom.xml +++ b/arbiter/pom.xml @@ -151,7 +151,7 @@ ${skipTestResourceEnforcement} - test-nd4j-native,test-nd4j-cuda-10.2 + test-nd4j-native,test-nd4j-cuda-11.0 false @@ -333,11 +333,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${nd4j.version} test diff --git a/change-cuda-versions.sh b/change-cuda-versions.sh index 7b354d68b..cbe75d830 100755 --- a/change-cuda-versions.sh +++ b/change-cuda-versions.sh @@ -20,7 +20,7 @@ set -e -VALID_VERSIONS=( 9.2 10.0 10.1 10.2 ) +VALID_VERSIONS=( 9.2 10.0 10.1 10.2 11.0 ) usage() { echo "Usage: $(basename $0) [-h|--help] @@ -47,6 +47,10 @@ check_cuda_version() { check_cuda_version "$VERSION" case $VERSION in + 11.0) + VERSION2="8.0" + VERSION3="1.5.4-SNAPSHOT" + ;; 10.2) VERSION2="7.6" VERSION3="1.5.3" diff --git a/datavec/datavec-api/pom.xml b/datavec/datavec-api/pom.xml index 3c3eec86e..0b01863f6 100644 --- a/datavec/datavec-api/pom.xml +++ b/datavec/datavec-api/pom.xml @@ -126,7 +126,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-arrow/pom.xml b/datavec/datavec-arrow/pom.xml index 60409bc53..eb61221c8 100644 --- a/datavec/datavec-arrow/pom.xml +++ b/datavec/datavec-arrow/pom.xml @@ -62,7 +62,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-data/datavec-data-audio/pom.xml b/datavec/datavec-data/datavec-data-audio/pom.xml index 3b9674cd9..0c67ae396 100644 --- a/datavec/datavec-data/datavec-data-audio/pom.xml +++ b/datavec/datavec-data/datavec-data-audio/pom.xml @@ -79,7 +79,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-data/datavec-data-codec/pom.xml b/datavec/datavec-data/datavec-data-codec/pom.xml index 7ef25d5a3..58eda4820 100644 --- a/datavec/datavec-data/datavec-data-codec/pom.xml +++ b/datavec/datavec-data/datavec-data-codec/pom.xml @@ -66,7 +66,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-data/datavec-data-image/pom.xml b/datavec/datavec-data/datavec-data-image/pom.xml index aef66381a..c88c89208 100644 --- a/datavec/datavec-data/datavec-data-image/pom.xml +++ b/datavec/datavec-data/datavec-data-image/pom.xml @@ -128,7 +128,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-data/datavec-data-nlp/pom.xml b/datavec/datavec-data/datavec-data-nlp/pom.xml index fb30b93e7..dd5860f9a 100644 --- a/datavec/datavec-data/datavec-data-nlp/pom.xml +++ b/datavec/datavec-data/datavec-data-nlp/pom.xml @@ -81,7 +81,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-data/datavec-geo/pom.xml b/datavec/datavec-data/datavec-geo/pom.xml index f88bc84d1..792265434 100644 --- a/datavec/datavec-data/datavec-geo/pom.xml +++ b/datavec/datavec-data/datavec-geo/pom.xml @@ -56,7 +56,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-data/datavec-hadoop/pom.xml b/datavec/datavec-data/datavec-hadoop/pom.xml index fb7eee69c..4ab64d8b5 100644 --- a/datavec/datavec-data/datavec-hadoop/pom.xml +++ b/datavec/datavec-data/datavec-hadoop/pom.xml @@ -74,7 +74,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-data/pom.xml b/datavec/datavec-data/pom.xml index e40d96149..2510924b8 100644 --- a/datavec/datavec-data/pom.xml +++ b/datavec/datavec-data/pom.xml @@ -54,7 +54,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-excel/pom.xml b/datavec/datavec-excel/pom.xml index 49dc26db8..dc6d3d6b7 100644 --- a/datavec/datavec-excel/pom.xml +++ b/datavec/datavec-excel/pom.xml @@ -65,7 +65,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-jdbc/pom.xml b/datavec/datavec-jdbc/pom.xml index 6ef9b0441..612fb05e7 100644 --- a/datavec/datavec-jdbc/pom.xml +++ b/datavec/datavec-jdbc/pom.xml @@ -72,7 +72,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-local/pom.xml b/datavec/datavec-local/pom.xml index 3adc0e011..06fd13fe2 100644 --- a/datavec/datavec-local/pom.xml +++ b/datavec/datavec-local/pom.xml @@ -95,7 +95,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-python/pom.xml b/datavec/datavec-python/pom.xml index 526b8238a..dae915909 100644 --- a/datavec/datavec-python/pom.xml +++ b/datavec/datavec-python/pom.xml @@ -78,7 +78,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml index 3b564b1b3..ff8bdb853 100644 --- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml +++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml @@ -60,7 +60,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml index bac20d42e..68a78450d 100644 --- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml +++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml @@ -59,7 +59,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml index 0c05f327b..331c58a8c 100644 --- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml +++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml @@ -178,7 +178,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-spark-inference-parent/pom.xml b/datavec/datavec-spark-inference-parent/pom.xml index cc3a6b0c1..5e98a1aad 100644 --- a/datavec/datavec-spark-inference-parent/pom.xml +++ b/datavec/datavec-spark-inference-parent/pom.xml @@ -38,7 +38,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/datavec-spark/pom.xml b/datavec/datavec-spark/pom.xml index 345b774c3..2c547499c 100644 --- a/datavec/datavec-spark/pom.xml +++ b/datavec/datavec-spark/pom.xml @@ -144,7 +144,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/datavec/pom.xml b/datavec/pom.xml index 1c49960d6..8e403ea1e 100644 --- a/datavec/pom.xml +++ b/datavec/pom.xml @@ -108,7 +108,7 @@ ${skipTestResourceEnforcement} - test-nd4j-native,test-nd4j-cuda-10.2 + test-nd4j-native,test-nd4j-cuda-11.0 false @@ -361,11 +361,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${nd4j.version} test diff --git a/deeplearning4j/deeplearning4j-common-tests/pom.xml b/deeplearning4j/deeplearning4j-common-tests/pom.xml index 5a4ba921d..d19aa85c4 100644 --- a/deeplearning4j/deeplearning4j-common-tests/pom.xml +++ b/deeplearning4j/deeplearning4j-common-tests/pom.xml @@ -62,11 +62,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-common/pom.xml b/deeplearning4j/deeplearning4j-common/pom.xml index 1928eff91..7e2e0ebb6 100644 --- a/deeplearning4j/deeplearning4j-common/pom.xml +++ b/deeplearning4j/deeplearning4j-common/pom.xml @@ -40,7 +40,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-core/pom.xml b/deeplearning4j/deeplearning4j-core/pom.xml index 90c88d4c3..f8dc0123b 100644 --- a/deeplearning4j/deeplearning4j-core/pom.xml +++ b/deeplearning4j/deeplearning4j-core/pom.xml @@ -180,11 +180,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-cuda/pom.xml b/deeplearning4j/deeplearning4j-cuda/pom.xml index 30373db3a..e76e905df 100644 --- a/deeplearning4j/deeplearning4j-cuda/pom.xml +++ b/deeplearning4j/deeplearning4j-cuda/pom.xml @@ -16,7 +16,7 @@ 4.0.0 - deeplearning4j-cuda-10.2 + deeplearning4j-cuda-11.0 deeplearning4j-cuda org.deeplearning4j @@ -26,9 +26,9 @@ - 10.2 - 7.6 - 1.5.3 + 11.0 + 8.0 + 1.5.4-SNAPSHOT @@ -112,7 +112,7 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-cuda/src/main/java/org/deeplearning4j/cuda/convolution/CudnnConvolutionHelper.java b/deeplearning4j/deeplearning4j-cuda/src/main/java/org/deeplearning4j/cuda/convolution/CudnnConvolutionHelper.java index a4dddb759..91e3c4829 100644 --- a/deeplearning4j/deeplearning4j-cuda/src/main/java/org/deeplearning4j/cuda/convolution/CudnnConvolutionHelper.java +++ b/deeplearning4j/deeplearning4j-cuda/src/main/java/org/deeplearning4j/cuda/convolution/CudnnConvolutionHelper.java @@ -254,17 +254,34 @@ public class CudnnConvolutionHelper extends BaseCudnnHelper implements Convoluti throw new IllegalArgumentException("Unknown BwdDataAlgo: " + bwdDataAlgo); } } else { + /* code = cudnnGetConvolutionBackwardFilterAlgorithm(cudnnContext, cudnnContext.srcTensorDesc, cudnnContext.deltaTensorDesc, cudnnContext.convDesc, cudnnContext.filterDesc, mode == AlgoMode.NO_WORKSPACE ? CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE : CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST, 0, algo1); + */ + val fa = new cudnnConvolutionBwdFilterAlgoPerf_t(); + val counts = new int[1]; + code = cudnnFindConvolutionBackwardFilterAlgorithm(cudnnContext, cudnnContext.srcTensorDesc, + cudnnContext.deltaTensorDesc, cudnnContext.convDesc, cudnnContext.filterDesc, 1, counts, fa); + algo1[0] = fa.algo(); + checkCudnn(false, "cudnnGetConvolutionBackwardFilterAlgorithm", code, input, weights, null, delta, kernel, strides, pad, mode, null, bwdFilterAlgo, bwdDataAlgo, convolutionMode, dilation); + + /* code = cudnnGetConvolutionBackwardDataAlgorithm(cudnnContext, cudnnContext.filterDesc, cudnnContext.deltaTensorDesc, cudnnContext.convDesc, cudnnContext.srcTensorDesc, mode == AlgoMode.NO_WORKSPACE ? CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE : CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, algo2); + */ + + val da = new cudnnConvolutionBwdDataAlgoPerf_t(); + code = cudnnFindConvolutionBackwardDataAlgorithm(cudnnContext, cudnnContext.filterDesc, + cudnnContext.deltaTensorDesc, cudnnContext.convDesc, cudnnContext.srcTensorDesc, 1, counts, da); + + algo2[0] = da.algo(); checkCudnn(false, "cudnnGetConvolutionBackwardDataAlgorithm", code, input, weights, null, delta, kernel, strides, pad, mode, null, bwdFilterAlgo, bwdDataAlgo, convolutionMode, dilation); } @@ -461,11 +478,17 @@ public class CudnnConvolutionHelper extends BaseCudnnHelper implements Convoluti throw new IllegalArgumentException("Unknown FwdAlgo: " + fwdAlgo); } } else { - code = cudnnGetConvolutionForwardAlgorithm(cudnnContext, cudnnContext.srcTensorDesc, + /* + code = cudnnGetConvolutionForwardAlgorithm_v7(cudnnContext, cudnnContext.srcTensorDesc, cudnnContext.filterDesc, cudnnContext.convDesc, cudnnContext.dstTensorDesc, mode == AlgoMode.NO_WORKSPACE - ? CUDNN_CONVOLUTION_FWD_NO_WORKSPACE : CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, + ? CUDNN_CONVOLUTION_FWD_ : CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, algo); + */ + + val cdf = new cudnnConvolutionFwdAlgoPerf_t(); + val count = new int[1]; + code = cudnnFindConvolutionForwardAlgorithm(cudnnContext, cudnnContext.srcTensorDesc, cudnnContext.filterDesc, cudnnContext.convDesc, cudnnContext.dstTensorDesc, 1, count, cdf); if(code != CUDNN_STATUS_SUCCESS){ //If CuDNN can't infer algorithm - try IMPLICIT_GEMM @@ -477,6 +500,8 @@ public class CudnnConvolutionHelper extends BaseCudnnHelper implements Convoluti fwdAlgo = FwdAlgo.IMPLICIT_GEMM; algo[0] = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM; } + + algo[0] = cdf.algo(); } if(log.isTraceEnabled()){ diff --git a/deeplearning4j/deeplearning4j-cuda/src/test/java/org/deeplearning4j/cuda/lstm/ValidateCudnnLSTM.java b/deeplearning4j/deeplearning4j-cuda/src/test/java/org/deeplearning4j/cuda/lstm/ValidateCudnnLSTM.java index 8d5ece6ff..636071a28 100644 --- a/deeplearning4j/deeplearning4j-cuda/src/test/java/org/deeplearning4j/cuda/lstm/ValidateCudnnLSTM.java +++ b/deeplearning4j/deeplearning4j-cuda/src/test/java/org/deeplearning4j/cuda/lstm/ValidateCudnnLSTM.java @@ -29,6 +29,7 @@ import org.deeplearning4j.nn.multilayer.MultiLayerNetwork; import org.junit.Test; import org.nd4j.linalg.activations.Activation; import org.nd4j.linalg.api.buffer.DataType; +import org.nd4j.linalg.api.concurrency.AffinityManager; import org.nd4j.linalg.api.ndarray.INDArray; import org.nd4j.linalg.dataset.DataSet; import org.nd4j.linalg.factory.Nd4j; @@ -269,7 +270,7 @@ public class ValidateCudnnLSTM extends BaseDL4JTest { assertTrue(f.get(l0) instanceof CudnnLSTMHelper); assertTrue(f.get(l1) instanceof CudnnLSTMHelper); - Random r = new Random(12345); + Random r = new Random(123456); for (int x = 0; x < 1; x++) { INDArray input = Nd4j.rand(new int[] {minibatch, inputSize, timeSeriesLength}); INDArray labels = Nd4j.zeros(minibatch, nOut, timeSeriesLength); @@ -284,7 +285,6 @@ public class ValidateCudnnLSTM extends BaseDL4JTest { mln2.fit(ds); } - assertEquals(mln1.params(), mln2.params()); } diff --git a/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/pom.xml b/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/pom.xml index b1adbe93e..5a48cfc2d 100644 --- a/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/pom.xml +++ b/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/pom.xml @@ -51,7 +51,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-data/deeplearning4j-datavec-iterators/pom.xml b/deeplearning4j/deeplearning4j-data/deeplearning4j-datavec-iterators/pom.xml index c87a94b37..2ce302370 100644 --- a/deeplearning4j/deeplearning4j-data/deeplearning4j-datavec-iterators/pom.xml +++ b/deeplearning4j/deeplearning4j-data/deeplearning4j-datavec-iterators/pom.xml @@ -46,7 +46,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/pom.xml b/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/pom.xml index 7806bab88..103a1a8e7 100644 --- a/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/pom.xml +++ b/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/pom.xml @@ -43,7 +43,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-data/pom.xml b/deeplearning4j/deeplearning4j-data/pom.xml index ca29f35b7..bcf4d3355 100644 --- a/deeplearning4j/deeplearning4j-data/pom.xml +++ b/deeplearning4j/deeplearning4j-data/pom.xml @@ -38,7 +38,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml b/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml index ed0160ccb..15c3c0715 100644 --- a/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml +++ b/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml @@ -116,11 +116,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-graph/pom.xml b/deeplearning4j/deeplearning4j-graph/pom.xml index 645b4eca2..876470377 100644 --- a/deeplearning4j/deeplearning4j-graph/pom.xml +++ b/deeplearning4j/deeplearning4j-graph/pom.xml @@ -64,7 +64,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/pom.xml b/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/pom.xml index 7ebb82e75..86c3e23d6 100644 --- a/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/pom.xml +++ b/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/pom.xml @@ -69,7 +69,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-manifold/pom.xml b/deeplearning4j/deeplearning4j-manifold/pom.xml index 921ee9653..6e1a8cd42 100644 --- a/deeplearning4j/deeplearning4j-manifold/pom.xml +++ b/deeplearning4j/deeplearning4j-manifold/pom.xml @@ -41,7 +41,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml b/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml index 383eb1c8c..13b7ee45d 100644 --- a/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml +++ b/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml @@ -302,11 +302,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-modelimport/pom.xml b/deeplearning4j/deeplearning4j-modelimport/pom.xml index 3dcdcc720..4bbb32806 100644 --- a/deeplearning4j/deeplearning4j-modelimport/pom.xml +++ b/deeplearning4j/deeplearning4j-modelimport/pom.xml @@ -135,11 +135,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml index 911432cf0..9820c29f2 100644 --- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml +++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml @@ -125,11 +125,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-client/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-client/pom.xml index 0886e8d5b..eed007f32 100644 --- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-client/pom.xml +++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-client/pom.xml @@ -49,7 +49,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-model/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-model/pom.xml index bfd004c41..902a67ae7 100644 --- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-model/pom.xml +++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbors-model/pom.xml @@ -53,7 +53,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml index fbe0ddccf..6987dc556 100644 --- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml +++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml @@ -89,11 +89,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/pom.xml index 23d5d225d..70778f67d 100644 --- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/pom.xml +++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/pom.xml @@ -44,7 +44,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-chinese/pom.xml b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-chinese/pom.xml index 23d863cdc..3b0fd8944 100644 --- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-chinese/pom.xml +++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-chinese/pom.xml @@ -72,7 +72,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-japanese/pom.xml b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-japanese/pom.xml index a4fea6b07..c85e18cdd 100644 --- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-japanese/pom.xml +++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-japanese/pom.xml @@ -75,7 +75,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml index c0fcdb84a..be02f45b6 100644 --- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml +++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml @@ -67,7 +67,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/pom.xml b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/pom.xml index e5ee63ea0..7ec64d395 100644 --- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/pom.xml +++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/pom.xml @@ -110,7 +110,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/pom.xml b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/pom.xml index 668c728ae..6595a3a1e 100644 --- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/pom.xml +++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/pom.xml @@ -91,7 +91,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nlp-parent/pom.xml b/deeplearning4j/deeplearning4j-nlp-parent/pom.xml index 6c4eea3fd..61627d8a9 100644 --- a/deeplearning4j/deeplearning4j-nlp-parent/pom.xml +++ b/deeplearning4j/deeplearning4j-nlp-parent/pom.xml @@ -42,7 +42,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-nn/pom.xml b/deeplearning4j/deeplearning4j-nn/pom.xml index b817c0dc6..268a70cd9 100644 --- a/deeplearning4j/deeplearning4j-nn/pom.xml +++ b/deeplearning4j/deeplearning4j-nn/pom.xml @@ -128,7 +128,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml b/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml index 338d0173f..ca4c49e9e 100644 --- a/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml +++ b/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml @@ -107,14 +107,14 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 false org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-remote/pom.xml b/deeplearning4j/deeplearning4j-remote/pom.xml index 4329a1554..1816689a4 100644 --- a/deeplearning4j/deeplearning4j-remote/pom.xml +++ b/deeplearning4j/deeplearning4j-remote/pom.xml @@ -27,7 +27,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml index 2c7a94de8..18392dfc0 100644 --- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml @@ -109,11 +109,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/pom.xml b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/pom.xml index 3c083d40d..36a77391e 100644 --- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/pom.xml @@ -104,7 +104,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-scaleout/pom.xml b/deeplearning4j/deeplearning4j-scaleout/pom.xml index 539aa3ef7..d0a199d76 100644 --- a/deeplearning4j/deeplearning4j-scaleout/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/pom.xml @@ -38,7 +38,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml index 3eafbb9e2..693fbcf7a 100644 --- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml @@ -80,7 +80,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/pom.xml index c4e8dc7ab..ffd7a4b0f 100644 --- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/pom.xml @@ -79,7 +79,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml index 1198ae733..53297cb13 100644 --- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml @@ -86,7 +86,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/pom.xml index b7b20d161..9b399fa22 100644 --- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/pom.xml @@ -105,7 +105,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml index 8a4fb02d5..8bafabc38 100644 --- a/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml +++ b/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml @@ -181,7 +181,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-components/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-components/pom.xml index 8f83b803e..09f5bb084 100644 --- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-components/pom.xml +++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-components/pom.xml @@ -77,7 +77,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-model/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-model/pom.xml index 3c755eac8..6e9cdad17 100644 --- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-model/pom.xml +++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-model/pom.xml @@ -113,7 +113,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-standalone/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-standalone/pom.xml index 1b85a1d87..c807b7a46 100644 --- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-standalone/pom.xml +++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-standalone/pom.xml @@ -32,7 +32,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui/pom.xml index 44b868ae6..f24bf9109 100644 --- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui/pom.xml +++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui/pom.xml @@ -67,7 +67,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-vertx/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-vertx/pom.xml index a66b85ece..835e77fe0 100644 --- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-vertx/pom.xml +++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-vertx/pom.xml @@ -457,7 +457,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 \ No newline at end of file diff --git a/deeplearning4j/deeplearning4j-ui-parent/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/pom.xml index 70c32b984..947a28783 100644 --- a/deeplearning4j/deeplearning4j-ui-parent/pom.xml +++ b/deeplearning4j/deeplearning4j-ui-parent/pom.xml @@ -49,7 +49,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/deeplearning4j-zoo/pom.xml b/deeplearning4j/deeplearning4j-zoo/pom.xml index bec71ec04..ea431ebd9 100644 --- a/deeplearning4j/deeplearning4j-zoo/pom.xml +++ b/deeplearning4j/deeplearning4j-zoo/pom.xml @@ -85,7 +85,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/deeplearning4j/dl4j-integration-tests/pom.xml b/deeplearning4j/dl4j-integration-tests/pom.xml index 43e6bfa60..e8d958cf1 100644 --- a/deeplearning4j/dl4j-integration-tests/pom.xml +++ b/deeplearning4j/dl4j-integration-tests/pom.xml @@ -120,7 +120,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 \ No newline at end of file diff --git a/deeplearning4j/pom.xml b/deeplearning4j/pom.xml index 17c89f931..e121e305d 100644 --- a/deeplearning4j/pom.xml +++ b/deeplearning4j/pom.xml @@ -225,7 +225,7 @@ ${skipBackendChoice} - test-nd4j-native,test-nd4j-cuda-10.2 + test-nd4j-native,test-nd4j-cuda-11.0 false @@ -500,7 +500,7 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 false @@ -513,7 +513,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${nd4j.version} test diff --git a/libnd4j/include/array/NDArray.h b/libnd4j/include/array/NDArray.h index 7b32b7d49..67452dda2 100644 --- a/libnd4j/include/array/NDArray.h +++ b/libnd4j/include/array/NDArray.h @@ -628,17 +628,17 @@ namespace sd { * keepDims - if true then put unities in place of reduced dimensions */ - NDArray reduceAlongDimension(sd::reduce::FloatOps op, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; - NDArray reduceAlongDimension(sd::reduce::FloatOps op, const std::initializer_list& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; + NDArray reduceAlongDimension(sd::reduce::FloatOps op, const std::vector& dimensions, const bool keepDims = false) const; + NDArray reduceAlongDimension(sd::reduce::FloatOps op, const std::initializer_list& dimensions, const bool keepDims = false) const; - NDArray reduceAlongDimension(sd::reduce::SameOps op, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; - NDArray reduceAlongDimension(sd::reduce::SameOps op, const std::initializer_list& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; + NDArray reduceAlongDimension(sd::reduce::SameOps op, const std::vector& dimensions, const bool keepDims = false) const; + NDArray reduceAlongDimension(sd::reduce::SameOps op, const std::initializer_list& dimensions, const bool keepDims = false) const; - NDArray reduceAlongDimension(sd::reduce::BoolOps op, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; - NDArray reduceAlongDimension(sd::reduce::BoolOps op, const std::initializer_list& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; + NDArray reduceAlongDimension(sd::reduce::BoolOps op, const std::vector& dimensions, const bool keepDims = false) const; + NDArray reduceAlongDimension(sd::reduce::BoolOps op, const std::initializer_list& dimensions, const bool keepDims = false) const; - NDArray reduceAlongDimension(sd::reduce::LongOps op, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; - NDArray reduceAlongDimension(sd::reduce::LongOps op, const std::initializer_list& dimensions, const bool keepDims = false, const bool supportOldShapes = false) const; + NDArray reduceAlongDimension(sd::reduce::LongOps op, const std::vector& dimensions, const bool keepDims = false) const; + NDArray reduceAlongDimension(sd::reduce::LongOps op, const std::initializer_list& dimensions, const bool keepDims = false) const; /** * method reduces array by excluding its shapes along dimensions present in given dimensions vector @@ -647,10 +647,10 @@ namespace sd { * keepDims - if true then put unities in place of reduced dimensions * extras - extra parameters */ - void reduceAlongDimension(sd::reduce::FloatOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false, const bool checkTargetShape = true) const; - void reduceAlongDimension(sd::reduce::SameOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false, const bool checkTargetShape = true) const; - void reduceAlongDimension(sd::reduce::BoolOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false, const bool checkTargetShape = true) const; - void reduceAlongDimension(sd::reduce::LongOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool supportOldShapes = false, const bool checkTargetShape = true) const; + void reduceAlongDimension(sd::reduce::FloatOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool checkTargetShape = true) const; + void reduceAlongDimension(sd::reduce::SameOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool checkTargetShape = true) const; + void reduceAlongDimension(sd::reduce::BoolOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool checkTargetShape = true) const; + void reduceAlongDimension(sd::reduce::LongOps op, NDArray& target, const std::vector& dimensions, const bool keepDims = false, const bool checkTargetShape = true) const; /** * return variance of array elements set diff --git a/libnd4j/include/array/NDArray.hXX b/libnd4j/include/array/NDArray.hXX index eefe169cf..cfd910343 100644 --- a/libnd4j/include/array/NDArray.hXX +++ b/libnd4j/include/array/NDArray.hXX @@ -1353,80 +1353,80 @@ void* NDArray::bufferWithOffset(Nd4jLong offset) { ////////////////////////////////////////////////////////////////////////// // eventually method reduces array by excluding its shapes along axes present in dimensions vector -NDArray NDArray::reduceAlongDimension(sd::reduce::FloatOps op, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes) const { +NDArray NDArray::reduceAlongDimension(sd::reduce::FloatOps op, const std::vector& dimensions, const bool keepDims) const { std::vector copy(dimensions); - auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, isR() ? dataType() : Environment::getInstance().defaultFloatDataType(), keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, isR() ? dataType() : Environment::getInstance().defaultFloatDataType(), keepDims, false, getContext()->getWorkspace()); NDArray result(newShape, true, getContext()); - this->reduceAlongDimension(op, result, copy, keepDims, supportOldShapes, false); + this->reduceAlongDimension(op, result, copy, keepDims, false); return result; } ////////////////////////////////////////////////////////////////////////// -NDArray NDArray::reduceAlongDimension(sd::reduce::SameOps op, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes) const { +NDArray NDArray::reduceAlongDimension(sd::reduce::SameOps op, const std::vector& dimensions, const bool keepDims) const { std::vector copy(dimensions); - auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, keepDims, false, getContext()->getWorkspace()); NDArray result(newShape, true, getContext()); - reduceAlongDimension(op, result, copy, keepDims, supportOldShapes, false); + reduceAlongDimension(op, result, copy, keepDims, false); return result; } ////////////////////////////////////////////////////////////////////////// -NDArray NDArray::reduceAlongDimension(sd::reduce::BoolOps op, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes) const { +NDArray NDArray::reduceAlongDimension(sd::reduce::BoolOps op, const std::vector& dimensions, const bool keepDims) const { std::vector copy(dimensions); - auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, DataType::BOOL, keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, DataType::BOOL, keepDims, false, getContext()->getWorkspace()); NDArray result(newShape, true, getContext()); - reduceAlongDimension(op, result, copy, keepDims, supportOldShapes, false); + reduceAlongDimension(op, result, copy, keepDims, false); return result; } ////////////////////////////////////////////////////////////////////////// -NDArray NDArray::reduceAlongDimension(sd::reduce::LongOps op, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes) const { +NDArray NDArray::reduceAlongDimension(sd::reduce::LongOps op, const std::vector& dimensions, const bool keepDims) const { std::vector copy(dimensions); - auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, DataType::INT64, keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo('c', copy, *this, DataType::INT64, keepDims, false, getContext()->getWorkspace()); NDArray result(newShape, true, getContext()); - reduceAlongDimension(op, result, copy, keepDims, supportOldShapes, false); + reduceAlongDimension(op, result, copy, keepDims, false); return result; } ////////////////////////////////////////////////////////////////////////// // method reduces array by excluding its shapes along axes present in dimensions vector -NDArray NDArray::reduceAlongDimension(sd::reduce::FloatOps op, const std::initializer_list& dimensions, const bool keepDims, const bool supportOldShapes) const { - return reduceAlongDimension(op, std::vector(dimensions), keepDims, supportOldShapes); +NDArray NDArray::reduceAlongDimension(sd::reduce::FloatOps op, const std::initializer_list& dimensions, const bool keepDims) const { + return reduceAlongDimension(op, std::vector(dimensions), keepDims); } ////////////////////////////////////////////////////////////////////////// -NDArray NDArray::reduceAlongDimension(sd::reduce::SameOps op, const std::initializer_list& dimensions, const bool keepDims, const bool supportOldShapes) const { - return reduceAlongDimension(op, std::vector(dimensions), keepDims, supportOldShapes); +NDArray NDArray::reduceAlongDimension(sd::reduce::SameOps op, const std::initializer_list& dimensions, const bool keepDims) const { + return reduceAlongDimension(op, std::vector(dimensions), keepDims); } ////////////////////////////////////////////////////////////////////////// -NDArray NDArray::reduceAlongDimension(sd::reduce::BoolOps op, const std::initializer_list& dimensions, const bool keepDims, const bool supportOldShapes) const { - return reduceAlongDimension(op, std::vector(dimensions), keepDims, supportOldShapes); +NDArray NDArray::reduceAlongDimension(sd::reduce::BoolOps op, const std::initializer_list& dimensions, const bool keepDims) const { + return reduceAlongDimension(op, std::vector(dimensions), keepDims); } ////////////////////////////////////////////////////////////////////////// -NDArray NDArray::reduceAlongDimension(sd::reduce::LongOps op, const std::initializer_list& dimensions, const bool keepDims, const bool supportOldShapes) const { - return reduceAlongDimension(op, std::vector(dimensions), keepDims, supportOldShapes); +NDArray NDArray::reduceAlongDimension(sd::reduce::LongOps op, const std::initializer_list& dimensions, const bool keepDims) const { + return reduceAlongDimension(op, std::vector(dimensions), keepDims); } ////////////////////////////////////////////////////////////////////////// @@ -4240,7 +4240,7 @@ NDArray NDArray::applyAllReduce3(sd::reduce3::Ops op, const NDArray& other, cons ////////////////////////////////////////////////////////////////////////// // method reduces array by excluding its shapes along axes present in dimensions vector -void NDArray::reduceAlongDimension(sd::reduce::FloatOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes, const bool checkTargetShape) const { +void NDArray::reduceAlongDimension(sd::reduce::FloatOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool checkTargetShape) const { if (isS()) throw std::runtime_error("NDArray::reduceAlongDimension FloatOps: you can't use this method on String array!"); @@ -4250,7 +4250,7 @@ void NDArray::reduceAlongDimension(sd::reduce::FloatOps op, NDArray& target, con std::vector copy(dimensions); if(checkTargetShape) { - auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, false, getContext()->getWorkspace()); if(!shape::shapeEquals(newShape, target.shapeInfo())) throw std::runtime_error("NDArray::reduceAlongDimension FloatOps: wrong target shape!"); } @@ -4261,8 +4261,18 @@ void NDArray::reduceAlongDimension(sd::reduce::FloatOps op, NDArray& target, con NativeOpExecutioner::execReduceFloatScalar(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(),nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo()); } else { - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(shapeInfo(), copy); - NativeOpExecutioner::execReduceFloat(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo(), copy.data(), copy.size(), packX.platformShapeInfo(), packX.platformOffsets()); + const Nd4jLong* zShapeInfoH = target.shapeInfo(); + const Nd4jLong* zShapeInfoD = target.specialShapeInfo(); + + if(rankOf() - dimensions.size() != target.rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(target.shapeInfo(), copy, target.getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = ShapeUtils::evalDimsForReduceOp(rankOf(), copy); + NativeOpExecutioner::execReduceFloat(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), zShapeInfoH, target.specialBuffer(), zShapeInfoD, dims.data(), dims.size()); + } synchronize("NDArray::reduceAlongDimension FloatOps"); @@ -4271,7 +4281,7 @@ void NDArray::reduceAlongDimension(sd::reduce::FloatOps op, NDArray& target, con ////////////////////////////////////////////////////////////////////////// // method reduces array by excluding its shapes along axes present in dimensions vector -void NDArray::reduceAlongDimension(sd::reduce::SameOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes, const bool checkTargetShape) const { +void NDArray::reduceAlongDimension(sd::reduce::SameOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool checkTargetShape) const { if (isS()) throw std::runtime_error("NDArray::reduceAlongDimension SameOps: you can't use this method on String array!"); @@ -4281,7 +4291,7 @@ void NDArray::reduceAlongDimension(sd::reduce::SameOps op, NDArray& target, cons std::vector copy(dimensions); if(checkTargetShape) { - auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, false, getContext()->getWorkspace()); if(!shape::shapeEquals(newShape, target.shapeInfo())) throw std::runtime_error("NDArray::reduceAlongDimension SameOps: wrong target shape!"); } @@ -4291,10 +4301,19 @@ void NDArray::reduceAlongDimension(sd::reduce::SameOps op, NDArray& target, cons if(rankOf() == copy.size() || copy.empty()) { NativeOpExecutioner::execReduceSameScalar(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo()); } - else { //if (!isEmpty()) { - auto pDims = sd::Environment::getInstance().isCPU() ? copy.data() : nullptr; - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(this->shapeInfo(), copy); - NativeOpExecutioner::execReduceSame(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo(), pDims, copy.size(), packX.platformShapeInfo(), packX.platformOffsets()); + else { + + const Nd4jLong* zShapeInfoH = target.shapeInfo(); + const Nd4jLong* zShapeInfoD = target.specialShapeInfo(); + + if(rankOf() - dimensions.size() != target.rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(target.shapeInfo(), copy, target.getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = ShapeUtils::evalDimsForReduceOp(rankOf(), copy); + NativeOpExecutioner::execReduceSame(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), zShapeInfoH, target.specialBuffer(), zShapeInfoD, dims.data(), dims.size()); } synchronize("NDArray::reduceAlongDimension SameOps"); @@ -4303,7 +4322,7 @@ void NDArray::reduceAlongDimension(sd::reduce::SameOps op, NDArray& target, cons ////////////////////////////////////////////////////////////////////////// // method reduces array by excluding its shapes along axes present in dimensions vector -void NDArray::reduceAlongDimension(sd::reduce::LongOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes, const bool checkTargetShape) const { +void NDArray::reduceAlongDimension(sd::reduce::LongOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool checkTargetShape) const { if (isS()) throw std::runtime_error("NDArray::reduceAlongDimension LongOps: you can't use this method on String array!"); @@ -4313,7 +4332,7 @@ void NDArray::reduceAlongDimension(sd::reduce::LongOps op, NDArray& target, cons std::vector copy(dimensions); if(checkTargetShape) { - auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, false, getContext()->getWorkspace()); if(!shape::shapeEquals(newShape, target.shapeInfo())) throw std::runtime_error("NDArray::reduceAlongDimension LongOps: wrong target shape!"); } @@ -4324,9 +4343,17 @@ void NDArray::reduceAlongDimension(sd::reduce::LongOps op, NDArray& target, cons NativeOpExecutioner::execReduceLongScalar(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo()); } else { - auto pDims = sd::Environment::getInstance().isCPU() ? copy.data() : nullptr; - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(this->shapeInfo(), copy); - NativeOpExecutioner::execReduceLong(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo(), pDims, copy.size(), packX.platformShapeInfo(), packX.platformOffsets()); + const Nd4jLong* zShapeInfoH = target.shapeInfo(); + const Nd4jLong* zShapeInfoD = target.specialShapeInfo(); + + if(rankOf() - dimensions.size() != target.rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(target.shapeInfo(), copy, target.getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = ShapeUtils::evalDimsForReduceOp(rankOf(), copy); + NativeOpExecutioner::execReduceLong(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), zShapeInfoH, target.specialBuffer(), zShapeInfoD, dims.data(), dims.size()); } synchronize("NDArray::reduceAlongDimension LongOps"); @@ -4335,7 +4362,7 @@ void NDArray::reduceAlongDimension(sd::reduce::LongOps op, NDArray& target, cons ////////////////////////////////////////////////////////////////////////// // method reduces array by excluding its shapes along axes present in dimensions vector -void NDArray::reduceAlongDimension(sd::reduce::BoolOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool supportOldShapes, const bool checkTargetShape) const { +void NDArray::reduceAlongDimension(sd::reduce::BoolOps op, NDArray& target, const std::vector& dimensions, const bool keepDims, const bool checkTargetShape) const { if (isS()) throw std::runtime_error("NDArray::reduceAlongDimension BoolOps cuda: you can't use this method on String array!"); @@ -4345,7 +4372,7 @@ void NDArray::reduceAlongDimension(sd::reduce::BoolOps op, NDArray& target, cons std::vector copy(dimensions); if(checkTargetShape) { - auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, supportOldShapes, getContext()->getWorkspace()); + auto newShape = ShapeUtils::evalReduceShapeInfo(target.ordering(), copy, *this, keepDims, false, getContext()->getWorkspace()); if(!shape::shapeEquals(newShape, target.shapeInfo())) throw std::runtime_error("NDArray::reduceAlongDimension BoolOps cuda: wrong target shape!"); } @@ -4356,9 +4383,17 @@ void NDArray::reduceAlongDimension(sd::reduce::BoolOps op, NDArray& target, cons NativeOpExecutioner::execReduceBoolScalar(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo()); } else { - auto pDims = sd::Environment::getInstance().isCPU() ? copy.data() : nullptr; - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(this->shapeInfo(), copy); - NativeOpExecutioner::execReduceBool(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo(), pDims, copy.size(), packX.platformShapeInfo(), packX.platformOffsets()); + const Nd4jLong* zShapeInfoH = target.shapeInfo(); + const Nd4jLong* zShapeInfoD = target.specialShapeInfo(); + + if(rankOf() - dimensions.size() != target.rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(target.shapeInfo(), copy, target.getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = ShapeUtils::evalDimsForReduceOp(rankOf(), copy); + NativeOpExecutioner::execReduceBool(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, target.buffer(), zShapeInfoH, target.specialBuffer(), zShapeInfoD, dims.data(), dims.size()); } synchronize("NDArray::reduceAlongDimension LongOps"); diff --git a/libnd4j/include/helpers/ConstantShapeHelper.h b/libnd4j/include/helpers/ConstantShapeHelper.h index 25440e05c..65c3bcb99 100644 --- a/libnd4j/include/helpers/ConstantShapeHelper.h +++ b/libnd4j/include/helpers/ConstantShapeHelper.h @@ -46,11 +46,13 @@ namespace sd { static ConstantShapeHelper & getInstance(); - ConstantShapeBuffer& bufferForShapeInfo(sd::DataType dataType, char order, const std::vector &shape); - ConstantShapeBuffer& bufferForShapeInfo(const ShapeDescriptor &descriptor); - ConstantShapeBuffer& bufferForShapeInfo(const Nd4jLong *shapeInfo); - ConstantShapeBuffer& bufferForShapeInfo(sd::DataType dataType, char order, int rank, const Nd4jLong* shape); - ConstantShapeBuffer& createShapeInfoWithUnitiesForBroadcast(const Nd4jLong* maxShapeInfo, const Nd4jLong* minShapeInfo, sd::memory::Workspace* workspace = nullptr, const std::vector &dimensions = {}); + ConstantShapeBuffer& bufferForShapeInfo(sd::DataType dataType, char order, const std::vector &shape); + ConstantShapeBuffer& bufferForShapeInfo(const ShapeDescriptor &descriptor); + ConstantShapeBuffer& bufferForShapeInfo(const Nd4jLong *shapeInfo); + ConstantShapeBuffer& bufferForShapeInfo(sd::DataType dataType, char order, int rank, const Nd4jLong* shape); + ConstantShapeBuffer& createShapeInfoWithUnitiesForBroadcast(const Nd4jLong* maxShapeInfo, const Nd4jLong* minShapeInfo, sd::memory::Workspace* workspace = nullptr, const std::vector &dimensions = {}); + ConstantShapeBuffer& createShapeInfoWithNoUnitiesForReduce(const Nd4jLong* maxShapeInfo, const std::vector &dimsWithUnities, sd::memory::Workspace* workspace = nullptr); + ConstantShapeBuffer& createSubArrShapeInfo(const Nd4jLong* inShapeInfo, const int* dims, const int dimsSize, sd::memory::Workspace* workspace = nullptr); const Nd4jLong* emptyShapeInfo(sd::DataType dataType); diff --git a/libnd4j/include/helpers/Loops.h b/libnd4j/include/helpers/Loops.h index 9bf3daede..325fa3505 100644 --- a/libnd4j/include/helpers/Loops.h +++ b/libnd4j/include/helpers/Loops.h @@ -41,43 +41,43 @@ namespace sd { public: template - static FORCEINLINE void loopReduce(const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, E* extraParams, int64_t start, int64_t stop); + static FORCEINLINE void loopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, E* extraParams); }; template class ReductionFloatLoops : public ReductionLoops { public: - static void wrapper(int opNum, const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, Z* extraParams, int64_t start, int64_t stop); + static void wrapper(int opNum, sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, Z* extraParams); template - static void innerloopReduce(const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, Z* extraParams, int64_t start, int64_t stop); + static void innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, Z* extraParams); }; template class ND4J_EXPORT ReductionBoolLoops : public ReductionLoops { public: - static void wrapper(int opNum, const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop); + static void wrapper(int opNum, sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, X* extraParams); template - static void innerloopReduce(const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop); + static void innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, X* extraParams); }; template class ND4J_EXPORT ReductionLongLoops : public ReductionLoops { public: - static void wrapper(int opNum, const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop); + static void wrapper(int opNum, sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, X* extraParams); template - static void innerloopReduce(const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop); + static void innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, X* extraParams); }; template class ND4J_EXPORT ReductionSameLoops : public ReductionLoops { public: - static void wrapper(int opNum, const X* x, const Nd4jLong* xShapeInfo, X* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop); + static void wrapper(int opNum, sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, X* z, const Nd4jLong *zShapeInfo, const int* dims, X* extraParams); template - static void innerloopReduce(const X* x, const Nd4jLong* xShapeInfo, X* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop); + static void innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, X* z, const Nd4jLong *zShapeInfo, const int* dims, X* extraParams); }; @@ -122,372 +122,613 @@ namespace sd { static void innerloopReduce3All(const X* x, const Nd4jLong* xShapeInfo, const X* y, const Nd4jLong* yShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* xTadShapeInfo, const Nd4jLong* xTadOffsets, const Nd4jLong* yTadShapeInfo, const Nd4jLong* yTadOffsets, Z* extraParams, int64_t start, int64_t stop); }; +////////////////////////////////////////////////////////////////////////// +template +static void reduceExec21(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const uint xAxis0 = shape::sizeAt(xShapeInfo, dims[0]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, dims[0]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, 0); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, dims[1]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, dims[1]); + + auto func = PRAGMA_THREADS_FOR { + + for (auto i0 = start; i0 < stop; ++i0) { + + auto x0 = x + i0 * xStrd0; + auto z0 = z + i0 * zStrd0; + + auto s = OpType::startingValue(x0); + + if(xStrd1 == 1) + for (uint i1 = 0; i1 < xAxis1; ++i1) + s = OpType::update(s, OpType::op(x0[i1], extraParams), extraParams); + else + for (uint i1 = 0; i1 < xAxis1; ++i1) + s = OpType::update(s, OpType::op(x0[i1 * xStrd1], extraParams), extraParams); + + *z0 = OpType::postProcess(s, static_cast(xAxis1), extraParams); + } + }; + + samediff::Threads::parallel_for(func, 0,xAxis0); +} + +////////////////////////////////////////////////////////////////////////// +template +static void reduceExec31(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const uint xAxis0 = shape::sizeAt(xShapeInfo, dims[0]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, dims[0]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, 0); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, dims[1]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, dims[1]); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, dims[2]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, dims[2]); + + const Nd4jLong tadLen = static_cast(xAxis1 * xAxis2); + + auto func = PRAGMA_THREADS_FOR { + + for (auto i0 = start; i0 < stop; ++i0) { + + auto x0 = x + i0 * xStrd0; + auto z0 = z + i0 * zStrd0; + + auto s = OpType::startingValue(x0); + + if(xStrd1 == 1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i1 = 0; i1 < xAxis1; ++i1) + s = OpType::update(s, OpType::op(x0[i1 + i2*xStrd2], extraParams), extraParams); + else if(xStrd2 == 1) + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2], extraParams), extraParams); + else + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2*xStrd2], extraParams), extraParams); + + *z0 = OpType::postProcess(s, tadLen, extraParams); + } + }; + + samediff::Threads::parallel_for(func, 0,xAxis0); +} + +////////////////////////////////////////////////////////////////////////// +template +void reduceExec32(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const uint xAxis0 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[1]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[1]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 1); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[0]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[0]); + const Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 0); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, dims[2]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, dims[2]); + auto func = PRAGMA_THREADS_FOR_2D { - /* - ////////////////////////////////////////////////////////////////////////////// - template - void Loops::loopXYZ(const X* x, const Nd4jLong* xShapeInfo, - const Y* y, const Nd4jLong* yShapeInfo, - Z* z, const Nd4jLong* zShapeInfo, - Z* extraParams, - std::function op) { + for (auto i0 = start_x; i0 < stop_x; ++i0) { + for (auto i1 = start_y; i1 < stop_y; ++i1) { - const LoopKind::Kind kindOfLoop = LoopKind::deduceKindOfLoopXYZ(xShapeInfo, yShapeInfo, zShapeInfo); + auto x1 = x + i0 * xStrd0 + i1 * xStrd1; + auto z1 = z + i0 * zStrd0 + i1 * zStrd1; - const Nd4jLong* xShape = shape::shapeOf(xShapeInfo); - const Nd4jLong* xStride = shape::stride(xShapeInfo); - const Nd4jLong* yStride = shape::stride(yShapeInfo); - const Nd4jLong* zStride = shape::stride(zShapeInfo); + auto s = OpType::startingValue(x1); - const Nd4jLong len = shape::length(xShapeInfo); + if(xStrd2 == 1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x1[i2], extraParams), extraParams); + else + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x1[i2 * xStrd2], extraParams), extraParams); - OmpLaunchHelper threadsInfo(len); + *z1 = OpType::postProcess(s, static_cast(xAxis2), extraParams); + } + } + }; - switch (kindOfLoop) { + samediff::Threads::parallel_for(func, 0,xAxis0,1, 0,xAxis1,1); +} - case LoopKind::EWS1: { - PRAGMA_OMP_PARALLEL_THREADS(threadsInfo._numThreads) - { - const auto threadNum = omp_get_thread_num(); - const auto threadOffset = threadsInfo.getThreadOffset(threadNum); - const auto lenPerThread = static_cast(threadsInfo.getItersPerThread(threadNum)); +////////////////////////////////////////////////////////////////////////// +template +void reduceExec41(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { - const auto xi = x + threadOffset; - const auto yi = y + threadOffset; - auto zi = z + threadOffset; + const uint xAxis0 = shape::sizeAt(xShapeInfo, dims[0]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, dims[0]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, 0); - PRAGMA_OMP_SIMD - for (uint i = 0; i < lenPerThread; i++) - zi[i] = op(xi[i], yi[i], extraParams); + const uint xAxis1 = shape::sizeAt(xShapeInfo, dims[1]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, dims[1]); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, dims[2]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, dims[2]); + + const uint xAxis3 = shape::sizeAt(xShapeInfo, dims[3]); + const Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, dims[3]); + + const Nd4jLong tadLen = static_cast(xAxis1 * xAxis2 * xAxis3); + + auto func = PRAGMA_THREADS_FOR { + + for (auto i0 = start; i0 < stop; ++i0) { + + auto x0 = x + i0 * xStrd0; + auto z0 = z + i0 * zStrd0; + + auto s = OpType::startingValue(x0); + + if(xStrd1 == 1) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i1 = 0; i1 < xAxis1; ++i1) + s = OpType::update(s, OpType::op(x0[i1 + i2*xStrd2 + i3*xStrd3], extraParams), extraParams); + else if(xStrd2 == 1) + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2 + i3*xStrd3], extraParams), extraParams); + else if(xStrd3 == 1) + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2*xStrd2 + i3], extraParams), extraParams); + else + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2*xStrd2 + i3*xStrd3], extraParams), extraParams); + + *z0 = OpType::postProcess(s, tadLen, extraParams); + } + }; + + samediff::Threads::parallel_for(func, 0,xAxis0); +} + +////////////////////////////////////////////////////////////////////////// +template +void reduceExec42(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const uint xAxis0 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[1]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[1]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 1); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[0]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[0]); + const Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 0); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, dims[2]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, dims[2]); + + const uint xAxis3 = shape::sizeAt(xShapeInfo, dims[3]); + const Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, dims[3]); + + const Nd4jLong tadLen = static_cast(xAxis2 * xAxis3); + + auto func = PRAGMA_THREADS_FOR_2D { + + for (auto i0 = start_x; i0 < stop_x; ++i0) { + for (auto i1 = start_y; i1 < stop_y; ++i1) { + + auto x1 = x + i0 * xStrd0 + i1 * xStrd1; + auto z1 = z + i0 * zStrd0 + i1 * zStrd1; + + auto s = OpType::startingValue(x1); + + if(xStrd2 == 1) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x1[i2 + i3*xStrd3], extraParams), extraParams); + else if(xStrd3 == 1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x1[i2*xStrd2 + i3], extraParams), extraParams); + else + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x1[i2*xStrd2 + i3*xStrd3], extraParams), extraParams); + + *z1 = OpType::postProcess(s, tadLen, extraParams); + } + } + }; + + samediff::Threads::parallel_for(func, 0,xAxis0,1, 0,xAxis1,1); +} + +////////////////////////////////////////////////////////////////////////// +template +void reduceExec43(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const uint xAxis0 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[2]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[2]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 2); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, dims[1]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, dims[1]); + const Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, 1); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[2] : dims[0]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[2] : dims[0]); + const Nd4jLong zStrd2 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 0); + + const uint xAxis3 = shape::sizeAt(xShapeInfo, dims[3]); + const Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, dims[3]); + + auto func = PRAGMA_THREADS_FOR_3D { + + for (auto i0 = start_x; i0 < stop_x; ++i0) { + for (auto i1 = start_y; i1 < stop_y; ++i1) { + for (auto i2 = start_z; i2 < stop_z; ++i2) { + + auto x2 = x + i0 * xStrd0 + i1 * xStrd1 + i2 * xStrd2; + auto z2 = z + i0 * zStrd0 + i1 * zStrd1 + i2 * zStrd2; + + auto s = OpType::startingValue(x2); + + if(xStrd3 == 1) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x2[i3], extraParams), extraParams); + else + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x2[i3*xStrd3], extraParams), extraParams); + + *z2 = OpType::postProcess(s, static_cast(xAxis3), extraParams); } } - break; + } + }; - case LoopKind::EWSNONZERO: { - const uint xEws = shape::elementWiseStride(xShapeInfo); - const uint yEws = shape::elementWiseStride(yShapeInfo); - const uint zEws = shape::elementWiseStride(zShapeInfo); + samediff::Threads::parallel_for(func, 0,xAxis0,1, 0,xAxis1,1, 0,xAxis2,1); +} - PRAGMA_OMP_PARALLEL_THREADS(threadsInfo._numThreads) - { - const auto threadNum = omp_get_thread_num(); - const auto threadOffset = threadsInfo.getThreadOffset(threadNum); - const auto lenPerThread = static_cast(threadsInfo.getItersPerThread(threadNum)); - const auto xi = x + threadOffset * xEws; - const auto yi = y + threadOffset * yEws; - auto zi = z + threadOffset * zEws; +////////////////////////////////////////////////////////////////////////// +template +void reduceExec51(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { - PRAGMA_OMP_SIMD - for (uint i = 0; i < lenPerThread; i++) - zi[i*zEws] = op(xi[i*xEws], yi[i*yEws], extraParams); + const uint xAxis0 = shape::sizeAt(xShapeInfo, dims[0]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, dims[0]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, 0); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, dims[1]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, dims[1]); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, dims[2]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, dims[2]); + + const uint xAxis3 = shape::sizeAt(xShapeInfo, dims[3]); + const Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, dims[3]); + + const uint xAxis4 = shape::sizeAt(xShapeInfo, dims[4]); + const Nd4jLong xStrd4 = shape::strideAt(xShapeInfo, dims[4]); + + const Nd4jLong tadLen = static_cast(xAxis1 * xAxis2 * xAxis3 * xAxis4); + + auto func = PRAGMA_THREADS_FOR { + + for (auto i0 = start; i0 < stop; ++i0) { + + auto x0 = x + i0 * xStrd0; + auto z0 = z + i0 * zStrd0; + + auto s = OpType::startingValue(x0); + + if(xStrd1 == 1) + for (uint i4 = 0; i4 < xAxis4; ++i4) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i1 = 0; i1 < xAxis1; ++i1) + s = OpType::update(s, OpType::op(x0[i1 + i2*xStrd2 + i3*xStrd3 + i4*xStrd4], extraParams), extraParams); + else if(xStrd2 == 1) + for (uint i4 = 0; i4 < xAxis4; ++i4) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2 + i3*xStrd3 + i4*xStrd4], extraParams), extraParams); + else if(xStrd3 == 1) + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i4 = 0; i4 < xAxis4; ++i4) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2*xStrd2 + i3 + i4*xStrd4], extraParams), extraParams); + else if(xStrd4 == 1) + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2*xStrd2 + i3*xStrd3 + i4], extraParams), extraParams); + else + for (uint i1 = 0; i1 < xAxis1; ++i1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x0[i1*xStrd1 + i2*xStrd2 + i3*xStrd3 + i4*xStrd4], extraParams), extraParams); + + *z0 = OpType::postProcess(s, tadLen, extraParams); + } + }; + + samediff::Threads::parallel_for(func, 0,xAxis0); +} + +////////////////////////////////////////////////////////////////////////// +template +void reduceExec52(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const uint xAxis0 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[1]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[1]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 1); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[0]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[0]); + const Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 0); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, dims[2]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, dims[2]); + + const uint xAxis3 = shape::sizeAt(xShapeInfo, dims[3]); + const Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, dims[3]); + + const uint xAxis4 = shape::sizeAt(xShapeInfo, dims[4]); + const Nd4jLong xStrd4 = shape::strideAt(xShapeInfo, dims[4]); + + const Nd4jLong tadLen = static_cast(xAxis2 * xAxis3 * xAxis4); + + auto func = PRAGMA_THREADS_FOR_2D { + + for (auto i0 = start_x; i0 < stop_x; ++i0) { + for (auto i1 = start_y; i1 < stop_y; ++i1) { + + auto x1 = x + i0 * xStrd0 + i1 * xStrd1; + auto z1 = z + i0 * zStrd0 + i1 * zStrd1; + + auto s = OpType::startingValue(x1); + + if(xStrd2 == 1) + for (uint i4 = 0; i4 < xAxis4; ++i4) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i2 = 0; i2 < xAxis2; ++i2) + s = OpType::update(s, OpType::op(x1[i2 + i3*xStrd3 + i4*xStrd4], extraParams), extraParams); + else if(xStrd3 == 1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i4 = 0; i4 < xAxis4; ++i4) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x1[i2*xStrd2 + i3 + i4*xStrd4], extraParams), extraParams); + else if(xStrd4 == 1) + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x1[i2*xStrd2 + i3*xStrd3 + i4], extraParams), extraParams); + else + for (uint i2 = 0; i2 < xAxis2; ++i2) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x1[i2*xStrd2 + i3*xStrd3 + i4*xStrd4], extraParams), extraParams); + + *z1 = OpType::postProcess(s, tadLen, extraParams); + } + } + }; + + samediff::Threads::parallel_for(func, 0,xAxis0,1, 0,xAxis1,1); +} + +////////////////////////////////////////////////////////////////////////// +template +void reduceExec53(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const uint xAxis0 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[2]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[2]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 2); + + const uint xAxis1 = shape::sizeAt(xShapeInfo, dims[1]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, dims[1]); + const Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, 1); + + const uint xAxis2 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[2] : dims[0]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[2] : dims[0]); + const Nd4jLong zStrd2 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 0); + + const uint xAxis3 = shape::sizeAt(xShapeInfo, dims[3]); + const Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, dims[3]); + + const uint xAxis4 = shape::sizeAt(xShapeInfo, dims[4]); + const Nd4jLong xStrd4 = shape::strideAt(xShapeInfo, dims[4]); + + const Nd4jLong tadLen = static_cast(xAxis3 * xAxis4); + + auto func = PRAGMA_THREADS_FOR_3D { + + for (auto i0 = start_x; i0 < stop_x; ++i0) { + for (auto i1 = start_y; i1 < stop_y; ++i1) { + for (auto i2 = start_z; i2 < stop_z; ++i2) { + + auto x2 = x + i0 * xStrd0 + i1 * xStrd1 + i2 * xStrd2; + auto z2 = z + i0 * zStrd0 + i1 * zStrd1 + i2 * zStrd2; + + auto s = OpType::startingValue(x2); + + if(xStrd3 == 1) + for (uint i4 = 0; i4 < xAxis4; ++i4) + for (uint i3 = 0; i3 < xAxis3; ++i3) + s = OpType::update(s, OpType::op(x2[i3 + i4*xStrd4], extraParams), extraParams); + else if(xStrd4 == 1) + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x2[i3*xStrd3 + i4], extraParams), extraParams); + else + for (uint i3 = 0; i3 < xAxis3; ++i3) + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x2[i3*xStrd3 + i4*xStrd4], extraParams), extraParams); + + *z2 = OpType::postProcess(s, tadLen, extraParams); } } - break; + } + }; - case LoopKind::RANK1: { - PRAGMA_OMP_PARALLEL_FOR - for (uint i0 = 0; i0 < len; ++i0) - z[i0 * zStride[0]] = op(x[i0 * xStride[0]], y[i0 * yStride[0]], extraParams); - } - break; + samediff::Threads::parallel_for(func, 0,xAxis0,1, 0,xAxis1,1, 0,xAxis2,1); +} - case LoopKind::RANK2: { - PRAGMA_OMP_PARALLEL_FOR_SIMD - for (uint i0 = 0; i0 < xShape[0]; ++i0) - for (uint i1 = 0; i1 < xShape[1]; ++i1) - z[i0 * zStride[0] + i1 * zStride[1]] = op(x[i0 * xStride[0] + i1 * xStride[1]], y[i0 * yStride[0] + i1 * yStride[1]], extraParams); - } - break; +////////////////////////////////////////////////////////////////////////// +template +void reduceExec54(const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { - case LoopKind::RANK3: { - PRAGMA_OMP_PARALLEL_FOR_SIMD_COLLAPSE(2) - for (uint i0 = 0; i0 < xShape[0]; ++i0) - for (uint i1 = 0; i1 < xShape[1]; ++i1) - for (uint i2 = 0; i2 < xShape[2]; ++i2) - z[i0*zStride[0]+i1*zStride[1]+i2*zStride[2]] = op(x[i0*xStride[0]+i1*xStride[1]+i2*xStride[2]], y[i0*yStride[0]+i1*yStride[1]+i2*yStride[2]], extraParams); - } - break; + const uint xAxis0 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[3]); + const Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[0] : dims[3]); + const Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 3); - case LoopKind::RANK4: { - PRAGMA_OMP_PARALLEL_FOR_SIMD_COLLAPSE(3) - for (uint i0 = 0; i0 < xShape[0]; ++i0) - for (uint i1 = 0; i1 < xShape[1]; ++i1) - for (uint i2 = 0; i2 < xShape[2]; ++i2) - for (uint i3 = 0; i3 < xShape[3]; ++i3) - z[i0*zStride[0]+i1*zStride[1]+i2*zStride[2]+i3*zStride[3]] = op(x[i0*xStride[0]+i1*xStride[1]+i2*xStride[2]+i3*xStride[3]], y[i0*yStride[0]+i1*yStride[1]+i2*yStride[2]+i3*yStride[3]], extraParams); - } - break; + const uint xAxis1 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[2]); + const Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[1] : dims[2]); + const Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 2); - case LoopKind::RANK5: { - PRAGMA_OMP_PARALLEL_FOR_SIMD_COLLAPSE(4) - for (uint i0 = 0; i0 < xShape[0]; ++i0) - for (uint i1 = 0; i1 < xShape[1]; ++i1) - for (uint i2 = 0; i2 < xShape[2]; ++i2) - for (uint i3 = 0; i3 < xShape[3]; ++i3) - for (uint i4 = 0; i4 < xShape[4]; ++i4) - z[i0*zStride[0]+i1*zStride[1]+i2*zStride[2]+i3*zStride[3]+i4*zStride[4]] = op(x[i0*xStride[0]+i1*xStride[1]+i2*xStride[2]+i3*xStride[3]+i4*xStride[4]], y[i0*yStride[0]+i1*yStride[1]+i2*yStride[2]+i3*yStride[3]+i4*yStride[4]], extraParams); - } - break; + const uint xAxis2 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[2] : dims[1]); + const Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[2] : dims[1]); + const Nd4jLong zStrd2 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 1); - default: { - uint xShapeInfoCast[MAX_RANK]; - uint yShapeInfoCast[MAX_RANK]; - uint zShapeInfoCast[MAX_RANK]; + const uint xAxis3 = shape::sizeAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[3] : dims[0]); + const Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? dims[3] : dims[0]); + const Nd4jLong zStrd3 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 0); - bool canCastX = DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast); - bool canCastY = DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast); - bool canCastZ = DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast); + const uint xAxis4 = shape::sizeAt(xShapeInfo, dims[4]); + const Nd4jLong xStrd4 = shape::strideAt(xShapeInfo, dims[4]); - PRAGMA_OMP_PARALLEL_THREADS(threadsInfo._numThreads) - { - auto threadNum = omp_get_thread_num(); - auto threadOffset = threadsInfo.getThreadOffset(threadNum); - auto lenPerThread = static_cast(threadsInfo.getItersPerThread(threadNum)); - PRAGMA_OMP_SIMD - for (uint i = 0; i < lenPerThread; i++) { - auto xOffset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, canCastX); - auto yOffset = shape::indexOffset(i + threadOffset, yShapeInfo, yShapeInfoCast, canCastY); - auto zOffset = shape::indexOffset(i + threadOffset, zShapeInfo, zShapeInfoCast, canCastZ); - z[zOffset] = op(x[xOffset], y[yOffset], extraParams); + auto func = PRAGMA_THREADS_FOR_3D { + + for (auto i0 = start_x; i0 < stop_x; ++i0) { + for (auto i1 = start_y; i1 < stop_y; ++i1) { + for (auto i2 = start_z; i2 < stop_z; ++i2) { + for (auto i3 = 0; i3 < xAxis3; ++i3) { + + auto x3 = x + i0 * xStrd0 + i1 * xStrd1 + i2 * xStrd2 + i3 * xStrd3; + auto z3 = z + i0 * zStrd0 + i1 * zStrd1 + i2 * zStrd2 + i3 * zStrd3; + + auto s = OpType::startingValue(x3); + + if(xStrd4 == 1) + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x3[i4], extraParams), extraParams); + else + for (uint i4 = 0; i4 < xAxis4; ++i4) + s = OpType::update(s, OpType::op(x3[i4*xStrd4], extraParams), extraParams); + + *z3 = OpType::postProcess(s, static_cast(xAxis4), extraParams); } } } } + }; + + samediff::Threads::parallel_for(func, 0,xAxis0,1, 0,xAxis1,1, 0,xAxis2,1); +} + + +//////////////////////////////////////////////////////////////////////// +template +void reduceDefault(sd::memory::Workspace* workspace, const X *x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int *dims, E* extraParams) { + + const int zRank = shape::rank(zShapeInfo); + const int tadRank = shape::rank(xShapeInfo) - zRank; + + Nd4jLong* outerXTadShapeInfo = sd::ShapeBuilders::createSubArrShapeInfo(xShapeInfo, dims, zRank); + Nd4jLong* innerXTadShapeInfo = sd::ShapeBuilders::createSubArrShapeInfo(xShapeInfo, dims+zRank, tadRank); + + const bool sameOffsets1 = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo); + const bool sameOffsets2 = shape::haveSameShapeAndStrides(zShapeInfo, innerXTadShapeInfo); + + const Nd4jLong zLen = shape::length(zShapeInfo); + const Nd4jLong tadLen = shape::length(innerXTadShapeInfo); + + Nd4jLong* zOffsets = nullptr; + ALLOCATE(zOffsets, workspace, zLen, Nd4jLong); + shape::calcOffsets(zShapeInfo, zOffsets); + + Nd4jLong* outerXTadOffsets = zOffsets; + if(!sameOffsets1) { + ALLOCATE(outerXTadOffsets, workspace, zLen, Nd4jLong); + shape::calcOffsets(outerXTadShapeInfo, outerXTadOffsets); } - */ - - - ////////////////////////////////////////////////////////////////////////////// - template - template - void sd::ReductionLoops::loopReduce(const X* x, const Nd4jLong* xShapeInfo, - Z* z, const Nd4jLong* zShapeInfo, - const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, - E* extraParams, - int64_t start, int64_t stop) { - - const LoopKind::Kind kindOfLoop = LoopKind::deduceKindOfLoopTadXZ(xShapeInfo, zShapeInfo, tadShapeInfo); - - const Nd4jLong zLen = shape::length(zShapeInfo); - const Nd4jLong tadLen = shape::length(tadShapeInfo); - - const uint tadEws = shape::elementWiseStride(tadShapeInfo); - const uint zEws = shape::elementWiseStride(zShapeInfo); - - const Nd4jLong* tadShape = shape::shapeOf(tadShapeInfo); - const Nd4jLong* tadStride = shape::stride(tadShapeInfo); - - int numThreads = OmpLaunchHelper::tadThreads(tadLen, zLen); - - switch (kindOfLoop) { - - //*********************************************// - // case LoopKind::SMALLARR2DX: { - // shape::printShapeInfoLinear(xShapeInfo); - // shape::printShapeInfoLinear(zShapeInfo); - // const auto xLen = zLen * tadLen; - // for (uint i = 0; i < xLen; ++i) { - // const auto zOffset = shape::subArrayOffset(i, xShapeInfo, zShapeInfo, dimsToExclude, dimsLen); - // const uint tadInd = (i / tadEws) % tadLen; - // auto startVal = tadInd ? z[zOffset] : static_cast(OpType::startingValue(x)); - // z[zOffset] = OpType::update(startVal, OpType::op(x[i], extraParams), extraParams); - // if(tadInd == tadLen - 1) - // z[zOffset] = OpType::postProcess(z[zOffset], tadLen, extraParams); - // printf("%u - %lld\n", i, zOffset); - // } - // } - case LoopKind::SMALLARR2DX: { - const auto uTadLen = static_cast(tadLen); - const auto uZLenMinusOne = static_cast(zLen - 1); - const auto xLen = static_cast(zLen * uTadLen); - const auto sv = static_cast(OpType::startingValue(x)); - - for (uint i = 0; i <= uZLenMinusOne; i++) - z[i] = OpType::startingValue(x); - - uint zOffset = 0; - for (uint i = 0; i < xLen; ++i) { - z[zOffset] = OpType::update(z[zOffset], OpType::op(x[i], extraParams), extraParams); - zOffset = zOffset == uZLenMinusOne ? 0 : zOffset + 1; - } - - for (uint i = 0; i <= uZLenMinusOne; i++) - z[i] = OpType::postProcess(z[i], tadLen, extraParams); - } - break; - - //*********************************************// - case LoopKind::EWS1: { - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong j = 0; j < tadLen; j++) - s = OpType::update(s, OpType::op(tad[j], extraParams), extraParams); - - z[i] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::EWSNONZERO: { - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong j = 0; j < tadLen; j++) - s = OpType::update(s, OpType::op(tad[j * tadEws], extraParams), extraParams); - - z[i * zEws] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::RANK1: { - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong i0 = 0; i0 < tadLen; ++i0) - s = OpType::update(s, OpType::op(tad[i0 * tadStride[0]], extraParams), extraParams); - - z[i] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::RANK2: { - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) - for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) - s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1]], extraParams), extraParams); - - z[i] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::RANK3: { - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) - for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) - for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) - s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2]], extraParams), extraParams); - - z[i] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::RANK4: { - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) - for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) - for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) - for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) - s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2] + i3 * tadStride[3]], extraParams), extraParams); - - z[i] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::RANK5: { - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) - for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) - for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) - for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) - for (Nd4jLong i4 = 0; i4 < tadShape[4]; ++i4) - s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2] + i3 * tadStride[3] + i4 * tadStride[4]], extraParams), extraParams); - - z[i] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::X_EWSNONZERO: { - uint castZShapeInfo[MAX_RANK]; - const bool canCastZ = sd::DataTypeUtils::castShapeInfo(zShapeInfo, castZShapeInfo); - - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong j = 0; j < tadLen; j++) - s = OpType::update(s, OpType::op(tad[j * tadEws], extraParams), extraParams); - - auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, canCastZ); - z[zOffset] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - case LoopKind::Z_EWSNONZERO: { - uint castTadShapeInfo[MAX_RANK]; - const bool canCastTad = sd::DataTypeUtils::castShapeInfo(tadShapeInfo, castTadShapeInfo); - - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong j = 0; j < tadLen; j++) { - auto tadOffset = shape::indexOffset(j, tadShapeInfo, castTadShapeInfo, canCastTad); - s = OpType::update(s, OpType::op(tad[tadOffset], extraParams), extraParams); - } - - z[i * zEws] = OpType::postProcess(s, tadLen, extraParams); - }; - } - break; - - //*********************************************// - default: { - auto innertadOffsets = new Nd4jLong[tadLen]; - shape::calcOffsets(tadShapeInfo, innertadOffsets); - - uint castZShapeInfo[MAX_RANK]; - const bool canCastZ = sd::DataTypeUtils::castShapeInfo(zShapeInfo, castZShapeInfo); - - for (auto i = start; i < stop; i++) { - auto tad = x + tadOffsets[i]; - auto s = OpType::startingValue(tad); - - for (Nd4jLong j = 0; j < tadLen; j++) - s = OpType::update(s, OpType::op(tad[innertadOffsets[j]], extraParams), extraParams); - - auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, canCastZ); - z[zOffset] = OpType::postProcess(s, tadLen, extraParams); - }; - - delete[] innertadOffsets; - } - } + Nd4jLong* innerXTadOffsets = zOffsets; + if(!sameOffsets2) { + ALLOCATE(innerXTadOffsets, workspace, tadLen, Nd4jLong); + shape::calcOffsets(innerXTadShapeInfo, innerXTadOffsets); } + auto func = PRAGMA_THREADS_FOR{ + + for (auto i = start; i < stop; ++i) { + + const auto tad = x + outerXTadOffsets[i]; + auto s = OpType::startingValue(tad); + + for (Nd4jLong j = 0; j < tadLen; j++) + s = OpType::update(s, OpType::op(tad[innerXTadOffsets[j]], extraParams), extraParams); + + z[zOffsets[i]] = OpType::postProcess(s, tadLen, extraParams); + } + }; + + samediff::Threads::parallel_for(func, 0, shape::length(zShapeInfo)); + + RELEASE(outerXTadShapeInfo, workspace); + RELEASE(innerXTadShapeInfo, workspace); + RELEASE(zOffsets, workspace); + if(!sameOffsets1) + RELEASE(outerXTadOffsets, workspace); + if(!sameOffsets2) + RELEASE(innerXTadOffsets, workspace); +} + +////////////////////////////////////////////////////////////////////////////// +template +template +void sd::ReductionLoops::loopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong *xShapeInfo, Z* z, const Nd4jLong *zShapeInfo, const int* dims, E* extraParams) { + + const int xRank = shape::rank(xShapeInfo); + const int zRank = shape::rank(zShapeInfo); + + // shape::printShapeInfoLinear(xShapeInfo); + // shape::printShapeInfoLinear(zShapeInfo); + // shape::printIntArray(dims, shape::rank(xShapeInfo)); + + if(xRank == 2 && zRank == 1) + reduceExec21(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 3 && zRank == 1) + reduceExec31(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 3 && zRank == 2) + reduceExec32(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 4 && zRank == 1) + reduceExec41(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 4 && zRank == 2) + reduceExec42(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 4 && zRank == 3) + reduceExec43(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 5 && zRank == 1) + reduceExec51(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 5 && zRank == 2) + reduceExec52(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 5 && zRank == 3) + reduceExec53(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else if(xRank == 5 && zRank == 4) + reduceExec54(x, xShapeInfo, z, zShapeInfo, dims, extraParams); + else + reduceDefault(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +} + ////////////////////////////////////////////////////////////////////////////// diff --git a/libnd4j/include/helpers/ShapeBuilders.h b/libnd4j/include/helpers/ShapeBuilders.h index e2c29a280..14726d5e6 100644 --- a/libnd4j/include/helpers/ShapeBuilders.h +++ b/libnd4j/include/helpers/ShapeBuilders.h @@ -52,11 +52,9 @@ namespace sd { static Nd4jLong* copyShapeInfoAndType(const Nd4jLong* inShapeInfo, const Nd4jLong* shapeInfoToGetTypeFrom, const bool copyStrides, memory::Workspace* workspace = nullptr); /** - * allocates memory for new shapeInfo and copy all information from inShapeInfo to new shapeInfo except dimensions in dimsToExclude (unit dimensions) and corresponding strides - * for example inShapeInfo is {3, 2,1,3,1,4, 12,12,4,4,1, 16384,1,99}, dimsToExclude = {2,3}, dimsSize = 2 - * then outShapeInfo will contain {3, 2,3,4, 12,4,1, 16384,1,99} + * allocates memory for sub-array shapeInfo and copy shape and strides at axes(positions) stored in dims */ - static Nd4jLong* copyShapeInfoWithoutUnites(const Nd4jLong* inShapeInfo, const int dimsSize, const int* dimsToExclude, memory::Workspace* workspace = nullptr); + static Nd4jLong* createSubArrShapeInfo(const Nd4jLong* inShapeInfo, const int* dims, const int dimsSize, memory::Workspace* workspace = nullptr); static Nd4jLong* emptyShapeInfo(const sd::DataType dataType, memory::Workspace* workspace = nullptr); diff --git a/libnd4j/include/helpers/ShapeUtils.h b/libnd4j/include/helpers/ShapeUtils.h index cb2faa43d..bd30d9225 100644 --- a/libnd4j/include/helpers/ShapeUtils.h +++ b/libnd4j/include/helpers/ShapeUtils.h @@ -40,6 +40,12 @@ namespace sd { static const Nd4jLong* evalReduceShapeInfo(const char order, std::vector& dimensions, const NDArray& arr, const bool keepDims = false, const bool supportOldShapes = false, sd::memory::Workspace* workspace = nullptr); static const Nd4jLong* evalReduceShapeInfo(const char order, std::vector& dimensions, const Nd4jLong* shapeInfo, const bool keepDims = false, const bool supportOldShapes = false, sd::memory::Workspace* workspace = nullptr); + + // for example + // if rank = 3 and dimsToExclude = {0,2} then output = {1,0,2}, if rank = 3 and dimsToExclude = {2} then output = {0,1,2} + // if rank = 3 and dimsToExclude = {0} then output = {1,2,0}, if rank = 4 and dimsToExclude = {0,3} then output = {1,2,0,3} + static std::vector evalDimsForReduceOp(const int rank, const std::vector& dimsToExclude); + /** * evaluate output shape for reduce operation when input shape is empty * behavior is analogous to tf diff --git a/libnd4j/include/helpers/benchmark/ReductionBenchmark.h b/libnd4j/include/helpers/benchmark/ReductionBenchmark.h index d87c20d3c..c48030542 100644 --- a/libnd4j/include/helpers/benchmark/ReductionBenchmark.h +++ b/libnd4j/include/helpers/benchmark/ReductionBenchmark.h @@ -94,9 +94,9 @@ namespace sd { auto tadOffsets = Environment::getInstance().isCPU() ? pack.primaryOffsets() : pack.specialOffsets(); if (_opType == 0) - NativeOpExecutioner::execReduceFloat(LaunchContext::defaultContext(), _opNum, _x->buffer(), _x->shapeInfo(), _x->specialBuffer(), _x->specialShapeInfo(), nullptr, _z->buffer(), _z->shapeInfo(), _z->specialBuffer(), _z->specialShapeInfo(), nullptr, _axis.size(), tadOnlyShapeInfo, tadOffsets); + NativeOpExecutioner::execReduceFloat(LaunchContext::defaultContext(), _opNum, _x->buffer(), _x->shapeInfo(), _x->specialBuffer(), _x->specialShapeInfo(), nullptr, _z->buffer(), _z->shapeInfo(), _z->specialBuffer(), _z->specialShapeInfo(), nullptr, _axis.size()); else - NativeOpExecutioner::execReduceSame(LaunchContext::defaultContext(), _opNum, _x->buffer(), _x->shapeInfo(), _x->specialBuffer(), _x->specialShapeInfo(), nullptr, _z->buffer(), _z->shapeInfo(), _z->specialBuffer(), _z->specialShapeInfo(), nullptr, _axis.size(), tadOnlyShapeInfo, tadOffsets); + NativeOpExecutioner::execReduceSame(LaunchContext::defaultContext(), _opNum, _x->buffer(), _x->shapeInfo(), _x->specialBuffer(), _x->specialShapeInfo(), nullptr, _z->buffer(), _z->shapeInfo(), _z->specialBuffer(), _z->specialShapeInfo(), nullptr, _axis.size()); } manager.synchronize(); diff --git a/libnd4j/include/helpers/cpu/ConstantShapeHelper.cpp b/libnd4j/include/helpers/cpu/ConstantShapeHelper.cpp index 528527f36..f12616688 100644 --- a/libnd4j/include/helpers/cpu/ConstantShapeHelper.cpp +++ b/libnd4j/include/helpers/cpu/ConstantShapeHelper.cpp @@ -184,6 +184,43 @@ ConstantShapeBuffer& ConstantShapeHelper::createShapeInfoWithUnitiesForBroadcast return bufferForShapeInfo(descriptor); } -} // namespace sd + + +//////////////////////////////////////////////////////////////////////// +ConstantShapeBuffer& ConstantShapeHelper::createShapeInfoWithNoUnitiesForReduce(const Nd4jLong* inShapeInfo, const std::vector &dimsWithUnities, sd::memory::Workspace* workspace) { + + Nd4jLong* newShapeInfo = nullptr; + ALLOCATE(newShapeInfo, workspace, shape::shapeInfoLength(shape::rank(inShapeInfo) - dimsWithUnities.size()), Nd4jLong); + + int temp; + if(dimsWithUnities.size() == 1 && shape::isCommonVector(inShapeInfo, temp) && temp == dimsWithUnities[0]) { + auto dims = ShapeUtils::evalDimsToExclude(shape::rank(inShapeInfo), {temp}); + shape::excludeUnitiesFromShapeInfo(inShapeInfo, dims.data(), dims.size(), newShapeInfo); + } else { + shape::excludeUnitiesFromShapeInfo(inShapeInfo, dimsWithUnities.data(), dimsWithUnities.size(), newShapeInfo); + } + + ShapeDescriptor descriptor(newShapeInfo); + + RELEASE(newShapeInfo, workspace); + + return bufferForShapeInfo(descriptor); +} + +//////////////////////////////////////////////////////////////////////// +ConstantShapeBuffer& ConstantShapeHelper::createSubArrShapeInfo(const Nd4jLong* inShapeInfo, const int* dims, const int dimsSize, sd::memory::Workspace* workspace) { + + Nd4jLong* newShapeInfo = ShapeBuilders::createSubArrShapeInfo(inShapeInfo, dims, dimsSize, workspace); + + ShapeDescriptor descriptor(newShapeInfo); + + RELEASE(newShapeInfo, workspace); + + return bufferForShapeInfo(descriptor); +} + + + +} #endif \ No newline at end of file diff --git a/libnd4j/include/helpers/cpu/MmulHelper.cpp b/libnd4j/include/helpers/cpu/MmulHelper.cpp index 437eebe1d..91467758b 100644 --- a/libnd4j/include/helpers/cpu/MmulHelper.cpp +++ b/libnd4j/include/helpers/cpu/MmulHelper.cpp @@ -443,17 +443,17 @@ static void batchedGemm(const NDArray* vA, const NDArray* vB, NDArray* vC, // calculate index of current batch Nd4jLong batchInd; if(cRank > 2) - batchInd = shape::coords2index(cShapeInfo, cCoords.data(), cRank - 2, cBatchDims); + batchInd = shape::coords2index(cShapeInfo, cBatchDims, cRank - 2, cCoords.data()); // evaluate A coordinates if(aRank > 2) - shape::index2coords(batchInd, aShapeInfo, aCoords.data(), aRank - 2, aBatchDims); + shape::index2coords(batchInd, aShapeInfo, aBatchDims, aRank - 2, aCoords.data()); aCoords[aMaxis] = cCoords[cMaxis]; aCoords[aKaxis] = 0; // evaluate B coordinates if(bRank > 2) - shape::index2coords(batchInd, bShapeInfo, bCoords.data(), bRank - 2, bBatchDims); + shape::index2coords(batchInd, bShapeInfo, bBatchDims, bRank - 2, bCoords.data()); bCoords[bKaxis] = 0; bCoords[bNaxis] = cCoords[cNaxis]; diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp index e122717fc..04e6e4eb5 100644 --- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp +++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp @@ -26,20 +26,19 @@ namespace sd { template template - void ReductionBoolLoops::innerloopReduce(const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop) { + void ReductionBoolLoops::innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const int* dims, X* extraParams) { #ifndef INLINE_LOOPS - ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop); + ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); #endif } template - void ReductionBoolLoops::wrapper(const int opNum, + void ReductionBoolLoops::wrapper(const int opNum, sd::memory::Workspace* workspace, const X *x, const Nd4jLong *xShapeInfo, Y *z, const Nd4jLong *zShapeInfo, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets, - X *extraParams, int64_t start, int64_t stop) { + const int *dims, X *extraParams) { #ifndef INLINE_LOOPS - DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop), REDUCE_BOOL_OPS); + DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams), REDUCE_BOOL_OPS); #endif } diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float.hpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float.hpp index c7ed544b2..2ab71b34a 100644 --- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float.hpp +++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float.hpp @@ -28,20 +28,19 @@ namespace sd { template template - void ReductionFloatLoops::innerloopReduce(const X * x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, Z* extraParams, int64_t start, int64_t stop) { + void ReductionFloatLoops::innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong* xShapeInfo, Z* z, const Nd4jLong* zShapeInfo, const int* dims, Z* extraParams) { #ifndef INLINE_LOOPS - ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop); + ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); #endif } template - void ReductionFloatLoops::wrapper(const int opNum, const X *x, const Nd4jLong *xShapeInfo, + void ReductionFloatLoops::wrapper(const int opNum, sd::memory::Workspace* workspace, + const X *x, const Nd4jLong *xShapeInfo, Y *z, const Nd4jLong *zShapeInfo, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets, - Y *extraParams, - int64_t start, int64_t stop) { + const int *dims, Y *extraParams) { #ifndef INLINE_LOOPS - DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop), REDUCE_FLOAT_OPS); + DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams), REDUCE_FLOAT_OPS); #endif } diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp index be6cb28bd..820091f09 100644 --- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp +++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp @@ -33,18 +33,19 @@ namespace sd { template template - void ReductionLongLoops::innerloopReduce(const X * x, const Nd4jLong* xShapeInfo, Z *z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop) { + void ReductionLongLoops::innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong* xShapeInfo, Z *z, const Nd4jLong* zShapeInfo, const int* dims, X* extraParams) { #ifndef INLINE_LOOPS - ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop); + ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop); #endif } template - void ReductionLongLoops::wrapper(const int opNum, const X *x, const Nd4jLong *xShapeInfo, Y *z, - const Nd4jLong *zShapeInfo, const Nd4jLong *tadShapeInfo, - const Nd4jLong *tadOffsets, X *extraParams, int64_t start, int64_t stop) { + void ReductionLongLoops::wrapper(const int opNum, sd::memory::Workspace* workspace, + const X *x, const Nd4jLong *xShapeInfo, + Y *z, const Nd4jLong *zShapeInfo, + const int *dims, X *extraParams) { #ifndef INLINE_LOOPS - DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop), REDUCE_LONG_OPS); + DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams), REDUCE_LONG_OPS); #endif } diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp index 53725de83..2544a3c03 100644 --- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp +++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp @@ -26,23 +26,23 @@ namespace sd { template template - void ReductionSameLoops::innerloopReduce(const X* x, const Nd4jLong* xShapeInfo, X* z, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop) { + void ReductionSameLoops::innerloopReduce(sd::memory::Workspace* workspace, const X* x, const Nd4jLong* xShapeInfo, X* z, const Nd4jLong* zShapeInfo, const int* dims, X* extraParams) { #ifndef INLINE_LOOPS - ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop); + ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); #endif } template - void ReductionSameLoops::wrapper(const int opNum, const X *vx, const Nd4jLong *xShapeInfo, X *vz, - const Nd4jLong *zShapeInfo, const Nd4jLong *tadShapeInfo, - const Nd4jLong *tadOffsets, - X *vextraParams, int64_t start, int64_t stop) { + void ReductionSameLoops::wrapper(const int opNum, sd::memory::Workspace* workspace, + const X *vx, const Nd4jLong *xShapeInfo, + X *z, const Nd4jLong *zShapeInfo, + const int *dims, X *vextraParams) { #ifndef INLINE_LOOPS auto x = reinterpret_cast(vx); auto z = reinterpret_cast(vz); auto extraParams = reinterpret_cast(vextraParams); - DISPATCH_BY_OPNUM_T(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams, start, stop), REDUCE_SAME_OPS); + DISPATCH_BY_OPNUM_T(innerloopReduce, PARAMS(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams), REDUCE_SAME_OPS); #endif } diff --git a/libnd4j/include/helpers/cuda/ConstantShapeHelper.cu b/libnd4j/include/helpers/cuda/ConstantShapeHelper.cu index 35ba60ca9..fb093e7b7 100644 --- a/libnd4j/include/helpers/cuda/ConstantShapeHelper.cu +++ b/libnd4j/include/helpers/cuda/ConstantShapeHelper.cu @@ -24,6 +24,7 @@ #include #include #include +#include #include #include @@ -187,4 +188,38 @@ ConstantShapeBuffer& ConstantShapeHelper::createShapeInfoWithUnitiesForBroadcas return bufferForShapeInfo(descriptor); } +//////////////////////////////////////////////////////////////////////// +ConstantShapeBuffer& ConstantShapeHelper::createShapeInfoWithNoUnitiesForReduce(const Nd4jLong* inShapeInfo, const std::vector &dimsWithUnities, sd::memory::Workspace* workspace) { + + Nd4jLong* newShapeInfo = nullptr; + ALLOCATE(newShapeInfo, workspace, shape::shapeInfoLength(shape::rank(inShapeInfo) - dimsWithUnities.size()), Nd4jLong); + + int temp; + if(dimsWithUnities.size() == 1 && shape::isCommonVector(inShapeInfo, temp) && temp == dimsWithUnities[0]) { + auto dims = ShapeUtils::evalDimsToExclude(shape::rank(inShapeInfo), {temp}); + shape::excludeUnitiesFromShapeInfo(inShapeInfo, dims.data(), dims.size(), newShapeInfo); + } else { + shape::excludeUnitiesFromShapeInfo(inShapeInfo, dimsWithUnities.data(), dimsWithUnities.size(), newShapeInfo); + } + + ShapeDescriptor descriptor(newShapeInfo); + + RELEASE(newShapeInfo, workspace); + + return bufferForShapeInfo(descriptor); +} + +//////////////////////////////////////////////////////////////////////// +ConstantShapeBuffer& ConstantShapeHelper::createSubArrShapeInfo(const Nd4jLong* inShapeInfo, const int* dims, const int dimsSize, sd::memory::Workspace* workspace) { + + Nd4jLong* newShapeInfo = ShapeBuilders::createSubArrShapeInfo(inShapeInfo, dims, dimsSize, workspace); + + ShapeDescriptor descriptor(newShapeInfo); + + RELEASE(newShapeInfo, workspace); + + return bufferForShapeInfo(descriptor); +} + + } \ No newline at end of file diff --git a/libnd4j/include/helpers/cuda_off/MmulHelper.cu b/libnd4j/include/helpers/cuda_off/MmulHelper.cu index d1122d794..36f48184a 100644 --- a/libnd4j/include/helpers/cuda_off/MmulHelper.cu +++ b/libnd4j/include/helpers/cuda_off/MmulHelper.cu @@ -571,17 +571,17 @@ static __global__ void batchedCudaGemm(const void* vA, const Nd4jLong* aShapeInf // calculate index of current batch Nd4jLong batchInd; if(cBatchDims != nullptr) - batchInd = shape::coords2index(cShapeInfo, cCoords, cRank - 2, cBatchDims); + batchInd = shape::coords2index(cShapeInfo, cBatchDims, cRank - 2, cCoords); // evaluate A coordinates if(aBatchDims != nullptr) - shape::index2coords(batchInd, aShapeInfo, aCoords, aRank - 2, aBatchDims); + shape::index2coords(batchInd, aShapeInfo, aBatchDims, aRank - 2, aCoords); aCoords[aMaxis] = cCoords[cMaxis]; aCoords[aKaxis] = 0; // evaluate B coordinates if(bBatchDims != nullptr) - shape::index2coords(batchInd, bShapeInfo, bCoords, bRank - 2, bBatchDims); + shape::index2coords(batchInd, bShapeInfo, bBatchDims, bRank - 2, bCoords); bCoords[bKaxis] = 0; bCoords[bNaxis] = cCoords[cNaxis]; diff --git a/libnd4j/include/helpers/impl/ShapeBuilders.cpp b/libnd4j/include/helpers/impl/ShapeBuilders.cpp index 7c0c7fed6..dbcf6dac0 100644 --- a/libnd4j/include/helpers/impl/ShapeBuilders.cpp +++ b/libnd4j/include/helpers/impl/ShapeBuilders.cpp @@ -140,14 +140,26 @@ namespace sd { } //////////////////////////////////////////////////////////////////////////////// -Nd4jLong* ShapeBuilders::copyShapeInfoWithoutUnites(const Nd4jLong* inShapeInfo, const int dimsSize, const int* dimsToExclude, memory::Workspace* workspace) { +Nd4jLong* ShapeBuilders::createSubArrShapeInfo(const Nd4jLong* inShapeInfo, const int* dims, const int dimsSize, memory::Workspace* workspace) { - Nd4jLong *outShapeInfo = nullptr; - ALLOCATE(outShapeInfo, workspace, shape::shapeInfoLength(inShapeInfo[0] - dimsSize), Nd4jLong); + Nd4jLong *subArrShapeInfo = nullptr; + ALLOCATE(subArrShapeInfo, workspace, shape::shapeInfoLength(dimsSize), Nd4jLong); - shape::excludeUnitiesFromShapeInfo(inShapeInfo, dimsSize, dimsToExclude, outShapeInfo); + subArrShapeInfo[0] = dimsSize; // rank + sd::ArrayOptions::copyDataType(subArrShapeInfo, inShapeInfo); // type + subArrShapeInfo[2*dimsSize + 3] = shape::order(inShapeInfo); // order - return outShapeInfo; + Nd4jLong* shape = shape::shapeOf(subArrShapeInfo); + Nd4jLong* strides = shape::stride(subArrShapeInfo); + + for(int i = 0; i < dimsSize; ++i) { + shape[i] = shape::sizeAt(inShapeInfo, dims[i]); + strides[i] = shape::strideAt(inShapeInfo, dims[i]); + } + + shape::checkStridesEwsAndOrder(subArrShapeInfo); + + return subArrShapeInfo; } } \ No newline at end of file diff --git a/libnd4j/include/helpers/impl/ShapeUtils.cpp b/libnd4j/include/helpers/impl/ShapeUtils.cpp index 2c189cff1..998df7728 100644 --- a/libnd4j/include/helpers/impl/ShapeUtils.cpp +++ b/libnd4j/include/helpers/impl/ShapeUtils.cpp @@ -1062,6 +1062,17 @@ bool ShapeUtils::areShapesEqual(const Nd4jLong* shapeInfo, const std::vector ShapeUtils::evalDimsForReduceOp(const int rank, const std::vector& dimsToExclude) { + + std::vector output = ShapeUtils::evalDimsToExclude(rank, dimsToExclude); + + for(uint j = 0; j < dimsToExclude.size(); ++j) + output.emplace_back(dimsToExclude[j]); + + return output; +} + //////////////////////////////////////////////////////////////////////////////// /* bool ShapeUtils::isSubArrayCase(const NDArray& arr1, const NDArray& arr2, std::vector& sameDims) { diff --git a/libnd4j/include/helpers/shape.h b/libnd4j/include/helpers/shape.h index 719b086cb..ca6054482 100644 --- a/libnd4j/include/helpers/shape.h +++ b/libnd4j/include/helpers/shape.h @@ -901,6 +901,16 @@ namespace shape { ND4J_EXPORT _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const Nd4jLong *coords, Nd4jLong baseOffset = 0); ND4J_EXPORT _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const int *coords, Nd4jLong baseOffset = 0); ND4J_EXPORT _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const uint *coords, Nd4jLong baseOffset = 0); + ND4J_EXPORT _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const int *coords, const int* dims); // length of dims is equal to rank of shapeInfo + + // all three arrays should have same rank + // all three arrays should have same dimensions or some of them are 1 (that is satisfy broadcasting principle), strides may be different + // shapeInfo1 - first array should have max length compared to rest of two arrays + ND4J_EXPORT _CUDA_HD void getOffsetBroadcast(const Nd4jLong& startInd, const Nd4jLong ind, + const Nd4jLong* shapeInfo1, const Nd4jLong* shapeInfo2, const Nd4jLong* shapeInfo3, + const bool sameOffsets12, const bool sameOffsets13, + int* coords, + Nd4jLong& offset1, Nd4jLong& offset2, Nd4jLong& offset3); ND4J_EXPORT _CUDA_HD Nd4jLong* createShapeInfo(Nd4jLong *shape, Nd4jLong *stride, int rank); @@ -918,11 +928,12 @@ namespace shape { ND4J_EXPORT _CUDA_HD void index2coordsCPU(const Nd4jLong& startIndex, const Nd4jLong& index, const Nd4jLong *shapeInfo, Nd4jLong *coords); ND4J_EXPORT _CUDA_HD void index2coordsCPU(const Nd4jLong& startIndex, const Nd4jLong& index, const Nd4jLong *shapeInfo, int *coords); + // ND4J_EXPORT _CUDA_HD void index2coordsCPU(const Nd4jLong& startIndex, const Nd4jLong& index, const Nd4jLong *shapeInfo, const int* dims, Nd4jLong *coords); /** * take into account only dimensions stored in tadDims, tadDims must be sorted in increasing order! */ - ND4J_EXPORT _CUDA_HD void index2coords(Nd4jLong index, const Nd4jLong *shapeInfo, int *coords, const int dimsSize, const int* tadDims); + ND4J_EXPORT _CUDA_HD void index2coords(Nd4jLong index, const Nd4jLong *shapeInfo, const int* dims, const int dimsLen, int *coords); /** * Convert coordinates to the corresponding linear index (sequence number in other words) @@ -935,7 +946,7 @@ namespace shape { /** * take into account only dimensions stored in tadDims, tadDims must be sorted in increasing order! */ - ND4J_EXPORT _CUDA_HD Nd4jLong coords2index(const Nd4jLong *shapeInfo, const int *coords, const int dimsSize, const int* tadDims); + ND4J_EXPORT _CUDA_HD Nd4jLong coords2index(const Nd4jLong *shapeInfo, const int* dims, const int dimsSize, const int *coords); /** * increment n-dimensional array by one iteration by changing coord appropriately @@ -951,7 +962,7 @@ namespace shape { ND4J_EXPORT _CUDA_HD Nd4jLong getIndexOffset(Nd4jLong index, const Nd4jLong *shapeInfo); ND4J_EXPORT _CUDA_HD Nd4jLong indexOffset(Nd4jLong index, const Nd4jLong* lShapeInfo, const uint* uShapeInfo, const bool useUnsigned); - ND4J_EXPORT _CUDA_HD void printShapeInfo(Nd4jLong *shapeInfo); + ND4J_EXPORT _CUDA_HD void printShapeInfo(const Nd4jLong *shapeInfo); ND4J_EXPORT _CUDA_HD void printShapeInfoLinear(const Nd4jLong *shapeInfo); @@ -1057,10 +1068,10 @@ namespace shape { ND4J_EXPORT _CUDA_HD int excludeUnitiesFromShapeInfo(const Nd4jLong* inShapeInfo, Nd4jLong*& shapeNoUnities, Nd4jLong*& stridesNoUnities); /** - * for example inShapeInfo is {3, 2,1,3,1,4, 12,12,4,4,1, 16384,1,99}, dimsToExclude = {1,3}, dimsSize = 2 + * for example inShapeInfo is {3, 2,1,3,1,4, 12,12,4,4,1, 16384,1,99}, dimsToExclude(points on unity dimensions) = {1,3}, dimsSize = 2 * then outShapeInfo will contain {3, 2,3,4, 12,4,1, 16384,1,99} */ - INLINEDEF _CUDA_HD void excludeUnitiesFromShapeInfo(const Nd4jLong* inShapeInfo, const int dimsSize, const int* dimsToExclude, Nd4jLong* outShapeInfo); + INLINEDEF _CUDA_HD void excludeUnitiesFromShapeInfo(const Nd4jLong* inShapeInfo, const int* dimsToExclude, const int dimsSize, Nd4jLong* outShapeInfo); /** * get stride over contiguous axis (contiguous axis must have stride = 1) @@ -1847,13 +1858,13 @@ INLINEDEF _CUDA_HD Nd4jLong coords2index(const int rank, const Nd4jLong *shape, return index; } -INLINEDEF _CUDA_HD Nd4jLong coords2index(const Nd4jLong *shapeInfo, const int *coords, const int dimsSize, const int* tadDims) { +INLINEDEF _CUDA_HD Nd4jLong coords2index(const Nd4jLong *shapeInfo, const int* dims, const int dimsLen, const int *coords) { Nd4jLong index, shift = 1;; - index = coords[tadDims[dimsSize - 1]]; - for(uint i = dimsSize - 1; i >= 1; --i) { - shift *= shapeInfo[tadDims[i]]; + index = coords[dims[dimsLen - 1]]; + for(uint i = dimsLen - 1; i >= 1; --i) { + shift *= shapeInfo[dims[i]]; index += shift * coords[i - 1]; } @@ -3324,6 +3335,18 @@ INLINEDEF _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const Nd4jLong return offset; } +////////////////////////////////////////////////////////////////////////// +INLINEDEF _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const uint *coords, Nd4jLong baseOffset) { + + Nd4jLong offset = baseOffset; + + for(uint i = 1; i <= shapeInfo[0]; ++i) + if(shapeInfo[i] != 1) + offset += coords[i - 1] * shapeInfo[shapeInfo[0] + i]; + + return offset; +} + ////////////////////////////////////////////////////////////////////////// INLINEDEF _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const int *coords, Nd4jLong baseOffset) { @@ -3337,17 +3360,78 @@ INLINEDEF _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const int *coor } ////////////////////////////////////////////////////////////////////////// -INLINEDEF _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const uint *coords, Nd4jLong baseOffset) { +INLINEDEF _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const int *coords, const int* dims) { - Nd4jLong offset = baseOffset; + Nd4jLong offset = 0; for(uint i = 1; i <= shapeInfo[0]; ++i) if(shapeInfo[i] != 1) - offset += coords[i - 1] * shapeInfo[shapeInfo[0] + i]; + offset += coords[dims[i - 1]] * shapeInfo[shapeInfo[0] + i]; return offset; } +////////////////////////////////////////////////////////////////////// +INLINEDEF _CUDA_HD void getOffsetBroadcast(const Nd4jLong& startInd, const Nd4jLong ind, + const Nd4jLong* shapeInfo1, const Nd4jLong* shapeInfo2, const Nd4jLong* shapeInfo3, + const bool sameOffsets12, const bool sameOffsets13, + int* coords, + Nd4jLong& offset1, Nd4jLong& offset2, Nd4jLong& offset3) { + + const Nd4jLong* shape1 = shape::shapeOf(shapeInfo1); + const Nd4jLong* strides1 = shape::stride(shapeInfo1); + const Nd4jLong* shape2 = shape::shapeOf(shapeInfo2); + const Nd4jLong* strides2 = shape::stride(shapeInfo2); + const Nd4jLong* shape3 = shape::shapeOf(shapeInfo3); + const Nd4jLong* strides3 = shape::stride(shapeInfo3); + + if(startInd == ind) { + + if(shape::rank(shapeInfo1) == 0) { + offset1 = offset2 = offset3 = 0; + return; + } + + shape::index2coords(ind, shapeInfo1, coords); + offset1 = shape::getOffset(shapeInfo1, coords); + + if(sameOffsets12) + offset2 = offset1; + else + offset2 = shape::getOffset(shapeInfo2, coords); + + if(sameOffsets13) + offset3 = offset1; + else + offset3 = shape::getOffset(shapeInfo3, coords); + + return; + } + + int axis = shapeInfo1[0] - 1; + while(coords[axis] == shape1[axis] - 1) { + if(!sameOffsets12 && shape2[axis] != 1) + offset2 -= (shape2[axis] - 1) * strides2[axis]; + if(!sameOffsets13 && shape3[axis] != 1) + offset3 -= (shape3[axis] - 1) * strides3[axis]; + if(shape1[axis] != 1) + offset1 -= (shape1[axis] - 1) * strides1[axis]; + coords[axis--] = 0; + } + + ++coords[axis]; + offset1 += strides1[axis]; + + if(!sameOffsets12 && shape2[axis] != 1) + offset2 += strides2[axis]; + if(!sameOffsets13 && shape3[axis] != 1) + offset3 += strides3[axis]; + + if(sameOffsets12) + offset2 = offset1; + if(sameOffsets13) + offset3 = offset1; +} /** * Returns the tensor along dimension @@ -3443,7 +3527,7 @@ INLINEDEF _CUDA_HD Nd4jLong getOffset(const Nd4jLong *shapeInfo, const uint *coo printf("\n"); } - INLINEDEF _CUDA_HD void printShapeInfo(Nd4jLong *shapeInfo) { + INLINEDEF _CUDA_HD void printShapeInfo(const Nd4jLong *shapeInfo) { int rank = shape::rank(shapeInfo); Nd4jLong *shape = shape::shapeOf(shapeInfo); printf("Rank %d\n",rank); @@ -4583,89 +4667,92 @@ INLINEDEF void calcOffsets(const Nd4jLong* shapeInfo, Nd4jLong* offsets, const c ////////////////////////////////////////////////////////////////////// INLINEDEF void calcOffsets(const int rank, const Nd4jLong* shape, const Nd4jLong* strides, Nd4jLong* offsets, const char order) { - // if(false) { // tests showed that this code did calculation notably slower even for big N - // Nd4jLong indexes[MAX_RANK]; - // PRAGMA_OMP_PARALLEL_FOR_ARGS(private(indexes)) - // for (Nd4jLong i = 0; i < N; ++i) { - // shape::index2coords(rank, shape, i, indexes); - // subArrOffsets[i] = 0; - // for (int j = 0; j < rank; ++j) - // if(shape[j] != 1) - // subArrOffsets[i] += indexes[j] * strides[j]; - // } - // return; - // } + const uint64_t len = shape::prodLong(shape, rank); // set offset for first sub-array, it is equal to zero always offsets[0] = 0; - Nd4jLong * idx = new Nd4jLong[rank]; - Nd4jLong* offsetPerDim = new Nd4jLong[rank]; - memset(idx, 0, sizeof(Nd4jLong) * rank); + uint coords[MAX_RANK]; + memset(coords, 0, sizeof(uint) * rank); - PRAGMA_OMP_SIMD - for (int k = 0; k < rank; ++k) - offsetPerDim[k] = (shape[k] - 1) * strides[k]; - - Nd4jLong init = 0, i = 1; - // nested loops - calculation of sub-array offsets if(order == 'c') { - Nd4jLong rankMinusOne = rank - 1, j = rankMinusOne; - - while(j >= 0) { - - if(shape[j] == 1) { --j; continue; } // ignore dimensions equal to unity - - if(j == rankMinusOne) { // last dimension - for(int l = 1; l < shape[j]; ++l) { - offsets[i] = offsets[i - 1] + strides[j]; - i++; - } - --j; - } - else if(idx[j] < shape[j] - 1) { - init += strides[j]; - offsets[i++] = init; - ++idx[j]; - j = rankMinusOne; - } - else { - init -= offsetPerDim[j]; - idx[j--] = 0; + for (uint64_t i = 1; i < len; ++i) { + int axis = rank - 1; + offsets[i] = 0; + while(coords[axis] == shape[axis] - 1) { + offsets[i] -= (shape[axis] - 1) * strides[axis]; + coords[axis--] = 0; } + ++coords[axis]; + offsets[i] += offsets[i-1] + strides[axis]; } - } - else { + } else { - Nd4jLong j = 0; - - while(j < rank) { - - if(shape[j] == 1) { ++j; continue; } // ignore dimensions equal to unity - - if(j == 0) { // last dimension - for(int l = 1; l < shape[j]; ++l) { - offsets[i] = offsets[i - 1] + strides[j]; - i++; - } - ++j; - } - else if(idx[j] < shape[j] - 1) { - init += strides[j]; - offsets[i++] = init; - ++idx[j]; - j = 0; - } - else { - init -= offsetPerDim[j]; - idx[j++] = 0; + for (uint64_t i = 1; i < len; ++i) { + int axis = 0; + offsets[i] = 0; + while(coords[axis] == shape[axis] - 1) { + offsets[i] -= (shape[axis] - 1) * strides[axis]; + coords[axis++] = 0; } + ++coords[axis]; + offsets[i] += offsets[i-1] + strides[axis]; } } - delete []idx; - delete []offsetPerDim; + // Nd4jLong init = 0, i = 1; + // // nested loops - calculation of sub-array offsets + // if(order == 'c') { + + // int rankMinusOne = rank - 1, j = rankMinusOne; + + // while(j >= 0) { + + // if(shape[j] == 1) { --j; continue; } // ignore dimensions equal to unity + + // if(j == rankMinusOne) { // last dimension + // for(uint l = 1; l < shape[j]; ++l) + // offsets[i++] = offsets[i - 1] + strides[j]; + // --j; + // } + // else if(coords[j] < shape[j] - 1) { + // init += strides[j]; + // offsets[i++] = init; + // ++coords[j]; + // j = rankMinusOne; + // } + // else { + // init -= (shape[j] - 1) * strides[j]; + // coords[j--] = 0; + // } + // } + // } + // else { + + // int j = 0; + + // while(j < rank) { + + // if(shape[j] == 1) { ++j; continue; } // ignore dimensions equal to unity + + // if(j == 0) { // last dimension + // for(uint l = 1; l < shape[j]; ++l) + // offsets[i++] = offsets[i - 1] + strides[j]; + // ++j; + // } + // else if(coords[j] < shape[j] - 1) { + // init += strides[j]; + // offsets[i++] = init; + // ++coords[j]; + // j = 0; + // } + // else { + // init -= (shape[j] - 1) * strides[j]; + // coords[j++] = 0; + // } + // } + // } } ////////////////////////////////////////////////////////////////////// @@ -4884,13 +4971,14 @@ INLINEDEF void _CUDA_HD index2coords(Nd4jLong index, const int rank, const Nd4jL } ////////////////////////////////////////////////////////////////////// -INLINEDEF _CUDA_HD void index2coords(Nd4jLong index, const Nd4jLong *shapeInfo, int *coords, const int dimsSize, const int* tadDims) { +INLINEDEF _CUDA_HD void index2coords(Nd4jLong index, const Nd4jLong *shapeInfo, const int* dims, const int dimsLen, int *coords) { - for(uint i = dimsSize - 1; i > 0; --i) { - coords[tadDims[i]] = index % shapeInfo[1 + tadDims[i]]; - index /= shapeInfo[1 + tadDims[i]]; + for(uint i = dimsLen - 1; i > 0; --i) { + const auto ind = dims[i]; + coords[ind] = index % shapeInfo[1 + ind]; + index /= shapeInfo[1 + ind]; } - coords[tadDims[0]] = index; // last iteration + coords[dims[0]] = index; // last iteration } ////////////////////////////////////////////////////////////////////// @@ -4921,6 +5009,64 @@ INLINEDEF _CUDA_HD void index2coordsCPU(const Nd4jLong& startIndex, const Nd4jLo } } +////////////////////////////////////////////////////////////////////// +INLINEDEF _CUDA_HD int excludeUnitiesFromShapeInfo(const Nd4jLong* inShapeInfo, Nd4jLong*& shapeNoUnities, Nd4jLong*& stridesNoUnities) { + + const int rank = shape::rank(inShapeInfo); + const int numOfNonUnities = shape::numOfNonUnitDims(rank, shape::shapeOf(inShapeInfo)); + + if(numOfNonUnities == rank) { // no unities in shape, no copy procedure + shapeNoUnities = const_cast(inShapeInfo) + 1; + stridesNoUnities = const_cast(inShapeInfo) + 1 + rank; + return numOfNonUnities; + } + + for(uint j = 0, i = 0; i < rank; ++i) { + if(shape::shapeOf(inShapeInfo)[i] != 1) { + shapeNoUnities[j] = shape::shapeOf(inShapeInfo)[i]; + shapeNoUnities[numOfNonUnities + j++] = shape::stride(inShapeInfo)[i]; + } + } + + stridesNoUnities = shapeNoUnities + numOfNonUnities; + + return numOfNonUnities; +} + +////////////////////////////////////////////////////////////////////// +INLINEDEF _CUDA_HD void excludeUnitiesFromShapeInfo(const Nd4jLong* inShapeInfo, const int* dimsToExclude, const int dimsSize, Nd4jLong* outShapeInfo) { + + outShapeInfo[0] = inShapeInfo[0] - dimsSize; + + for(uint j = 0, k = 0, i = 0; i < inShapeInfo[0]; ++i) { + if(j < dimsSize && i == dimsToExclude[j]) { + ++j; + continue; + } + + shape::shapeOf(outShapeInfo)[k] = shape::shapeOf(inShapeInfo)[i]; + shape::stride(outShapeInfo)[k++] = shape::stride(inShapeInfo)[i]; + } + + sd::ArrayOptions::copyDataType(outShapeInfo, inShapeInfo); // type + *shape::ews(outShapeInfo) = shape::elementWiseStride(inShapeInfo); // ews + outShapeInfo[2 * outShapeInfo[0] + 3] = shape::order(inShapeInfo); // order +} + +////////////////////////////////////////////////////////////////////// +// INLINEDEF _CUDA_HD void index2coordsCPU(const Nd4jLong& startIndex, const Nd4jLong& index, const Nd4jLong *shapeInfo, const int* dims, const int dimsLen, int *coords) { + +// if(startIndex == index) { +// shape::index2coords(index, shapeInfo, dims, dimsLen, coords); +// } +// else { +// int i = dimsLen - 1; +// while(coords[dims[i]] == shape::sizeAt(shapeInfo, dims[i]) - 1) +// coords[dims[i--]] = 0; +// ++coords[dims[i]]; +// } +// } + ////////////////////////////////////////////////////////////////////// // INLINEDEF _CUDA_HD void calcOffsets(const Nd4jLong *xShapeInfo, Nd4jLong*& xOffsets, const Nd4jLong *yShapeInfo, Nd4jLong*& yOffsets, const Nd4jLong* zShapeInfo, Nd4jLong*& zOffsets, const char order) { @@ -5111,50 +5257,6 @@ INLINEDEF _CUDA_HD void index2coordsCPU(const Nd4jLong& startIndex, const Nd4jLo // } // } -////////////////////////////////////////////////////////////////////// -INLINEDEF _CUDA_HD int excludeUnitiesFromShapeInfo(const Nd4jLong* inShapeInfo, Nd4jLong*& shapeNoUnities, Nd4jLong*& stridesNoUnities) { - - const int rank = shape::rank(inShapeInfo); - const int numOfNonUnities = shape::numOfNonUnitDims(rank, shape::shapeOf(inShapeInfo)); - - if(numOfNonUnities == rank) { // no unities in shape, no copy procedure - shapeNoUnities = const_cast(inShapeInfo) + 1; - stridesNoUnities = const_cast(inShapeInfo) + 1 + rank; - return numOfNonUnities; - } - - for(uint j = 0, i = 0; i < rank; ++i) { - if(shape::shapeOf(inShapeInfo)[i] != 1) { - shapeNoUnities[j] = shape::shapeOf(inShapeInfo)[i]; - shapeNoUnities[numOfNonUnities + j++] = shape::stride(inShapeInfo)[i]; - } - } - - stridesNoUnities = shapeNoUnities + numOfNonUnities; - - return numOfNonUnities; -} - -////////////////////////////////////////////////////////////////////// -INLINEDEF _CUDA_HD void excludeUnitiesFromShapeInfo(const Nd4jLong* inShapeInfo, const int dimsSize, const int* dimsToExclude, Nd4jLong* outShapeInfo) { - - outShapeInfo[0] = inShapeInfo[0] - dimsSize; - - for(uint j = 0, k = 0, i = 0; i < inShapeInfo[0]; ++i) { - if(j < dimsSize && i == dimsToExclude[j]) { - ++j; - continue; - } - - shape::shapeOf(outShapeInfo)[k] = shape::shapeOf(inShapeInfo)[i]; - shape::stride(outShapeInfo)[k++] = shape::stride(inShapeInfo)[i]; - } - - sd::ArrayOptions::copyDataType(outShapeInfo, inShapeInfo); // type - *shape::ews(outShapeInfo) = shape::elementWiseStride(inShapeInfo); // ews - outShapeInfo[2 * outShapeInfo[0] + 3] = shape::order(inShapeInfo); // order -} - ////////////////////////////////////////////////////////////////////// // INLINEDEF _CUDA_HD Nd4jLong strideOverContigAxis(const int axis, const Nd4jLong* inShapeInfo) { diff --git a/libnd4j/include/legacy/NativeOpExecutioner.h b/libnd4j/include/legacy/NativeOpExecutioner.h index 84ab886c4..20f0b6532 100644 --- a/libnd4j/include/legacy/NativeOpExecutioner.h +++ b/libnd4j/include/legacy/NativeOpExecutioner.h @@ -470,8 +470,7 @@ static void execTransformBool(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + int *dimension, int dimensionLength); static void execReduceSame(sd::LaunchContext *lc, int opNum, @@ -480,8 +479,7 @@ static void execTransformBool(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + int *dimension, int dimensionLength); static void execReduceBool(sd::LaunchContext *lc, int opNum, @@ -490,8 +488,7 @@ static void execTransformBool(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + int *dimension, int dimensionLength); static void execReduceLong(sd::LaunchContext *lc, int opNum, @@ -500,8 +497,7 @@ static void execTransformBool(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + int *dimension, int dimensionLength); /** * diff --git a/libnd4j/include/legacy/cpu/NativeOpExecutioner.cpp b/libnd4j/include/legacy/cpu/NativeOpExecutioner.cpp index 6b6c51a13..be8a0fbb3 100644 --- a/libnd4j/include/legacy/cpu/NativeOpExecutioner.cpp +++ b/libnd4j/include/legacy/cpu/NativeOpExecutioner.cpp @@ -585,8 +585,7 @@ void NativeOpExecutioner::execReduceFloat(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { + int *dimension, int dimensionLength) { @@ -597,13 +596,7 @@ void NativeOpExecutioner::execReduceFloat(sd::LaunchContext *lc, if (shape::isEmpty(hZShapeInfo)) return; - auto func = PRAGMA_THREADS_FOR { - BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceFloatFunction, ::exec(opNum, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, start, stop), LIBND4J_TYPES, FLOAT_TYPES); - }; - - const sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopTadXZ(hXShapeInfo, hZShapeInfo, tadShapeInfo); - - samediff::Threads::parallel_tad(func, 0, shape::length(hZShapeInfo), 1, kindOfLoop == sd::LoopKind::Kind::SMALLARR2DX ? 1 : sd::Environment::getInstance().maxMasterThreads()); + BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceFloatFunction, ::exec(opNum, lc ? lc->getWorkspace() : nullptr, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension), LIBND4J_TYPES, FLOAT_TYPES); } //////////////////////////////////////////////////////////////////////// @@ -614,24 +607,16 @@ void NativeOpExecutioner::execReduceSame(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { + int *dimension, int dimensionLength) { auto xType = sd::ArrayOptions::dataType(hXShapeInfo); - auto zType = sd::ArrayOptions::dataType(hZShapeInfo); // nothing to do here if result is empty if (shape::isEmpty(hZShapeInfo)) return; - auto func = PRAGMA_THREADS_FOR { - BUILD_SINGLE_SELECTOR(xType, functions::reduce::ReduceSameFunction, ::exec(opNum, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, start, stop), LIBND4J_TYPES); - }; - - const sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopTadXZ(hXShapeInfo, hZShapeInfo, tadShapeInfo); - - samediff::Threads::parallel_tad(func, 0, shape::length(hZShapeInfo), 1, kindOfLoop == sd::LoopKind::Kind::SMALLARR2DX ? 1 : sd::Environment::getInstance().maxMasterThreads()); + BUILD_SINGLE_SELECTOR(xType, functions::reduce::ReduceSameFunction, ::exec(opNum, lc ? lc->getWorkspace() : nullptr, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension), LIBND4J_TYPES); } //////////////////////////////////////////////////////////////////////// @@ -642,8 +627,7 @@ void NativeOpExecutioner::execReduceBool(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { + int *dimension, int dimensionLength) { auto xType = sd::ArrayOptions::dataType(hXShapeInfo); @@ -653,13 +637,7 @@ void NativeOpExecutioner::execReduceBool(sd::LaunchContext *lc, if (shape::isEmpty(hZShapeInfo)) return; - auto func = PRAGMA_THREADS_FOR { - BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceBoolFunction, ::exec(opNum, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, start, stop), LIBND4J_TYPES, BOOL_TYPES); - }; - - const sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopTadXZ(hXShapeInfo, hZShapeInfo, tadShapeInfo); - - samediff::Threads::parallel_tad(func, 0, shape::length(hZShapeInfo), 1, kindOfLoop == sd::LoopKind::Kind::SMALLARR2DX ? 1 : sd::Environment::getInstance().maxMasterThreads()); + BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceBoolFunction, ::exec(opNum, lc ? lc->getWorkspace() : nullptr, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension), LIBND4J_TYPES, BOOL_TYPES); } //////////////////////////////////////////////////////////////////////// @@ -670,8 +648,7 @@ void NativeOpExecutioner::execReduceLong(sd::LaunchContext *lc, void *extraParams, void *hZ, const Nd4jLong *hZShapeInfo, void *dZ, const Nd4jLong *dZShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { + int *dimension, int dimensionLength) { auto xType = sd::ArrayOptions::dataType(hXShapeInfo); @@ -681,13 +658,7 @@ void NativeOpExecutioner::execReduceLong(sd::LaunchContext *lc, if (shape::isEmpty(hZShapeInfo)) return; - auto func = PRAGMA_THREADS_FOR { - BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceLongFunction, ::exec(opNum, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, start, stop), LIBND4J_TYPES, LONG_TYPES); - }; - - const sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopTadXZ(hXShapeInfo, hZShapeInfo, tadShapeInfo); - - samediff::Threads::parallel_tad(func, 0, shape::length(hZShapeInfo), 1, kindOfLoop == sd::LoopKind::Kind::SMALLARR2DX ? 1 : sd::Environment::getInstance().maxMasterThreads()); + BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceLongFunction, ::exec(opNum, lc ? lc->getWorkspace() : nullptr, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, dimension), LIBND4J_TYPES, LONG_TYPES); } //////////////////////////////////////////////////////////////////////// diff --git a/libnd4j/include/legacy/cpu/NativeOps.cpp b/libnd4j/include/legacy/cpu/NativeOps.cpp index f9e3f669c..463adc17e 100644 --- a/libnd4j/include/legacy/cpu/NativeOps.cpp +++ b/libnd4j/include/legacy/cpu/NativeOps.cpp @@ -447,28 +447,26 @@ void execReduceFloat2(Nd4jPointer *extraPointers, OpaqueDataBuffer *dbZ, const Nd4jLong *hZShapeInfo, const Nd4jLong *dZShapeInfo, OpaqueDataBuffer *dbDimension, const Nd4jLong *hDimensionShape, const Nd4jLong *dDimensionShape) { try { + auto dimension = reinterpret_cast(dbDimension->primary()); auto dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPackX = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, dimension, dimensionLength); + const auto zLen = shape::length(hZShapeInfo); - auto hTADShapeInfo = tadPackX.primaryShapeInfo(); - auto hTADOffsets = tadPackX.primaryOffsets(); + std::vector dimensions(dimension, dimension + dimensionLength); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + const Nd4jLong* zShapeInfoD = dZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo) && zLen != 1) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); + NativeOpExecutioner::execReduceFloat(nullptr, opNum, dbX->primary(), hXShapeInfo, dbX->special(), dXShapeInfo, extraParams, dbZ->primary(), zShapeInfoH, dbZ->special(), zShapeInfoD, dims.data(), dims.size()); - NativeOpExecutioner::execReduceFloat(nullptr, opNum, - dbX->primary(), - hXShapeInfo, - dbX->special(), - dXShapeInfo, - extraParams, - dbZ->primary(), - hZShapeInfo, - dbZ->special(), - dZShapeInfo, - dimension, - dimensionLength, - hTADShapeInfo, - hTADOffsets); } catch (std::exception &e) { sd::LaunchContext::defaultContext()->errorReference()->setErrorCode(1); sd::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what()); @@ -481,30 +479,27 @@ void execReduceBool2(Nd4jPointer *extraPointers, void *extraParams, OpaqueDataBuffer *dbZ, const Nd4jLong *hZShapeInfo, const Nd4jLong *dZShapeInfo, OpaqueDataBuffer *dbDimension, const Nd4jLong *hDimensionShape, const Nd4jLong *dDimensionShape) { + try { auto dimension = reinterpret_cast(dbDimension->primary()); auto dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, dimension, - dimensionLength); + std::vector dimensions(dimension, dimension + dimensionLength); - auto hTADShapeInfo = tadPack.primaryShapeInfo(); - auto hTADOffsets = tadPack.primaryOffsets(); + const auto zLen = shape::length(hZShapeInfo); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + const Nd4jLong* zShapeInfoD = dZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo)) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); + NativeOpExecutioner::execReduceBool(nullptr, opNum, dbX->primary(), hXShapeInfo, dbX->special(), dXShapeInfo, extraParams, dbZ->primary(), zShapeInfoH, dbZ->special(), zShapeInfoD, dims.data(), dims.size()); - NativeOpExecutioner::execReduceBool(nullptr, opNum, - dbX->primary(), - hXShapeInfo, - dbX->special(), - dXShapeInfo, - extraParams, - dbZ->primary(), - hZShapeInfo, - dbZ->special(), - dZShapeInfo, - dimension, - dimensionLength, - hTADShapeInfo, - hTADOffsets); } catch (std::exception &e) { sd::LaunchContext::defaultContext()->errorReference()->setErrorCode(1); sd::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what()); @@ -521,26 +516,22 @@ void execReduceSame2(Nd4jPointer *extraPointers, auto dimension = reinterpret_cast(dbDimension->primary()); int dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, dimension, - dimensionLength); + std::vector dimensions(dimension, dimension + dimensionLength); - auto hTADShapeInfo = tadPack.primaryShapeInfo(); - auto hTADOffsets = tadPack.primaryOffsets(); + const auto zLen = shape::length(hZShapeInfo); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + const Nd4jLong* zShapeInfoD = dZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo) && zLen != 1) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); + NativeOpExecutioner::execReduceSame(nullptr, opNum, dbX->primary(), hXShapeInfo, dbX->special(), dXShapeInfo, extraParams, dbZ->primary(), zShapeInfoH, dbZ->special(), zShapeInfoD, dims.data(), dims.size()); - NativeOpExecutioner::execReduceSame(nullptr, opNum, - dbX->primary(), - hXShapeInfo, - dbX->special(), - dXShapeInfo, - extraParams, - dbZ->primary(), - hZShapeInfo, - dbZ->special(), - dZShapeInfo, - dimension, - dimensionLength, - hTADShapeInfo, - hTADOffsets); } catch (std::exception &e) { sd::LaunchContext::defaultContext()->errorReference()->setErrorCode(1); sd::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what()); @@ -557,25 +548,22 @@ void execReduceLong2(Nd4jPointer *extraPointers, auto dimension = reinterpret_cast(dbDimension->primary()); int dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, dimension, dimensionLength); + std::vector dimensions(dimension, dimension + dimensionLength); - auto hTADShapeInfo = tadPack.primaryShapeInfo(); - auto hTADOffsets = tadPack.primaryOffsets(); + const auto zLen = shape::length(hZShapeInfo); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + const Nd4jLong* zShapeInfoD = dZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo) && zLen != 1) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); + NativeOpExecutioner::execReduceLong(nullptr, opNum, dbX->primary(), hXShapeInfo, dbX->special(), dXShapeInfo, extraParams, dbZ->primary(), zShapeInfoH, dbZ->special(), zShapeInfoD, dims.data(), dims.size()); - NativeOpExecutioner::execReduceLong(nullptr, opNum, - dbX->primary(), - hXShapeInfo, - dbX->special(), - dXShapeInfo, - extraParams, - dbZ->primary(), - hZShapeInfo, - dbZ->special(), - dZShapeInfo, - dimension, - dimensionLength, - hTADShapeInfo, - hTADOffsets); } catch (std::exception &e) { sd::LaunchContext::defaultContext()->errorReference()->setErrorCode(1); sd::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what()); diff --git a/libnd4j/include/legacy/cuda/NativeOpExecutioner.cu b/libnd4j/include/legacy/cuda/NativeOpExecutioner.cu index 14cbf306a..cb3c78238 100644 --- a/libnd4j/include/legacy/cuda/NativeOpExecutioner.cu +++ b/libnd4j/include/legacy/cuda/NativeOpExecutioner.cu @@ -210,7 +210,7 @@ void NativeOpExecutioner::execSummaryStatsScalar(sd::LaunchContext *lc, auto stream = lc->getCudaStream(); auto reductionPointer = lc->getReductionPointer(); - dim3 launchDims = dim3(256, 256, 32768); + dim3 launchDims = dim3(256, CUDA_BLOCK_SIZE, 1024); auto xType = sd::ArrayOptions::dataType(hXShapeInfo); auto zType = sd::ArrayOptions::dataType(hZShapeInfo); @@ -577,8 +577,7 @@ void NativeOpExecutioner::execReduceSame(sd::LaunchContext *lc, void *extraParams, void *hZ, Nd4jLong const* hZShapeInfo, void *dZ, Nd4jLong const* dZShapeInfo, - int *dimension, int dimensionLength, - Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets) { + int *dimension, int dimensionLength) { auto stream = lc->getCudaStream(); auto reductionPointer = lc->getReductionPointer(); @@ -588,15 +587,14 @@ void NativeOpExecutioner::execReduceSame(sd::LaunchContext *lc, auto xType = sd::ArrayOptions::dataType(hXShapeInfo); auto zType = sd::ArrayOptions::dataType(hZShapeInfo); - auto xRank = shape::rank(hXShapeInfo); if (zType != xType) throw datatype_exception::build("NativeOpExecutioner::execReduceSame requires both X & Z operands to have same type", xType, zType); auto numBlocks = shape::length(hZShapeInfo); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 8192); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, CUDA_BLOCK_SIZE, 1024); - BUILD_SINGLE_SELECTOR(xType, functions::reduce::ReduceSameFunction, ::execReduceXD(launchDims, stream, opNum, xRank, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), LIBND4J_TYPES); + BUILD_SINGLE_SELECTOR(xType, functions::reduce::ReduceSameFunction, ::execReduceXD(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, reductionPointer, dZ, dZShapeInfo, hZShapeInfo, dimension), LIBND4J_TYPES); // TODO: remove after the release auto res = cudaStreamSynchronize(*stream); @@ -612,8 +610,7 @@ void NativeOpExecutioner::execReduceLong(sd::LaunchContext *lc, void *extraParams, void *hZ, Nd4jLong const* hZShapeInfo, void *dZ, Nd4jLong const* dZShapeInfo, - int *dimension,int dimensionLength, - Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets) { + int *dimension,int dimensionLength) { auto stream = lc->getCudaStream(); auto reductionPointer = lc->getReductionPointer(); @@ -627,11 +624,10 @@ void NativeOpExecutioner::execReduceLong(sd::LaunchContext *lc, if (zType != sd::DataType::INT64) throw datatype_exception::build("NativeOpExecutioner::execReduceLong wrong Z data type", sd::DataType::INT64, zType); - auto xRank = shape::rank(hXShapeInfo); auto numBlocks = shape::length(hZShapeInfo); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, CUDA_BLOCK_SIZE, 1024); - BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceLongFunction, ::execReduceXD(launchDims, stream, opNum, xRank, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), LIBND4J_TYPES, LONG_TYPES); + BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceLongFunction, ::execReduceXD(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, reductionPointer, dZ, dZShapeInfo, hZShapeInfo, dimension), LIBND4J_TYPES, LONG_TYPES); // TODO: remove after the release auto res = cudaStreamSynchronize(*stream); @@ -648,8 +644,7 @@ void NativeOpExecutioner::execReduceBool(sd::LaunchContext *lc, void *extraParams, void *hZ, Nd4jLong const* hZShapeInfo, void *dZ, Nd4jLong const* dZShapeInfo, - int *dimension, int dimensionLength, - Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets) { + int *dimension, int dimensionLength) { auto stream = lc->getCudaStream(); auto reductionPointer = lc->getReductionPointer(); @@ -663,11 +658,10 @@ void NativeOpExecutioner::execReduceBool(sd::LaunchContext *lc, if (zType != sd::DataType::BOOL) throw std::runtime_error("NativeOpExecutioner::execReduceBool requires Z operand to have BOOL type"); - auto xRank = shape::rank(hXShapeInfo); auto numBlocks = shape::length(hZShapeInfo); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, CUDA_BLOCK_SIZE, 1024); - BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceBoolFunction, ::execReduceXD(launchDims, stream, opNum, xRank, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), LIBND4J_TYPES, BOOL_TYPES); + BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceBoolFunction, ::execReduceXD(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, reductionPointer, dZ, dZShapeInfo, hZShapeInfo, dimension), LIBND4J_TYPES, BOOL_TYPES); // TODO: remove after the release auto res = cudaStreamSynchronize(*stream); @@ -675,6 +669,45 @@ void NativeOpExecutioner::execReduceBool(sd::LaunchContext *lc, throw cuda_exception::build("execReduceBool failed", res); } +//////////////////////////////////////////////////////////////////////// +/** + * + * @param opNum + * @param dX + * @param dXShapeInfo + * @param extraParams + * @param dZ + * @param dZShapeInfo + */ +void NativeOpExecutioner::execReduceFloat(sd::LaunchContext *lc, + int opNum, + const void *hX, const Nd4jLong *hXShapeInfo, + const void *dX, const Nd4jLong *dXShapeInfo, + void *extraParams, + void *hZ, const Nd4jLong *hZShapeInfo, + void *dZ, const Nd4jLong *dZShapeInfo, + int *dimension, int dimensionLength) { + + auto stream = lc->getCudaStream(); + auto reductionPointer = lc->getReductionPointer(); + + if (sd::Environment::getInstance().isDebugAndVerbose()) + printf("F8 opNum:[%i]\n", opNum); + + auto xType = sd::ArrayOptions::dataType(hXShapeInfo); + auto zType = sd::ArrayOptions::dataType(hZShapeInfo); + + auto numBlocks = shape::length(hZShapeInfo); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 32768); + + BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceFloatFunction, ::execReduceXD(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, reductionPointer, dZ, dZShapeInfo, hZShapeInfo, dimension), LIBND4J_TYPES, FLOAT_TYPES); + + // TODO: remove after the release + auto res = cudaStreamSynchronize(*stream); + if (res != 0) + throw cuda_exception::build("execReduceFloat failed", res); +} + //////////////////////////////////////////////////////////////////////// /** * @@ -707,7 +740,8 @@ void NativeOpExecutioner::execIndexReduce(sd::LaunchContext *lc, auto xType = sd::ArrayOptions::dataType(hXShapeInfo); auto zType = sd::ArrayOptions::dataType(hZShapeInfo); auto numBlocks = shape::length(hZShapeInfo); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 32768); + auto tadLength = shape::length(hXShapeInfo) / numBlocks; + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, tadLength < CUDA_BLOCK_SIZE ? tadLength : CUDA_BLOCK_SIZE, 1024); if (zType != sd::DataType::INT64 && zType != sd::DataType::INT32) throw datatype_exception::build("NativeOpExecutioner::execIndexReduce requires Z operand to have INT32/INT64 type", zType); @@ -722,46 +756,6 @@ void NativeOpExecutioner::execIndexReduce(sd::LaunchContext *lc, throw cuda_exception::build("execIndexReduce failed", res); } -//////////////////////////////////////////////////////////////////////// -/** - * - * @param opNum - * @param dX - * @param dXShapeInfo - * @param extraParams - * @param dZ - * @param dZShapeInfo - */ -void NativeOpExecutioner::execReduceFloat(sd::LaunchContext *lc, - int opNum, - void const* hX, Nd4jLong const* hXShapeInfo, - void const* dX, Nd4jLong const* dXShapeInfo, - void *extraParams, - void *hZ, Nd4jLong const* hZShapeInfo, - void *dZ, Nd4jLong const* dZShapeInfo, - int *dimension,int dimensionLength, - Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets) { - - auto stream = lc->getCudaStream(); - auto reductionPointer = lc->getReductionPointer(); - - if (sd::Environment::getInstance().isDebugAndVerbose()) - printf("F8 opNum:[%i]\n", opNum); - - auto xType = sd::ArrayOptions::dataType(hXShapeInfo); - auto zType = sd::ArrayOptions::dataType(hZShapeInfo); - - auto xRank = shape::rank(hXShapeInfo); - auto numBlocks = shape::length(hZShapeInfo); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 32768); - - BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceFloatFunction, ::execReduceXD(launchDims, stream, opNum, xRank, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), LIBND4J_TYPES, FLOAT_TYPES); - - // TODO: remove after the release - auto res = cudaStreamSynchronize(*stream); - if (res != 0) - throw cuda_exception::build("execReduceFloat failed", res); -} /** @@ -790,7 +784,7 @@ void NativeOpExecutioner::execIndexReduceScalar(sd::LaunchContext *lc, auto xLength = shape::length(hXShapeInfo); auto blockWidth = 256; auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, CUDA_BLOCK_SIZE, 1024); if (sd::Environment::getInstance().isDebugAndVerbose() && launchDims.x == 1) printf("AF1 opNum:[%i]\n", opNum); @@ -840,7 +834,7 @@ void NativeOpExecutioner::execReduceFloatScalar(sd::LaunchContext *lc, auto xLength = shape::length(hXShapeInfo); auto blockWidth = 256; auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, CUDA_BLOCK_SIZE, 1024); BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceFloatFunction, ::execReduceScalar(launchDims, stream, opNum, dX,dXShapeInfo, hXShapeInfo, extraParams, dZ,dZShapeInfo, hZShapeInfo, nullptr, 0, reductionPointer, nullptr), LIBND4J_TYPES, FLOAT_TYPES); @@ -870,9 +864,9 @@ void NativeOpExecutioner::execReduceBoolScalar(sd::LaunchContext *lc, throw std::runtime_error("NativeOpExecutioner::execReduceBoolScalar requires Z operand to have BOOL type"); auto xLength = shape::length(hXShapeInfo); - auto blockWidth = 256; + auto blockWidth = CUDA_BLOCK_SIZE; auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 1024); BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceBoolFunction, ::execReduceScalar(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, nullptr, 0, reductionPointer, nullptr), LIBND4J_TYPES, BOOL_TYPES); @@ -901,9 +895,9 @@ void NativeOpExecutioner::execReduceSameScalar(sd::LaunchContext *lc, throw datatype_exception::build("NativeOpExecutioner::execReduceSameScalar requires both X & Z operands to have same type", xType, zType); auto xLength = shape::length(hXShapeInfo); - auto blockWidth = 256; + auto blockWidth = CUDA_BLOCK_SIZE; auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 1024); BUILD_SINGLE_SELECTOR(xType, functions::reduce::ReduceSameFunction, ::execReduceScalar(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, nullptr, 0, reductionPointer, nullptr), LIBND4J_TYPES); @@ -932,9 +926,9 @@ void NativeOpExecutioner::execReduceLongScalar(sd::LaunchContext *lc, throw datatype_exception::build("NativeOpExecutioner::execReduceLongScalar wrong Z data type", sd::DataType::INT64, zType); auto xLength = shape::length(hXShapeInfo); - auto blockWidth = 256; + auto blockWidth = CUDA_BLOCK_SIZE; auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 1024); BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceLongFunction, ::execReduceScalar(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, nullptr, 0, reductionPointer, nullptr), LIBND4J_TYPES, LONG_TYPES); @@ -1128,7 +1122,7 @@ void NativeOpExecutioner::execSummaryStats(sd::LaunchContext *lc, auto stream = lc->getCudaStream(); auto reductionPointer = lc->getReductionPointer(); - dim3 launchDims = dim3(256, 256, 32768); + dim3 launchDims = dim3(256, CUDA_BLOCK_SIZE, 1024); auto xType = sd::ArrayOptions::dataType(hXShapeInfo); auto zType = sd::ArrayOptions::dataType(hZShapeInfo); @@ -1158,7 +1152,7 @@ void NativeOpExecutioner::execSummaryStats(sd::LaunchContext *lc, auto stream = lc->getCudaStream(); auto reductionPointer = lc->getReductionPointer(); - dim3 launchDims = dim3(256, 256, 32768); + dim3 launchDims = dim3(256, CUDA_BLOCK_SIZE, 1024); auto xType = sd::ArrayOptions::dataType(hXShapeInfo); auto zType = sd::ArrayOptions::dataType(hZShapeInfo); @@ -1194,9 +1188,9 @@ void NativeOpExecutioner::execReduce3(sd::LaunchContext *lc, auto yType = sd::ArrayOptions::dataType(hYShapeInfo); auto zType = sd::ArrayOptions::dataType(hZShapeInfo); - auto blockWidth = 256; + auto blockWidth = CUDA_BLOCK_SIZE; auto numBlocks = CudaLaunchHelper::getReductionBlocks(shape::length(hXShapeInfo), blockWidth); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 1024); if (xType != yType) throw sd::datatype_exception::build("NativeOpExecutioner::execReduce3 requires Y operand to have X type", xType, yType); @@ -1246,7 +1240,7 @@ void NativeOpExecutioner::execReduce3(sd::LaunchContext *lc, auto numBlocks = shape::length(hZShapeInfo); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, CUDA_BLOCK_SIZE, 1024); BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce3::Reduce3, ::exec(launchDims, stream, opNum, dX, dXShapeInfo, @@ -1286,9 +1280,9 @@ void NativeOpExecutioner::execReduce3Scalar(sd::LaunchContext *lc, auto zType = sd::ArrayOptions::dataType(hZShapeInfo); auto xLength = shape::length(hXShapeInfo); - auto blockWidth = 256; + auto blockWidth = CUDA_BLOCK_SIZE; auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, blockWidth, 1024); if (xType != yType) throw sd::datatype_exception::build("NativeOpExecutioner::execReduce3Scalar requires Y operand to have X type", xType, yType); @@ -1652,7 +1646,7 @@ void NativeOpExecutioner::execReduce3All(sd::LaunchContext *lc, if (sd::Environment::getInstance().isDebugAndVerbose()) printf("D119 opNum:[%i]\n", opNum); - dim3 launchDims(shape::length(hZShapeInfo), 256, 32768); + dim3 launchDims(shape::length(hZShapeInfo), CUDA_BLOCK_SIZE / 2, 1024); if (sd::Environment::getInstance().isVerbose() && launchDims.x == 1) printf("AD119 opNum:[%i]\n", opNum); @@ -1706,7 +1700,7 @@ void NativeOpExecutioner::execReduce3TAD(sd::LaunchContext *lc, throw sd::datatype_exception::build("NativeOpExecutioner::execReduce3TAD requires Z operand to have floating point data type", zType); auto numBlocks = shape::length(hZShapeInfo); - dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, 256, 32768); + dim3 launchDims(numBlocks == 0 ? 1 : numBlocks, CUDA_BLOCK_SIZE, 1024); BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce3::Reduce3, ::exec(launchDims, stream, opNum, dX, dXShapeInfo, dY, dYShapeInfo, extraParams, dZ, dZShapeInfo, dimension, dimensionLength, 1, allocationPointer, tadShapeInfo, tadOffsets, yTadShapeInfo, yTadOffsets), LIBND4J_TYPES, FLOAT_TYPES); diff --git a/libnd4j/include/legacy/cuda/NativeOps.cu b/libnd4j/include/legacy/cuda/NativeOps.cu index 1ccc2c7d5..186b3a9cb 100755 --- a/libnd4j/include/legacy/cuda/NativeOps.cu +++ b/libnd4j/include/legacy/cuda/NativeOps.cu @@ -454,17 +454,24 @@ void execReduceSame2(Nd4jPointer *extraPointers, auto dimension = reinterpret_cast(dbDimension->primary()); int dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, - dimension, - shape::length(hDimensionShape)); + const auto zLen = shape::length(hZShapeInfo); + std::vector dimensions(dimension, dimension + dimensionLength); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo) && zLen != 1) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]); NativeOpExecutioner::execReduceSame(&lc, opNum, dbX->primary(), hXShapeInfo, dbX->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hXShapeInfo).special(), extraParams, - dbZ->primary(), hZShapeInfo, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hZShapeInfo).special(), - dimension, dimensionLength, - tadPack.specialShapeInfo(), tadPack.specialOffsets()); + dbZ->primary(), zShapeInfoH, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(zShapeInfoH).special(), + dims.data(), dims.size()); InteropDataBuffer::registerSpecialUse({dbZ}, {dbX}); } catch (std::exception &e) { @@ -487,17 +494,25 @@ void execReduceLong2(Nd4jPointer *extraPointers, auto dimension = reinterpret_cast(dbDimension->primary()); int dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, - dimension, - shape::length(hDimensionShape)); + const auto zLen = shape::length(hZShapeInfo); + + std::vector dimensions(dimension, dimension + dimensionLength); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo) && zLen != 1) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]); NativeOpExecutioner::execReduceLong(&lc, opNum, dbX->primary(), hXShapeInfo, dbX->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hXShapeInfo).special(), extraParams, - dbZ->primary(), hZShapeInfo, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hZShapeInfo).special(), - dimension, dimensionLength, - tadPack.specialShapeInfo(), tadPack.specialOffsets()); + dbZ->primary(), zShapeInfoH, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(zShapeInfoH).special(), + dims.data(), dims.size()); InteropDataBuffer::registerSpecialUse({dbZ}, {dbX}); } catch (std::exception &e) { @@ -562,17 +577,25 @@ void execReduceBool2(Nd4jPointer *extraPointers, auto dimension = reinterpret_cast(dbDimension->primary()); int dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, - dimension, - shape::length(hDimensionShape)); + const auto zLen = shape::length(hZShapeInfo); + + std::vector dimensions(dimension, dimension + dimensionLength); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo) && zLen != 1) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]); NativeOpExecutioner::execReduceBool(&lc, opNum, dbX->primary(), hXShapeInfo, dbX->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hXShapeInfo).special(), extraParams, - dbZ->primary(), hZShapeInfo, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hZShapeInfo).special(), - dimension, dimensionLength, - tadPack.specialShapeInfo(), tadPack.specialOffsets()); + dbZ->primary(), zShapeInfoH, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(zShapeInfoH).special(), + dims.data(), dims.size()); InteropDataBuffer::registerSpecialUse({dbZ}, {dbX}); } catch (std::exception &e) { @@ -690,17 +713,25 @@ void execReduceFloat2(Nd4jPointer *extraPointers, auto dimension = reinterpret_cast(dbDimension->primary()); int dimensionLength = static_cast(shape::length(hDimensionShape)); - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(hXShapeInfo, - dimension, - shape::length(hDimensionShape)); + const auto zLen = shape::length(hZShapeInfo); + + std::vector dimensions(dimension, dimension + dimensionLength); + + const Nd4jLong* zShapeInfoH = hZShapeInfo; + + if(shape::rank(hXShapeInfo) - dimensionLength != shape::rank(hZShapeInfo) && zLen != 1) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(hZShapeInfo, dimensions); + zShapeInfoH = reinterpret_cast(zPack.primary()); + } + + std::vector dims = (zLen != 1) ? ShapeUtils::evalDimsForReduceOp(shape::rank(hXShapeInfo), dimensions) : std::vector(); LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]); NativeOpExecutioner::execReduceFloat(&lc, opNum, dbX->primary(), hXShapeInfo, dbX->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hXShapeInfo).special(), extraParams, - dbZ->primary(), hZShapeInfo, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(hZShapeInfo).special(), - dimension, dimensionLength, - tadPack.specialShapeInfo(), tadPack.specialOffsets()); + dbZ->primary(), zShapeInfoH, dbZ->special(), ConstantShapeHelper::getInstance().bufferForShapeInfo(zShapeInfoH).special(), + dims.data(), dims.size()); InteropDataBuffer::registerSpecialUse({dbZ}, {dbX}); } catch (std::exception &e) { diff --git a/libnd4j/include/loops/cpu/broadcasting.hpp b/libnd4j/include/loops/cpu/broadcasting.hpp index 4c59de0ec..51b76df73 100644 --- a/libnd4j/include/loops/cpu/broadcasting.hpp +++ b/libnd4j/include/loops/cpu/broadcasting.hpp @@ -784,24 +784,14 @@ static void execDefault(const X *x, const Nd4jLong *xShapeInfo, const Y *y, cons const bool xzSameOffsets = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo); const bool yzSameOffsets = shape::haveSameShapeAndStrides(yShapeInfo, zShapeInfo); - const int rank = shape::rank(zShapeInfo); // xRank = yRank = zRank - auto func = PRAGMA_THREADS_FOR{ - int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK]; + int coords[MAX_RANK]; + Nd4jLong xOffset, yOffset, zOffset; for (auto i = start; i < stop; ++i) { - shape::index2coordsCPU(start, i, zShapeInfo, zCoords); - - for (uint j = 0; j < rank; ++j) { - xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j]; - yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j]; - } - - const auto zOffset = shape::getOffset(zShapeInfo, zCoords); - const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords); - const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords); + shape::getOffsetBroadcast(start, i, zShapeInfo, xShapeInfo, yShapeInfo, xzSameOffsets, yzSameOffsets, coords, zOffset, xOffset, yOffset); z[zOffset] = OpType::op(x[xOffset], y[yOffset]); } diff --git a/libnd4j/include/loops/cpu/broadcasting_bool.hpp b/libnd4j/include/loops/cpu/broadcasting_bool.hpp index a15935124..22f30e40a 100644 --- a/libnd4j/include/loops/cpu/broadcasting_bool.hpp +++ b/libnd4j/include/loops/cpu/broadcasting_bool.hpp @@ -665,24 +665,14 @@ static void execDefault(const X *x, const Nd4jLong *xShapeInfo, const X *y, cons const bool xzSameOffsets = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo); const bool yzSameOffsets = shape::haveSameShapeAndStrides(yShapeInfo, zShapeInfo); - const int rank = shape::rank(zShapeInfo); // xRank = yRank = zRank - auto func = PRAGMA_THREADS_FOR{ - int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK]; + int coords[MAX_RANK]; + Nd4jLong xOffset, yOffset, zOffset; for (auto i = start; i < stop; ++i) { - shape::index2coordsCPU(start, i, zShapeInfo, zCoords); - - for (uint j = 0; j < rank; ++j) { - xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j]; - yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j]; - } - - const auto zOffset = shape::getOffset(zShapeInfo, zCoords); - const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords); - const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords); + shape::getOffsetBroadcast(start, i, zShapeInfo, xShapeInfo, yShapeInfo, xzSameOffsets, yzSameOffsets, coords, zOffset, xOffset, yOffset); z[zOffset] = OpType::op(x[xOffset], y[yOffset], extraParams); } diff --git a/libnd4j/include/loops/cpu/broadcasting_int.hpp b/libnd4j/include/loops/cpu/broadcasting_int.hpp index 39b251594..5b95d963f 100644 --- a/libnd4j/include/loops/cpu/broadcasting_int.hpp +++ b/libnd4j/include/loops/cpu/broadcasting_int.hpp @@ -651,24 +651,14 @@ static void execDefault(const X *x, const Nd4jLong *xShapeInfo, const X *y, cons const bool xzSameOffsets = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo); const bool yzSameOffsets = shape::haveSameShapeAndStrides(yShapeInfo, zShapeInfo); - const int rank = shape::rank(zShapeInfo); // xRank = yRank = zRank - auto func = PRAGMA_THREADS_FOR{ - int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK]; + int coords[MAX_RANK]; + Nd4jLong xOffset, yOffset, zOffset; for (auto i = start; i < stop; ++i) { - shape::index2coordsCPU(start, i, zShapeInfo, zCoords); - - for (uint j = 0; j < rank; ++j) { - xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j]; - yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j]; - } - - const auto zOffset = shape::getOffset(zShapeInfo, zCoords); - const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords); - const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords); + shape::getOffsetBroadcast(start, i, zShapeInfo, xShapeInfo, yShapeInfo, xzSameOffsets, yzSameOffsets, coords, zOffset, xOffset, yOffset); z[zOffset] = OpType::op(x[xOffset], y[yOffset]); } diff --git a/libnd4j/include/loops/cpu/reduce/reduce_bool.cpp b/libnd4j/include/loops/cpu/reduce/reduce_bool.cpp index 94e156705..754002eaa 100644 --- a/libnd4j/include/loops/cpu/reduce/reduce_bool.cpp +++ b/libnd4j/include/loops/cpu/reduce/reduce_bool.cpp @@ -114,71 +114,6 @@ namespace functions { DISPATCH_BY_OPNUM_TT(execScalar, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo), REDUCE_BOOL_OPS); } - template - void ReduceBoolFunction::exec(const int opNum, - const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop) { - DISPATCH_BY_OPNUM_TT(exec, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffset, start, stop), REDUCE_BOOL_OPS); - } - - template - template - void _CUDA_H ReduceBoolFunction::exec(const void *vx, const Nd4jLong *xShapeInfo, - void *vextraParams, - void *vresult, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, int64_t start, int64_t stop) { - - auto x = reinterpret_cast(vx); - auto z = reinterpret_cast(vresult); - auto extraParams = reinterpret_cast(vextraParams); - - auto resultLength = shape::length(zShapeInfo); - - if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { - if(sd::ArrayOptions::arrayType(zShapeInfo) == sd::ArrayType::EMPTY) - return; - const auto startingVal = OpType::startingValue(x); - - for (Nd4jLong i = 0; i < resultLength; i++) - z[i] = startingVal; - return; - } - - //pre squeezed: this is for keeping the pointer to the original - //shape information for tad offset - //the squeezed information doesn't render the right strides for - //tad offset - // || tad.wholeThing - if (resultLength == 1 || dimension == nullptr || dimensionLength == shape::rank(xShapeInfo)) { - z[0] = execScalar(x, xShapeInfo, extraParams); - return; - } - - auto tadOnlyShapeInfo = tadShapeInfo; - auto tadOffsets = tadOffset; - - if (tadOnlyShapeInfo == nullptr || tadOffsets == nullptr) { - if (dimensionLength < 1) - return; - - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(xShapeInfo, dimension, dimensionLength); - tadOnlyShapeInfo = tadPack.primaryShapeInfo(); - tadOffsets = tadPack.primaryOffsets(); - } - -#ifdef INLINE_LOOPS - sd::ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#else - sd::ReductionBoolLoops::template innerloopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#endif - } - - template template void _CUDA_H ReduceBoolFunction::exec(const void *x, const Nd4jLong *xShapeInfo, @@ -220,7 +155,51 @@ namespace functions { return OpType::postProcess(intermediate[0], length, extraParams); } +//////////////////////////////////////////////////////////////////////// +template +template +void _CUDA_H ReduceBoolFunction::exec(sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int* dims) { - BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReduceBoolFunction, , LIBND4J_TYPES, BOOL_TYPES); + const X* x = reinterpret_cast(vx); + Z* z = reinterpret_cast(vz); + X* extraParams = reinterpret_cast(vextraParams); + + const int xRank = shape::rank(xShapeInfo); + const int zRank = shape::rank(zShapeInfo); + + if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { + + const auto startingVal = OpType::startingValue(x); + const auto zLen = shape::length(zShapeInfo); + + for (Nd4jLong i = 0; i < zLen; i++) + z[i] = startingVal; + return; } -} \ No newline at end of file + + if (shape::length(zShapeInfo) == 1) { + z[0] = execScalar(x, xShapeInfo, extraParams); + return; + } + + +#ifdef INLINE_LOOPS + sd::ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#else + sd::ReductionBoolLoops::template innerloopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#endif + +} + +//////////////////////////////////////////////////////////////////////// +template +void ReduceBoolFunction::exec(const int opNum, sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int *dims) { + + DISPATCH_BY_OPNUM_TT(exec, PARAMS(workspace, vx, xShapeInfo, vextraParams, vz, zShapeInfo, dims), REDUCE_BOOL_OPS); +} + + +BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReduceBoolFunction, , LIBND4J_TYPES, BOOL_TYPES); +} +} + diff --git a/libnd4j/include/loops/cpu/reduce/reduce_float.hpp b/libnd4j/include/loops/cpu/reduce/reduce_float.hpp index 6be93b1c4..352fa2200 100644 --- a/libnd4j/include/loops/cpu/reduce/reduce_float.hpp +++ b/libnd4j/include/loops/cpu/reduce/reduce_float.hpp @@ -26,11 +26,13 @@ #include #include #include +#include using namespace simdOps; namespace functions { - namespace reduce { +namespace reduce { + template template void _CUDA_H ReduceFloatFunction::execScalar(const void *vx, const Nd4jLong *xShapeInfo, @@ -133,86 +135,6 @@ namespace functions { DISPATCH_BY_OPNUM_TT(execScalar, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo), REDUCE_FLOAT_OPS); } - template - void ReduceFloatFunction::exec(const int opNum, - const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop) { - DISPATCH_BY_OPNUM_TT(exec, PARAMS(x, - xShapeInfo, - extraParams, - z, - zShapeInfo, - dimension, - dimensionLength, - tadShapeInfo, - tadOffset, start, stop), - REDUCE_FLOAT_OPS); - } - - template - template - void _CUDA_H ReduceFloatFunction::exec(const void *vx, const Nd4jLong *xShapeInfo, - void *vextraParams, - void *vresult, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop) { - - auto x = reinterpret_cast(vx); - auto z = reinterpret_cast(vresult); - auto extraParams = reinterpret_cast(vextraParams); - - auto resultLength = shape::length(zShapeInfo); - - if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { - if(sd::ArrayOptions::arrayType(zShapeInfo) == sd::ArrayType::EMPTY) - return; - const auto startingVal = std::is_same>::value ? sd::DataTypeUtils::nanOrZero() : static_cast(OpType::startingValue(x)); - - for (Nd4jLong i = 0; i < resultLength; i++) - z[i] = startingVal; - return; - } - - //pre squeezed: this is for keeping the pointer to the original - //shape information for tad offset - //the squeezed information doesn't render the right strides for - //tad offset - // || tad.wholeThing - if (resultLength == 1 || dimension == nullptr || dimensionLength == shape::rank(xShapeInfo)) { - z[0] = execScalar(x, xShapeInfo, extraParams); - return; - } - - if (OpType::requiresSpecialAccumulation) { - OpType::execSpecial(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffset); - return; - } - - auto tadOnlyShapeInfo = tadShapeInfo; - auto tadOffsets = tadOffset; - - if (tadOnlyShapeInfo == nullptr || tadOffsets == nullptr) { - if (dimensionLength < 0) - return; - - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(xShapeInfo, dimension, dimensionLength); - tadOnlyShapeInfo = tadPack.primaryShapeInfo(); - tadOffsets = tadPack.primaryOffsets(); - } - -#ifdef INLINE_LOOPS - sd::ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#else - sd::ReductionFloatLoops::template innerloopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#endif - } - - template template void _CUDA_H ReduceFloatFunction::exec(const void *x, const Nd4jLong *xShapeInfo, @@ -255,5 +177,54 @@ namespace functions { // return result return OpType::postProcess(intermediate[0], length, extraParams); } + + +//////////////////////////////////////////////////////////////////////// +template +template +void _CUDA_H ReduceFloatFunction::exec(sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int* dims) { + + const X* x = reinterpret_cast(vx); + Z* z = reinterpret_cast(vz); + Z* extraParams = reinterpret_cast(vextraParams); + + const int xRank = shape::rank(xShapeInfo); + const int zRank = shape::rank(zShapeInfo); + + if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { + + const auto startingVal = std::is_same>::value ? sd::DataTypeUtils::nanOrZero() : static_cast(OpType::startingValue(x)); + const auto zLen = shape::length(zShapeInfo); + + for (Nd4jLong i = 0; i < zLen; i++) + z[i] = startingVal; + return; } + + if (shape::length(zShapeInfo) == 1) { + z[0] = execScalar(x, xShapeInfo, extraParams); + return; + } + + if (OpType::requiresSpecialAccumulation) { + OpType::execSpecial(x, xShapeInfo, extraParams, z, zShapeInfo, const_cast(dims)+zRank, xRank-zRank, nullptr, nullptr); + return; + } + +#ifdef INLINE_LOOPS + sd::ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#else + sd::ReductionFloatLoops::template innerloopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#endif + +} + +//////////////////////////////////////////////////////////////////////// +template +void ReduceFloatFunction::exec(const int opNum, sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int *dims) { + + DISPATCH_BY_OPNUM_TT(exec, PARAMS(workspace, vx, xShapeInfo, vextraParams, vz, zShapeInfo, dims), REDUCE_FLOAT_OPS); +} + +} } \ No newline at end of file diff --git a/libnd4j/include/loops/cpu/reduce/reduce_long.cpp b/libnd4j/include/loops/cpu/reduce/reduce_long.cpp index a4fae3228..e908f1fc7 100644 --- a/libnd4j/include/loops/cpu/reduce/reduce_long.cpp +++ b/libnd4j/include/loops/cpu/reduce/reduce_long.cpp @@ -129,76 +129,6 @@ namespace functions { DISPATCH_BY_OPNUM_TT(execScalar, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo), REDUCE_LONG_OPS); } - template - void ReduceLongFunction::exec(const int opNum, - const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop) { - DISPATCH_BY_OPNUM_TT(exec, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffset, start, stop), REDUCE_LONG_OPS); - } - - template - template - void _CUDA_H ReduceLongFunction::exec(const void *vx, const Nd4jLong *xShapeInfo, - void *vextraParams, - void *vresult, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop) { - - auto x = reinterpret_cast(vx); - auto z = reinterpret_cast(vresult); - auto extraParams = reinterpret_cast(vextraParams); - - auto resultLength = shape::length(zShapeInfo); - - if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { - if(sd::ArrayOptions::arrayType(zShapeInfo) == sd::ArrayType::EMPTY) - return; - const auto startingVal = OpType::startingValue(x); - - for (Nd4jLong i = 0; i < resultLength; i++) - z[i] = startingVal; - return; - } - - //pre squeezed: this is for keeping the pointer to the original - //shape information for tad offset - //the squeezed information doesn't render the right strides for - //tad offset - // || tad.wholeThing - if (resultLength == 1 || dimension == nullptr || dimensionLength == shape::rank(xShapeInfo)) { - z[0] = execScalar(x, xShapeInfo, extraParams); - return; - } - - if (OpType::requiresSpecialAccumulation) { - OpType::execSpecial(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffset); - return; - } - - auto tadOnlyShapeInfo = tadShapeInfo; - auto tadOffsets = tadOffset; - - if (tadOnlyShapeInfo == nullptr || tadOffsets == nullptr) { - if (dimensionLength < 1) - return; - - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(xShapeInfo, dimension, dimensionLength); - tadOnlyShapeInfo = tadPack.primaryShapeInfo(); - tadOffsets = tadPack.primaryOffsets(); - } - -#ifdef INLINE_LOOPS - sd::ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#else - sd::ReductionLongLoops::template innerloopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#endif - } - template template @@ -243,6 +173,56 @@ namespace functions { } - BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReduceLongFunction, , LIBND4J_TYPES, LONG_TYPES); +//////////////////////////////////////////////////////////////////////// +template +template +void _CUDA_H ReduceLongFunction::exec(sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int* dims) { + + const X* x = reinterpret_cast(vx); + Z* z = reinterpret_cast(vz); + X* extraParams = reinterpret_cast(vextraParams); + + const int xRank = shape::rank(xShapeInfo); + const int zRank = shape::rank(zShapeInfo); + + if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { + + const auto startingVal = OpType::startingValue(x); + const auto zLen = shape::length(zShapeInfo); + + for (Nd4jLong i = 0; i < zLen; i++) + z[i] = startingVal; + return; } -} \ No newline at end of file + + if (shape::length(zShapeInfo) == 1) { + z[0] = execScalar(x, xShapeInfo, extraParams); + return; + } + + if (OpType::requiresSpecialAccumulation) { + OpType::execSpecial(x, xShapeInfo, extraParams, z, zShapeInfo, const_cast(dims)+zRank, xRank-zRank, nullptr, nullptr); + return; + } + +#ifdef INLINE_LOOPS + sd::ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#else + sd::ReductionLongLoops::template innerloopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#endif + +} + +//////////////////////////////////////////////////////////////////////// +template +void ReduceLongFunction::exec(const int opNum, sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int *dims) { + + DISPATCH_BY_OPNUM_TT(exec, PARAMS(workspace, vx, xShapeInfo, vextraParams, vz, zShapeInfo, dims), REDUCE_LONG_OPS); +} + + + +BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReduceLongFunction, , LIBND4J_TYPES, LONG_TYPES); +} +} + diff --git a/libnd4j/include/loops/cpu/reduce/reduce_same.cpp b/libnd4j/include/loops/cpu/reduce/reduce_same.cpp index 10607fb6d..78cce19a0 100644 --- a/libnd4j/include/loops/cpu/reduce/reduce_same.cpp +++ b/libnd4j/include/loops/cpu/reduce/reduce_same.cpp @@ -129,85 +129,6 @@ namespace functions { DISPATCH_BY_OPNUM_T(execScalar, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo), REDUCE_SAME_OPS); } - template - void ReduceSameFunction::exec(const int opNum, - const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop) { - DISPATCH_BY_OPNUM_T(exec, PARAMS(x, - xShapeInfo, - extraParams, - z, - zShapeInfo, - dimension, - dimensionLength, - tadShapeInfo, - tadOffset, start, stop), - REDUCE_SAME_OPS); - } - - template - template - void _CUDA_H ReduceSameFunction::exec(const void *vx, const Nd4jLong *xShapeInfo, - void *vextraParams, - void *vz, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop) { - - auto x = reinterpret_cast(vx); - auto z = reinterpret_cast(vz); - auto extraParams = reinterpret_cast(vextraParams); - - auto zLength = shape::length(zShapeInfo); - - if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { - if(sd::ArrayOptions::arrayType(zShapeInfo) == sd::ArrayType::EMPTY) - return; - const auto startingVal = OpType::startingValue(x); - - for (Nd4jLong i = 0; i < zLength; i++) - z[i] = startingVal; - return; - } - - //pre squeezed: this is for keeping the pointer to the original - //shape information for tad offset - //the squeezed information doesn't render the right strides for - //tad offset - // || tad.wholeThing - if (zLength == 1 || dimension == nullptr || dimensionLength == shape::rank(xShapeInfo)) { - z[0] = execScalar(x, xShapeInfo, extraParams); - return; - } - - if (OpType::requiresSpecialAccumulation) { - OpType::execSpecial(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffset); - return; - } - - auto tadOnlyShapeInfo = tadShapeInfo; - auto tadOffsets = tadOffset; - - if (tadOnlyShapeInfo == nullptr || tadOffsets == nullptr) { - if (dimensionLength < 1) - return; - - auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(xShapeInfo, dimension, dimensionLength); - tadOnlyShapeInfo = tadPack.primaryShapeInfo(); - tadOffsets = tadPack.primaryOffsets(); - } - -#ifdef INLINE_LOOPS - sd::ReductionLoops::template loopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#else - sd::ReductionSameLoops::template innerloopReduce(x, xShapeInfo, z, zShapeInfo, tadOnlyShapeInfo, tadOffsets, extraParams, start, stop); -#endif - } - template template @@ -251,7 +172,57 @@ namespace functions { return OpType::postProcess(intermediate[0], length, extraParams); } +//////////////////////////////////////////////////////////////////////// +template +template +void _CUDA_H ReduceSameFunction::exec(sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int* dims) { - BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT ReduceSameFunction, , LIBND4J_TYPES); + const X* x = reinterpret_cast(vx); + X* z = reinterpret_cast(vz); + X* extraParams = reinterpret_cast(vextraParams); + + const int xRank = shape::rank(xShapeInfo); + const int zRank = shape::rank(zShapeInfo); + + if(sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) { + + const auto startingVal = OpType::startingValue(x); + const auto zLen = shape::length(zShapeInfo); + + for (Nd4jLong i = 0; i < zLen; i++) + z[i] = startingVal; + return; } -} \ No newline at end of file + + if (shape::length(zShapeInfo) == 1) { + z[0] = execScalar(x, xShapeInfo, extraParams); + return; + } + + if (OpType::requiresSpecialAccumulation) { + OpType::execSpecial(x, xShapeInfo, extraParams, z, zShapeInfo, const_cast(dims)+zRank, xRank-zRank, nullptr, nullptr); + return; + } + +#ifdef INLINE_LOOPS + sd::ReductionLoops::template loopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#else + sd::ReductionSameLoops::template innerloopReduce(workspace, x, xShapeInfo, z, zShapeInfo, dims, extraParams); +#endif + +} + +//////////////////////////////////////////////////////////////////////// +template +void ReduceSameFunction::exec(const int opNum, sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, void *vextraParams, void *vz, const Nd4jLong *zShapeInfo, const int *dims) { + + DISPATCH_BY_OPNUM_T(exec, PARAMS(workspace, vx, xShapeInfo, vextraParams, vz, zShapeInfo, dims), REDUCE_SAME_OPS); +} + + + +BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT ReduceSameFunction, , LIBND4J_TYPES); +} +} + + diff --git a/libnd4j/include/loops/cuda/broadcasting.chpp b/libnd4j/include/loops/cuda/broadcasting.chpp index 4b5c7833f..80db91782 100644 --- a/libnd4j/include/loops/cuda/broadcasting.chpp +++ b/libnd4j/include/loops/cuda/broadcasting.chpp @@ -279,20 +279,15 @@ __device__ void Broadcast::transformCuda( const auto tid = blockIdx.x * blockDim.x + threadIdx.x; - int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK]; + int coords[MAX_RANK]; for (int i = tid; i < zLen; i += blockDim.x * gridDim.x) { - shape::index2coords(i, zShapeInfo, zCoords); + shape::index2coords(i, zShapeInfo, coords); - for (uint j = 0; j < rank; ++j) { - xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j]; - yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j]; - } - - const auto zOffset = shape::getOffset(zShapeInfo, zCoords); - const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords); - const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords); + const auto zOffset = shape::getOffset(zShapeInfo, coords); + const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, coords); + const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, coords); z[zOffset] = OpType::op(x[xOffset], y[yOffset]); } diff --git a/libnd4j/include/loops/cuda/broadcasting_bool.cu b/libnd4j/include/loops/cuda/broadcasting_bool.cu index bed00a20f..b2d94def8 100644 --- a/libnd4j/include/loops/cuda/broadcasting_bool.cu +++ b/libnd4j/include/loops/cuda/broadcasting_bool.cu @@ -295,20 +295,15 @@ __device__ void BroadcastBool::transformCuda(const void *vx, const Nd4jLong const auto tid = blockIdx.x * blockDim.x + threadIdx.x; - int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK]; + int coords[MAX_RANK]; for (int i = tid; i < zLen; i += blockDim.x * gridDim.x) { - shape::index2coords(i, zShapeInfo, zCoords); + shape::index2coords(i, zShapeInfo, coords); - for (uint j = 0; j < rank; ++j) { - xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j]; - yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j]; - } - - const auto zOffset = shape::getOffset(zShapeInfo, zCoords); - const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords); - const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords); + const auto zOffset = shape::getOffset(zShapeInfo, coords); + const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, coords); + const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, coords); z[zOffset] = OpType::op(x[xOffset], y[yOffset], extraParams); } diff --git a/libnd4j/include/loops/cuda/broadcasting_int.cu b/libnd4j/include/loops/cuda/broadcasting_int.cu index 37cbf3eba..1d3c0375a 100644 --- a/libnd4j/include/loops/cuda/broadcasting_int.cu +++ b/libnd4j/include/loops/cuda/broadcasting_int.cu @@ -275,20 +275,15 @@ __device__ void BroadcastInt::transformCuda(const void *vx, const Nd4jLong co const auto tid = blockIdx.x * blockDim.x + threadIdx.x; - int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK]; + int coords[MAX_RANK]; for (int i = tid; i < zLen; i += blockDim.x * gridDim.x) { - shape::index2coords(i, zShapeInfo, zCoords); + shape::index2coords(i, zShapeInfo, coords); - for (uint j = 0; j < rank; ++j) { - xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j]; - yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j]; - } - - const auto zOffset = shape::getOffset(zShapeInfo, zCoords); - const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords); - const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords); + const auto zOffset = shape::getOffset(zShapeInfo, coords); + const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, coords); + const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, coords); z[zOffset] = OpType::op(x[xOffset], y[yOffset]); } diff --git a/libnd4j/include/loops/cuda/indexreduce.cu b/libnd4j/include/loops/cuda/indexreduce.cu index dbe03a9bf..23f0be7cd 100644 --- a/libnd4j/include/loops/cuda/indexreduce.cu +++ b/libnd4j/include/loops/cuda/indexreduce.cu @@ -120,12 +120,11 @@ namespace functions { template template - __device__ void IndexReduce::aggregatePartials(IndexValue **sPartialsRef, Nd4jLong tid, Nd4jLong numElements, void *vextraParams) { + __device__ void IndexReduce::aggregatePartials(IndexValue *sPartials, Nd4jLong tid, Nd4jLong numElements, void *vextraParams) { // start the shared memory loop on the next power of 2 less // than the block size. If block size is not a power of 2, // accumulate the intermediate sums in the remainder range. auto extraParams = static_cast(vextraParams); - IndexValue *sPartials = *sPartialsRef; Nd4jLong floorPow2 = blockDim.x; if (floorPow2 & (floorPow2 - 1)) { @@ -191,12 +190,7 @@ namespace functions { __shared__ volatile bool resultScalar; //shared memory space for storing intermediate results - __shared__ IndexValue* sPartials; - if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast*>(shmem); - } - __syncthreads(); + __shared__ IndexValue sPartials[CUDA_BLOCK_SIZE]; sPartials[threadIdx.x] = OpType::startingIndexValue(dx); @@ -261,7 +255,7 @@ namespace functions { } __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength),extraParams); + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength),extraParams); __syncthreads(); if (threadIdx.x == 0) { @@ -282,7 +276,7 @@ namespace functions { } __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength),extraParams); + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength),extraParams); __syncthreads(); if (threadIdx.x == 0) { @@ -313,7 +307,7 @@ namespace functions { sPartials[threadIdx.x] = reduction; __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, (int) n),extraParams); + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, (int) n),extraParams); __syncthreads(); if (gridDim.x > 1) { @@ -345,7 +339,7 @@ namespace functions { } __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, sd::math::nd4j_min(gridDim.x, blockDim.x),extraParams); + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(gridDim.x, blockDim.x),extraParams); __syncthreads(); if (tid == 0) { diff --git a/libnd4j/include/loops/cuda/reduce/reduce_bool.cu b/libnd4j/include/loops/cuda/reduce/reduce_bool.cu index de854416d..0c81334a6 100644 --- a/libnd4j/include/loops/cuda/reduce/reduce_bool.cu +++ b/libnd4j/include/loops/cuda/reduce/reduce_bool.cu @@ -33,14 +33,10 @@ using namespace simdOps; //////////////////////////////////////////////////////////////////////// template -__global__ void simpleReduce(const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - void *reductionBuffer, - const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets) { +__global__ void simpleReduce(const void *x, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *extraParams, void *vreductionBuffer, void *z, const Nd4jLong *zShapeInfo) { - functions::reduce::ReduceBoolFunction::template transformCudaXD(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo, tadOffsets); + functions::reduce::ReduceBoolFunction::template transformCudaXD(x, outerXTadShapeInfo, innerXTadShapeInfo, vreductionBuffer, extraParams, z, zShapeInfo); } //////////////////////////////////////////////////////////////////////// @@ -55,7 +51,6 @@ __global__ void simpleScalar(const void *x, const Nd4jLong *xShapeInfo, functions::reduce::ReduceBoolFunction::template execScalarCuda(x, xShapeInfo, extraParams, z, zShapeInfo, reductionBuffer, tadOnlyShapeInfo); } - namespace functions { namespace reduce { @@ -95,53 +90,49 @@ __device__ void ReduceBoolFunction::aggregatePartials(void *vsPartials, Nd4 //////////////////////////////////////////////////////////////////////// template template -__device__ void ReduceBoolFunction::transformCudaXD(const void *vx, const Nd4jLong *xShapeInfo, - void *vextraParams, - void *vz, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - void *vreductionBuffer, - const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets) { +__device__ void ReduceBoolFunction::transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *vextraParams, void *vreductionBuffer, + void *vz, const Nd4jLong *zShapeInfo) { auto x = reinterpret_cast(vx); auto z = reinterpret_cast(vz); auto extraParams = reinterpret_cast(vextraParams); - auto reductionBuffer = reinterpret_cast(vreductionBuffer); //shared memory space for storing intermediate results - __shared__ Z* sPartials; - __shared__ int tadLength, numTads; - __shared__ bool isPlainOutput; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; + __shared__ int tadLen, numTads; + __shared__ bool sameOffsets; if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); + sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo); - isPlainOutput = shape::order(zShapeInfo) == 'c' && shape::elementWiseStride(zShapeInfo) == 1; - - tadLength = shape::length(tadOnlyShapeInfo); //tadLength(xShapeInfo, dimension, dimensionLength); - numTads = shape::length(xShapeInfo) / tadLength; + tadLen = shape::length(innerXTadShapeInfo); + numTads = shape::length(outerXTadShapeInfo); } __syncthreads(); + int coords[MAX_RANK]; + for (int r = blockIdx.x; r < numTads; r += gridDim.x) { - Nd4jLong tadOffsetForBlock = tadOffsets[r]; - sPartials[threadIdx.x] = OpType::startingValue(x + tadOffsetForBlock); + shape::index2coords(r, outerXTadShapeInfo, coords); + const auto outerOffset = shape::getOffset(outerXTadShapeInfo, coords); + const auto zOffset = sameOffsets ? outerOffset : shape::getOffset(zShapeInfo, coords); - for (int i = threadIdx.x; i < tadLength; i += blockDim.x) { + const X* xTad = x + outerOffset; + sPartials[threadIdx.x] = OpType::startingValue(xTad); - auto xOffset = tadOffsetForBlock + shape::getIndexOffset(i, tadOnlyShapeInfo); - sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(x[xOffset], extraParams), extraParams); - } + for (int i = threadIdx.x; i < tadLen; i += blockDim.x) + sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(xTad[shape::getIndexOffset(i, innerXTadShapeInfo)], extraParams), extraParams); __syncthreads(); - // aggregate. do NOT reduce for elements > tadLength - aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); + // aggregate. do NOT reduce for elements > tadLen + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLen), extraParams); __syncthreads(); if (threadIdx.x == 0) - z[isPlainOutput ? r : shape::getIndexOffset(r, zShapeInfo)] = OpType::postProcess(sPartials[threadIdx.x], tadLength, extraParams); + z[zOffset] = OpType::postProcess(sPartials[threadIdx.x], tadLen, extraParams); } } @@ -162,13 +153,11 @@ __device__ void ReduceBoolFunction::execScalarCuda(const void *vx, const Nd auto tid = blockDim.x * blockIdx.x + threadIdx.x; //shared memory space for storing intermediate results - __shared__ Z* sPartials; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; __shared__ Nd4jLong xEws; __shared__ Nd4jLong len; if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); xEws = shape::elementWiseStride(xShapeInfo); len = shape::length(xShapeInfo); } @@ -237,12 +226,9 @@ __device__ void ReduceBoolFunction::execScalarCuda(const void *vx, const Nd template template __host__ void ReduceBoolFunction::intermediateXD(dim3 launchDims, cudaStream_t *stream, - const void *x, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, - int *dimension, int dimensionLength, - void *reductionPointer, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims) { if(shape::isEmpty(hXShapeInfo)) { if(shape::isEmpty(hZShapeInfo)) @@ -257,11 +243,17 @@ __host__ void ReduceBoolFunction::intermediateXD(dim3 launchDims, cudaStrea auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer(); // scalar assign - functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, zShapeInfo, hZShapeInfo, z, zShapeInfo, hZShapeInfo, ptr, nullptr); + functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, dZShapeInfo, hZShapeInfo, z, dZShapeInfo, hZShapeInfo, ptr, nullptr); sd::DebugHelper::checkErrorCode(stream, "reduceBoolDim empty(...) failed"); } else { - simpleReduce<<>>(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets); + const int zRank = shape::rank(hZShapeInfo); + const int tadRank = shape::rank(hXShapeInfo) - zRank; + + auto outerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims, zRank); + auto innerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims+zRank, tadRank); + + simpleReduce<<>>(x, reinterpret_cast(outerPack.special()), reinterpret_cast(innerPack.special()), extraParams, vreductionBuffer, z, dZShapeInfo); sd::DebugHelper::checkErrorCode(stream, "reduceBoolDim(...) failed"); } } @@ -314,16 +306,16 @@ _CUDA_H void ReduceBoolFunction::execReduceScalar(dim3 launchDims, cudaStre //////////////////////////////////////////////////////////////////////// template -_CUDA_H void ReduceBoolFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, - const int opNum, - const int rank, const void *x, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, - int *dimension, int dimensionLength, - void *reductionPointer, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { - - DISPATCH_BY_OPNUM_TT(intermediateXD, PARAMS(launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z, zShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_BOOL_OPS)); +_CUDA_H void ReduceBoolFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, const int opNum, + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims) { + if(shape::length(hZShapeInfo) == 1) { + ReduceBoolFunction::execReduceScalar(launchDims, stream, opNum, x, dXShapeInfo, hXShapeInfo, extraParams, z, dZShapeInfo, hZShapeInfo, nullptr, 0, vreductionBuffer, nullptr); + } + else { + DISPATCH_BY_OPNUM_TT(intermediateXD, PARAMS(launchDims, stream, x, dXShapeInfo, hXShapeInfo, extraParams, vreductionBuffer, z, dZShapeInfo, hZShapeInfo, dims), OPS_A(REDUCE_BOOL_OPS)); + } DEBUG_KERNEL(stream, opNum); } diff --git a/libnd4j/include/loops/cuda/reduce/reduce_float.chpp b/libnd4j/include/loops/cuda/reduce/reduce_float.chpp index 71f5d03da..d4882d6c0 100644 --- a/libnd4j/include/loops/cuda/reduce/reduce_float.chpp +++ b/libnd4j/include/loops/cuda/reduce/reduce_float.chpp @@ -35,14 +35,10 @@ using namespace simdOps; //////////////////////////////////////////////////////////////////////// template -__global__ void simpleReduce(const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - void *reductionBuffer, - const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets) { +__global__ void simpleReduce(const void *x, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *extraParams, void *vreductionBuffer, void *z, const Nd4jLong *zShapeInfo) { - functions::reduce::ReduceFloatFunction::template transformCudaXD(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo, tadOffsets); + functions::reduce::ReduceFloatFunction::template transformCudaXD(x, outerXTadShapeInfo, innerXTadShapeInfo, extraParams, vreductionBuffer, z, zShapeInfo); } //////////////////////////////////////////////////////////////////////// @@ -96,51 +92,48 @@ __device__ void ReduceFloatFunction::aggregatePartials(void *vsPartials, Nd //////////////////////////////////////////////////////////////////////// template template -__device__ void ReduceFloatFunction::transformCudaXD(const void *vx, const Nd4jLong *xShapeInfo, - void *vextraParams, - void *vz, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - void *vreductionBuffer, - const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets) { +__device__ void ReduceFloatFunction::transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *vextraParams, void *vreductionBuffer, + void *vz, const Nd4jLong *zShapeInfo) { auto x = reinterpret_cast(vx); auto z = reinterpret_cast(vz); auto extraParams = reinterpret_cast(vextraParams); - auto reductionBuffer = reinterpret_cast(vreductionBuffer); //shared memory space for storing intermediate results - __shared__ Z* sPartials; - __shared__ int tadLength, numTads; - __shared__ bool isPlainOutput; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; + __shared__ int tadLen, numTads; + __shared__ bool sameOffsets; if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); + sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo); - isPlainOutput = shape::order(zShapeInfo) == 'c' && shape::elementWiseStride(zShapeInfo) == 1; - - tadLength = shape::length(tadOnlyShapeInfo); - numTads = shape::length(xShapeInfo) / tadLength; + tadLen = shape::length(innerXTadShapeInfo); + numTads = shape::length(outerXTadShapeInfo); } __syncthreads(); + int coords[MAX_RANK]; + for (int r = blockIdx.x; r < numTads; r += gridDim.x) { - auto tadOffsetForBlock = tadOffsets[r]; - sPartials[threadIdx.x] = OpType::startingValue(x + tadOffsetForBlock); + shape::index2coords(r, outerXTadShapeInfo, coords); + const auto outerOffset = shape::getOffset(outerXTadShapeInfo, coords); + const auto zOffset = sameOffsets ? outerOffset : shape::getOffset(zShapeInfo, coords); - for (int i = threadIdx.x; i < tadLength; i += blockDim.x) { - auto xOffset = tadOffsetForBlock + shape::getIndexOffset(i, tadOnlyShapeInfo); - sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(x[xOffset], extraParams), extraParams); - } + const X* xTad = x + outerOffset; + sPartials[threadIdx.x] = OpType::startingValue(xTad); + + for (int i = threadIdx.x; i < tadLen; i += blockDim.x) + sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(xTad[shape::getIndexOffset(i, innerXTadShapeInfo)], extraParams), extraParams); __syncthreads(); - // aggregate. do NOT reduce for elements > tadLength - aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); + // aggregate. do NOT reduce for elements > tadLen + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLen), extraParams); __syncthreads(); if (threadIdx.x == 0) - z[isPlainOutput ? r : shape::getIndexOffset(r, zShapeInfo)] = OpType::postProcess(sPartials[threadIdx.x], tadLength, extraParams); + z[zOffset] = OpType::postProcess(sPartials[threadIdx.x], tadLen, extraParams); } } @@ -161,13 +154,11 @@ __device__ void ReduceFloatFunction::execScalarCuda(const void *vx, const N auto tid = blockDim.x * blockIdx.x + threadIdx.x; //shared memory space for storing intermediate results - __shared__ Z* sPartials; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; __shared__ Nd4jLong xEws; __shared__ Nd4jLong len; if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); xEws = shape::elementWiseStride(xShapeInfo); len = shape::length(xShapeInfo); } @@ -236,12 +227,9 @@ __device__ void ReduceFloatFunction::execScalarCuda(const void *vx, const N template template __host__ void ReduceFloatFunction::intermediateXD(dim3 launchDims, cudaStream_t *stream, - const void *x, const Nd4jLong *xShape, const Nd4jLong *hXShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShape, const Nd4jLong *hZShapeInfo, - int *dimension, int dimensionLength, - void *reductionPointer, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims) { if(shape::isEmpty(hXShapeInfo)) { @@ -256,10 +244,17 @@ __host__ void ReduceFloatFunction::intermediateXD(dim3 launchDims, cudaStre auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer(); // scalar assign - functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, zShape, hZShapeInfo, z, zShape, hZShapeInfo, ptr, nullptr); + functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, dZShapeInfo, hZShapeInfo, z, dZShapeInfo, hZShapeInfo, ptr, nullptr); } else { - simpleReduce<<>>(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets); + + const int zRank = shape::rank(hZShapeInfo); + const int tadRank = shape::rank(hXShapeInfo) - zRank; + + auto outerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims, zRank); + auto innerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims+zRank, tadRank); + + simpleReduce<<>>(x, reinterpret_cast(outerPack.special()), reinterpret_cast(innerPack.special()), extraParams, vreductionBuffer, z, dZShapeInfo); } } @@ -269,7 +264,7 @@ template __host__ void ReduceFloatFunction::intermediateScalar(dim3 launchDims, cudaStream_t *stream, const void *x, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, - void *z, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo) { @@ -286,7 +281,7 @@ __host__ void ReduceFloatFunction::intermediateScalar(dim3 launchDims, cuda throw sd::cuda_exception::build("ReduceFloatFunction::intermediateScalar: failed to copy resulting scalar", res); } else { - simpleScalar <<>>(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo); + simpleScalar <<>>(x, xShapeInfo, extraParams, z, dZShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo); } } @@ -296,27 +291,28 @@ _CUDA_H void ReduceFloatFunction::execReduceScalar(dim3 launchDims, cudaStr const int opNum, const void *x, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, - void *z, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo) { - DISPATCH_BY_OPNUM_TT(intermediateScalar, PARAMS(launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z, zShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_FLOAT_OPS)); + DISPATCH_BY_OPNUM_TT(intermediateScalar, PARAMS(launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z, dZShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_FLOAT_OPS)); sd::DebugHelper::checkErrorCode(stream, "execReduceScalarFloat(...) failed"); } //////////////////////////////////////////////////////////////////////// template -_CUDA_H void ReduceFloatFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, - const int opNum, - const int rank, const void *x, const Nd4jLong *xShape, const Nd4jLong *hXShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShape, const Nd4jLong *hZShapeInfo, - int *dimension, int dimensionLength, - void *reductionPointer, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { +_CUDA_H void ReduceFloatFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, const int opNum, + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims) { - DISPATCH_BY_OPNUM_TT(intermediateXD, PARAMS(launchDims, stream, x, xShape, hXShapeInfo, extraParams, z, zShape, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_FLOAT_OPS)); + if(shape::length(hZShapeInfo) == 1) { + ReduceFloatFunction::execReduceScalar(launchDims, stream, opNum, x, dXShapeInfo, hXShapeInfo, extraParams, z, dZShapeInfo, hZShapeInfo, nullptr, 0, vreductionBuffer, nullptr); + } + else { + DISPATCH_BY_OPNUM_TT(intermediateXD, PARAMS(launchDims, stream, x, dXShapeInfo, hXShapeInfo, extraParams, vreductionBuffer, z, dZShapeInfo, hZShapeInfo, dims), OPS_A(REDUCE_FLOAT_OPS)); + } DEBUG_KERNEL(stream, opNum); } diff --git a/libnd4j/include/loops/cuda/reduce/reduce_long.cu b/libnd4j/include/loops/cuda/reduce/reduce_long.cu index 1beac5330..e80afa6b2 100644 --- a/libnd4j/include/loops/cuda/reduce/reduce_long.cu +++ b/libnd4j/include/loops/cuda/reduce/reduce_long.cu @@ -33,14 +33,10 @@ using namespace simdOps; //////////////////////////////////////////////////////////////////////// template -__device__ void reduceSimpleGeneric(const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - void *reductionBuffer, - const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets) { +__global__ void simpleReduce(const void *x, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *extraParams, void *vreductionBuffer, void *z, const Nd4jLong *zShapeInfo) { - functions::reduce::ReduceLongFunction::template transformCudaXD(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo, tadOffsets); + functions::reduce::ReduceLongFunction::template transformCudaXD(x, outerXTadShapeInfo, innerXTadShapeInfo, extraParams, vreductionBuffer, z, zShapeInfo); } //////////////////////////////////////////////////////////////////////// @@ -55,17 +51,6 @@ __device__ void reduceScalarGeneric(const void *x, const Nd4jLong *xShapeInfo, functions::reduce::ReduceLongFunction::template execScalarCuda(x, xShapeInfo, extraParams, z, zShapeInfo, reductionBuffer, tadOnlyShapeInfo); } -//////////////////////////////////////////////////////////////////////// -template -__global__ void simpleReduce(const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - void *reductionBuffer, - const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets) { - - reduceSimpleGeneric(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo, tadOffsets); -} //////////////////////////////////////////////////////////////////////// template @@ -118,51 +103,48 @@ __device__ void ReduceLongFunction::aggregatePartials(void *vsPartials, Nd4 //////////////////////////////////////////////////////////////////////// template template -__device__ void ReduceLongFunction::transformCudaXD(const void *vx, const Nd4jLong *xShapeInfo, - void *vextraParams, - void *vz, const Nd4jLong *zShapeInfo, - int *dimension, int dimensionLength, - void *vreductionBuffer, - const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets) { +__device__ void ReduceLongFunction::transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *vextraParams, void *vreductionBuffer, + void *vz, const Nd4jLong *zShapeInfo) { auto x = reinterpret_cast(vx); auto z = reinterpret_cast(vz); auto extraParams = reinterpret_cast(vextraParams); - auto reductionBuffer = reinterpret_cast(vreductionBuffer); //shared memory space for storing intermediate results - __shared__ Z* sPartials; - __shared__ int tadLength, numTads; - __shared__ bool isPlainOutput; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; + __shared__ int tadLen, numTads; + __shared__ bool sameOffsets; if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); + sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo); - isPlainOutput = shape::order(zShapeInfo) == 'c' && shape::elementWiseStride(zShapeInfo) == 1; - - tadLength = shape::length(tadOnlyShapeInfo); - numTads = shape::length(xShapeInfo) / tadLength; + tadLen = shape::length(innerXTadShapeInfo); + numTads = shape::length(outerXTadShapeInfo); } __syncthreads(); + int coords[MAX_RANK]; + for (int r = blockIdx.x; r < numTads; r += gridDim.x) { - Nd4jLong tadOffsetForBlock = tadOffsets[r]; - sPartials[threadIdx.x] = OpType::startingValue(x + tadOffsetForBlock); + shape::index2coords(r, outerXTadShapeInfo, coords); + const auto outerOffset = shape::getOffset(outerXTadShapeInfo, coords); + const auto zOffset = sameOffsets ? outerOffset : shape::getOffset(zShapeInfo, coords); - for (int i = threadIdx.x; i < tadLength; i += blockDim.x) { - auto xOffset = tadOffsetForBlock + shape::getIndexOffset(i, tadOnlyShapeInfo); - sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(x[xOffset], extraParams), extraParams); - } + const X* xTad = x + outerOffset; + sPartials[threadIdx.x] = OpType::startingValue(xTad); + + for (int i = threadIdx.x; i < tadLen; i += blockDim.x) + sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(xTad[shape::getIndexOffset(i, innerXTadShapeInfo)], extraParams), extraParams); __syncthreads(); - // aggregate. do NOT reduce for elements > tadLength - aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); + // aggregate. do NOT reduce for elements > tadLen + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLen), extraParams); __syncthreads(); if (threadIdx.x == 0) - z[isPlainOutput ? r : shape::getIndexOffset(r, zShapeInfo)] = OpType::postProcess(sPartials[threadIdx.x], tadLength, extraParams); + z[zOffset] = OpType::postProcess(sPartials[threadIdx.x], tadLen, extraParams); } } @@ -183,13 +165,11 @@ __device__ void ReduceLongFunction::execScalarCuda(const void *vx, const Nd auto tid = blockDim.x * blockIdx.x + threadIdx.x; //shared memory space for storing intermediate results - __shared__ Z* sPartials; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; __shared__ Nd4jLong xEws; __shared__ Nd4jLong len; if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); xEws = shape::elementWiseStride(xShapeInfo); len = shape::length(xShapeInfo); } @@ -257,12 +237,9 @@ __device__ void ReduceLongFunction::execScalarCuda(const void *vx, const Nd template template __host__ void ReduceLongFunction::intermediateXD(dim3 launchDims, cudaStream_t *stream, - const void *x, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, - int *dimension, int dimensionLength, - void *reductionPointer, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims) { if(shape::isEmpty(hXShapeInfo)) { @@ -278,10 +255,16 @@ __host__ void ReduceLongFunction::intermediateXD(dim3 launchDims, cudaStrea auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer(); // scalar assign - functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, zShapeInfo, hXShapeInfo, z, zShapeInfo, hZShapeInfo, ptr, nullptr); + functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, dZShapeInfo, hXShapeInfo, z, dZShapeInfo, hZShapeInfo, ptr, nullptr); } else { - simpleReduce<<>>(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets); + const int zRank = shape::rank(hZShapeInfo); + const int tadRank = shape::rank(hXShapeInfo) - zRank; + + auto outerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims, zRank); + auto innerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims+zRank, tadRank); + + simpleReduce<<>>(x, reinterpret_cast(outerPack.special()), reinterpret_cast(innerPack.special()), extraParams, vreductionBuffer, z, dZShapeInfo); } } @@ -329,16 +312,17 @@ _CUDA_H void ReduceLongFunction::execReduceScalar(dim3 launchDims, cudaStre //////////////////////////////////////////////////////////////////////// template -_CUDA_H void ReduceLongFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, - const int opNum, - int rank, const void *x, const Nd4jLong *xShapeInfo, const Nd4jLong* hXShapeInfo, - void *extraParams, - void *z, const Nd4jLong *zShapeInfo, const Nd4jLong* hZShapeInfo, - int *dimension, int dimensionLength, - void *reductionPointer, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets) { +_CUDA_H void ReduceLongFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, const int opNum, + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims) { - DISPATCH_BY_OPNUM_TT(intermediateXD, PARAMS(launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z, zShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_LONG_OPS)); + if(shape::length(hZShapeInfo) == 1) { + ReduceLongFunction::execReduceScalar(launchDims, stream, opNum, x, dXShapeInfo, hXShapeInfo, extraParams, z, dZShapeInfo, hZShapeInfo, nullptr, 0, vreductionBuffer, nullptr); + } + else { + DISPATCH_BY_OPNUM_TT(intermediateXD, PARAMS(launchDims, stream, x, dXShapeInfo, hXShapeInfo, extraParams, vreductionBuffer, z, dZShapeInfo, hZShapeInfo, dims), OPS_A(REDUCE_LONG_OPS)); + } DEBUG_KERNEL(stream, opNum); } diff --git a/libnd4j/include/loops/cuda/reduce/reduce_same.cu b/libnd4j/include/loops/cuda/reduce/reduce_same.cu index c1947314e..0ae76eb51 100644 --- a/libnd4j/include/loops/cuda/reduce/reduce_same.cu +++ b/libnd4j/include/loops/cuda/reduce/reduce_same.cu @@ -31,17 +31,12 @@ using namespace simdOps; - //////////////////////////////////////////////////////////////////////// template -__global__ void simpleReduce(void const* x, Nd4jLong const* xShapeInfo, - void *extraParams, - void *z, Nd4jLong const* zShapeInfo, - int *dimension, int dimensionLength, - void *reductionBuffer, - Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets) { +__global__ void simpleReduce(const void *x, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *extraParams, void *vreductionBuffer, void *z, const Nd4jLong *zShapeInfo) { - functions::reduce::ReduceSameFunction::template transformCudaXD(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo, tadOffsets); + functions::reduce::ReduceSameFunction::template transformCudaXD(x, outerXTadShapeInfo, innerXTadShapeInfo, extraParams, vreductionBuffer, z, zShapeInfo); } //////////////////////////////////////////////////////////////////////// @@ -95,61 +90,54 @@ __device__ void ReduceSameFunction::aggregatePartials(void *vsPartials, Nd4jL //////////////////////////////////////////////////////////////////////// template template -__device__ void ReduceSameFunction::transformCudaXD( void const* vx, Nd4jLong const* xShapeInfo, - void *vextraParams, - void *vz, Nd4jLong const* zShapeInfo, - int *dimension, int dimensionLength, - void *vreductionBuffer, - Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets) { +__device__ void ReduceSameFunction::transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, + void *vextraParams, void *vreductionBuffer, + void *vz, const Nd4jLong *zShapeInfo) { auto x = reinterpret_cast(vx); auto z = reinterpret_cast(vz); auto extraParams = reinterpret_cast(vextraParams); auto reductionBuffer = reinterpret_cast(vreductionBuffer); - if (OpType::requiresSpecialAccumulation) { - OpType::execSpecialCuda(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo, tadOffsets); - return; - } + // if (OpType::requiresSpecialAccumulation) { + // OpType::execSpecialCuda(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo, tadOffsets); + // return; + // } //shared memory space for storing intermediate results - __shared__ X* sPartials; - - __shared__ int tadLength, tadRank, numTads; - __shared__ Nd4jLong *tadShape, *tadStride; - __shared__ bool isPlainOutput; + __shared__ X sPartials[CUDA_BLOCK_SIZE]; + __shared__ int tadLen, numTads; + __shared__ bool sameOffsets; if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); + sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo); - isPlainOutput = shape::order(zShapeInfo) == 'c' && shape::elementWiseStride(zShapeInfo) == 1; - - tadLength = shape::length(tadOnlyShapeInfo); - tadRank = shape::rank(tadOnlyShapeInfo); - numTads = shape::length(xShapeInfo) / tadLength; - tadShape = shape::shapeOf(tadOnlyShapeInfo); - tadStride = shape::stride(tadOnlyShapeInfo); + tadLen = shape::length(innerXTadShapeInfo); + numTads = shape::length(outerXTadShapeInfo); } __syncthreads(); + int coords[MAX_RANK]; + for (int r = blockIdx.x; r < numTads; r += gridDim.x) { - Nd4jLong tadOffsetForBlock = tadOffsets[r]; - sPartials[threadIdx.x] = OpType::startingValue(x + tadOffsetForBlock); + shape::index2coords(r, outerXTadShapeInfo, coords); + const auto outerOffset = shape::getOffset(outerXTadShapeInfo, coords); + const auto zOffset = sameOffsets ? outerOffset : shape::getOffset(zShapeInfo, coords); - for (int i = threadIdx.x; i < tadLength; i += blockDim.x) { - auto xOffset = tadOffsetForBlock + shape::getIndexOffset(i, tadOnlyShapeInfo); - sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(x[xOffset], extraParams), extraParams); - } + const X* xTad = x + outerOffset; + sPartials[threadIdx.x] = OpType::startingValue(xTad); + + for (int i = threadIdx.x; i < tadLen; i += blockDim.x) + sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(xTad[shape::getIndexOffset(i, innerXTadShapeInfo)], extraParams), extraParams); __syncthreads(); - // aggregate. do NOT reduce for elements > tadLength - aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); + // aggregate. do NOT reduce for elements > tadLen + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLen), extraParams); __syncthreads(); if (threadIdx.x == 0) - z[isPlainOutput ? r : shape::getIndexOffset(r, zShapeInfo)] = OpType::postProcess(sPartials[threadIdx.x], tadLength, extraParams); + z[zOffset] = OpType::postProcess(sPartials[threadIdx.x], tadLen, extraParams); } } @@ -179,13 +167,11 @@ __device__ void ReduceSameFunction::execScalarCuda(void const* vx, Nd4jLong c auto tid = blockDim.x * blockIdx.x + threadIdx.x; //shared memory space for storing intermediate results - __shared__ X* sPartials; + __shared__ X sPartials[CUDA_BLOCK_SIZE]; __shared__ Nd4jLong xEws; __shared__ Nd4jLong len; if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); xEws = shape::elementWiseStride(xShapeInfo); len = shape::length(xShapeInfo); } @@ -251,7 +237,10 @@ __device__ void ReduceSameFunction::execScalarCuda(void const* vx, Nd4jLong c //////////////////////////////////////////////////////////////////////// template template -__host__ void ReduceSameFunction::intermediateXD(dim3 launchDims, cudaStream_t *stream, void const* x, Nd4jLong const* xShapeInfo, Nd4jLong const* hXShapeInfo, void *extraParams, void *z, Nd4jLong const* zShapeInfo, Nd4jLong const* hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets) { +__host__ void ReduceSameFunction::intermediateXD(dim3 launchDims, cudaStream_t *stream, + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims) { if(shape::isEmpty(hXShapeInfo)) { @@ -267,10 +256,16 @@ __host__ void ReduceSameFunction::intermediateXD(dim3 launchDims, cudaStream_ auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer(); // scalar assign - functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, zShapeInfo, hXShapeInfo, z, zShapeInfo, hZShapeInfo, ptr, nullptr); + functions::scalar::ScalarTransform::executeCudaShaped(launchDims, stream, 14, z, dZShapeInfo, hXShapeInfo, z, dZShapeInfo, hZShapeInfo, ptr, nullptr); } else { - simpleReduce<<>>(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets); + + const int zRank = shape::rank(hZShapeInfo); + const int tadRank = shape::rank(hXShapeInfo) - zRank; + + auto outerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims, zRank); + auto innerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(hXShapeInfo, dims+zRank, tadRank); + simpleReduce<<>>(x, reinterpret_cast(outerPack.special()), reinterpret_cast(innerPack.special()), extraParams, vreductionBuffer, z, dZShapeInfo); } } @@ -305,9 +300,17 @@ _CUDA_H void ReduceSameFunction::execReduceScalar(dim3 launchDims, cudaStream //////////////////////////////////////////////////////////////////////// template -_CUDA_H void ReduceSameFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, void const* x, Nd4jLong const* xShapeInfo, Nd4jLong const* hXShapeInfo, void *extraParams, void *z, Nd4jLong const* zShapeInfo, Nd4jLong const* hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets) { +_CUDA_H void ReduceSameFunction::execReduceXD(dim3 launchDims, cudaStream_t *stream, const int opNum, + const void *x, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, + void *extraParams, void *vreductionBuffer, + void *z, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims) { - DISPATCH_BY_OPNUM_T(intermediateXD, PARAMS(launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z, zShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), REDUCE_SAME_OPS); + if(shape::length(hZShapeInfo) == 1) { + ReduceSameFunction::execReduceScalar(launchDims, stream, opNum, x, dXShapeInfo, hXShapeInfo, extraParams, z, dZShapeInfo, hZShapeInfo, nullptr, 0, vreductionBuffer, nullptr); + } + else { + DISPATCH_BY_OPNUM_T(intermediateXD, PARAMS(launchDims, stream, x, dXShapeInfo, hXShapeInfo, extraParams, vreductionBuffer, z, dZShapeInfo, hZShapeInfo, dims), REDUCE_SAME_OPS); + } DEBUG_KERNEL(stream, opNum); } diff --git a/libnd4j/include/loops/cuda/reduce3.chpp b/libnd4j/include/loops/cuda/reduce3.chpp index 2a301b817..799ddda6a 100644 --- a/libnd4j/include/loops/cuda/reduce3.chpp +++ b/libnd4j/include/loops/cuda/reduce3.chpp @@ -122,12 +122,9 @@ __device__ void Reduce3::execScalarCuda( void const* vx, Nd4jLong const* xS auto z = reinterpret_cast(vz); __shared__ Z extraZ[3]; - __shared__ Z* sPartials; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); - extraZ[0] = (Z) 0.0f; extraZ[1] = (Z) 0.0f; @@ -250,16 +247,11 @@ __device__ void Reduce3::transformAll( void const* vx, Nd4jLong const* xSha auto z = reinterpret_cast(vz); // initialize partials first - __shared__ Z* sPartials; - if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); - } - __syncthreads(); + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; Z startingVal = OpType::startingValue(dx); sPartials[threadIdx.x] = startingVal; - X *tempX = reinterpret_cast(sPartials) + blockDim.x; + auto tempX = reinterpret_cast(sPartials) + blockDim.x; const int maxBlock = blockDim.x; @@ -365,7 +357,7 @@ __device__ void Reduce3::transform(void const* vx, Nd4jLong const* xShapeIn __shared__ Z extraZ[OpType::extraParamsLen > 0 ? OpType::extraParamsLen : 1]; - __shared__ Z* sPartials; + __shared__ Z sPartials[CUDA_BLOCK_SIZE]; __shared__ int tadLen; __shared__ Nd4jLong zLen; __shared__ Nd4jLong xTadEws; @@ -375,9 +367,6 @@ __device__ void Reduce3::transform(void const* vx, Nd4jLong const* xShapeIn __shared__ char yTadOrder; if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast(shmem); - tadLen = shape::length(tadOnlyShapeInfo); zLen = shape::length(zShapeInfo); xTadEws = shape::elementWiseStride(tadOnlyShapeInfo); diff --git a/libnd4j/include/loops/cuda/summarystatsreduce.cu b/libnd4j/include/loops/cuda/summarystatsreduce.cu index 521ac5b06..bdab3d743 100644 --- a/libnd4j/include/loops/cuda/summarystatsreduce.cu +++ b/libnd4j/include/loops/cuda/summarystatsreduce.cu @@ -52,12 +52,12 @@ void _CUDA_G summaryStatsReduceT(int op, void const* dx, Nd4jLong const* xShapeI */ template template - _CUDA_D void SummaryStatsReduce::aggregatePartials(SummaryStatsData **sPartialsRef, Nd4jLong tid, Nd4jLong numElements, void *vextraParams) { + _CUDA_D void SummaryStatsReduce::aggregatePartials(SummaryStatsData *sPartials, Nd4jLong tid, Nd4jLong numElements, void *vextraParams) { // start the shared memory loop on the next power of 2 less // than the block size. If block size is not a power of 2, // accumulate the intermediate sums in the remainder range. + auto extraParams = static_cast(vextraParams); - SummaryStatsData *sPartials = *sPartialsRef; Nd4jLong floorPow2 = blockDim.x; if (floorPow2 & (floorPow2 - 1)) { @@ -123,12 +123,7 @@ void _CUDA_G summaryStatsReduceT(int op, void const* dx, Nd4jLong const* xShapeI int numElements = blockDim.x; //shared memory space for storing intermediate results - __shared__ SummaryStatsData *sPartials; - if(threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sPartials = reinterpret_cast*>(shmem); - } - __syncthreads(); + __shared__ SummaryStatsData sPartials[CUDA_BLOCK_SIZE]; Z startingVal = startingValue(dx); @@ -211,7 +206,7 @@ void _CUDA_G summaryStatsReduceT(int op, void const* dx, Nd4jLong const* xShapeI sPartials[threadIdx.x] = update(sPartials[threadIdx.x], OpType::op(indexVal2, extraParams), extraParams); } __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); __syncthreads(); if (threadIdx.x == 0) { @@ -237,7 +232,7 @@ void _CUDA_G summaryStatsReduceT(int op, void const* dx, Nd4jLong const* xShapeI } __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); + aggregatePartials(sPartials, threadIdx.x, sd::math::nd4j_min(blockDim.x, tadLength), extraParams); __syncthreads(); if (threadIdx.x == 0) { @@ -274,7 +269,7 @@ void _CUDA_G summaryStatsReduceT(int op, void const* dx, Nd4jLong const* xShapeI sPartials[threadIdx.x] = reduction; __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, blockDim.x, extraParams); + aggregatePartials(sPartials, threadIdx.x, blockDim.x, extraParams); __syncthreads(); if (gridDim.x > 1) { @@ -311,7 +306,7 @@ void _CUDA_G summaryStatsReduceT(int op, void const* dx, Nd4jLong const* xShapeI } __syncthreads(); - aggregatePartials(&sPartials, threadIdx.x, gridDim.x, extraParams); + aggregatePartials(sPartials, threadIdx.x, gridDim.x, extraParams); __syncthreads(); if (tid == 0) { diff --git a/libnd4j/include/loops/indexreduce.h b/libnd4j/include/loops/indexreduce.h index 2e8bc33d2..173c79b64 100755 --- a/libnd4j/include/loops/indexreduce.h +++ b/libnd4j/include/loops/indexreduce.h @@ -22,7 +22,7 @@ #ifndef INDEXREDUCE_H_ #define INDEXREDUCE_H_ -#include "../helpers/shape.h" +#include #ifdef _OPENMP #include #endif @@ -63,7 +63,7 @@ namespace functions { const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset); template - static __device__ void aggregatePartials(IndexValue **sPartialsRef, Nd4jLong tid, Nd4jLong numElements, void *extraParams); + static __device__ void aggregatePartials(IndexValue *sPartialsRef, Nd4jLong tid, Nd4jLong numElements, void *extraParams); template diff --git a/libnd4j/include/loops/reduce_bool.h b/libnd4j/include/loops/reduce_bool.h index a74d53033..170b29991 100644 --- a/libnd4j/include/loops/reduce_bool.h +++ b/libnd4j/include/loops/reduce_bool.h @@ -29,6 +29,7 @@ #include #include #include +#include #pragma once #ifdef __CUDACC__ @@ -61,17 +62,17 @@ namespace functions { static __device__ void execScalarCuda(const void *vx, const Nd4jLong *xShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); template - static __device__ void transformCudaXD(const void *vx, const Nd4jLong *xShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets); + static __device__ void transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *zShapeInfo); template static __host__ void intermediateScalar(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); template - static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims); static __host__ void execReduceScalar(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong* hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong* hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); - static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong* hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong* hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims); #else /** @@ -98,13 +99,11 @@ namespace functions { void *extraParams, void *z, const Nd4jLong *zShapeInfo); - static void exec(int opNum, - const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *result, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + static void exec(int opNum, sd::memory::Workspace* workspace, + const void *vx, const Nd4jLong *xShapeInfo, + void *vextraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int *dims); /** * Execute on the cpu @@ -118,14 +117,12 @@ namespace functions { * @param dimensionLength the length of the dimension buffer */ - template - static void _CUDA_H exec(const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *result, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + static void _CUDA_H exec(sd::memory::Workspace* workspace, + const void *vx, const Nd4jLong *xShapeInfo, + void *vextraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int *dims); /** * CPU implementation diff --git a/libnd4j/include/loops/reduce_float.h b/libnd4j/include/loops/reduce_float.h index c78082f8e..bc78df2d9 100644 --- a/libnd4j/include/loops/reduce_float.h +++ b/libnd4j/include/loops/reduce_float.h @@ -29,6 +29,7 @@ #include #include #include +#include #pragma once #ifdef __CUDACC__ @@ -52,28 +53,28 @@ namespace functions { */ template class ReduceFloatFunction { - + public: -#ifdef __CUDACC__ +#ifdef __CUDACC__ template static __device__ void aggregatePartials(void *sPartials, Nd4jLong tid, Nd4jLong numItems, void *extraParams); template static __device__ void execScalarCuda(const void *vx, const Nd4jLong *xShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); - template - static __device__ void transformCudaXD(const void *vx, const Nd4jLong *xShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets); + template + static __device__ void transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *zShapeInfo); template static __host__ void intermediateScalar(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); template - static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShape, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims); static __host__ void execReduceScalar(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); - static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShape, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims); #else /** @@ -102,13 +103,11 @@ namespace functions { void *extraParams, void *vz, const Nd4jLong *zShapeInfo); - static void exec(int opNum, + static void exec(int opNum, sd::memory::Workspace* workspace, const void *vx, const Nd4jLong *xShapeInfo, - void *extraParams, - void *vz, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + void *vextraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int *dims); /** * Execute on the cpu @@ -124,12 +123,11 @@ namespace functions { template - static void _CUDA_H exec(const void *vx, const Nd4jLong *xShapeInfo, - void *extraParams, - void *vz, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + static void _CUDA_H exec(sd::memory::Workspace* workspace, + const void *vx, const Nd4jLong *xShapeInfo, + void *extraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int* dims); /** * CPU implementation diff --git a/libnd4j/include/loops/reduce_long.h b/libnd4j/include/loops/reduce_long.h index 45ede2985..5ee0cce3b 100644 --- a/libnd4j/include/loops/reduce_long.h +++ b/libnd4j/include/loops/reduce_long.h @@ -29,6 +29,7 @@ #include #include #include +#include #pragma once #ifdef __CUDACC__ @@ -60,17 +61,17 @@ namespace functions { static __device__ void execScalarCuda(const void *vx, const Nd4jLong *xShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); template - static __device__ void transformCudaXD(const void *vx, const Nd4jLong *xShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo, const Nd4jLong *tadOffsets); + static __device__ void transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *zShapeInfo); template static __host__ void intermediateScalar(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); template - static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong *hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims); static __host__ void execReduceScalar(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong* hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong* hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, const Nd4jLong *tadOnlyShapeInfo); - static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, const void *vx, const Nd4jLong *xShapeInfo, const Nd4jLong* hXShapeInfo, void *extraParams, void *vz, const Nd4jLong *zShapeInfo, const Nd4jLong* hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffsets); + static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void *vreductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims); #else @@ -100,13 +101,11 @@ namespace functions { void *extraParams, void *z, const Nd4jLong *zShapeInfo); - static void exec(int opNum, - const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *result, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + static void exec(int opNum, sd::memory::Workspace* workspace, + const void *vx, const Nd4jLong *xShapeInfo, + void *vextraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int *dims); /** * Execute on the cpu @@ -120,14 +119,12 @@ namespace functions { * @param dimensionLength the length of the dimension buffer */ - template - static void _CUDA_H exec(const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *result, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + static void _CUDA_H exec(sd::memory::Workspace* workspace, + const void *vx, const Nd4jLong *xShapeInfo, + void *vextraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int *dims); /** * CPU implementation diff --git a/libnd4j/include/loops/reduce_same.h b/libnd4j/include/loops/reduce_same.h index 5f3622f39..f28409bc6 100644 --- a/libnd4j/include/loops/reduce_same.h +++ b/libnd4j/include/loops/reduce_same.h @@ -29,6 +29,7 @@ #include #include #include +#include #pragma once #ifdef __CUDACC__ @@ -63,17 +64,17 @@ namespace functions { static __device__ void execScalarCudaLegacy(int opNum, void const* vx, Nd4jLong const* xShapeInfo, void *extraParams, void *vz, Nd4jLong const* zShapeInfo, void *reductionBuffer, Nd4jLong const* tadOnlyShapeInfo); template - static __device__ void transformCudaXD( void const* vx, Nd4jLong const* xShapeInfo, void *extraParams, void *vz, Nd4jLong const* zShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets); + static __device__ void transformCudaXD(const void *vx, const Nd4jLong *outerXTadShapeInfo, const Nd4jLong *innerXTadShapeInfo, void *extraParams, void* reductionBuffer, void *vz, const Nd4jLong *zShapeInfo); template static __host__ void intermediateScalar(dim3 launchDims, cudaStream_t *stream, void const* vx, Nd4jLong const* xShapeInfo, Nd4jLong const* hXShapeInfo, void *extraParams, void *vz, Nd4jLong const* zShapeInfo, Nd4jLong const* hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, Nd4jLong const* tadOnlyShapeInfo); template - static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, void const* vx, Nd4jLong const* xShapeInfo, Nd4jLong const* hXShapeInfo, void *extraParams, void *vz, Nd4jLong const* zShapeInfo, Nd4jLong const* hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets); + static __host__ void intermediateXD(dim3 launchDims, cudaStream_t *stream, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void* reductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int* dims); static __host__ void execReduceScalar(dim3 launchDims, cudaStream_t *stream, int opNum, void const* vx, Nd4jLong const* xShapeInfo, Nd4jLong const* hXShapeInfo, void *extraParams, void *vz, Nd4jLong const* zShapeInfo, Nd4jLong const* hZShapeInfo, int *dimension, int dimensionLength, void *reductionBuffer, Nd4jLong const* tadOnlyShapeInfo); - static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, void const* vx, Nd4jLong const* xShapeInfo, Nd4jLong const* hXShapeInfo, void *extraParams, void *vz, Nd4jLong const* zShapeInfo, Nd4jLong const* hZShapeInfo, int *dimension, int dimensionLength, void *reductionPointer, Nd4jLong const* tadShapeInfo, Nd4jLong const* tadOffsets); + static __host__ void execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx, const Nd4jLong *dXShapeInfo, const Nd4jLong *hXShapeInfo, void *extraParams, void* reductionBuffer, void *vz, const Nd4jLong *dZShapeInfo, const Nd4jLong *hZShapeInfo, const int *dims); #else /** @@ -103,13 +104,11 @@ namespace functions { void *extraParams, void *z, const Nd4jLong *zShapeInfo); - static void exec(int opNum, - const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *result, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + static void exec(int opNum, sd::memory::Workspace* workspace, + const void *vx, const Nd4jLong *xShapeInfo, + void *vextraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int *dims); /** * Execute on the cpu @@ -123,14 +122,12 @@ namespace functions { * @param dimensionLength the length of the dimension buffer */ - template - static void _CUDA_H exec(const void *x, const Nd4jLong *xShapeInfo, - void *extraParams, - void *result, const Nd4jLong *resultShapeInfoBuffer, - int *dimension, int dimensionLength, - const Nd4jLong *tadShapeInfo, const Nd4jLong *tadOffset, - int64_t start, int64_t stop); + static void _CUDA_H exec(sd::memory::Workspace* workspace, + const void *vx, const Nd4jLong *xShapeInfo, + void *vextraParams, + void *vz, const Nd4jLong *zShapeInfo, + const int *dims); /** * CPU implementation diff --git a/libnd4j/include/loops/summarystatsreduce.h b/libnd4j/include/loops/summarystatsreduce.h index 1ab06a11b..12aea687a 100755 --- a/libnd4j/include/loops/summarystatsreduce.h +++ b/libnd4j/include/loops/summarystatsreduce.h @@ -275,7 +275,7 @@ namespace functions { } template - static _CUDA_D void aggregatePartials(SummaryStatsData **sPartialsRef, Nd4jLong tid, Nd4jLong numElements, void *extraParams); + static _CUDA_D void aggregatePartials(SummaryStatsData *sPartials, Nd4jLong tid, Nd4jLong numElements, void *extraParams); template diff --git a/libnd4j/include/ops/declarable/generic/loss/absoluteDifference.cpp b/libnd4j/include/ops/declarable/generic/loss/absoluteDifference.cpp index 0d5d1d011..a542ae8e4 100644 --- a/libnd4j/include/ops/declarable/generic/loss/absoluteDifference.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/absoluteDifference.cpp @@ -184,7 +184,7 @@ CUSTOM_OP_IMPL(absolute_difference_loss_grad, 3, 3, false, 0, 1) { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -210,7 +210,7 @@ CUSTOM_OP_IMPL(absolute_difference_loss_grad, 3, 3, false, 0, 1) { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -238,7 +238,7 @@ CUSTOM_OP_IMPL(absolute_difference_loss_grad, 3, 3, false, 0, 1) { dLdw->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/cosineDistance.cpp b/libnd4j/include/ops/declarable/generic/loss/cosineDistance.cpp index 99cf2e3c1..9317aed32 100644 --- a/libnd4j/include/ops/declarable/generic/loss/cosineDistance.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/cosineDistance.cpp @@ -216,7 +216,7 @@ CUSTOM_OP_IMPL(cosine_distance_loss_grad, 3, 3, false, 0, 2) { else { if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -249,7 +249,7 @@ CUSTOM_OP_IMPL(cosine_distance_loss_grad, 3, 3, false, 0, 2) { if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -284,7 +284,7 @@ CUSTOM_OP_IMPL(cosine_distance_loss_grad, 3, 3, false, 0, 2) { if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeights; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/hingeLoss.cpp b/libnd4j/include/ops/declarable/generic/loss/hingeLoss.cpp index 71e7489ea..0766cf600 100644 --- a/libnd4j/include/ops/declarable/generic/loss/hingeLoss.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/hingeLoss.cpp @@ -193,7 +193,7 @@ namespace sd { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -222,7 +222,7 @@ namespace sd { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -251,7 +251,7 @@ namespace sd { dLdw->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/huberLoss.cpp b/libnd4j/include/ops/declarable/generic/loss/huberLoss.cpp index 2d0b44b3c..5f7c94c88 100644 --- a/libnd4j/include/ops/declarable/generic/loss/huberLoss.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/huberLoss.cpp @@ -208,7 +208,7 @@ DECLARE_SHAPE_FN(huber_loss) { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -237,7 +237,7 @@ DECLARE_SHAPE_FN(huber_loss) { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -266,7 +266,7 @@ DECLARE_SHAPE_FN(huber_loss) { dLdw->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/logLoss.cpp b/libnd4j/include/ops/declarable/generic/loss/logLoss.cpp index ab0c8923e..3d1332302 100644 --- a/libnd4j/include/ops/declarable/generic/loss/logLoss.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/logLoss.cpp @@ -197,7 +197,7 @@ CUSTOM_OP_IMPL(log_loss_grad, 3, 3, false, 1, 1) { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -228,7 +228,7 @@ CUSTOM_OP_IMPL(log_loss_grad, 3, 3, false, 1, 1) { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -256,7 +256,7 @@ CUSTOM_OP_IMPL(log_loss_grad, 3, 3, false, 1, 1) { dLdw->assign(E.reduceNumber(reduce::Sum) / numOfNonZeroWeights); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/log_poisson_loss.cpp b/libnd4j/include/ops/declarable/generic/loss/log_poisson_loss.cpp index 5cc6b60ab..f1b7d5f41 100644 --- a/libnd4j/include/ops/declarable/generic/loss/log_poisson_loss.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/log_poisson_loss.cpp @@ -201,7 +201,7 @@ namespace ops { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -230,7 +230,7 @@ namespace ops { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -259,7 +259,7 @@ namespace ops { dLdw->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/meanPairWsSqErr.cpp b/libnd4j/include/ops/declarable/generic/loss/meanPairWsSqErr.cpp index f36fa3c62..e604a3da8 100644 --- a/libnd4j/include/ops/declarable/generic/loss/meanPairWsSqErr.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/meanPairWsSqErr.cpp @@ -274,7 +274,7 @@ namespace sd { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -301,7 +301,7 @@ namespace sd { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -329,7 +329,7 @@ namespace sd { dLdw->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/meanSqErr.cpp b/libnd4j/include/ops/declarable/generic/loss/meanSqErr.cpp index 6c54706c4..50fdb46e1 100644 --- a/libnd4j/include/ops/declarable/generic/loss/meanSqErr.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/meanSqErr.cpp @@ -192,7 +192,7 @@ CUSTOM_OP_IMPL(mean_sqerr_loss_grad, 3, 3, false, 0, 1) { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -219,7 +219,7 @@ CUSTOM_OP_IMPL(mean_sqerr_loss_grad, 3, 3, false, 0, 1) { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -247,7 +247,7 @@ CUSTOM_OP_IMPL(mean_sqerr_loss_grad, 3, 3, false, 0, 1) { dLdw->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/sigmCrossEntropy.cpp b/libnd4j/include/ops/declarable/generic/loss/sigmCrossEntropy.cpp index ddd28d43d..372a93388 100644 --- a/libnd4j/include/ops/declarable/generic/loss/sigmCrossEntropy.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/sigmCrossEntropy.cpp @@ -212,7 +212,7 @@ CUSTOM_OP_IMPL(sigm_cross_entropy_loss_grad, 3, 3, false, 1, 1) { dLdw->assign(E.reduceNumber(reduce::Sum)); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -241,7 +241,7 @@ CUSTOM_OP_IMPL(sigm_cross_entropy_loss_grad, 3, 3, false, 1, 1) { *dLdw = 0.; else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum * sum)); @@ -269,7 +269,7 @@ CUSTOM_OP_IMPL(sigm_cross_entropy_loss_grad, 3, 3, false, 1, 1) { dLdw->assign(E.reduceNumber(reduce::Sum) / numOfNonZeroWeightsScalar); else if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeightsScalar; } else diff --git a/libnd4j/include/ops/declarable/generic/loss/softmaxCrossEntropy.cpp b/libnd4j/include/ops/declarable/generic/loss/softmaxCrossEntropy.cpp index 79d46e448..c0bb78015 100644 --- a/libnd4j/include/ops/declarable/generic/loss/softmaxCrossEntropy.cpp +++ b/libnd4j/include/ops/declarable/generic/loss/softmaxCrossEntropy.cpp @@ -258,7 +258,7 @@ CUSTOM_OP_IMPL(softmax_cross_entropy_loss_grad, 3, 3, false, 1, 1) { if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign(E); @@ -294,7 +294,7 @@ CUSTOM_OP_IMPL(softmax_cross_entropy_loss_grad, 3, 3, false, 1, 1) { if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + ((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); } else dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)); @@ -331,7 +331,7 @@ CUSTOM_OP_IMPL(softmax_cross_entropy_loss_grad, 3, 3, false, 1, 1) { if(weights != weightsBroad) { std::vector axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo()); - E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false); + E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true); *dLdw /= numOfNonZeroWeights; } else diff --git a/libnd4j/include/ops/declarable/generic/nn/recurrent/sru.cpp b/libnd4j/include/ops/declarable/generic/nn/recurrent/sru.cpp index ba4e3d52f..918b33d4b 100644 --- a/libnd4j/include/ops/declarable/generic/nn/recurrent/sru.cpp +++ b/libnd4j/include/ops/declarable/generic/nn/recurrent/sru.cpp @@ -278,8 +278,8 @@ CUSTOM_OP_IMPL(sru_bp, 8, 4, true, 0, 0) { gradX->applyBroadcast(broadcast::Multiply, {0,1}, *mask, *gradX); // apply mask // gradB - auto temp3 = gradBias->reduceAlongDimension(reduce::Sum, {0,2}, false, true); // [1 x 2K] - gradB->assign(temp3); + auto gradB2 = gradB->reshape(gradB->ordering(), {2*K}); + gradBias->reduceAlongDimension(reduce::Sum, gradB2, {0,2}); // [1 x 2K] // gradW [bS x 3K x K] x->permutei({0, 2, 1}); // [bS x N x K] diff --git a/libnd4j/include/ops/declarable/helpers/cuda/activations.cu b/libnd4j/include/ops/declarable/helpers/cuda/activations.cu index e675342d9..205929119 100644 --- a/libnd4j/include/ops/declarable/helpers/cuda/activations.cu +++ b/libnd4j/include/ops/declarable/helpers/cuda/activations.cu @@ -193,11 +193,9 @@ __device__ void softMaxForVectorCuda(const void *vx, const Nd4jLong *xShapeInfo, __shared__ Nd4jLong len; __shared__ int numOfIters; - __shared__ T* shmem; + __shared__ T shmem[CUDA_BLOCK_SIZE]; if (threadIdx.x == 0) { - extern __shared__ char shared[]; - shmem = reinterpret_cast(shared); len = shape::length(xShapeInfo); numOfIters = (len + blockDim.x - 1) / blockDim.x; // ceil (len / blockDim.x) } @@ -274,7 +272,7 @@ __global__ void softMaxForVectorCudaGlobal(const void *vx, const Nd4jLong *xShap template linkage void softMaxForVectorCudaLauncher(const cudaStream_t* stream, const void *vx, const Nd4jLong *xShapeInfo, void *vz, const Nd4jLong *zShapeInfo) { - softMaxForVectorCudaGlobal<<<1, MAX_NUM_THREADS / 4 , (MAX_NUM_THREADS / 4) * sizeof(T) + 512, *stream>>>(vx, xShapeInfo, vz, zShapeInfo); + softMaxForVectorCudaGlobal<<<1, CUDA_BLOCK_SIZE, 1024, *stream>>>(vx, xShapeInfo, vz, zShapeInfo); } /////////////////////////////////////////////////////////////////// @@ -324,9 +322,9 @@ void softmax(sd::LaunchContext * context, const NDArray& input, NDArray& output, auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(input.shapeInfo(), {dimension}); auto packZ = sd::ConstantTadHelper::getInstance().tadForDimensions(output.shapeInfo(), {dimension}); - const int threadsPerBlock = MAX_NUM_THREADS / 4; + const int threadsPerBlock = CUDA_BLOCK_SIZE; const int blocksPerGrid = packZ.numberOfTads(); - const int sharedMem = input.sizeOfT() * threadsPerBlock + 512; + const int sharedMem = 1024; NDArray::prepareSpecialUse({&output}, {&input}); BUILD_SINGLE_SELECTOR(input.dataType(), softMaxCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), input.specialBuffer(), packX.specialShapeInfo(), packX.specialOffsets(), output.specialBuffer(), packZ.specialShapeInfo(), packZ.specialOffsets()), FLOAT_TYPES); @@ -356,11 +354,9 @@ __global__ void logSoftMaxForVectorCuda(const void *vx, const Nd4jLong *xzShape __shared__ Nd4jLong len; __shared__ int numOfIters; - __shared__ T* shmem; + __shared__ T shmem[CUDA_BLOCK_SIZE]; if (threadIdx.x == 0) { - extern __shared__ char shared[]; - shmem = reinterpret_cast(shared); len = shape::length(xzShapeInfo); numOfIters = (len + blockDim.x - 1) / blockDim.x; // ceil (len / blockDim.x) } @@ -430,7 +426,7 @@ __global__ void logSoftMaxForVectorCuda(const void *vx, const Nd4jLong *xzShape template linkage void logSoftMaxForVectorCudaLauncher(const cudaStream_t* stream, const void *vx, const Nd4jLong *xzShapeInfo, void *vz) { - logSoftMaxForVectorCuda<<<1, MAX_NUM_THREADS, MAX_NUM_THREADS * sizeof(T) + 512, *stream>>>(vx, xzShapeInfo, vz); + logSoftMaxForVectorCuda<<<1, CUDA_BLOCK_SIZE, 1024, *stream>>>(vx, xzShapeInfo, vz); } ////////////////////////////////////////////////////////////////////////// @@ -475,11 +471,9 @@ __global__ linkage void softMaxDerivForVectorCuda(const void *vx, const Nd4jLong __shared__ Nd4jLong len; __shared__ int numOfIters; - __shared__ T* shmem; + __shared__ T shmem[CUDA_BLOCK_SIZE]; if (threadIdx.x == 0) { - extern __shared__ char shared[]; - shmem = reinterpret_cast(shared); len = shape::length(xzShapeInfo); numOfIters = (len + blockDim.x - 1) / blockDim.x; // ceil (len / blockDim.x) } @@ -550,7 +544,7 @@ __global__ linkage void softMaxDerivForVectorCuda(const void *vx, const Nd4jLong template linkage void softMaxDerivForVectorCudaLauncher(const cudaStream_t* stream, const void *vx, const Nd4jLong *xzShapeInfo, void *vz) { - softMaxDerivForVectorCuda<<<1, MAX_NUM_THREADS, MAX_NUM_THREADS * sizeof(T) + 512, *stream>>>(vx, xzShapeInfo, vz); + softMaxDerivForVectorCuda<<<1, CUDA_BLOCK_SIZE, 1024, *stream>>>(vx, xzShapeInfo, vz); } /////////////////////////////////////////////////////////////////// diff --git a/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu b/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu index e3fdd9411..67847518d 100644 --- a/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu +++ b/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu @@ -30,13 +30,7 @@ namespace sd { auto y = reinterpret_cast(vy); auto z = reinterpret_cast(vz); - __shared__ Nd4jLong *shared; - - if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - shared = reinterpret_cast(shmem); - } - __syncthreads(); + __shared__ Nd4jLong shared[CUDA_BLOCK_SIZE]; // we want to nullify temporary memory before accumulating intermediate results shared[threadIdx.x] = 0; @@ -82,7 +76,7 @@ namespace sd { template static void _hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &z) { - _hammingKernel<<<256, 256, 256 * sizeof(Nd4jLong) + 256, *context->getCudaStream()>>>(x.specialBuffer(), x.specialShapeInfo(), y.specialBuffer(), y.specialShapeInfo(), z.specialBuffer(), nullptr, x.lengthOf()); + _hammingKernel<<<256, CUDA_BLOCK_SIZE, 1024, *context->getCudaStream()>>>(x.specialBuffer(), x.specialShapeInfo(), y.specialBuffer(), y.specialShapeInfo(), z.specialBuffer(), nullptr, x.lengthOf()); } void hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &output) { diff --git a/libnd4j/include/ops/declarable/helpers/cuda/pad.cu b/libnd4j/include/ops/declarable/helpers/cuda/pad.cu index 842a41ced..361cae22c 100644 --- a/libnd4j/include/ops/declarable/helpers/cuda/pad.cu +++ b/libnd4j/include/ops/declarable/helpers/cuda/pad.cu @@ -251,7 +251,12 @@ namespace sd { sd::DebugHelper::checkErrorCode(stream, "helpers::mirrorPadLinearKernel(...) failed"); } else { - mirrorPadKernel<<<256, 256, 8192, *stream>>>(input.specialBuffer(), input.specialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), outLen, paddings.specialBuffer(), paddings.specialShapeInfo(), reflBorder); + + const int threadsPerBlock = MAX_NUM_THREADS / 2; + const int blocksPerGrid = (outLen + threadsPerBlock - 1) / threadsPerBlock; + const int sharedMem = threadsPerBlock * sizeof(Nd4jLong) * input.rankOf() + 256; + + mirrorPadKernel<<>>(input.specialBuffer(), input.specialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), outLen, paddings.specialBuffer(), paddings.specialShapeInfo(), reflBorder); sd::DebugHelper::checkErrorCode(stream, "helpers::mirrorPadKernel(...) failed"); } NDArray::registerSpecialUse({&output}, {&input, &paddings}); diff --git a/libnd4j/include/ops/declarable/helpers/cuda/top_k.cu b/libnd4j/include/ops/declarable/helpers/cuda/top_k.cu index 61aefa255..c6aba57bc 100644 --- a/libnd4j/include/ops/declarable/helpers/cuda/top_k.cu +++ b/libnd4j/include/ops/declarable/helpers/cuda/top_k.cu @@ -38,15 +38,12 @@ __global__ static void inTopKCuda(const void* vx, const Nd4jLong* xShapeInfo, const auto y = reinterpret_cast(vy); auto z = reinterpret_cast(vz); - __shared__ uint* sharedMem; + __shared__ uint sharedMem[CUDA_BLOCK_SIZE]; __shared__ X elemToCompare; __shared__ const X* xTad; __shared__ Nd4jLong idx, xTadLen; if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sharedMem = reinterpret_cast(shmem); - xTadLen = shape::length(xTadShapeInfo); xTad = reinterpret_cast(vx) + xTadOffsets[blockIdx.x]; @@ -93,9 +90,9 @@ int inTopKFunctor(sd::LaunchContext * context, const NDArray* predictions, const const auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(predictions->shapeInfo(), {1}); - const int threadsPerBlock = MAX_NUM_THREADS; + const int threadsPerBlock = CUDA_BLOCK_SIZE; const int blocksPerGrid = static_cast(packX.numberOfTads()); - const int sharedMem = sizeof(uint) * threadsPerBlock + 128; + const int sharedMem = 1024; const auto xType = predictions->dataType(); const auto yType = targets->dataType(); diff --git a/libnd4j/include/ops/declarable/helpers/cuda/transforms.cu b/libnd4j/include/ops/declarable/helpers/cuda/transforms.cu index 80e0e0858..f1b57f52c 100644 --- a/libnd4j/include/ops/declarable/helpers/cuda/transforms.cu +++ b/libnd4j/include/ops/declarable/helpers/cuda/transforms.cu @@ -92,15 +92,11 @@ __global__ static void traceCuda(const void* vx, const Nd4jLong* xShapeInfo, voi const auto x = reinterpret_cast(vx); auto z = reinterpret_cast(vz); - __shared__ T* sharedMem; - __shared__ int xRank, zRank, *coordsMem; // xRank = zRank + 2 + __shared__ T sharedMem[CUDA_BLOCK_SIZE]; + __shared__ int xRank, zRank; // xRank = zRank + 2 __shared__ Nd4jLong xLen, zLen; if (threadIdx.x == 0) { - extern __shared__ unsigned char shmem[]; - sharedMem = reinterpret_cast(shmem); - coordsMem = reinterpret_cast(shmem + blockDim.x * sizeof(T)); - xRank = shape::rank(xShapeInfo); zRank = shape::rank(zShapeInfo); xLen = shape::length(xShapeInfo); @@ -109,7 +105,7 @@ __global__ static void traceCuda(const void* vx, const Nd4jLong* xShapeInfo, voi } __syncthreads(); - auto coords = coordsMem + threadIdx.x * xRank; + Nd4jLong coords[MAX_RANK]; for (uint m = blockIdx.x; m < zLen; m += gridDim.x) { // one block per each element of z, that is per each matrix @@ -158,9 +154,9 @@ void trace(sd::LaunchContext* context, const NDArray& input, NDArray& output) { PointersManager manager(context, "trace"); const uint diagLen = input.sizeAt(-1) < input.sizeAt(-2) ? input.sizeAt(-1) : input.sizeAt(-2); - const int threadsPerBlock = MAX_NUM_THREADS / 4; + const int threadsPerBlock = CUDA_BLOCK_SIZE; const int blocksPerGrid = (output.lengthOf() + threadsPerBlock - 1) / threadsPerBlock; - const int sharedMem = threadsPerBlock * (sizeof(int) * input.rankOf() + input.sizeOfT()) + 128; + const int sharedMem = 1024; NDArray::prepareSpecialUse({&output}, {&input}); BUILD_SINGLE_SELECTOR(input.dataType(), traceCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), input.specialBuffer(), input.specialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), diagLen), LIBND4J_TYPES); @@ -177,13 +173,10 @@ __global__ static void triuBPCuda(const void* vx, const Nd4jLong* xShapeInfo, vo const auto x = reinterpret_cast(vx); // gradO auto z = reinterpret_cast(vz); // gradI - __shared__ int rank, areSameOffsets, *sharedMem; // xRank = zRank + __shared__ int rank, areSameOffsets; __shared__ Nd4jLong len, totalThreads; // xLen = zLen if (threadIdx.x == 0) { - - extern __shared__ unsigned char shmem[]; - sharedMem = reinterpret_cast(shmem); areSameOffsets = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo); rank = shape::rank(xShapeInfo); len = shape::length(zShapeInfo); @@ -192,7 +185,7 @@ __global__ static void triuBPCuda(const void* vx, const Nd4jLong* xShapeInfo, vo __syncthreads(); - auto coords = sharedMem + threadIdx.x * rank; + Nd4jLong coords[MAX_RANK]; const auto tid = blockIdx.x * blockDim.x + threadIdx.x; @@ -240,14 +233,10 @@ __global__ static void tileBPCuda(const void* vx, const Nd4jLong* xShapeInfo, vo const auto x = reinterpret_cast(vx); // gradO auto z = reinterpret_cast(vz); // gradI - __shared__ int xRank, zRank, *sharedMem; // xRank >= zRank + __shared__ int xRank, zRank; // xRank >= zRank __shared__ Nd4jLong numOfXOffsets, zLen, totalThreads; // xLen >= zLen if (threadIdx.x == 0) { - - extern __shared__ unsigned char shmem[]; - sharedMem = reinterpret_cast(shmem); - xRank = shape::rank(zShapeInfo); zLen = shape::length(zShapeInfo); numOfXOffsets = shape::length(xShapeInfo) / zLen; @@ -259,7 +248,7 @@ __global__ static void tileBPCuda(const void* vx, const Nd4jLong* xShapeInfo, vo const auto tid = blockIdx.x * blockDim.x + threadIdx.x; - auto memBuff = sharedMem + threadIdx.x * 2 * xRank; + int memBuff[MAX_RANK * 2]; auto xOffsets = globMem + tid * numOfXOffsets; for (Nd4jLong i = tid; i < zLen; i += totalThreads) { diff --git a/libnd4j/include/ops/declarable/impl/LegacyReduceBoolOp.cpp b/libnd4j/include/ops/declarable/impl/LegacyReduceBoolOp.cpp index e16e71619..48fc0d84c 100644 --- a/libnd4j/include/ops/declarable/impl/LegacyReduceBoolOp.cpp +++ b/libnd4j/include/ops/declarable/impl/LegacyReduceBoolOp.cpp @@ -75,15 +75,28 @@ namespace sd { REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + if(x->rankOf() - dims.size() != z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } - NativeOpExecutioner::execReduceBool(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), - extras.argumentsAsT(x->dataType()), - z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), - dims.data(), (int) dims.size(), reinterpret_cast(pTadShape), reinterpret_cast(pTadOffsets)); + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); + NativeOpExecutioner::execReduceBool(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), nullptr, z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, dims2.data(), dims2.size()); + + + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + // NativeOpExecutioner::execReduceBool(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), + // extras.argumentsAsT(x->dataType()), + // z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), + // dims.data(), (int) dims.size(), reinterpret_cast(pTadShape), reinterpret_cast(pTadOffsets)); } STORE_RESULT(*z); @@ -111,13 +124,25 @@ namespace sd { REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + if(x->rankOf() - dims.size() != z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } - NativeOpExecutioner::execReduceBool(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), extras.argumentsAsT(x->dataType()), - z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), dims.data(), (int) dims.size(), pTadShape, pTadOffsets); + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); + NativeOpExecutioner::execReduceBool(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), nullptr, z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, dims2.data(), dims2.size()); + + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + // NativeOpExecutioner::execReduceBool(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), extras.argumentsAsT(x->dataType()), + // z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), dims.data(), (int) dims.size(), pTadShape, pTadOffsets); } } diff --git a/libnd4j/include/ops/declarable/impl/LegacyReduceFloatOp.cpp b/libnd4j/include/ops/declarable/impl/LegacyReduceFloatOp.cpp index a0ff14858..1fa7ce351 100644 --- a/libnd4j/include/ops/declarable/impl/LegacyReduceFloatOp.cpp +++ b/libnd4j/include/ops/declarable/impl/LegacyReduceFloatOp.cpp @@ -76,14 +76,25 @@ namespace sd { REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); + + if(x->rankOf() == z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); NativeOpExecutioner::execReduceFloat(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), - extras.argumentsAsT(z->dataType()), z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), - dims.data(), (int) dims.size(), reinterpret_cast(pTadShape), reinterpret_cast(pTadOffsets)); + extras.argumentsAsT(z->dataType()), z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, + dims2.data(), (int) dims2.size()); } @@ -109,16 +120,25 @@ namespace sd { // TAD REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); + + if(x->rankOf() == z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } + + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); NativeOpExecutioner::execReduceFloat(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), - extras.argumentsAsT(z->dataType()), z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), - dims.data(), (int) dims.size(), pTadShape, pTadOffsets); - - + extras.argumentsAsT(z->dataType()), z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, + dims2.data(), (int) dims2.size()); } } diff --git a/libnd4j/include/ops/declarable/impl/LegacyReduceLongOp.cpp b/libnd4j/include/ops/declarable/impl/LegacyReduceLongOp.cpp index f5007ff03..2e6a9f78e 100644 --- a/libnd4j/include/ops/declarable/impl/LegacyReduceLongOp.cpp +++ b/libnd4j/include/ops/declarable/impl/LegacyReduceLongOp.cpp @@ -78,15 +78,27 @@ namespace sd { REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + if(x->rankOf() - dims.size() != z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } - NativeOpExecutioner::execReduceLong(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), - extras.argumentsAsT(x->dataType()), - z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), - dims.data(), (int) dims.size(), pTadShape, pTadOffsets); + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); + NativeOpExecutioner::execReduceLong(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), nullptr, z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, dims2.data(), dims2.size()); + + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + // NativeOpExecutioner::execReduceLong(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), + // extras.argumentsAsT(x->dataType()), + // z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), + // dims.data(), (int) dims.size(), pTadShape, pTadOffsets); } STORE_RESULT(*z); @@ -111,13 +123,26 @@ namespace sd { // TAD REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + if(x->rankOf() - dims.size() != z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } - NativeOpExecutioner::execReduceLong(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), extras.argumentsAsT(x->dataType()), - z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), dims.data(), (int) dims.size(), pTadShape, pTadOffsets); + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); + NativeOpExecutioner::execReduceLong(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), nullptr, z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, dims2.data(), dims2.size()); + + + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + // NativeOpExecutioner::execReduceLong(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), extras.argumentsAsT(x->dataType()), + // z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), dims.data(), (int) dims.size(), pTadShape, pTadOffsets); } } diff --git a/libnd4j/include/ops/declarable/impl/LegacyReduceSameOp.cpp b/libnd4j/include/ops/declarable/impl/LegacyReduceSameOp.cpp index 299d19f14..e6dff7a20 100644 --- a/libnd4j/include/ops/declarable/impl/LegacyReduceSameOp.cpp +++ b/libnd4j/include/ops/declarable/impl/LegacyReduceSameOp.cpp @@ -73,15 +73,27 @@ namespace sd { REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + if(x->rankOf() - dims.size() != z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } - NativeOpExecutioner::execReduceSame(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), - extras.argumentsAsT(z->dataType()), - z->buffer(), z->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), - dims.data(), (int) dims.size(), pTadShape, pTadOffsets); + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); + NativeOpExecutioner::execReduceSame(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), nullptr, z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, dims2.data(), dims2.size()); + + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + // NativeOpExecutioner::execReduceSame(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), + // extras.argumentsAsT(z->dataType()), + // z->buffer(), z->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), + // dims.data(), (int) dims.size(), pTadShape, pTadOffsets); } STORE_RESULT(*z); @@ -106,14 +118,26 @@ namespace sd { // TAD REQUIRE_TRUE(dims.size() > 0, 0, "Some dimensions required for reduction!"); - auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + const Nd4jLong* zShapeInfoH = z->shapeInfo(); + const Nd4jLong* zShapeInfoD = z->specialShapeInfo(); - auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); - auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + if(x->rankOf() - dims.size() != z->rankOf()) { + auto zPack = ConstantShapeHelper::getInstance().createShapeInfoWithNoUnitiesForReduce(z->shapeInfo(), dims, z->getContext()->getWorkspace()); + zShapeInfoH = reinterpret_cast(zPack.primary()); + zShapeInfoD = reinterpret_cast(zPack.special()); + } - NativeOpExecutioner::execReduceSame(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), - extras.argumentsAsT(z->dataType()), z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), - dims.data(), (int) dims.size(), pTadShape, pTadOffsets); + std::vector dims2 = ShapeUtils::evalDimsForReduceOp(x->rankOf(), dims); + NativeOpExecutioner::execReduceSame(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), nullptr, z->buffer(), zShapeInfoH, z->specialBuffer(), zShapeInfoD, dims2.data(), dims2.size()); + + // auto packX = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims); + + // auto pTadShape = Environment::getInstance().isCPU() ? packX.primaryShapeInfo() : packX.specialShapeInfo(); //(Nd4jLong *) manager.replicatePointer(tad.tadOnlyShapeInfo, shape::shapeInfoByteLength(tad.tadOnlyShapeInfo)); + // auto pTadOffsets = Environment::getInstance().isCPU() ? packX.primaryOffsets() : packX.specialOffsets(); //(Nd4jLong *) manager.replicatePointer(tad.tadOffsets, tad.numTads * sizeof(Nd4jLong)); + + // NativeOpExecutioner::execReduceSame(block.launchContext(), opNum, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), + // extras.argumentsAsT(z->dataType()), z->buffer(), z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(), + // dims.data(), (int) dims.size(), pTadShape, pTadOffsets); } } diff --git a/libnd4j/include/ops/declarable/platform/cudnn/conv2d.cu b/libnd4j/include/ops/declarable/platform/cudnn/conv2d.cu index a77faf6f7..e19751a0e 100644 --- a/libnd4j/include/ops/declarable/platform/cudnn/conv2d.cu +++ b/libnd4j/include/ops/declarable/platform/cudnn/conv2d.cu @@ -81,8 +81,12 @@ static void conv2dCUDNN(const LaunchContext* context, // algorithm description cudnnConvolutionFwdAlgo_t algo; - err = cudnnGetConvolutionForwardAlgorithm(*handle, x, w, conv, z, CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo); - if (err != 0) throw sd::cuda_exception::build("conv2dCUDNN: cudnnGetConvolutionForwardAlgorithm failed", err); + cudnnConvolutionFwdAlgoPerf_t algoPerf; + int count = 0; + //err = cudnnGetConvolutionForwardAlgorithm(*handle, x, w, conv, z, CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo); + err = cudnnFindConvolutionForwardAlgorithm(*handle, x, w, conv, z, 1, &count, &algoPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("conv2dCUDNN: cudnnGetConvolutionForwardAlgorithm failed", err); + algo = algoPerf.algo; // allocate auxiliary device memory, abbreviation ws means workspace @@ -188,13 +192,20 @@ static void conv2dBpCUDNN(const LaunchContext* context, // gradW algorithm description cudnnConvolutionBwdFilterAlgo_t algoGradW; - err = cudnnGetConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST, 0, &algoGradW); - if (err != 0) throw sd::cuda_exception::build("conv2dBpCUDNN: cudnnGetConvolutionBackwardFilterAlgorithm failed", err); + cudnnConvolutionBwdFilterAlgoPerf_t algoGradWPerf; + int count = 0; + //err = cudnnGetConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST, 0, &algoGradW); + err = cudnnFindConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, 1, &count, &algoGradWPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("conv2dBpCUDNN: cudnnGetConvolutionBackwardFilterAlgorithm failed", err); + algoGradW = algoGradWPerf.algo; // gradI algorithm description cudnnConvolutionBwdDataAlgo_t algoGradI; - err = cudnnGetConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &algoGradI); - if (err != 0) throw sd::cuda_exception::build("conv2dBpCUDNN: cudnnGetConvolutionBackwardDataAlgorithm failed", err); + cudnnConvolutionBwdDataAlgoPerf_t algoGradIPerf; + //err = cudnnGetConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &algoGradI); + err = cudnnFindConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, 1, &count, &algoGradIPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("conv2dBpCUDNN: cudnnGetConvolutionBackwardDataAlgorithm failed", err); + algoGradI = algoGradIPerf.algo; // allocate auxiliary device memory for gradW calculation, abbreviation ws means workspace size_t wsGradWSize; diff --git a/libnd4j/include/ops/declarable/platform/cudnn/conv3d.cu b/libnd4j/include/ops/declarable/platform/cudnn/conv3d.cu index 693ebeefa..f11a590c2 100644 --- a/libnd4j/include/ops/declarable/platform/cudnn/conv3d.cu +++ b/libnd4j/include/ops/declarable/platform/cudnn/conv3d.cu @@ -94,8 +94,13 @@ static void conv3dCUDNN(const LaunchContext* context, // algorithm description cudnnConvolutionFwdAlgo_t algo; - err = cudnnGetConvolutionForwardAlgorithm(*handle, x, w, conv, z, CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo); - if (err != 0) throw sd::cuda_exception::build("conv3dCUDNN: cudnnGetConvolutionForwardAlgorithm failed", err); + cudnnConvolutionFwdAlgoPerf_t algoPerf; + int count = 0; + //err = cudnnGetConvolutionForwardAlgorithm(*handle, x, w, conv, z, CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo); + err = cudnnFindConvolutionForwardAlgorithm(*handle, x, w, conv, z, 1, &count, &algoPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("conv3dCUDNN: cudnnGetConvolutionForwardAlgorithm failed", err); + algo = algoPerf.algo; + // allocate auxiliary device memory, abbreviation ws means workspace size_t wsSize; @@ -217,13 +222,20 @@ static void conv3dBpCUDNN(const LaunchContext* context, // gradW algorithm description cudnnConvolutionBwdFilterAlgo_t algoGradW; - err = cudnnGetConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST, 0, &algoGradW); - if (err != 0) throw sd::cuda_exception::build("conv3dBpCUDNN: cudnnGetConvolutionBackwardFilterAlgorithm failed", err); + cudnnConvolutionBwdFilterAlgoPerf_t algoGradWPerf; + int count = 0; + //err = cudnnGetConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST, 0, &algoGradW); + err = cudnnFindConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, 1, &count, &algoGradWPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("conv3dBpCUDNN: cudnnGetConvolutionBackwardFilterAlgorithm failed", err); + algoGradW = algoGradWPerf.algo; // gradI algorithm description cudnnConvolutionBwdDataAlgo_t algoGradI; - err = cudnnGetConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &algoGradI); - if (err != 0) throw sd::cuda_exception::build("conv3dBpCUDNN: cudnnGetConvolutionBackwardDataAlgorithm failed", err); + cudnnConvolutionBwdDataAlgoPerf_t algoGradIPerf; + //err = cudnnGetConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &algoGradI); + err = cudnnFindConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, 1, &count, &algoGradIPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("conv3dBpCUDNN: cudnnGetConvolutionBackwardDataAlgorithm failed", err); + algoGradI = algoGradIPerf.algo; // allocate auxiliary device memory for gradW calculation, abbreviation ws means workspace size_t wsGradWSize; diff --git a/libnd4j/include/ops/declarable/platform/cudnn/depthwiseConv2d.cu b/libnd4j/include/ops/declarable/platform/cudnn/depthwiseConv2d.cu index c268961ce..a1f408cc5 100644 --- a/libnd4j/include/ops/declarable/platform/cudnn/depthwiseConv2d.cu +++ b/libnd4j/include/ops/declarable/platform/cudnn/depthwiseConv2d.cu @@ -89,8 +89,12 @@ static void depthwiseConv2dCUDNN(const LaunchContext* context, // algorithm description cudnnConvolutionFwdAlgo_t algo; - err = cudnnGetConvolutionForwardAlgorithm(*handle, x, w, conv, z, CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo); - if (err != 0) throw sd::cuda_exception::build("depthwiseConv2dCUDNN: cudnnGetConvolutionForwardAlgorithm failed", err); + cudnnConvolutionFwdAlgoPerf_t algoPerf; + int count = 0; + //err = cudnnGetConvolutionForwardAlgorithm(*handle, x, w, conv, z, CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo); + err = cudnnFindConvolutionForwardAlgorithm(*handle, x, w, conv, z, 1, &count, &algoPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("depthwiseConv2dCUDNN: cudnnGetConvolutionForwardAlgorithm failed", err); + algo = algoPerf.algo; // allocate auxiliary device memory, abbreviation ws means workspace size_t wsSize; @@ -206,13 +210,20 @@ static void depthwiseConv2dBpCUDNN(const LaunchContext* context, // gradW algorithm description cudnnConvolutionBwdFilterAlgo_t algoGradW; - err = cudnnGetConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST, 0, &algoGradW); - if (err != 0) throw sd::cuda_exception::build("depthwiseConv2dBpCUDNN: cudnnGetConvolutionBackwardFilterAlgorithm failed", err); + cudnnConvolutionBwdFilterAlgoPerf_t algoGradWPerf; + int count = 0; + //err = cudnnGetConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST, 0, &algoGradW); + err = cudnnFindConvolutionBackwardFilterAlgorithm(*handle, x, dz, conv, dw, 1, &count, &algoGradWPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("depthwiseConv2dBpCUDNN: cudnnGetConvolutionBackwardFilterAlgorithm failed", err); + algoGradW = algoGradWPerf.algo; // gradI algorithm description cudnnConvolutionBwdDataAlgo_t algoGradI; - err = cudnnGetConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &algoGradI); - if (err != 0) throw sd::cuda_exception::build("depthwiseConv2dBpCUDNN: cudnnGetConvolutionBackwardDataAlgorithm failed", err); + cudnnConvolutionBwdDataAlgoPerf_t algoGradIPerf; + //err = cudnnGetConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &algoGradI); + err = cudnnFindConvolutionBackwardDataAlgorithm(*handle, dw, dz, conv, x, 1, &count, &algoGradIPerf); + if (err != 0 || count == 0) throw sd::cuda_exception::build("depthwiseConv2dBpCUDNN: cudnnGetConvolutionBackwardDataAlgorithm failed", err); + algoGradI = algoGradIPerf.algo; // allocate auxiliary device memory for gradW calculation, abbreviation ws means workspace size_t wsGradWSize; diff --git a/libnd4j/include/system/pointercast.h b/libnd4j/include/system/pointercast.h index 2c64d608e..778667402 100644 --- a/libnd4j/include/system/pointercast.h +++ b/libnd4j/include/system/pointercast.h @@ -24,6 +24,8 @@ #include #include +#define CUDA_BLOCK_SIZE 256 + typedef void* Nd4jPointer; typedef long long Nd4jLong; typedef uint64_t Nd4jULong; diff --git a/libnd4j/pom.xml b/libnd4j/pom.xml index d1d9944fa..1819d8112 100644 --- a/libnd4j/pom.xml +++ b/libnd4j/pom.xml @@ -55,8 +55,8 @@ - 10.2 - 7.6 + 11.0 + 8.0 release cpu ${javacpp.platform} diff --git a/libnd4j/tests_cpu/layers_tests/ConvolutionTests1.cpp b/libnd4j/tests_cpu/layers_tests/ConvolutionTests1.cpp index b87985458..8eb294e7c 100644 --- a/libnd4j/tests_cpu/layers_tests/ConvolutionTests1.cpp +++ b/libnd4j/tests_cpu/layers_tests/ConvolutionTests1.cpp @@ -2849,9 +2849,68 @@ TYPED_TEST(TypedConvolutionTests1, deconv2d_tf_test1) { ASSERT_EQ(Status::OK(), results.status()); ASSERT_TRUE(exp.isSameShape(output)); ASSERT_TRUE(exp.equalsTo(output)); - } +////////////////////////////////////////////////////////////////////// +TEST_F(ConvolutionTests1, conv2d_bp_7) { + int bS=2, iH=12,iW=12, iC=3,oC=3, kH=3,kW=3, sH=2,sW=2, pH=0,pW=0, dH=1,dW=1; + int oH=6,oW=6; + int paddingMode = 1; // 1-SAME, 0-VALID; + int dataFormat = 0; // 1-NHWC, 0-NCHW + + NDArray input('c', {bS, iC, iH, iW}, sd::DataType::FLOAT32); + NDArray weights('c', {kH, kW, iC, oC}, sd::DataType::FLOAT32); + NDArray bias('c', {oC}, {1,2,3}, sd::DataType::FLOAT32); + NDArray gradO('c', {bS, oC, oH, oW}, sd::DataType::FLOAT32); + NDArray gradI('c', {bS, iC, iH, iW}, sd::DataType::FLOAT32); + NDArray gradW('c', {kH, kW, iC, oC}, sd::DataType::FLOAT32); + NDArray gradB('c', {oC}, sd::DataType::FLOAT32); + input = 2.; + weights.linspace(0.1, 0.1); + gradO.linspace(0.01, 0.01); + + sd::ops::conv2d_bp op; + auto status = op.execute({&input, &weights, &bias, &gradO}, {&gradI, &gradW, &gradB}, {}, {kH,kW, sH,sW, pH,pW, dH,dW, paddingMode, dataFormat}, {}); + ASSERT_EQ(Status::OK(), status); +} + +////////////////////////////////////////////////////////////////////// +TEST_F(ConvolutionTests1, conv2d_ff_119_1) { + auto i = NDArrayFactory::create('c', {2, 3, 13, 13}); + auto w = NDArrayFactory::create('c', {3, 3, 3, 3}); + auto b = NDArrayFactory::create('c', {3}); + auto o = NDArrayFactory::create('c', {2, 3, 6, 6}); + + sd::ops::conv2d op_ff; + auto status = op_ff.execute({&i, &w, &b}, {&o}, {3,3, 2,2, 0,0, 1,1, 0, 0, 1}); + ASSERT_EQ(Status::OK(), status); + + auto gi = i.ulike(); + auto gw = w.ulike(); + + sd::ops::conv2d_bp op_bp; + status = op_bp.execute({&i, &w, &b, &o}, {&gi, &gw}, {3,3, 2,2, 0,0, 1,1, 0, 0, 1}); + ASSERT_EQ(Status::OK(), status); +} + +////////////////////////////////////////////////////////////////////// +TEST_F(ConvolutionTests1, conv2d_ff_119_2) { + auto i = NDArrayFactory::create('c', {2, 3, 17, 17}); + auto w = NDArrayFactory::create('c', {3, 3, 3, 3}); + auto b = NDArrayFactory::create('c', {3}); + auto o = NDArrayFactory::create('c', {2, 3, 8, 8}); + + sd::ops::conv2d op_ff; + auto status = op_ff.execute({&i, &w, &b}, {&o}, {3,3, 2,2, 0,0, 1,1, 0, 0, 1}); + ASSERT_EQ(Status::OK(), status); + + auto gi = i.ulike(); + auto gw = w.ulike(); + + sd::ops::conv2d_bp op_bp; + status = op_bp.execute({&i, &w, &b, &o}, {&gi, &gw}, {3,3, 2,2, 0,0, 1,1, 0, 0, 1}); + ASSERT_EQ(Status::OK(), status); +} #endif //LIBND4J_CONVOLUTIONTESTS1_H diff --git a/libnd4j/tests_cpu/layers_tests/CudaBasicsTests1.cu b/libnd4j/tests_cpu/layers_tests/CudaBasicsTests1.cu index 3d6886565..cb7806bd4 100644 --- a/libnd4j/tests_cpu/layers_tests/CudaBasicsTests1.cu +++ b/libnd4j/tests_cpu/layers_tests/CudaBasicsTests1.cu @@ -35,6 +35,8 @@ #include #include #include +#include +#include using namespace sd; using namespace sd::graph; @@ -46,7 +48,7 @@ public: ////////////////////////////////////////////////////////////////////////// -static cudaError_t allocateDeviceMem(LaunchContext& lc, std::vector& devicePtrs, const std::vector>& hostData) { +static cudaError_t allocateDeviceMem(LaunchContext& lc, std::vector& devicePtrs, const std::vector>& hostData) { if(devicePtrs.size() != hostData.size()) throw std::invalid_argument("prepareDataForCuda: two input sts::vectors should same sizes !"); @@ -63,9 +65,9 @@ static cudaError_t allocateDeviceMem(LaunchContext& lc, std::vector& devi cudaStream_t stream = *lc.getCudaStream(); for(int i = 0; i < devicePtrs.size(); ++i) { - + cudaResult = cudaMalloc(reinterpret_cast(&devicePtrs[i]), hostData[i].second); if(cudaResult != 0) return cudaResult; - cudaMemcpyAsync(devicePtrs[i], hostData[i].first, hostData[i].second, cudaMemcpyHostToDevice, stream); + cudaMemcpyAsync(devicePtrs[i], hostData[i].first, hostData[i].second, cudaMemcpyHostToDevice, stream); } return cudaResult; } @@ -99,7 +101,7 @@ TEST_F(CudaBasicsTests1, TestPairwise_1) { cudaMemcpyAsync(devShapePtrX, x.shapeInfo(), shape::shapeInfoByteLength(x.shapeInfo()), cudaMemcpyHostToDevice, *stream); res = cudaStreamSynchronize(*stream); ASSERT_EQ(0, res); - + LaunchContext lc(stream, nullptr, nullptr); NativeOpExecutioner::execPairwiseTransform(&lc, pairwise::Add, nullptr, x.shapeInfo(), devBufferPtrX, reinterpret_cast(devShapePtrX), nullptr, x.shapeInfo(), devBufferPtrX, reinterpret_cast(devShapePtrX), nullptr, z.shapeInfo(), devBufferPtrZ, reinterpret_cast(devShapePtrX), nullptr); res = cudaStreamSynchronize(*stream); @@ -131,21 +133,21 @@ TEST_F(CudaBasicsTests1, execIndexReduceScalar_1) { NDArray x1('c', {2,2}, {0, 1, 2, 3}, sd::DataType::INT32); NDArray x2('c', {2,2}, {0.5, 1.5, -4.5, 3.5}, sd::DataType::BFLOAT16); NDArray x3('c', {2,2}, {0, -1, 0, 1}, sd::DataType::BOOL); - + NDArray scalar('c', {}, std::vector{0}, sd::DataType::INT64); NDArray exp1('c', {}, std::vector{3}, sd::DataType::INT64); NDArray exp2('c', {}, std::vector{2}, sd::DataType::INT64); NDArray exp3('c', {}, std::vector{1}, sd::DataType::INT64); - void *dX1, *dX2, *dX3, *dZ; + void *dX1, *dX2, *dX3, *dZ; Nd4jLong *dX1ShapeInfo, *dX2ShapeInfo, *dX3ShapeInfo, *dZShapeInfo; cudaError_t cudaResult; cudaResult = cudaMalloc(reinterpret_cast(&dX1), x1.lengthOf() * x1.sizeOfT()); ASSERT_EQ(0, cudaResult); - cudaResult = cudaMalloc(reinterpret_cast(&dX2), x2.lengthOf() * x2.sizeOfT()); ASSERT_EQ(0, cudaResult); - cudaResult = cudaMalloc(reinterpret_cast(&dX3), x3.lengthOf() * x3.sizeOfT()); ASSERT_EQ(0, cudaResult); + cudaResult = cudaMalloc(reinterpret_cast(&dX2), x2.lengthOf() * x2.sizeOfT()); ASSERT_EQ(0, cudaResult); + cudaResult = cudaMalloc(reinterpret_cast(&dX3), x3.lengthOf() * x3.sizeOfT()); ASSERT_EQ(0, cudaResult); cudaResult = cudaMalloc(reinterpret_cast(&dZ), scalar.lengthOf() * scalar.sizeOfT()); ASSERT_EQ(0, cudaResult); cudaResult = cudaMalloc(reinterpret_cast(&dX1ShapeInfo), shape::shapeInfoByteLength(x1.shapeInfo())); ASSERT_EQ(0, cudaResult); cudaResult = cudaMalloc(reinterpret_cast(&dX2ShapeInfo), shape::shapeInfoByteLength(x2.shapeInfo())); ASSERT_EQ(0, cudaResult); @@ -153,14 +155,14 @@ TEST_F(CudaBasicsTests1, execIndexReduceScalar_1) { cudaResult = cudaMalloc(reinterpret_cast(&dZShapeInfo), shape::shapeInfoByteLength(scalar.shapeInfo())); ASSERT_EQ(0, cudaResult); cudaStream_t stream; - cudaResult = cudaStreamCreate(&stream); + cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); x1.syncToHost(); x2.syncToHost(); x3.syncToHost(); scalar.syncToHost(); - + cudaMemcpyAsync(dX1, x1.buffer(), x1.lengthOf() * x1.sizeOfT(), cudaMemcpyHostToDevice, stream); cudaMemcpyAsync(dX2, x2.buffer(), x2.lengthOf() * x2.sizeOfT(), cudaMemcpyHostToDevice, stream); cudaMemcpyAsync(dX3, x3.buffer(), x3.lengthOf() * x3.sizeOfT(), cudaMemcpyHostToDevice, stream); @@ -168,7 +170,7 @@ TEST_F(CudaBasicsTests1, execIndexReduceScalar_1) { cudaMemcpyAsync(dX2ShapeInfo, x2.shapeInfo(), shape::shapeInfoByteLength(x2.shapeInfo()), cudaMemcpyHostToDevice, stream); cudaMemcpyAsync(dX3ShapeInfo, x3.shapeInfo(), shape::shapeInfoByteLength(x3.shapeInfo()), cudaMemcpyHostToDevice, stream); cudaMemcpyAsync(dZShapeInfo, scalar.shapeInfo(), shape::shapeInfoByteLength(scalar.shapeInfo()), cudaMemcpyHostToDevice, stream); - + void* reductionPointer = nullptr; cudaResult = cudaMalloc(reinterpret_cast(&reductionPointer), 1024*1024); ASSERT_EQ(0, cudaResult); @@ -178,21 +180,21 @@ TEST_F(CudaBasicsTests1, execIndexReduceScalar_1) { LaunchContext lc(&stream, LaunchContext::defaultContext()->getReductionPointer(), LaunchContext::defaultContext()->getScalarPointer(), LaunchContext::defaultContext()->getAllocationPointer()); /***************************************/ - + NativeOpExecutioner::execIndexReduceScalar(&lc, sd::indexreduce::IndexAbsoluteMax, x1.buffer(), x1.shapeInfo(), - dX1, dX1ShapeInfo, - nullptr, + dX1, dX1ShapeInfo, + nullptr, scalar.buffer(), scalar.shapeInfo(), dZ, dZShapeInfo); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); cudaMemcpyAsync(scalar.buffer(), dZ, scalar.lengthOf() * scalar.sizeOfT(), cudaMemcpyDeviceToHost, stream); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); scalar.tickWriteHost(); @@ -200,55 +202,55 @@ TEST_F(CudaBasicsTests1, execIndexReduceScalar_1) { ASSERT_NEAR(exp1.e(0), scalar.e(0), 1e-5); /***************************************/ - + NativeOpExecutioner::execIndexReduceScalar(&lc, sd::indexreduce::IndexAbsoluteMax, nullptr, x2.shapeInfo(), - dX2, dX2ShapeInfo, - nullptr, + dX2, dX2ShapeInfo, + nullptr, nullptr, scalar.shapeInfo(), dZ, dZShapeInfo); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); cudaMemcpyAsync(scalar.buffer(), dZ, scalar.lengthOf() * scalar.sizeOfT(), cudaMemcpyDeviceToHost, stream); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); ASSERT_NEAR(exp2.e(0), scalar.e(0), 1e-5); // ************************************* - NativeOpExecutioner::execIndexReduceScalar(&lc, + NativeOpExecutioner::execIndexReduceScalar(&lc, sd::indexreduce::IndexAbsoluteMax, nullptr, x3.shapeInfo(), - dX3, dX3ShapeInfo, - nullptr, + dX3, dX3ShapeInfo, + nullptr, nullptr, scalar.shapeInfo(), dZ, dZShapeInfo); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); cudaMemcpyAsync(scalar.buffer(), dZ, scalar.lengthOf() * scalar.sizeOfT(), cudaMemcpyDeviceToHost, stream); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); ASSERT_NEAR(exp3.e(0), scalar.e(0), 1e-5); - + /***************************************/ cudaFree(dX1); cudaFree(dX2); cudaFree(dX3); cudaFree(dZ); - cudaFree(dX1ShapeInfo); cudaFree(dX2ShapeInfo); cudaFree(dX3ShapeInfo); cudaFree(dZShapeInfo); + cudaFree(dX1ShapeInfo); cudaFree(dX2ShapeInfo); cudaFree(dX3ShapeInfo); cudaFree(dZShapeInfo); - /***************************************/ + /***************************************/ - cudaResult = cudaStreamDestroy(stream); + cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); - + } //////////////////////////////////////////////////////////////////////////// @@ -264,11 +266,11 @@ TEST_F(CudaBasicsTests1, execReduce3Scalar_1) { NDArray exp1('c', {}, std::vector{-30.f}, sd::DataType::FLOAT32); NDArray exp2('c', {}, std::vector{15.}, sd::DataType::DOUBLE); - + NDArray scalar1('c', {}, std::vector{100.f}, sd::DataType::FLOAT32); NDArray scalar2('c', {}, std::vector{100.}, sd::DataType::DOUBLE); - void *dX1, *dX2, *dX3, *dX4, *dZ1, *dZ2; + void *dX1, *dX2, *dX3, *dX4, *dZ1, *dZ2; Nd4jLong *dX1ShapeInfo, *dX3ShapeInfo, *dZ1ShapeInfo, *dZ2ShapeInfo; cudaError_t cudaResult; @@ -285,7 +287,7 @@ TEST_F(CudaBasicsTests1, execReduce3Scalar_1) { cudaResult = cudaMalloc(reinterpret_cast(&dZ2ShapeInfo), shape::shapeInfoByteLength(scalar2.shapeInfo())); ASSERT_EQ(0, cudaResult); cudaStream_t stream; - cudaResult = cudaStreamCreate(&stream); + cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); x1.syncToHost(); @@ -294,7 +296,7 @@ TEST_F(CudaBasicsTests1, execReduce3Scalar_1) { x4.syncToHost(); scalar1.syncToHost(); scalar2.syncToHost(); - + cudaMemcpyAsync(dX1, x1.buffer(), x1.lengthOf() * x1.sizeOfT(), cudaMemcpyHostToDevice, stream); cudaMemcpyAsync(dX2, x2.buffer(), x2.lengthOf() * x2.sizeOfT(), cudaMemcpyHostToDevice, stream); cudaMemcpyAsync(dX3, x3.buffer(), x3.lengthOf() * x3.sizeOfT(), cudaMemcpyHostToDevice, stream); @@ -307,7 +309,7 @@ TEST_F(CudaBasicsTests1, execReduce3Scalar_1) { /***************************************/ void* reductionPointer = nullptr; - int* allocationPointer = nullptr; + int* allocationPointer = nullptr; cudaResult = cudaMalloc(reinterpret_cast(&reductionPointer), 1024*1024); ASSERT_EQ(0, cudaResult); cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); @@ -315,10 +317,10 @@ TEST_F(CudaBasicsTests1, execReduce3Scalar_1) { LaunchContext lc(&stream, reductionPointer, nullptr, allocationPointer); /***************************************/ - + NativeOpExecutioner::execReduce3Scalar(&lc, sd::reduce3::Dot,nullptr, x1.shapeInfo(),dX1, dX1ShapeInfo, nullptr, nullptr, x2.shapeInfo(),dX2, dX1ShapeInfo,nullptr, scalar1.shapeInfo(),dZ1, dZ1ShapeInfo); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); scalar1.tickWriteHost(); @@ -326,36 +328,36 @@ TEST_F(CudaBasicsTests1, execReduce3Scalar_1) { cudaMemcpyAsync(scalar1.buffer(), dZ1, scalar1.lengthOf() * scalar1.sizeOfT(), cudaMemcpyDeviceToHost, stream); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); ASSERT_NEAR(exp1.e(0), scalar1.e(0), 1e-5); /***************************************/ - + NativeOpExecutioner::execReduce3Scalar(&lc, sd::reduce3::Dot,nullptr, x3.shapeInfo(),dX3, dX3ShapeInfo, nullptr, nullptr, x4.shapeInfo(),dX4, dX3ShapeInfo,nullptr, scalar2.shapeInfo(),dZ2, dZ2ShapeInfo); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); cudaMemcpyAsync(scalar2.buffer(), dZ2, scalar2.lengthOf() * scalar2.sizeOfT(), cudaMemcpyDeviceToHost, stream); - cudaResult = cudaStreamSynchronize(stream); + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); ASSERT_NEAR(exp2.e(0), scalar2.e(0), 1e-5); - + /***************************************/ cudaFree(dX1); cudaFree(dX2); cudaFree(dX3); cudaFree(dX4); cudaFree(dZ1); cudaFree(dZ2); cudaFree(dX1ShapeInfo); cudaFree(dX3ShapeInfo); cudaFree(dZ1ShapeInfo); cudaFree(dZ2ShapeInfo); - /***************************************/ + /***************************************/ - cudaResult = cudaStreamDestroy(stream); + cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); } - + //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3_1) { @@ -372,9 +374,9 @@ TEST_F(CudaBasicsTests1, execReduce3_1) { y.syncToHost(); z.syncToHost(); - - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions + + std::vector> hostData; + hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions std::vector devicePtrs(hostData.size(), nullptr); cudaError_t cudaResult; @@ -382,27 +384,27 @@ TEST_F(CudaBasicsTests1, execReduce3_1) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), nullptr, nullptr, nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -411,18 +413,18 @@ TEST_F(CudaBasicsTests1, execReduce3_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3_2) { - + NDArray x('c', {2,2}, {1.5,1.5,1.5,1.5}, sd::DataType::DOUBLE); NDArray y('c', {2,2}, {1,2,3,4}, sd::DataType::DOUBLE); NDArray exp('c', {}, std::vector{15.}, sd::DataType::DOUBLE); NDArray z('c', {}, std::vector{100.}, sd::DataType::DOUBLE); - - std::vector dimensions = {0, 1}; - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions + std::vector dimensions = {0, 1}; + + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; + hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -431,28 +433,28 @@ TEST_F(CudaBasicsTests1, execReduce3_2) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - - // call cuda kernel which calculates result + + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), nullptr, nullptr, nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -460,16 +462,16 @@ TEST_F(CudaBasicsTests1, execReduce3_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3_3) { - + NDArray x('c', {2,3}, {1,2,3,4,5,6}, sd::DataType::INT32); NDArray y('c', {2,3}, {-6,-5,-4,-3,-2,-1}, sd::DataType::INT32); NDArray exp('c', {3}, {-18,-20,-18}, sd::DataType::FLOAT32); NDArray z('c', {3}, {100,100,100}, sd::DataType::FLOAT32); - + std::vector dimensions = {0}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); @@ -481,13 +483,13 @@ TEST_F(CudaBasicsTests1, execReduce3_3) { yTad.createTadOnlyShapeInfo(); yTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets hostData.emplace_back(yTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(yTad.tadOnlyShapeInfo));// 3 -- yTadShapeInfo - hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets + hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -496,29 +498,29 @@ TEST_F(CudaBasicsTests1, execReduce3_3) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - - // call cuda kernel which calculates result + + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], (Nd4jLong*)devicePtrs[3], (Nd4jLong*)devicePtrs[4]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -526,16 +528,16 @@ TEST_F(CudaBasicsTests1, execReduce3_3) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3_4) { - + NDArray x('c', {2,3}, {1,2,3,4,5,6}, sd::DataType::DOUBLE); NDArray y('c', {2,3}, {1.5,1.5,1.5,1.5,1.5,1.5}, sd::DataType::DOUBLE); NDArray exp('c', {2}, {9,22.5}, sd::DataType::DOUBLE); NDArray z('c', {2}, {100,100}, sd::DataType::DOUBLE); - + std::vector dimensions = {1}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); @@ -548,12 +550,12 @@ TEST_F(CudaBasicsTests1, execReduce3_4) { yTad.createOffsets(); // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets hostData.emplace_back(yTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(yTad.tadOnlyShapeInfo));// 3 -- yTadShapeInfo - hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets + hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -562,29 +564,29 @@ TEST_F(CudaBasicsTests1, execReduce3_4) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - - // call cuda kernel which calculates result + + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], (Nd4jLong*)devicePtrs[3], (Nd4jLong*)devicePtrs[4]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -592,16 +594,16 @@ TEST_F(CudaBasicsTests1, execReduce3_4) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3_5) { - + NDArray x('c', {2,2,3}, {1.5,1.5,1.5,1.5,1.5,1.5,1.5,1.5,1.5,1.5,1.5,1.5}, sd::DataType::FLOAT32); NDArray y('c', {2,2,3}, {1,2,3,4,5,6,7,8,9,10,11,12}, sd::DataType::FLOAT32); NDArray exp('c', {2,3}, {7.5, 10.5, 13.5, 25.5, 28.5, 31.5}, sd::DataType::FLOAT32); NDArray z('c', {2,3}, {100,100,100,100,100,100}, sd::DataType::FLOAT32); - + std::vector dimensions = {1}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); @@ -614,12 +616,12 @@ TEST_F(CudaBasicsTests1, execReduce3_5) { yTad.createOffsets(); // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets hostData.emplace_back(yTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(yTad.tadOnlyShapeInfo));// 3 -- yTadShapeInfo - hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets + hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -628,29 +630,29 @@ TEST_F(CudaBasicsTests1, execReduce3_5) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], (Nd4jLong*)devicePtrs[3], (Nd4jLong*)devicePtrs[4]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -658,16 +660,16 @@ TEST_F(CudaBasicsTests1, execReduce3_5) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3All_1) { - + NDArray x('c', {2,2}, {1,2,3,4}, sd::DataType::INT32); NDArray y('c', {2,3}, {-1,1,-1,1,-1,1}, sd::DataType::INT32); NDArray exp('c', {2,3}, {2,-2,2,2,-2,2}, sd::DataType::FLOAT32); NDArray z('c', {2,3}, {100,100,100,100,100,100}, sd::DataType::FLOAT32); - + std::vector dimensions = {0}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); @@ -679,13 +681,13 @@ TEST_F(CudaBasicsTests1, execReduce3All_1) { yTad.createTadOnlyShapeInfo(); yTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets hostData.emplace_back(yTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(yTad.tadOnlyShapeInfo));// 3 -- yTadShapeInfo - hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4 -- yTadOffsets + hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4 -- yTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -694,29 +696,29 @@ TEST_F(CudaBasicsTests1, execReduce3All_1) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3All(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], (Nd4jLong*)devicePtrs[3], (Nd4jLong*)devicePtrs[4]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -724,16 +726,16 @@ TEST_F(CudaBasicsTests1, execReduce3All_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3All_2) { - + NDArray x('c', {2,2}, {1,2,3,4}, sd::DataType::DOUBLE); NDArray y('c', {2,3}, {1.5,1.5,1.5,1.5,1.5,1.5}, sd::DataType::DOUBLE); NDArray exp('c', {2,3}, {6,6,6,9,9,9}, sd::DataType::DOUBLE); NDArray z('c', {2,3}, {100,100,100,100,100,100,},sd::DataType::DOUBLE); - + std::vector dimensions = {0}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); @@ -745,13 +747,13 @@ TEST_F(CudaBasicsTests1, execReduce3All_2) { yTad.createTadOnlyShapeInfo(); yTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets hostData.emplace_back(yTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(yTad.tadOnlyShapeInfo));// 3 -- yTadShapeInfo - hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets + hostData.emplace_back(yTad.tadOffsets, yTad.numTads * sizeof(Nd4jLong)); // 4-- yTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -760,29 +762,29 @@ TEST_F(CudaBasicsTests1, execReduce3All_2) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3All(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], (Nd4jLong*)devicePtrs[3], (Nd4jLong*)devicePtrs[4]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -790,22 +792,22 @@ TEST_F(CudaBasicsTests1, execReduce3All_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execIndexReduce_1) { - + NDArray x('c', {2,3}, {100,100,100,100,100,100}, sd::DataType::DOUBLE); x.linspace(-2.); x.syncToDevice(); NDArray exp('c', {2}, {2, 2}, sd::DataType::INT64); NDArray z('c', {2}, {100,100}, sd::DataType::INT64); - - std::vector dimensions = {1}; - // evaluate xTad data + std::vector dimensions = {1}; + + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets @@ -817,28 +819,31 @@ TEST_F(CudaBasicsTests1, execIndexReduce_1) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execIndexReduce(&lc, sd::indexreduce::IndexMax, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); - cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamSynchronize(stream); + if (cudaResult != 0) + throw sd::cuda_exception::build("execIndexReduce failed", cudaResult); + z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -846,7 +851,7 @@ TEST_F(CudaBasicsTests1, execIndexReduce_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execIndexReduce_2) { - + NDArray x('c', {2,3,4,5}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100, 100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100, 100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100, @@ -856,18 +861,18 @@ TEST_F(CudaBasicsTests1, execIndexReduce_2) { x.linspace(-2.f); x.syncToDevice(); NDArray exp('c', {2,5}, {11,11,11,11,11,11,11,11,11,11}, sd::DataType::INT64); NDArray z('c', {2,5}, {100,100,100,100,100,100,100,100,100,100}, sd::DataType::INT64); - - std::vector dimensions = {1,2}; - // evaluate xTad data + std::vector dimensions = {1,2}; + + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets @@ -879,28 +884,28 @@ TEST_F(CudaBasicsTests1, execIndexReduce_2) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execIndexReduce(&lc, sd::indexreduce::IndexMax, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -908,7 +913,7 @@ TEST_F(CudaBasicsTests1, execIndexReduce_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execIndexReduce_3) { - + NDArray x('c', {2,3,4,5}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100, 100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100, 100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100, @@ -918,50 +923,50 @@ TEST_F(CudaBasicsTests1, execIndexReduce_3) { x.linspace(-2.); x.syncToDevice(); NDArray exp('c', {3}, {39, 39, 39}, sd::DataType::INT64); NDArray z('c', {3}, {100,100,100}, sd::DataType::INT64); - + std::vector dimensions = {0,2,3}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execIndexReduce(&lc, sd::indexreduce::IndexMax, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -972,18 +977,18 @@ TEST_F(CudaBasicsTests1, execScalar_1) { if (!Environment::getInstance().isExperimentalBuild()) return; - + NDArray x('c', {2,3}, {0,1,2,3,4,5}, sd::DataType::INT64); NDArray exp('c',{2,3}, {0,0,1,1,2,2}, sd::DataType::INT64); NDArray scalar('c',{}, std::vector{2.f}, sd::DataType::FLOAT32); NDArray z('c', {2,3}, {100,100,100,100,100,100}, sd::DataType::INT64); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - + // call cuda kernel which calculates result NativeOpExecutioner::execScalar(&lc, sd::scalar::Divide, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), @@ -993,9 +998,9 @@ TEST_F(CudaBasicsTests1, execScalar_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1007,18 +1012,18 @@ TEST_F(CudaBasicsTests1, execScalar_2) { if (!Environment::getInstance().isExperimentalBuild()) return; - + NDArray x('c', {2,3}, {-1,-2,-3,-4,-5,-6}, sd::DataType::INT64); NDArray exp('c',{2,3}, {10,10,10,10,10,10}, sd::DataType::FLOAT32); NDArray scalar('c',{}, std::vector{10.f}, sd::DataType::FLOAT32); NDArray z('c', {2,3}, {100,100,100,100,100,100}, sd::DataType::FLOAT32); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - + // call cuda kernel which calculates result NativeOpExecutioner::execScalar(&lc, sd::scalar::CopyPws, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), @@ -1028,9 +1033,9 @@ TEST_F(CudaBasicsTests1, execScalar_2) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); @@ -1043,7 +1048,7 @@ TEST_F(CudaBasicsTests1, execScalar_3) { if (!Environment::getInstance().isExperimentalBuild()) return; - + NDArray x('c', {2,3,2}, {0,1,2,3,4,5,6,7,8,9,10,11}, sd::DataType::INT64); NDArray scalars('c',{2,2}, {1,2,3,4}, sd::DataType::FLOAT32); NDArray exp('c', {2,3,2}, {0,0,2,1,4,2, 2,1,2,2,3,2}, sd::DataType::INT64); @@ -1051,49 +1056,49 @@ TEST_F(CudaBasicsTests1, execScalar_3) { std::vector dimensions = {1}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo)); // 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execScalar(&lc, sd::scalar::Divide, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, scalars.shapeInfo(), scalars.specialBuffer(), scalars.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1101,18 +1106,18 @@ TEST_F(CudaBasicsTests1, execScalar_3) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execScalarBool_1) { - + NDArray x('c', {2,3}, {-1,-2,0,1,2,3}, sd::DataType::BFLOAT16); NDArray scalar('c',{}, std::vector{0}, sd::DataType::BFLOAT16); NDArray exp('c',{2,3}, {0,0,0,1,1,1}, sd::DataType::BOOL); NDArray z('c', {2,3}, {100,100,100,100,100,100,}, sd::DataType::BOOL); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - + // call cuda kernel which calculates result // call cuda kernel which calculates result NativeOpExecutioner::execScalarBool(&lc, sd::scalar::GreaterThan, @@ -1123,9 +1128,9 @@ TEST_F(CudaBasicsTests1, execScalarBool_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1134,7 +1139,7 @@ TEST_F(CudaBasicsTests1, execScalarBool_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execScalarBool_2) { - + NDArray x('c', {2,3}, {0,1,2,3,4,5}, sd::DataType::FLOAT32); NDArray scalars('c',{2}, {-1,4}, sd::DataType::FLOAT32); NDArray exp('c', {2,3}, {1,1,1,0,0,1}, sd::DataType::BOOL); @@ -1142,19 +1147,19 @@ TEST_F(CudaBasicsTests1, execScalarBool_2) { std::vector dimensions = {1}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo)); // 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1163,27 +1168,27 @@ TEST_F(CudaBasicsTests1, execScalarBool_2) { // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execScalarBool(&lc, sd::scalar::GreaterThan, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, scalars.shapeInfo(), scalars.specialBuffer(), scalars.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1194,7 +1199,7 @@ TEST_F(CudaBasicsTests1, execBroadcast_1) { if (!Environment::getInstance().isExperimentalBuild()) return; - + NDArray x('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100}, sd::DataType::INT32); NDArray y('c', {3}, {10, 20, 30}, sd::DataType::INT64); NDArray z('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100}, sd::DataType::INT32); @@ -1203,19 +1208,19 @@ TEST_F(CudaBasicsTests1, execBroadcast_1) { std::vector dimensions = {1}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo)); // 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1230,20 +1235,20 @@ TEST_F(CudaBasicsTests1, execBroadcast_1) { nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1254,7 +1259,7 @@ TEST_F(CudaBasicsTests1, execBroadcast_2) { if (!Environment::getInstance().isExperimentalBuild()) return; - + NDArray x('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100}, sd::DataType::INT32); NDArray y('c', {2,4}, {10,20,30,40,50,60,70,80}, sd::DataType::FLOAT32); NDArray z('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100}, sd::DataType::FLOAT32); @@ -1263,19 +1268,19 @@ TEST_F(CudaBasicsTests1, execBroadcast_2) { std::vector dimensions = {0,2}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo)); // 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1290,20 +1295,20 @@ TEST_F(CudaBasicsTests1, execBroadcast_2) { nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1311,7 +1316,7 @@ TEST_F(CudaBasicsTests1, execBroadcast_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execBroadcastBool_1) { - + NDArray x('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100}, sd::DataType::INT32); NDArray y('c', {3}, {2, 12, 22}, sd::DataType::INT32); NDArray z('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,}, sd::DataType::BOOL); @@ -1320,19 +1325,19 @@ TEST_F(CudaBasicsTests1, execBroadcastBool_1) { std::vector dimensions = {1}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo)); // 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1348,20 +1353,20 @@ TEST_F(CudaBasicsTests1, execBroadcastBool_1) { nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], - nullptr, nullptr); + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1369,7 +1374,7 @@ TEST_F(CudaBasicsTests1, execBroadcastBool_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execBroadcastBool_2) { - + NDArray x('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100},sd::DataType::FLOAT32); NDArray y('c', {2,4}, {1,10,10,15,20,20,20,24}, sd::DataType::FLOAT32); NDArray z('c', {2,3,4}, {100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100,100}, sd::DataType::BOOL); @@ -1378,20 +1383,20 @@ TEST_F(CudaBasicsTests1, execBroadcastBool_2) { std::vector dimensions = {0,2}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + // prepare input arrays for prepareDataForCuda function + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo)); // 1 -- xTadShapeInfo hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - + std::vector devicePtrs(hostData.size(), nullptr); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1407,20 +1412,20 @@ TEST_F(CudaBasicsTests1, execBroadcastBool_2) { nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], - nullptr, nullptr); + (int*)devicePtrs[0], dimensions.size(), + (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], + nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) + cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1431,7 +1436,7 @@ TEST_F(CudaBasicsTests1, execPairwiseTransform_1) { if (!Environment::getInstance().isExperimentalBuild()) return; - + NDArray x('c', {2,2,2}, {1,5,3,7,2,6,4,8}, sd::DataType::INT32); NDArray y('c', {4,2}, {0.1,0.2,0.3,0.4,1.5,0.6,0.7,1.8}, sd::DataType::DOUBLE); NDArray z('c', {8}, {100,100,100,100,100,100,100,100}, sd::DataType::INT32); @@ -1444,7 +1449,7 @@ TEST_F(CudaBasicsTests1, execPairwiseTransform_1) { cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - + // call cuda kernel which calculates result NativeOpExecutioner::execPairwiseTransform(&lc, sd::pairwise::Subtract, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), @@ -1454,30 +1459,30 @@ TEST_F(CudaBasicsTests1, execPairwiseTransform_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - + // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execPairwiseBoolTransform_1) { - + NDArray x('c', {2,2,2}, {1,5,3,7,2,6,4,8}, sd::DataType::INT64); NDArray y('c', {4,2}, {0,2,0,4,0,6,0,8}, sd::DataType::INT64); NDArray z('c', {8}, {100,100,100,100,100,100,100,100}, sd::DataType::BOOL); NDArray exp('c', {8}, {0,1,0,1,0,1,0,1}, sd::DataType::BOOL); x.permutei({2,1,0}); // -> {1,2,3,4,5,6,7,8} x.syncShape(); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); - LaunchContext lc(&stream); + LaunchContext lc(&stream); // call cuda kernel which calculates result NativeOpExecutioner::execPairwiseBoolTransform(&lc, sd::pairwise::EqualTo, @@ -1488,11 +1493,11 @@ TEST_F(CudaBasicsTests1, execPairwiseBoolTransform_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - + // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); } @@ -1500,13 +1505,13 @@ TEST_F(CudaBasicsTests1, execPairwiseBoolTransform_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformFloat_1) { - + NDArray x('c', {2,2}, {0, 6.25, 2.25, 12.25}, sd::DataType::DOUBLE); NDArray z('c', {4}, {100,100,100,100}, sd::DataType::FLOAT32); NDArray exp('c', {4}, {0, 1.5, 2.5, 3.5}, sd::DataType::FLOAT32); x.permutei({1,0}); x.syncShape(); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1521,9 +1526,9 @@ TEST_F(CudaBasicsTests1, execTransformFloat_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1532,17 +1537,17 @@ TEST_F(CudaBasicsTests1, execTransformFloat_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformFloat_2) { - + NDArray x('c', {1,4}, {0, 4, 9, 16}, sd::DataType::INT64); NDArray z('c', {2,2}, {100,100,100,100}, sd::DataType::DOUBLE); NDArray exp('c', {2,2}, {0, 2, 3, 4}, sd::DataType::DOUBLE); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - + // call cuda kernel which calculates result NativeOpExecutioner::execTransformFloat(&lc, sd::transform::Sqrt, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), @@ -1551,9 +1556,9 @@ TEST_F(CudaBasicsTests1, execTransformFloat_2) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1562,12 +1567,12 @@ TEST_F(CudaBasicsTests1, execTransformFloat_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformAny_1) { - + NDArray x('c', {2,2}, {0, 6.25, 2.25, 12.25}, sd::DataType::DOUBLE); NDArray z('c', {4,1}, {100,100,100,100}, sd::DataType::INT32); NDArray exp('c', {4,1}, {0, 2, 6, 12}, sd::DataType::INT32); x.permutei({1,0}); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1582,10 +1587,10 @@ TEST_F(CudaBasicsTests1, execTransformAny_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1593,7 +1598,7 @@ TEST_F(CudaBasicsTests1, execTransformAny_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformAny_2) { - + NDArray x('c', {1,4}, {0, 6.25, 2.25, 12.25}, sd::DataType::BFLOAT16); NDArray z('c', {2,2}, {100,100,100,100}, sd::DataType::FLOAT32); NDArray exp('c', {2,2}, {0, 6.25, 2.25, 12.25}, sd::DataType::FLOAT32); @@ -1612,9 +1617,9 @@ TEST_F(CudaBasicsTests1, execTransformAny_2) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1623,12 +1628,12 @@ TEST_F(CudaBasicsTests1, execTransformAny_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformStrict_1) { - + NDArray x('c', {2,3}, {0,2,4,1,3,5}, sd::DataType::DOUBLE); NDArray z('c', {3,2}, {100,100,100,100,100,100}, sd::DataType::DOUBLE); NDArray exp('c', {3,2}, {0, 3, 12, 27, 48, 75}, sd::DataType::DOUBLE); x.permutei({1,0}); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1643,9 +1648,9 @@ TEST_F(CudaBasicsTests1, execTransformStrict_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1654,11 +1659,11 @@ TEST_F(CudaBasicsTests1, execTransformStrict_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformStrict_2) { - + NDArray x('c', {6}, {0,1,2,3,4,5}, sd::DataType::FLOAT32); NDArray z('c', {3,2}, {100,100,100,100,100,100}, sd::DataType::FLOAT32); NDArray exp('c', {3,2}, {0, 3, 12, 27, 48, 75}, sd::DataType::FLOAT32); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1673,9 +1678,9 @@ TEST_F(CudaBasicsTests1, execTransformStrict_2) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1684,12 +1689,12 @@ TEST_F(CudaBasicsTests1, execTransformStrict_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformSame_1) { - + NDArray x('c', {2,3}, {0,2.5,4.5,1.5,3.5,5.5}, sd::DataType::DOUBLE); NDArray z('c', {1,6}, {100,100,100,100,100,100}, sd::DataType::DOUBLE); NDArray exp('c', {1,6}, {0,2.25,6.25,12.25,20.25,30.25}, sd::DataType::DOUBLE); x.permutei({1,0}); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1704,9 +1709,9 @@ TEST_F(CudaBasicsTests1, execTransformSame_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1715,17 +1720,17 @@ TEST_F(CudaBasicsTests1, execTransformSame_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformSame_2) { - + NDArray x('c', {6}, {0,1,2,3,4,5}, sd::DataType::INT32); NDArray z('c', {3,2}, {100,100,100,100,100,100}, sd::DataType::INT32); NDArray exp('c', {3,2}, {0,1,4,9,16,25}, sd::DataType::INT32); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - + // call cuda kernel which calculates result NativeOpExecutioner::execTransformSame(&lc, sd::transform::Square, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), @@ -1734,9 +1739,9 @@ TEST_F(CudaBasicsTests1, execTransformSame_2) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1745,12 +1750,12 @@ TEST_F(CudaBasicsTests1, execTransformSame_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformBool_1) { - + NDArray x('c', {2,3}, {0,2,4,-1,-3,-5}, sd::DataType::DOUBLE); NDArray z('c', {1,6}, {100,100,100,100,100,100}, sd::DataType::BOOL); NDArray exp('c', {1,6}, {0,0,1,0,1,0}, sd::DataType::BOOL); x.permutei({1,0}); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1765,9 +1770,9 @@ TEST_F(CudaBasicsTests1, execTransformBool_1) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -1776,11 +1781,11 @@ TEST_F(CudaBasicsTests1, execTransformBool_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execTransformBool_2) { - + NDArray x('c', {6}, {0,-1,2,-3,4,-5}, sd::DataType::INT32); NDArray z('c', {3,2}, {100,100,100,100,100,100}, sd::DataType::BOOL); NDArray exp('c', {3,2}, {0,0,1,0,1,0}, sd::DataType::BOOL); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -1795,10 +1800,10 @@ TEST_F(CudaBasicsTests1, execTransformBool_2) { cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); - + // verify results - for (int e = 0; e < z.lengthOf(); e++) - ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); + for (int e = 0; e < z.lengthOf(); e++) + ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -1806,449 +1811,264 @@ TEST_F(CudaBasicsTests1, execTransformBool_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceFloat_1) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::INT32); NDArray z('c', {3}, {100,100,100}, sd::DataType::FLOAT32); NDArray exp('c', {3}, {2.5, 6.5, 10.5}, sd::DataType::FLOAT32); x.permutei({2,1,0}); - + std::vector dimensions = {0,2}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceFloat(&lc, sd::reduce::Mean, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceFloat(&lc, sd::reduce::Mean, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceFloat_2) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::INT32); NDArray z('c', {2,4}, {100,100,100,100,100,100,100,100}, sd::DataType::DOUBLE); NDArray exp('c', {2,4}, {-1., 0., 1., 2.,11., 12., 13., 14.}, sd::DataType::DOUBLE); - + std::vector dimensions = {1}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceFloat(&lc, sd::reduce::Mean, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceFloat(&lc, sd::reduce::Mean, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceSame_1) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::INT32); NDArray z('c', {3}, {100,100,100}, sd::DataType::INT32); NDArray exp('c', {3}, {20, 52, 84}, sd::DataType::INT32); x.permutei({2,1,0}); - + std::vector dimensions = {0,2}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceSame(&lc, sd::reduce::Sum, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceSame(&lc, sd::reduce::Sum, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceSame_2) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::FLOAT32); NDArray z('c', {2,4}, {100,100,100,100,100,100,100,100}, sd::DataType::FLOAT32); NDArray exp('c', {2,4}, {-3., 0., 3., 6.,33., 36., 39., 42.}, sd::DataType::FLOAT32); - + std::vector dimensions = {1}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceSame(&lc, sd::reduce::Sum, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceSame(&lc, sd::reduce::Sum, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceBool_1) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,-7,-8,-9,-10,-11,-12,-13,-14,-15,-16,-17,-18}, sd::DataType::INT32); NDArray z('c', {3}, {100,100,100}, sd::DataType::BOOL); NDArray exp('c', {3}, {0, 1, 1}, sd::DataType::BOOL); x.permutei({2,1,0}); - std::vector dimensions = {0,2}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceBool(&lc, sd::reduce::IsPositive, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceBool(&lc, sd::reduce::IsPositive, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceBool_2) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,-7,-8,-9,-10,-11,-12,-13,-14,-15,-16,-17,-18}, sd::DataType::FLOAT32); NDArray z('c', {2,4}, {100,100,100,100,100,100,100,100}, sd::DataType::BOOL); NDArray exp('c', {2,4}, {1, 1, 1, 1, 0, 0, 0, 0}, sd::DataType::BOOL); - + std::vector dimensions = {1}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceBool(&lc, sd::reduce::IsPositive, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceBool(&lc, sd::reduce::IsPositive, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceLong_1) { - + NDArray x('c', {2,3,4}, {-5,0,-3,0,-1,0,1,2,3,4,5,6,7,0,9,10,11,0,13,14,0,16,0,18}, sd::DataType::INT32); NDArray z('c', {3}, {100,100,100}, sd::DataType::INT64); NDArray exp('c', {3}, {5,6,6}, sd::DataType::INT64); x.permutei({2,1,0}); - + std::vector dimensions = {0,2}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceLong(&lc, sd::reduce::CountNonZero, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceLong(&lc, sd::reduce::CountNonZero, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceLong_2) { - + NDArray x('c', {2,3,4}, {-5,0,-3,0,-1,0,1,2,3,4,5,6,7,0,9,10,11,0,13,14,0,16,0,18}, sd::DataType::FLOAT32); NDArray z('c', {2,4}, {100,100,100,100,100,100,100,100}, sd::DataType::INT64); NDArray exp('c', {2,4}, {3, 1, 3, 2, 2, 1, 2, 3}, sd::DataType::INT64); std::vector dimensions = {1}; - // evaluate xTad data - shape::TAD xTad; - xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); - xTad.createTadOnlyShapeInfo(); - xTad.createOffsets(); - - // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions - hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets - std::vector devicePtrs(hostData.size(), nullptr); - // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it - cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - // call cuda kernel which calculates result - NativeOpExecutioner::execReduceLong(&lc, sd::reduce::CountNonZero, - nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, - nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), - (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); + std::vector dims = sd::ShapeUtils::evalDimsForReduceOp(x.rankOf(), dimensions); + NativeOpExecutioner::execReduceLong(&lc, sd::reduce::CountNonZero, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), dims.data(), dims.size()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); - // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) - cudaFree(devicePtrs[i]); - // delete cuda stream - cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); + cudaResult = cudaStreamDestroy(stream); + ASSERT_EQ(0, cudaResult); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceFloatScalar_1) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::INT32); NDArray z('c', {}, std::vector{100}, sd::DataType::FLOAT32); NDArray exp('c', {}, std::vector{6.5}, sd::DataType::FLOAT32); x.permutei({2,1,0}); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2260,18 +2080,18 @@ TEST_F(CudaBasicsTests1, execReduceFloatScalar_1) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceFloatScalar(&lc, sd::reduce::Mean, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); - + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2280,11 +2100,11 @@ TEST_F(CudaBasicsTests1, execReduceFloatScalar_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceFloatScalar_2) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::INT32); NDArray z('c', {}, std::vector{100}, sd::DataType::DOUBLE); NDArray exp('c', {}, std::vector{6.5}, sd::DataType::DOUBLE); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2296,18 +2116,18 @@ TEST_F(CudaBasicsTests1, execReduceFloatScalar_2) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceFloatScalar(&lc, sd::reduce::Mean, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2316,12 +2136,12 @@ TEST_F(CudaBasicsTests1, execReduceFloatScalar_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceSameScalar_1) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::INT32); NDArray z('c', {}, std::vector{100}, sd::DataType::INT32); NDArray exp('c', {}, std::vector{156}, sd::DataType::INT32); x.permutei({2,1,0}); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2333,18 +2153,18 @@ TEST_F(CudaBasicsTests1, execReduceSameScalar_1) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceSameScalar(&lc, sd::reduce::Sum, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); - + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2353,11 +2173,11 @@ TEST_F(CudaBasicsTests1, execReduceSameScalar_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceSameScalar_2) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18}, sd::DataType::DOUBLE); NDArray z('c', {}, std::vector{100}, sd::DataType::DOUBLE); NDArray exp('c', {}, std::vector{156}, sd::DataType::DOUBLE); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2369,18 +2189,18 @@ TEST_F(CudaBasicsTests1, execReduceSameScalar_2) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceSameScalar(&lc, sd::reduce::Sum, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2389,13 +2209,13 @@ TEST_F(CudaBasicsTests1, execReduceSameScalar_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceBoolScalar_1) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,-7,-8,-9,-10,-11,-12,-13,-14,-15,-16,-17,-18}, sd::DataType::INT32); NDArray z('c', {}, std::vector{100}, sd::DataType::BOOL); NDArray exp('c', {}, std::vector{1}, sd::DataType::BOOL); x.permutei({2,1,0}); - x.syncShape(); - + x.syncShape(); + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2407,18 +2227,18 @@ TEST_F(CudaBasicsTests1, execReduceBoolScalar_1) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceBoolScalar(&lc, sd::reduce::IsPositive, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); - + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2427,11 +2247,11 @@ TEST_F(CudaBasicsTests1, execReduceBoolScalar_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceBoolScalar_2) { - + NDArray x('c', {2,3,4}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6,-7,-8,-9,-10,-11,-12,-13,-14,-15,-16,-17,-18}, sd::DataType::DOUBLE); NDArray z('c', {}, std::vector{100}, sd::DataType::BOOL); NDArray exp('c', {}, std::vector{1}, sd::DataType::BOOL); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2443,18 +2263,18 @@ TEST_F(CudaBasicsTests1, execReduceBoolScalar_2) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceBoolScalar(&lc, sd::reduce::IsPositive, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2463,13 +2283,13 @@ TEST_F(CudaBasicsTests1, execReduceBoolScalar_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceLongScalar_1) { - + NDArray x('c', {2,3,4}, {-5,0,-3,0,-1,0,1,2,3,4,5,6,7,0,9,10,11,0,13,14,0,16,0,18}, sd::DataType::INT32); NDArray z('c', {}, std::vector{100}, sd::DataType::INT64); NDArray exp('c', {}, std::vector{17}, sd::DataType::INT64); x.permutei({2,1,0}); - x.syncShape(); - + x.syncShape(); + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2481,18 +2301,18 @@ TEST_F(CudaBasicsTests1, execReduceLongScalar_1) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceLongScalar(&lc, sd::reduce::CountNonZero, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); - + cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2501,11 +2321,11 @@ TEST_F(CudaBasicsTests1, execReduceLongScalar_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduceLongScalar_2) { - + NDArray x('c', {2,3,4}, {-5,0,-3,0,-1,0,1,2,3,4,5,6,7,0,9,10,11,0,13,14,0,16,0,18}, sd::DataType::DOUBLE); NDArray z('c', {}, std::vector{100}, sd::DataType::INT64); NDArray exp('c', {}, std::vector{17}, sd::DataType::INT64); - + // create cuda stream and LaunchContext cudaError_t cudaResult; cudaStream_t stream; @@ -2517,18 +2337,18 @@ TEST_F(CudaBasicsTests1, execReduceLongScalar_2) { cudaResult = cudaMalloc(reinterpret_cast(&allocationPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); lc.setAllocationPointer(allocationPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduceLongScalar(&lc, sd::reduce::CountNonZero, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo()); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2537,12 +2357,12 @@ TEST_F(CudaBasicsTests1, execReduceLongScalar_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3TAD_1) { - + NDArray x('c', {2,2,3}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6}, sd::DataType::FLOAT32); NDArray y('c', {2,2}, {1,2,3,4}, sd::DataType::FLOAT32); NDArray exp('c', {3}, {10,20,30}, sd::DataType::DOUBLE); NDArray z('c', {3}, {100,100,100}, sd::DataType::DOUBLE); - + std::vector dimensions = {0,1}; auto packX = ConstantTadHelper::getInstance().tadForDimensions(x.shapeInfo(), dimensions); LaunchContext* context = x.getContext(); @@ -2553,7 +2373,7 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_1) { // call cuda kernel which calculates result NativeOpExecutioner::execReduce3TAD(context, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, dimensions.size(), @@ -2563,32 +2383,32 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_1) { z.tickWriteDevice(); // z.printIndexedBuffer("OutputReduce3TAD"); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); } //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3TAD_2) { - + NDArray x('c', {2,2,3}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6}, sd::DataType::INT64); NDArray y('c', {2,3}, {1,2,3,4,5,6}, sd::DataType::INT64); NDArray exp('c', {2}, {10,73}, sd::DataType::FLOAT32); NDArray z('c', {2}, {100,100}, sd::DataType::FLOAT32); - + std::vector dimensions = {0,2}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -2597,28 +2417,28 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_2) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3TAD(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], nullptr, nullptr); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -2626,25 +2446,25 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_2) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3TAD_3) { - + NDArray x('c', {2,2,3}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6}, sd::DataType::INT64); NDArray y('c', {3}, {1,2,3}, sd::DataType::INT64); NDArray exp('c', {2,2}, {-22,-4,14,32}, sd::DataType::FLOAT32); NDArray z('c', {2,2}, {100,100,100,100}, sd::DataType::FLOAT32); - + std::vector dimensions = {2}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -2653,28 +2473,28 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_3) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3TAD(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -2682,7 +2502,7 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_3) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execReduce3TAD_4) { - + NDArray x('c', {2,2,3}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6}, sd::DataType::DOUBLE); NDArray y('c', {2,2,3}, {10,20,30,40,50,60,70,80,90,100,110,120}, sd::DataType::DOUBLE); NDArray exp('c', {}, std::vector{1820}, sd::DataType::FLOAT32); @@ -2690,17 +2510,17 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_4) { std::vector dimensions = {0,1,2}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -2709,27 +2529,27 @@ TEST_F(CudaBasicsTests1, execReduce3TAD_4) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execReduce3TAD(&lc, sd::reduce3::Dot, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, y.shapeInfo(), y.specialBuffer(), y.specialShapeInfo(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2]); cudaResult = cudaStreamSynchronize(stream); ASSERT_EQ(0, cudaResult); z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -2751,13 +2571,13 @@ TEST_F(CudaBasicsTests1, execSummaryStats_1) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); void* reductionPointer; - cudaResult = cudaMalloc(reinterpret_cast(&reductionPointer), 1024*1024); ASSERT_EQ(0, cudaResult); + cudaResult = cudaMalloc(reinterpret_cast(&reductionPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execSummaryStats(&lc, sd::variance::SummaryStatsStandardDeviation, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), true); @@ -2765,7 +2585,7 @@ TEST_F(CudaBasicsTests1, execSummaryStats_1) { z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2774,24 +2594,24 @@ TEST_F(CudaBasicsTests1, execSummaryStats_1) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execSummaryStats_2) { - + NDArray x('c', {2,2,3}, {-5,-4,-3,-20,-1,0,1,2,3,4,5,6}, sd::DataType::DOUBLE); NDArray exp('c', {2}, {3.405877, 9.715966}, sd::DataType::FLOAT32); NDArray z('c', {2}, {100,100}, sd::DataType::FLOAT32); std::vector dimensions = {0,2}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -2800,15 +2620,15 @@ TEST_F(CudaBasicsTests1, execSummaryStats_2) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execSummaryStats(&lc, sd::variance::SummaryStatsStandardDeviation, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - (int*)devicePtrs[0], dimensions.size(), + (int*)devicePtrs[0], dimensions.size(), (Nd4jLong*)devicePtrs[1], (Nd4jLong*)devicePtrs[2], true); @@ -2816,11 +2636,11 @@ TEST_F(CudaBasicsTests1, execSummaryStats_2) { z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -2828,24 +2648,24 @@ TEST_F(CudaBasicsTests1, execSummaryStats_2) { /* //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execSummaryStats_3) { - + NDArray x('c', {2,2,3}, {-5,-4,-3,-20,-1,0,1,2,3,4,5,6}, sd::DataType::DOUBLE); NDArray exp('c', {2}, {10.606602, 2.121320}, sd::DataType::FLOAT32); NDArray z('c', {2}, {100,100}, sd::DataType::FLOAT32); std::vector dimensions = {1}; - // evaluate xTad data + // evaluate xTad data shape::TAD xTad; xTad.init(x.shapeInfo(), dimensions.data(), dimensions.size()); xTad.createTadOnlyShapeInfo(); xTad.createOffsets(); // prepare input arrays for prepareDataForCuda function - std::vector> hostData; + std::vector> hostData; hostData.emplace_back(dimensions.data(), dimensions.size() * sizeof(int)); // 0 -- dimensions hostData.emplace_back(xTad.tadOnlyShapeInfo, shape::shapeInfoByteLength(xTad.tadOnlyShapeInfo));// 1 -- xTadShapeInfo - hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets + hostData.emplace_back(xTad.tadOffsets, xTad.numTads * sizeof(Nd4jLong)); // 2 -- xTadOffsets std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -2854,9 +2674,9 @@ TEST_F(CudaBasicsTests1, execSummaryStats_3) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execSummaryStats(&lc, sd::variance::SummaryStatsStandardDeviation, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), @@ -2870,11 +2690,11 @@ TEST_F(CudaBasicsTests1, execSummaryStats_3) { z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -2883,7 +2703,7 @@ TEST_F(CudaBasicsTests1, execSummaryStats_3) { //////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execSummaryStatsScalar_1) { - + NDArray x('c', {2,2,3}, {-5,-4,-3,-2,-1,0,1,2,3,4,5,6}, sd::DataType::INT64); NDArray exp('c', {}, std::vector{3.605551}, sd::DataType::FLOAT32); NDArray z('c', {}, std::vector{100}, sd::DataType::FLOAT32); @@ -2894,13 +2714,13 @@ TEST_F(CudaBasicsTests1, execSummaryStatsScalar_1) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); void* reductionPointer; - cudaResult = cudaMalloc(reinterpret_cast(&reductionPointer), 1024*1024); ASSERT_EQ(0, cudaResult); + cudaResult = cudaMalloc(reinterpret_cast(&reductionPointer), 1024*1024); ASSERT_EQ(0, cudaResult); lc.setReductionPointer(reductionPointer); - + // call cuda kernel which calculates result NativeOpExecutioner::execSummaryStatsScalar(&lc, sd::variance::SummaryStatsStandardDeviation, nullptr, x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), - nullptr, + nullptr, nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), true); @@ -2908,7 +2728,7 @@ TEST_F(CudaBasicsTests1, execSummaryStatsScalar_1) { z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // delete cuda stream @@ -2917,7 +2737,7 @@ TEST_F(CudaBasicsTests1, execSummaryStatsScalar_1) { ////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execRandom_1) { - + // NDArray z('c', {10}, {100,0,0,0,0,0,0,0,0,0}, sd::DataType::DOUBLE); NDArray z('c', {10}, {100,0,0,0,0,0,0,0,0,100}, sd::DataType::FLOAT32); NDArray exp('c', {10}, {0.050942, -0.183229, -0.093921, 0.075469, 0.257166, -0.254838, 0.342227, -0.682188, -0.004345, 0.464633}, sd::DataType::FLOAT32); @@ -2971,14 +2791,14 @@ TEST_F(CudaBasicsTests1, execRandom_1) { ////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execRandom_2) { - + NDArray x('c', {10}, {0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1}, sd::DataType::DOUBLE); NDArray z('c', {2,5}, {100,100,100,100,100,100,100,100,100,100}, sd::DataType::DOUBLE); NDArray exp('c', {10}, {0., 0., 0.3, 0., 0.5, 0., 0.7, 0., 0., 1.}, sd::DataType::DOUBLE); - + ExtraArguments extraArguments({0.7}); sd::graph::RandomGenerator gen(119,5); - + // // prepare input arrays for prepareDataForCuda function // std::vector> hostData; // hostData.emplace_back(extraArguments.data(), extraArguments.size() * sizeof(double)); // 0 -- dimensions @@ -2990,9 +2810,9 @@ TEST_F(CudaBasicsTests1, execRandom_2) { // cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext* lc = x.getContext(); //(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it // cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execRandom(lc, sd::random::DropOut, &gen, @@ -3004,7 +2824,7 @@ TEST_F(CudaBasicsTests1, execRandom_2) { z.tickWriteDevice(); z.syncToHost(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory @@ -3016,16 +2836,16 @@ TEST_F(CudaBasicsTests1, execRandom_2) { ////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execRandom_3) { - + NDArray z('c', {10}, {100,100,100,100,100,100,100,100,100,100}, sd::DataType::DOUBLE); NDArray exp('c', {10}, {2.373649, 2.239791, 1.887353, 2.488636, 2.068904, 2.281399, 1.828228, 2.228222, 2.490847, 1.669537}, sd::DataType::DOUBLE); - + std::vector extraArguments = {1.5, 2.5}; sd::graph::RandomGenerator gen(119,5); - + // prepare input arrays for prepareDataForCuda function - std::vector> hostData; - hostData.emplace_back(extraArguments.data(), extraArguments.size() * sizeof(double)); // 0 -- dimensions + std::vector> hostData; + hostData.emplace_back(extraArguments.data(), extraArguments.size() * sizeof(double)); // 0 -- dimensions std::vector devicePtrs(hostData.size(), nullptr); // create cuda stream and LaunchContext @@ -3034,9 +2854,9 @@ TEST_F(CudaBasicsTests1, execRandom_3) { cudaResult = cudaStreamCreate(&stream); ASSERT_EQ(0, cudaResult); LaunchContext lc(&stream); - // allocate required amount of global device memory and copy host data to it + // allocate required amount of global device memory and copy host data to it cudaResult = allocateDeviceMem(lc, devicePtrs, hostData); ASSERT_EQ(0, cudaResult); - + // call cuda kernel which calculates result NativeOpExecutioner::execRandom(&lc, sd::random::UniformDistribution, &gen, @@ -3047,11 +2867,11 @@ TEST_F(CudaBasicsTests1, execRandom_3) { z.tickWriteDevice(); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory - for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); + for(int i = 0; i < devicePtrs.size(); ++i) cudaFree(devicePtrs[i]); // delete cuda stream cudaResult = cudaStreamDestroy(stream); ASSERT_EQ(0, cudaResult); @@ -3059,14 +2879,14 @@ TEST_F(CudaBasicsTests1, execRandom_3) { ////////////////////////////////////////////////////////////////////////// TEST_F(CudaBasicsTests1, execRandom_4) { - + NDArray z('c', {2,5}, {1,2,3,4,5,6,7,8,9,10}, sd::DataType::FLOAT32); NDArray exp('c', {10}, {2.373649, 2.281399, 2.239791, 1.828228, 1.887353, 2.228222, 2.488636, 2.490847, 2.068904, 1.669537}, sd::DataType::FLOAT32); - z.permutei({1,0}); - + z.permutei({1,0}); + ExtraArguments extraArguments({1.5, 2.5}); sd::graph::RandomGenerator gen(119,5); - + // // prepare input arrays for prepareDataForCuda function // std::vector> hostData; // hostData.emplace_back(extraArguments.data(), extraArguments.size() * sizeof(double)); // 0 -- dimensions @@ -3092,7 +2912,7 @@ TEST_F(CudaBasicsTests1, execRandom_4) { z.tickWriteDevice(); // z.printIndexedBuffer("Output Uniform4"); // verify results - for (int e = 0; e < z.lengthOf(); e++) + for (int e = 0; e < z.lengthOf(); e++) ASSERT_NEAR(exp.e(e), z.e(e), 1e-5); // free allocated global device memory diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp index 2a099230e..81b869457 100644 --- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp +++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp @@ -186,7 +186,7 @@ TEST_F(DeclarableOpsTests3, Test_Norm_1) { auto result0 = op.evaluate({&x}, {0.}, {}); auto z0 = result0.at(0); - auto exp0 = x.reduceAlongDimension(reduce::NormFrobenius, empty, false, false); + auto exp0 = x.reduceAlongDimension(reduce::NormFrobenius, empty); ASSERT_TRUE(exp0.isSameShape(z0)); ASSERT_TRUE(exp0.equalsTo(z0)); @@ -194,7 +194,7 @@ TEST_F(DeclarableOpsTests3, Test_Norm_1) { ASSERT_EQ(result1.status(), ND4J_STATUS_OK); auto z1 = result1.at(0); // z1->printIndexedBuffer("Z1"); - auto exp1 = x.reduceAlongDimension(reduce::Norm2, dims, false, false); + auto exp1 = x.reduceAlongDimension(reduce::Norm2, dims); // exp1.printIndexedBuffer("EXP1"); // z1->printShapeInfo("Z1 shape"); // exp1.printShapeInfo("EXP1 shape"); @@ -204,7 +204,7 @@ TEST_F(DeclarableOpsTests3, Test_Norm_1) { auto result4 = op.evaluate({&x}, {4.}, {1}); auto z4 = result4.at(0); - auto exp4= x.reduceAlongDimension(reduce::NormMax, dims, false, false); + auto exp4= x.reduceAlongDimension(reduce::NormMax, dims); ASSERT_TRUE(exp4.isSameShape(z4)); ASSERT_TRUE(exp4.equalsTo(z4)); } @@ -222,7 +222,7 @@ TEST_F(DeclarableOpsTests3, Test_Norm_2) { auto result0 = op.evaluate({&x}, {0}, {}); auto z0 = result0.at(0); - auto exp0 = x.reduceAlongDimension(reduce::NormFrobenius, empty, false, false); + auto exp0 = x.reduceAlongDimension(reduce::NormFrobenius, empty); ASSERT_TRUE(exp0.isSameShape(z0)); ASSERT_TRUE(exp0.equalsTo(z0)); @@ -231,14 +231,14 @@ TEST_F(DeclarableOpsTests3, Test_Norm_2) { auto result1 = op.evaluate({&x, &axis}, {1}, {}); auto z1 = result1.at(0); - auto exp1 = x.reduceAlongDimension(reduce::Norm2, dims, false, false); + auto exp1 = x.reduceAlongDimension(reduce::Norm2, dims); ASSERT_TRUE(exp1.isSameShape(z1)); ASSERT_TRUE(exp1.equalsTo(z1)); auto result4 = op.evaluate({&x, &axis}, {4}, {}); auto z4 = result4.at(0); - auto exp4= x.reduceAlongDimension(reduce::NormMax, dims, false, false); + auto exp4= x.reduceAlongDimension(reduce::NormMax, dims); ASSERT_TRUE(exp4.isSameShape(z4)); ASSERT_TRUE(exp4.equalsTo(z4)); diff --git a/libnd4j/tests_cpu/layers_tests/LegacyOpsTests.cpp b/libnd4j/tests_cpu/layers_tests/LegacyOpsTests.cpp index fe9c5a7a0..385e042f8 100644 --- a/libnd4j/tests_cpu/layers_tests/LegacyOpsTests.cpp +++ b/libnd4j/tests_cpu/layers_tests/LegacyOpsTests.cpp @@ -70,7 +70,7 @@ TEST_F(LegacyOpsTests, TransformTests_2) { ASSERT_TRUE(exp.equalsTo(z)); - + } TEST_F(LegacyOpsTests, Reciprocal_1) { @@ -126,7 +126,7 @@ TEST_F(LegacyOpsTests, PWT_Tests_2) { //z->printBuffer("Z"); ASSERT_TRUE(exp.equalsTo(z)); - + } TEST_F(LegacyOpsTests, Scalar_Test_1) { @@ -157,7 +157,7 @@ TEST_F(LegacyOpsTests, Scalar_Test_2) { auto z = result.at(0); ASSERT_TRUE(exp.equalsTo(z)); - + } @@ -176,7 +176,7 @@ TEST_F(LegacyOpsTests, ReduceTests_1) { ASSERT_TRUE(z->isScalar()); ASSERT_NEAR(x.sumNumber().e(0), z->e(0), 1e-5f); - + } @@ -197,7 +197,7 @@ TEST_F(LegacyOpsTests, ReduceTests_2) { ASSERT_TRUE(exp.isSameShape(z)); ASSERT_TRUE(exp.equalsTo(z)); - + } @@ -217,7 +217,7 @@ TEST_F(LegacyOpsTests, ReduceTests_3) { ASSERT_TRUE(exp.isSameShape(z)); ASSERT_TRUE(exp.equalsTo(z)); - + } @@ -238,7 +238,7 @@ TEST_F(LegacyOpsTests, ReduceTests_4) { ASSERT_TRUE(exp.isSameShape(z)); ASSERT_TRUE(exp.equalsTo(z)); - + } TEST_F(LegacyOpsTests, ReduceTests_5) { @@ -256,7 +256,7 @@ TEST_F(LegacyOpsTests, ReduceTests_5) { ASSERT_TRUE(z->isScalar()); ASSERT_NEAR(x.meanNumber().e(0), z->e(0), 1e-5f); - + } @@ -277,7 +277,7 @@ TEST_F(LegacyOpsTests, ReduceTests_6) { ASSERT_TRUE(exp.isSameShape(z)); ASSERT_TRUE(exp.equalsTo(z)); - + } @@ -297,7 +297,7 @@ TEST_F(LegacyOpsTests, ReduceTests_7) { ASSERT_TRUE(exp.isSameShape(z)); ASSERT_TRUE(exp.equalsTo(z)); - + } @@ -319,7 +319,7 @@ TEST_F(LegacyOpsTests, ReduceTests_8) { ASSERT_TRUE(exp.isSameShape(z)); ASSERT_TRUE(exp.equalsTo(z)); - + } @@ -338,7 +338,7 @@ TEST_F(LegacyOpsTests, IndexReduceTests_1) { ASSERT_TRUE(z->isScalar()); ASSERT_EQ(24, z->e(0)); - + } @@ -362,7 +362,7 @@ TEST_F(LegacyOpsTests, IndexReduceTests_2) { //ASSERT_EQ(4, z->e(3)); //ASSERT_EQ(4, z->e(4)); - + } TEST_F(LegacyOpsTests, BroadcastingTests_1) { @@ -707,7 +707,7 @@ TEST_F(LegacyOpsTests, test_legacy_reduce_empty_1) { x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), - &dim, 1, x.platformShapeInfo(), nullptr); + &dim, 1); ASSERT_EQ(e, z); } @@ -720,7 +720,7 @@ TEST_F(LegacyOpsTests, test_legacy_reduce_empty_2) { int dim = 1; - NativeOpExecutioner::execReduceSame(LaunchContext::defaultContext(), reduce::SameOps::Min, x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), &dim, 1, x.platformShapeInfo(), nullptr); + NativeOpExecutioner::execReduceSame(LaunchContext::defaultContext(), reduce::SameOps::Min, x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), &dim, 1); ASSERT_EQ(e, z); } @@ -733,7 +733,7 @@ TEST_F(LegacyOpsTests, test_legacy_reduce_empty_3) { int dim = 1; - NativeOpExecutioner::execReduceSame(LaunchContext::defaultContext(), reduce::SameOps::Max, x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), &dim, 1, x.platformShapeInfo(), nullptr); + NativeOpExecutioner::execReduceSame(LaunchContext::defaultContext(), reduce::SameOps::Max, x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo(), nullptr, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), &dim, 1); ASSERT_EQ(e, z); } diff --git a/libnd4j/tests_cpu/layers_tests/NDArrayTests.cpp b/libnd4j/tests_cpu/layers_tests/NDArrayTests.cpp index 8150976e1..807de9fed 100644 --- a/libnd4j/tests_cpu/layers_tests/NDArrayTests.cpp +++ b/libnd4j/tests_cpu/layers_tests/NDArrayTests.cpp @@ -462,7 +462,7 @@ TEST_F(NDArrayTest, TestTranspose2) { } ////////////////////////////////////////////////////////////////////// -TEST_F(NDArrayTest, TestSumAlongDimension1) { +TEST_F(NDArrayTest, TestReduceAlongDimension1) { NDArray array('c', {2,2}, {1,2,3,4}, sd::DataType::FLOAT32); @@ -475,23 +475,7 @@ TEST_F(NDArrayTest, TestSumAlongDimension1) { } ////////////////////////////////////////////////////////////////////// -TEST_F(NDArrayTest, TestSumAlongDimension2) { - float *c = new float[4] {1, 2, 3, 4}; - auto array = new NDArray(c, cShape); - - auto res = array->reduceAlongDimension(reduce::Sum, {1}); - - ASSERT_EQ(2, res.lengthOf()); - - ASSERT_EQ(3.0f, res.e(0)); - ASSERT_EQ(7.0f, res.e(1)); - - delete[] c; - delete array; -} - -////////////////////////////////////////////////////////////////////// -TEST_F(NDArrayTest, TestReduceAlongDimension1) { +TEST_F(NDArrayTest, TestReduceAlongDimension2) { float *c = new float[4] {1, 2, 3, 4}; auto array = new NDArray(c, cShape); diff --git a/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp b/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp index e07a0496d..b55f971d4 100644 --- a/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp +++ b/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp @@ -1132,6 +1132,26 @@ TEST_F(PlaygroundTests, lstmLayerCellBp_1) { const bool isGradCorrect = GradCheck::checkGrad(opFF, opBP, argsHolderFF, argsHolderBP, {true, true, true, true, true, true, true}); } +TEST_F(PlaygroundTests, my) { + + const int N = 40; + + NDArray x('c', {256,256,128,128}, sd::DataType::FLOAT32); + NDArray z1('c', {256,2,128}, sd::DataType::DOUBLE); + NDArray z = z1({0,0,0,1,0,0}); + z.printShapeInfo(); + + auto timeStart = std::chrono::system_clock::now(); + for (int i = 0; i < N; ++i) { + // x.reduceAlongDimension(sd::reduce::Mean, z, {1,3}); + x.applyBroadcast(sd::broadcast::Ops::Add, {1,3}, z, x); + } + auto timeEnd = std::chrono::system_clock::now(); + auto time = std::chrono::duration_cast ((timeEnd - timeStart) / N).count(); + + printf("old %ld\n", time); +} + /////////////////////////////////////////////////////////////////// TEST_F(DeclarableOpsTests13, lstmLayer_bp_1) { @@ -1657,52 +1677,4 @@ TEST_F(DeclarableOpsTests15, gru_bp_1) { const bool isGradCorrect = GradCheck::checkGrad(opFF, opBP, argsHolderFF, argsHolderBP); } -#include -////////////////////////////////////////////////////////////////////// -TEST_F(PlaygroundTests, my) { - - const int N = 10; - - NDArray input('c', {8000000}, sd::DataType::INT32); - input.linspace(1); - NDArray output = input.dup(); - - - sd::graph::RandomGenerator rng; - - sd::ops::helpers::randomShuffle(input.getContext(), input, output, rng, true); - - // auto timeStart = std::chrono::system_clock::now(); - // for (int i = 0; i < N; ++i) - // sd::ops::helpers::randomShuffle(input.getContext(), input, output, rng, true); - // auto timeEnd = std::chrono::system_clock::now(); - // auto time = std::chrono::duration_cast ((timeEnd - timeStart) / N).count(); - // printf("time: %i \n", time); - - // bool hasDublicates = false; - // for(int i = 0; i < output.lengthOf() - 1; ++i) - // for(int j = i+1; j < output.lengthOf(); ++j) - // if(output.t(i) == output.t(j)) { - // hasDublicates = true; - // i = output.lengthOf(); - // break; - // } - - ASSERT_TRUE(!input.equalsTo(output)); - - bool hasDublicates = false; - for(int i = 0; i < input.lengthOf() - 1; ++i) - for(int j = i+1; j < input.lengthOf(); ++j) - if(input.t(i) == input.t(j)) { - hasDublicates = true; - i = input.lengthOf(); - break; - } - ASSERT_TRUE(!hasDublicates); -} - - -} - */ - diff --git a/libnd4j/tests_cpu/layers_tests/RNGTests.cpp b/libnd4j/tests_cpu/layers_tests/RNGTests.cpp index a2c33374a..dfc23f559 100644 --- a/libnd4j/tests_cpu/layers_tests/RNGTests.cpp +++ b/libnd4j/tests_cpu/layers_tests/RNGTests.cpp @@ -30,13 +30,11 @@ using namespace sd; class RNGTests : public testing::Test { private: - //Nd4jLong *_bufferA; - //Nd4jLong *_bufferB; + public: long _seed = 119L; - //sd::random::RandomBuffer *_rngA; - //sd::random::RandomBuffer *_rngB; + sd::graph::RandomGenerator _rngA; sd::graph::RandomGenerator _rngB; @@ -45,10 +43,7 @@ public: NDArray* nexp2 = NDArrayFactory::create_('c', {10, 10}); RNGTests() { - //_bufferA = new Nd4jLong[100000]; - //_bufferB = new Nd4jLong[100000]; - //_rngA = (sd::random::RandomBuffer *) initRandom(nullptr, _seed, 100000, (Nd4jPointer) _bufferA); - //_rngB = (sd::random::RandomBuffer *) initRandom(nullptr, _seed, 100000, (Nd4jPointer) _bufferB); + _rngA.setStates(_seed * 0xDEADBEEF * 13, _seed * 0xDEADBEEF * 7); _rngB.setStates(_seed * 0xDEADBEEF * 13, _seed * 0xDEADBEEF * 7); nexp0->assign(-1.0f); @@ -57,10 +52,6 @@ public: } ~RNGTests() { - //destroyRandom(_rngA); - //destroyRandom(_rngB); - //delete[] _bufferA; - //delete[] _bufferB; delete nexp0; delete nexp1; @@ -103,7 +94,6 @@ TEST_F(RNGTests, TestGenerator_SGA_1) { array.r(idx) = x; } auto minimum = array.reduceNumber(reduce::AMin); - minimum.printBuffer("Randomly float min on 1M array"); ASSERT_EQ(123, generator.rootState()); ASSERT_EQ(456, generator.nodeState()); } @@ -118,13 +108,11 @@ TEST_F(RNGTests, Test_Dropout_1) { float prob[] = {0.5f}; - //x0.applyRandom(random::DropOut, _rngA, nullptr, &x0, prob); - //x1.applyRandom(random::DropOut, _rngB, nullptr, &x1, prob); + RandomLauncher::applyDropOut(LaunchContext::defaultContext(), _rngA, &x0, 0.5); RandomLauncher::applyDropOut(LaunchContext::defaultContext(), _rngB, &x1, 0.5); ASSERT_TRUE(x0.equalsTo(&x1)); - //x0.printIndexedBuffer("Dropout"); - // this check is required to ensure we're calling wrong signature + ASSERT_FALSE(x0.equalsTo(nexp0)); ASSERT_FALSE(x0.equalsTo(nexp1)); ASSERT_FALSE(x0.equalsTo(nexp2)); @@ -139,13 +127,11 @@ TEST_F(RNGTests, Test_DropoutInverted_1) { float prob[] = {0.5f}; - //x0.template applyRandom>(_rngA, nullptr, &x0, prob); - //x1.template applyRandom>(_rngB, nullptr, &x1, prob); + RandomLauncher::applyInvertedDropOut(LaunchContext::defaultContext(), _rngA, &x0, 0.5); RandomLauncher::applyInvertedDropOut(LaunchContext::defaultContext(), _rngB, &x1, 0.5); ASSERT_TRUE(x0.equalsTo(&x1)); - //x0.printIndexedBuffer("DropoutInverted"); - // this check is required to ensure we're calling wrong signature + ASSERT_FALSE(x0.equalsTo(nexp0)); ASSERT_FALSE(x0.equalsTo(nexp1)); ASSERT_FALSE(x0.equalsTo(nexp2)); @@ -189,7 +175,6 @@ TEST_F(RNGTests, Test_Launcher_3) { RandomLauncher::applyAlphaDropOut(LaunchContext::defaultContext(), _rngA, &x0, 0.5f, 0.2f, 0.1f, 0.3f); RandomLauncher::applyAlphaDropOut(LaunchContext::defaultContext(), _rngB, &x1, 0.5f, 0.2f, 0.1f, 0.3f); - //x1.printIndexedBuffer("x1"); ASSERT_TRUE(x0.equalsTo(&x1)); ASSERT_FALSE(x0.equalsTo(nexp0)); @@ -204,9 +189,6 @@ TEST_F(RNGTests, Test_Uniform_1) { RandomLauncher::fillUniform(LaunchContext::defaultContext(), _rngA, &x0, 1.0f, 2.0f); RandomLauncher::fillUniform(LaunchContext::defaultContext(), _rngB, &x1, 1.0f, 2.0f); - x0.printLinearBuffer(); - x1.printLinearBuffer(); - ASSERT_TRUE(x0.equalsTo(&x1)); ASSERT_FALSE(x0.equalsTo(nexp0)); @@ -215,7 +197,6 @@ TEST_F(RNGTests, Test_Uniform_1) { for (int e = 0; e < x0.lengthOf(); e++) { float v = x0.e(e); - nd4j_printf("%f\n", v); ASSERT_TRUE(v >= 1.0f && v <= 2.0f); } } @@ -253,8 +234,6 @@ TEST_F(RNGTests, Test_Uniform_11) { if (v > max) max = v; } - - nd4j_printf("Max value: %i\n", (int) max); } TEST_F(RNGTests, Test_Uniform_12) { @@ -269,7 +248,6 @@ TEST_F(RNGTests, Test_Uniform_12) { min = v; } - nd4j_printf("Max value: %.8f; Min value: %.8f\n", (float) max, (float) min); ASSERT_LT(max, 1.0f); ASSERT_GE(min, 0.0); } @@ -286,7 +264,6 @@ TEST_F(RNGTests, Test_Uniform_13) { min = v; } - nd4j_printf("Max value: %.8f; Min value: %.8f\n", (float) max, (float) min); ASSERT_LT(max, 1.0); ASSERT_GE(min, 0.0); } @@ -323,8 +300,6 @@ TEST_F(RNGTests, Test_Gaussian_1) { RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngA, &x0, 1.0f, 2.0f); RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngB, &x1, 1.0f, 2.0f); - //x0.printIndexedBuffer("x0"); - //x1.printIndexedBuffer("x1"); ASSERT_TRUE(x0.equalsTo(&x1)); ASSERT_FALSE(x0.equalsTo(nexp0)); @@ -339,8 +314,6 @@ TEST_F(RNGTests, Test_Gaussian_21) { RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngA, &x0, 0.0f, 1.0f); RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngB, &x1, 0.0f, 1.0f); -// x0.printIndexedBuffer("x0"); -// x1.printIndexedBuffer("x1"); ASSERT_TRUE(x0.equalsTo(&x1)); ASSERT_FALSE(x0.equalsTo(nexp0)); @@ -348,15 +321,10 @@ TEST_F(RNGTests, Test_Gaussian_21) { ASSERT_FALSE(x0.equalsTo(nexp2)); sd::ops::moments op; auto result = op.evaluate({&x0}, {}, {}); - //x0.printIndexedBuffer("X0 Normal"); - //x1.printIndexedBuffer("X1 Normal"); ASSERT_TRUE(result.status() == Status::OK()); auto mean = result.at(0); auto variance = result.at(1); - // mean->printIndexedBuffer("Mean"); - // variance->printIndexedBuffer("Variance"); - ASSERT_NEAR(sd::math::nd4j_abs(mean->e(0)), 0.f, 0.2f); ASSERT_NEAR(variance->e(0), 1.0f, 0.2f); @@ -371,8 +339,6 @@ TEST_F(RNGTests, Test_Gaussian_22) { RandomLauncher::fillGaussian(sd::LaunchContext::defaultContext(), _rngA, &x0, 0.0f, 1.0f); RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngB, &x1, 0.0f, 1.0f); - //x0.printIndexedBuffer("x0"); - //x1.printIndexedBuffer("x1"); ASSERT_TRUE(x0.equalsTo(&x1)); ASSERT_FALSE(x0.equalsTo(nexp0)); @@ -380,14 +346,11 @@ TEST_F(RNGTests, Test_Gaussian_22) { ASSERT_FALSE(x0.equalsTo(nexp2)); sd::ops::moments op; auto result = op.evaluate({&x0}, {}, {}); - //x0.printIndexedBuffer("X0 Normal"); - //x1.printIndexedBuffer("X1 Normal"); + ASSERT_TRUE(result.status() == Status::OK()); auto mean0 = result.at(0); auto variance0 = result.at(1); - //mean0->printIndexedBuffer("Mean"); - //variance0->printIndexedBuffer("Variance"); ASSERT_NEAR(sd::math::nd4j_abs(mean0->e(0)), 0.f, 1.0e-3f); ASSERT_NEAR(variance0->e(0), 1.0f, 1.e-3f); @@ -398,8 +361,8 @@ TEST_F(RNGTests, Test_Gaussian_3) { RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngA, &x0, 0.0, 1.0); - auto mean = x0.meanNumber(); //.e(0); - auto stdev = x0.varianceNumber(sd::variance::SummaryStatsStandardDeviation, false);//.e(0); + auto mean = x0.meanNumber(); + auto stdev = x0.varianceNumber(sd::variance::SummaryStatsStandardDeviation, false); auto meanExp = NDArrayFactory::create(0.); auto devExp = NDArrayFactory::create(1.); ASSERT_TRUE(meanExp.equalsTo(mean, 1.e-3)); @@ -435,15 +398,11 @@ TEST_F(RNGTests, Test_Truncated_1) { /* Check up distribution */ auto mean = x1.reduceNumber(reduce::Mean); - // mean.printIndexedBuffer("Mean 1.0"); auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation = x1.varianceNumber(variance::SummaryStatsStandardDeviation, false); - //deviation /= (double)x1.lengthOf(); - // deviation.printIndexedBuffer("Deviation should be 2.0"); - // x1.printIndexedBuffer("Distribution TN"); - } + TEST_F(RNGTests, Test_Truncated_2) { auto x0 = NDArrayFactory::create('c', {1000, 1000}); auto x1 = NDArrayFactory::create('c', {1000, 1000}); @@ -452,20 +411,10 @@ TEST_F(RNGTests, Test_Truncated_2) { RandomLauncher::fillTruncatedNormal(LaunchContext::defaultContext(), _rngB, &x1, 1.0f, 2.0f); ASSERT_TRUE(x0.equalsTo(&x1)); - - //ASSERT_FALSE(x0.equalsTo(nexp0)); - //ASSERT_FALSE(x0.equalsTo(nexp1)); - //ASSERT_FALSE(x0.equalsTo(nexp2)); - /* Check up distribution */ auto mean = x1.reduceNumber(reduce::Mean); - // mean.printIndexedBuffer("Mean 1.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation = x1.varianceNumber(variance::SummaryStatsStandardDeviation, false); - //deviation /= (double)x1.lengthOf(); - // deviation.printIndexedBuffer("Deviation should be 2.0"); - //x1.printIndexedBuffer("Distribution TN"); ASSERT_NEAR(mean.e(0), 1.f, 0.5); ASSERT_NEAR(deviation.e(0), 2.f, 0.5); } @@ -480,40 +429,24 @@ TEST_F(RNGTests, Test_Truncated_21) { ASSERT_TRUE(x0.equalsTo(&x1)); auto mean0 = x0.reduceNumber(reduce::Mean); - // mean0.printIndexedBuffer("0Mean 1.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation0 = x0.varianceNumber(variance::SummaryStatsStandardDeviation, false); - // deviation0.printIndexedBuffer("0Deviation should be 2.0"); - - //ASSERT_FALSE(x0.equalsTo(nexp0)); - //ASSERT_FALSE(x0.equalsTo(nexp1)); - //ASSERT_FALSE(x0.equalsTo(nexp2)); /* Check up distribution */ auto mean = x1.reduceNumber(reduce::Mean); - // mean.printIndexedBuffer("Mean 1.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation = x1.varianceNumber(variance::SummaryStatsStandardDeviation, false); - //deviation /= (double)x1.lengthOf(); - // deviation.printIndexedBuffer("Deviation should be 2.0"); - //x1.printIndexedBuffer("Distribution TN"); + ASSERT_NEAR(mean.e(0), 1.f, 0.002); ASSERT_NEAR(deviation.e(0), 2.f, 0.5); sd::ops::moments op; auto result = op.evaluate({&x0}, {}, {}, {}, {}, false); - // result.at(0)->printBuffer("MEAN"); - // result.at(1)->printBuffer("VARIANCE"); - sd::ops::reduce_min minOp; sd::ops::reduce_max maxOp; auto minRes = minOp.evaluate({&x1}, {}, {}, {}); auto maxRes = maxOp.evaluate({&x0}, {}, {}, {}); - // minRes->at(0)->printBuffer("MIN for Truncated"); - // maxRes->at(0)->printBuffer("MAX for Truncated"); } TEST_F(RNGTests, Test_Truncated_22) { @@ -526,25 +459,13 @@ TEST_F(RNGTests, Test_Truncated_22) { ASSERT_TRUE(x0.equalsTo(&x1)); auto mean0 = x0.reduceNumber(reduce::Mean); - // mean0.printIndexedBuffer("0Mean 2.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation0 = x0.varianceNumber(variance::SummaryStatsStandardDeviation, false); - // deviation0.printIndexedBuffer("0Deviation should be 4.0"); - - //ASSERT_FALSE(x0.equalsTo(nexp0)); - //ASSERT_FALSE(x0.equalsTo(nexp1)); - //ASSERT_FALSE(x0.equalsTo(nexp2)); /* Check up distribution */ auto mean = x1.reduceNumber(reduce::Mean); - // mean.printIndexedBuffer("Mean 2.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation = x1.varianceNumber(variance::SummaryStatsStandardDeviation, false); - //deviation /= (double)x1.lengthOf(); - // deviation.printIndexedBuffer("Deviation should be 4.0"); - //x1.printIndexedBuffer("Distribution TN"); ASSERT_NEAR(mean.e(0), 2.f, 0.01); ASSERT_NEAR(deviation.e(0), 4.f, 0.52); sd::ops::moments op; @@ -555,9 +476,6 @@ TEST_F(RNGTests, Test_Truncated_22) { auto minRes = minOp.evaluate({&x1}, {}, {}, {}); auto maxRes = maxOp.evaluate({&x0}, {}, {}, {}); - // minRes->at(0)->printBuffer("MIN for Truncated2"); - // maxRes->at(0)->printBuffer("MAX for Truncated2"); - } TEST_F(RNGTests, Test_Truncated_23) { @@ -570,38 +488,22 @@ TEST_F(RNGTests, Test_Truncated_23) { ASSERT_TRUE(x0.equalsTo(&x1)); auto mean0 = x0.reduceNumber(reduce::Mean); - // mean0.printIndexedBuffer("0Mean 2.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation0 = x0.varianceNumber(variance::SummaryStatsStandardDeviation, false); - // deviation0.printIndexedBuffer("0Deviation should be 4.0"); - - //ASSERT_FALSE(x0.equalsTo(nexp0)); - //ASSERT_FALSE(x0.equalsTo(nexp1)); - //ASSERT_FALSE(x0.equalsTo(nexp2)); /* Check up distribution */ auto mean = x1.reduceNumber(reduce::Mean); - // mean.printIndexedBuffer("Mean 2.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation = x1.varianceNumber(variance::SummaryStatsStandardDeviation, false); - //deviation /= (double)x1.lengthOf(); - // deviation.printIndexedBuffer("Deviation should be 4.0"); - //x1.printIndexedBuffer("Distribution TN"); ASSERT_NEAR(mean.e(0), 0.f, 0.01); ASSERT_NEAR(deviation.e(0), 1.f, 0.5); sd::ops::moments op; auto result = op.evaluate({&x0}); - // result->at(0)->printBuffer("MEAN"); - // result->at(1)->printBuffer("VARIANCE"); sd::ops::reduce_min minOp; sd::ops::reduce_max maxOp; auto minRes = minOp.evaluate({&x1}, {}, {}, {}); auto maxRes = maxOp.evaluate({&x0}, {}, {}, {}); - // minRes->at(0)->printBuffer("MIN for Truncated3"); - // maxRes->at(0)->printBuffer("MAX for Truncated3"); } @@ -616,8 +518,6 @@ TEST_F(RNGTests, Test_Truncated_3) { // Check up distribution auto mean = x1.reduceNumber(reduce::Mean); - // mean.printIndexedBuffer("Mean 1.0"); - //auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation = x1.varianceNumber(variance::SummaryStatsStandardDeviation, false); ASSERT_NEAR(mean.e(0), 1.f, 0.001); @@ -634,9 +534,6 @@ TEST_F(RNGTests, Test_Binomial_1) { ASSERT_TRUE(x0.equalsTo(&x1)); - //nexp2->printIndexedBuffer("nexp2"); - //x0.printIndexedBuffer("x0"); - ASSERT_FALSE(x0.equalsTo(nexp0)); ASSERT_FALSE(x0.equalsTo(nexp1)); ASSERT_FALSE(x0.equalsTo(nexp2)); @@ -669,7 +566,6 @@ TEST_F(RNGTests, Test_Uniform_SGA_3) { RandomLauncher::fillUniform(LaunchContext::defaultContext(), _rngB, &x1, -sd::DataTypeUtils::template max(), sd::DataTypeUtils::template max()); auto minimumU = x1.reduceNumber(reduce::AMin); - minimumU.printBuffer("\nMinimum"); } TEST_F(RNGTests, Test_Gaussian_2) { @@ -827,17 +723,14 @@ TEST_F(RNGTests, Test_ExponentialDistribution_1) { auto z = result.at(0); ASSERT_TRUE(exp0.isSameShape(z)); ASSERT_FALSE(exp0.equalsTo(z)); - // - z->printBuffer("\nExponential1"); + auto mean = z->reduceNumber(reduce::Mean); auto variance = z->varianceNumber(variance::SummaryStatsVariance, false); - mean.printBuffer("Mean for exponential with param 0.25 (4 exp) is"); - variance.printBuffer("Variance for exponential with param 0.25 (16 exp) is"); + ASSERT_FALSE(nexp0->equalsTo(z)); ASSERT_FALSE(nexp1->equalsTo(z)); ASSERT_FALSE(nexp2->equalsTo(z)); -// delete result; } TEST_F(RNGTests, Test_ExponentialDistribution_1_SGA) { @@ -852,12 +745,10 @@ TEST_F(RNGTests, Test_ExponentialDistribution_1_SGA) { auto z = result.at(0); ASSERT_TRUE(exp0.isSameShape(z)); ASSERT_FALSE(exp0.equalsTo(z)); - // - z->printBuffer("\nExponential2"); + auto mean = z->reduceNumber(reduce::Mean); auto variance = z->varianceNumber(variance::SummaryStatsVariance, false); - mean.printBuffer("Mean for exponential with param 1.0 (1 exp) is"); - variance.printBuffer("Variance for exponential with param 1. (1 exp) is"); + ASSERT_FALSE(nexp0->equalsTo(z)); ASSERT_FALSE(nexp1->equalsTo(z)); ASSERT_FALSE(nexp2->equalsTo(z)); @@ -878,60 +769,16 @@ TEST_F(RNGTests, Test_ExponentialDistribution_2_SGA) { auto z = result.at(0); ASSERT_TRUE(exp0.isSameShape(z)); ASSERT_FALSE(exp0.equalsTo(z)); - // -// z->printBuffer("\nExponential2+"); + auto mean = z->reduceNumber(reduce::Mean); auto variance = z->varianceNumber(variance::SummaryStatsVariance, false); - mean.printBuffer("Mean for exponential with param 1.0 (1 exp) is"); - variance.printBuffer("Variance for exponential with param 1. (1 exp) is"); + ASSERT_FALSE(nexp0->equalsTo(z)); ASSERT_FALSE(nexp1->equalsTo(z)); ASSERT_FALSE(nexp2->equalsTo(z)); mean = exp0.reduceNumber(reduce::Mean); variance = exp0.varianceNumber(variance::SummaryStatsVariance, false); - mean.printBuffer("Mean for exponential with param 2.0 (1/2 exp) is"); - variance.printBuffer("Variance for exponential with param 2. (1/2 exp) is"); -} -TEST_F(RNGTests, Test_ExponentialDistribution_3_SGA) { - auto x = NDArrayFactory::create('c', {2}, {1000, 1000}); - auto exp0 = NDArrayFactory::create('c', {1000, 1000}); - RandomGenerator oc(2716049175077475646L, -6182841917129177862L); - auto expMean = NDArrayFactory::create(0.5f); - auto expVar = NDArrayFactory::create(0.25f); - sd::ops::random_exponential op; - RandomLauncher::fillExponential(exp0.getContext(), oc, &exp0, 2.f); - - auto result = op.evaluate({&x}, {1.}); - ASSERT_EQ(Status::OK(), result.status()); - - auto z = result.at(0); - //ASSERT_TRUE(exp0.isSameShape(z)); - //ASSERT_FALSE(exp0.equalsTo(z)); - // -// z->printBuffer("\nExponential2+"); - auto mean = z->reduceNumber(reduce::Mean); - auto variance = z->varianceNumber(variance::SummaryStatsVariance, false); - mean.printBuffer("Mean"); - variance.printBuffer("Variance"); - ASSERT_NEAR(mean.e(0), 1.f, 1.e-2f); - ASSERT_NEAR(variance.e(0), 1.f, 1.e-2f); -// mean.printBuffer("Mean for exponential with param 1.0 (1 exp) is"); -// variance.printBuffer("Variance for exponential with param 1. (1 exp) is"); -// ASSERT_FALSE(nexp0->equalsTo(z)); -// ASSERT_FALSE(nexp1->equalsTo(z)); -// ASSERT_FALSE(nexp2->equalsTo(z)); - mean = exp0.reduceNumber(reduce::Mean); - variance = exp0.varianceNumber(variance::SummaryStatsVariance, false); - mean.printBuffer("Mean for exponential with param 2.0 (1/2 exp) is"); - variance.printBuffer("Variance for exponential with param 2. (1/4 exp) is"); - ASSERT_TRUE(mean.equalsTo(expMean, 1.e-3)); - ASSERT_TRUE(variance.equalsTo(expVar, 1.e-3)); - RandomLauncher::fillExponential(exp0.getContext(), oc, &exp0, 1.f); - mean = exp0.reduceNumber(reduce::Mean); - variance = exp0.varianceNumber(variance::SummaryStatsVariance, false); - mean.printBuffer("Mean for exponential with param 1.0 (1 exp) is"); - variance.printBuffer("Variance for exponential with param 1.0 (1 exp) is"); } TEST_F(RNGTests, Test_ExponentialDistribution_2) { @@ -971,11 +818,8 @@ TEST_F(RNGTests, Test_PoissonDistribution_1) { ASSERT_EQ(Status::OK(), result.status()); auto z = result.at(0); -// z->printIndexedBuffer("Poisson distribution"); ASSERT_TRUE(exp0.isSameShape(z)); ASSERT_FALSE(exp0.equalsTo(z)); - - } TEST_F(RNGTests, Test_GammaDistribution_1) { @@ -991,11 +835,9 @@ TEST_F(RNGTests, Test_GammaDistribution_1) { ASSERT_EQ(Status::OK(), result.status()); auto z = result.at(0); -// z->printIndexedBuffer("Gamma distribution"); + ASSERT_TRUE(exp0.isSameShape(z)); ASSERT_FALSE(exp0.equalsTo(z)); - - } TEST_F(RNGTests, Test_GammaDistribution_2) { @@ -1012,7 +854,7 @@ TEST_F(RNGTests, Test_GammaDistribution_2) { ASSERT_EQ(Status::OK(), result.status()); auto z = result.at(0); -// z->printIndexedBuffer("Gamma distribution"); + ASSERT_TRUE(exp0.isSameShape(z)); ASSERT_FALSE(exp0.equalsTo(z)); } @@ -1031,7 +873,6 @@ TEST_F(RNGTests, Test_GammaDistribution_3) { ASSERT_EQ(Status::OK(), result.status()); auto z = result.at(0); -// z->printIndexedBuffer("Gamma distribution"); ASSERT_TRUE(exp0.isSameShape(z)); ASSERT_FALSE(exp0.equalsTo(z)); @@ -1127,7 +968,6 @@ TEST_F(RNGTests, Test_UniformDistribution_05) { sd::ops::reduce_max checkOp; auto checkResult = checkOp.evaluate({z}); - checkResult[0]->printIndexedBuffer("Max on uniform with 0 to 1 on 100M cases is"); } namespace sd { @@ -1168,7 +1008,6 @@ TEST_F(RNGTests, Test_Reproducibility_1) { bool t = arrayE->equalsTo(arrayT); if (!t) { - // nd4j_printf("Failed at iteration [%i] for array [%i]\n", e, a); ASSERT_TRUE(false); } @@ -1200,19 +1039,15 @@ TEST_F(RNGTests, Test_Reproducibility_2) { bool t = arrayE->equalsTo(arrayT); if (!t) { - // nd4j_printf("Failed at iteration [%i] for array [%i]\n", e, a); for (Nd4jLong f = 0; f < arrayE->lengthOf(); f++) { double x = arrayE->e(f); double y = arrayT->e(f); if (sd::math::nd4j_re(x, y) > 0.1) { - // nd4j_printf("E[%lld] %f != T[%lld] %f\n", (long long) f, (float) x, (long long) f, (float) y); throw std::runtime_error("boom"); } } - - // just breaker, since test failed ASSERT_TRUE(false); } @@ -1235,8 +1070,6 @@ TEST_F(RNGTests, Test_Uniform_4) { auto sumA = x1 - mean; //.reduceNumber(reduce::Sum); auto deviation = x1.varianceNumber(variance::SummaryStatsVariance, false); - //deviation /= (double)x1.lengthOf(); - // deviation.printIndexedBuffer("Deviation should be 1/12 (0.083333)"); ASSERT_NEAR(mean.e(0), 1.5, 1e-3); ASSERT_NEAR(1/12., deviation.e(0), 1e-3); @@ -1251,8 +1084,6 @@ TEST_F(RNGTests, test_choice_1) { RandomGenerator rng(119, 256); NativeOpExecutioner::execRandom(sd::LaunchContext ::defaultContext(), random::Choice, &rng, x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(), prob->buffer(), prob->shapeInfo(), prob->specialBuffer(), prob->specialShapeInfo(), z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr); - // z.printIndexedBuffer("z"); - delete x; delete prob; } @@ -1367,7 +1198,7 @@ TEST_F(RNGTests, test_multinomial_5) { ASSERT_EQ(Status::OK(), op.execute(rng, { &probs, &samples }, { &output }, {}, { 1 }, {}, {}, false)); auto deviation = output.varianceNumber(variance::SummaryStatsStandardDeviation, false); auto mean = output.meanNumber(); - // printf("Var: %f Mean: %f \n", deviation.e(0), mean.e(0)); + // theoretical values for binomial ASSERT_NEAR(0.5, deviation.e(0), 4e-3); // 1000000 3e-3); ASSERT_NEAR(0.5, mean.e(0), 4e-3); // 1000000 3e-3); @@ -1383,7 +1214,7 @@ TEST_F(RNGTests, test_multinomial_5) { deviation = outputR->varianceNumber(variance::SummaryStatsStandardDeviation, false); mean = outputR->meanNumber(); - // printf("Random seed - Var: %f Mean: %f \n", deviation.e(0), mean.e(0)); + ASSERT_NEAR(0.5, deviation.e(0), 45e-3); // 1000000 35e-3); ASSERT_NEAR(0.5, mean.e(0), 45e-3); // 1000000 35e-3); @@ -1425,13 +1256,13 @@ TEST_F(RNGTests, test_multinomial_6) { for (int i = 0; i < countsR.lengthOf(); i++) { auto c = countsR.e(i); auto p = probExpect.e(i); - // printf("Get freq : %f Expect freq: %f \n", c / Samples, p); + ASSERT_NEAR((c / Samples), p, 45e-3); // 1000000 35e-3); } auto deviation = outputR->varianceNumber(variance::SummaryStatsStandardDeviation, false); auto mean = outputR->meanNumber(); - // printf("Var: %f Mean: %f \n", deviation.e(0), mean.e(0)); + ASSERT_NEAR(1.2175, deviation.e(0), 45e-3); // 1000000 35e-3); ASSERT_NEAR(2.906, mean.e(0), 45e-3); // 1000000 35e-3); @@ -1454,13 +1285,12 @@ TEST_F(RNGTests, test_multinomial_6) { for (int i = 0; i < counts.lengthOf(); i++) { auto c = counts.e(i); auto p = probExpect.e(i); - // printf("Get freq : %f Expect freq: %f \n", c / Samples, p); + ASSERT_NEAR((c / Samples), p, 4e-3); // 1000000 3e-3); } deviation = output.varianceNumber(variance::SummaryStatsStandardDeviation, false); mean = output.meanNumber(); - // printf("Var: %f Mean: %f \n", deviation.e(0), mean.e(0)); ASSERT_NEAR(1.2175, deviation.e(0), 5e-3); // 1000000 3e-3); ASSERT_NEAR(2.906, mean.e(0), 5e-3); // 1000000 3e-3); } diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/memory/abstracts/Nd4jWorkspace.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/memory/abstracts/Nd4jWorkspace.java index 45029fb1e..e133e171b 100644 --- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/memory/abstracts/Nd4jWorkspace.java +++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/memory/abstracts/Nd4jWorkspace.java @@ -124,6 +124,9 @@ public abstract class Nd4jWorkspace implements MemoryWorkspace { protected AtomicLong generationId = new AtomicLong(0); + // this field is used as alignment base for all allocations within this workspace + public final static int alignmentBase = 16; + // this memory manager implementation will be used to allocate real memory for this workspace public Nd4jWorkspace(@NonNull WorkspaceConfiguration configuration) { @@ -340,9 +343,9 @@ public abstract class Nd4jWorkspace implements MemoryWorkspace { long numElements = requiredMemory / Nd4j.sizeOfDataType(type); // we enforce 8 byte alignment to ensure CUDA doesn't blame us - long div = requiredMemory % 8; + long div = requiredMemory % alignmentBase; if (div != 0) - requiredMemory += (8 - div); + requiredMemory += (alignmentBase - div); // shortcut made to skip workspace if (!isUsed.get()) { diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda-platform/pom.xml b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda-platform/pom.xml index 344e77861..d86e0d353 100644 --- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda-platform/pom.xml +++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda-platform/pom.xml @@ -22,14 +22,14 @@ 4.0.0 - nd4j-cuda-10.2-platform + nd4j-cuda-11.0-platform nd4j-cuda-platform - 10.2 - 7.6 - 1.5.3 + 11.0 + 8.0 + 1.5.4-SNAPSHOT nd4j-cuda-${cuda.version} diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/pom.xml b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/pom.xml index c181d4328..ca03fbcae 100644 --- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/pom.xml +++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/pom.xml @@ -22,14 +22,14 @@ 4.0.0 - nd4j-cuda-10.2 + nd4j-cuda-11.0 nd4j-cuda - 10.2 - 7.6 - 1.5.3 + 11.0 + 8.0 + 1.5.4-SNAPSHOT diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/workspace/CudaWorkspace.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/workspace/CudaWorkspace.java index 0c4cf9ffa..a04dae5d5 100644 --- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/workspace/CudaWorkspace.java +++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/workspace/CudaWorkspace.java @@ -92,9 +92,15 @@ public class CudaWorkspace extends Nd4jWorkspace { AllocationsTracker.getInstance().markAllocated(AllocationKind.GENERAL, Nd4j.getAffinityManager().getDeviceForCurrentThread(), bytes + SAFETY_OFFSET); MemoryTracker.getInstance().incrementWorkspaceAllocatedAmount(Nd4j.getAffinityManager().getDeviceForCurrentThread(), bytes + SAFETY_OFFSET); - } - //log.info("Workspace [{}] initialized successfully", id); + // if base pointer isn't aligned to 16 bytes (128 bits) - adjust the offfset then + val addr = workspace.getDevicePointer().address(); + val div = addr % alignmentBase; + if (div != 0) { + deviceOffset.set(alignmentBase - div); + hostOffset.set(alignmentBase - div); + } + } } } @@ -134,8 +140,9 @@ public class CudaWorkspace extends Nd4jWorkspace { public PagedPointer alloc(long requiredMemory, MemoryKind kind, DataType type, boolean initialize) { long numElements = requiredMemory / Nd4j.sizeOfDataType(type); - if (requiredMemory % 8 != 0) - requiredMemory += 8 - (requiredMemory % 8); + // alignment + if (requiredMemory % alignmentBase != 0) + requiredMemory += alignmentBase - (requiredMemory % alignmentBase); if (!isUsed.get()) { if (disabledCounter.incrementAndGet() % 10 == 0) diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java index 38c7188f1..6c4d0921e 100644 --- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java +++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java @@ -1,4 +1,4 @@ -// Targeted by JavaCPP version 1.5.3: DO NOT EDIT THIS FILE +// Targeted by JavaCPP version 1.5.4-SNAPSHOT: DO NOT EDIT THIS FILE package org.nd4j.nativeblas; diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java index 6ac8e133a..db09d9625 100644 --- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java +++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java @@ -135,20 +135,16 @@ public class Nd4jCudaPresets implements LoadEnabled, InfoMapper { return; } int i = 0; - String[] libs = {"cudart", "cublasLt", "cublas", "cusolver", "cusparse", "cudnn"}; + String[] libs = {"cudart", "cublasLt", "cublas", "curand", "cusolver", "cusparse", "cudnn", + "cudnn_ops_infer", "cudnn_ops_train", "cudnn_adv_infer", + "cudnn_adv_train", "cudnn_cnn_infer", "cudnn_cnn_train"}; for (String lib : libs) { - switch (platform) { - case "linux-arm64": - case "linux-ppc64le": - case "linux-x86_64": - case "macosx-x86_64": - lib += lib.equals("cudnn") ? "@.7" : lib.equals("cudart") ? "@.10.2" : "@.10"; - break; - case "windows-x86_64": - lib += lib.equals("cudnn") ? "64_7" : lib.equals("cudart") ? "64_102" : "64_10"; - break; - default: - continue; // no CUDA + if (platform.startsWith("linux")) { + lib += lib.startsWith("cudnn") ? "@.8" : lib.equals("curand") || lib.equals("cusolver") ? "@.10" : lib.equals("cudart") ? "@.11.0" : "@.11"; + } else if (platform.startsWith("windows")) { + lib += lib.startsWith("cudnn") ? "64_8" : lib.equals("curand") || lib.equals("cusolver") ? "64_10" : lib.equals("cudart") ? "64_110" : "64_11"; + } else { + continue; // no CUDA } if (!preloads.contains(lib)) { preloads.add(i++, lib); diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java index 2926c06b9..5387cf43b 100644 --- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java +++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java @@ -1,4 +1,4 @@ -// Targeted by JavaCPP version 1.5.3: DO NOT EDIT THIS FILE +// Targeted by JavaCPP version 1.5.4-SNAPSHOT: DO NOT EDIT THIS FILE package org.nd4j.nativeblas; diff --git a/nd4j/nd4j-backends/nd4j-tests-tensorflow/pom.xml b/nd4j/nd4j-backends/nd4j-tests-tensorflow/pom.xml index 3e4367992..66efac0fe 100644 --- a/nd4j/nd4j-backends/nd4j-tests-tensorflow/pom.xml +++ b/nd4j/nd4j-backends/nd4j-tests-tensorflow/pom.xml @@ -127,12 +127,6 @@ ${tensorflow.javacpp.version} windows-x86_64-gpu - - org.bytedeco - tensorflow - ${tensorflow.javacpp.version} - macosx-x86_64-gpu - @@ -233,7 +227,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} @@ -248,12 +242,6 @@ ${tensorflow.javacpp.version} windows-x86_64-gpu - - org.bytedeco - tensorflow - ${tensorflow.javacpp.version} - macosx-x86_64-gpu - diff --git a/nd4j/nd4j-backends/nd4j-tests/pom.xml b/nd4j/nd4j-backends/nd4j-tests/pom.xml index c621330d4..823b38590 100644 --- a/nd4j/nd4j-backends/nd4j-tests/pom.xml +++ b/nd4j/nd4j-backends/nd4j-tests/pom.xml @@ -212,7 +212,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/BasicWorkspaceTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/BasicWorkspaceTests.java index 290b69723..4c3679df6 100644 --- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/BasicWorkspaceTests.java +++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/BasicWorkspaceTests.java @@ -221,11 +221,11 @@ public class BasicWorkspaceTests extends BaseNd4jTest { INDArray array2 = Nd4j.create(new double[] {1f, 2f, 3f, 4f, 5f}); long reqMemory = 5 * Nd4j.sizeOfDataType(DOUBLE); - assertEquals(reqMemory + reqMemory % 8, wsOne.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsOne.getPrimaryOffset()); array2.leverageTo("EXT"); - assertEquals((reqMemory + reqMemory % 8) * 2, wsOne.getPrimaryOffset()); + assertEquals((reqMemory + reqMemory % 16) * 2, wsOne.getPrimaryOffset()); } } } @@ -237,7 +237,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { INDArray array1 = Nd4j.create(new double[] {1f, 2f, 3f, 4f, 5f}); long reqMemory = 5 * Nd4j.sizeOfDataType(array1.dataType()); - assertEquals(reqMemory + reqMemory % 8, wsOne.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsOne.getPrimaryOffset()); INDArray array2; @@ -252,7 +252,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { INDArray array3 = Nd4j.create(new double[] {1f, 2f, 3f, 4f, 5f}); reqMemory = 5 * Nd4j.sizeOfDataType(array3.dataType()); - assertEquals((reqMemory + reqMemory % 8) * 2, wsOne.getPrimaryOffset()); + assertEquals((reqMemory + reqMemory % 16) * 2, wsOne.getPrimaryOffset()); array1.addi(array2); @@ -275,12 +275,12 @@ public class BasicWorkspaceTests extends BaseNd4jTest { assertEquals(0, wsOne.getPrimaryOffset()); long reqMemory = 5 * Nd4j.sizeOfDataType(array.dataType()); - assertEquals(reqMemory + reqMemory % 8, wsTwo.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsTwo.getPrimaryOffset()); INDArray copy = array.leverage(); - assertEquals(reqMemory + reqMemory % 8, wsTwo.getPrimaryOffset()); - assertEquals(reqMemory + reqMemory % 8, wsOne.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsTwo.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsOne.getPrimaryOffset()); assertNotEquals(null, copy); @@ -324,7 +324,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { array2.assign(array1); long reqMemory = 5 * Nd4j.sizeOfDataType(array1.dataType()); - assertEquals(reqMemory + reqMemory % 8, wsI.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsI.getPrimaryOffset()); assertEquals(array1, array2); INDArray array3 = Nd4j.createUninitializedDetached(DataType.FLOAT, new long[0]); @@ -348,17 +348,17 @@ public class BasicWorkspaceTests extends BaseNd4jTest { assertTrue(array.isAttached()); long reqMemory = 5 * Nd4j.sizeOfDataType(array.dataType()); - assertEquals(reqMemory + reqMemory % 8, wsI.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsI.getPrimaryOffset()); copy = array.detach(); assertTrue(array.isInScope()); assertTrue(array.isAttached()); - assertEquals(reqMemory + reqMemory % 8, wsI.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsI.getPrimaryOffset()); assertFalse(copy.isAttached()); assertTrue(copy.isInScope()); - assertEquals(reqMemory + reqMemory % 8, wsI.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, wsI.getPrimaryOffset()); } assertEquals(15.0f, copy.sumNumber().floatValue(), 0.01f); @@ -685,7 +685,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { workspace.initializeWorkspace(); long reqMemory = 11 * Nd4j.sizeOfDataType(arrayCold.dataType()); - assertEquals(reqMemory + reqMemory % 8, workspace.getCurrentSize()); + assertEquals(reqMemory + reqMemory % 16, workspace.getCurrentSize()); log.info("-----------------------"); @@ -700,11 +700,11 @@ public class BasicWorkspaceTests extends BaseNd4jTest { long reqMem = 10 * Nd4j.sizeOfDataType(array.dataType()); - assertEquals(reqMem + reqMem % 8, workspace.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, workspace.getPrimaryOffset()); array.addi(1.0); - assertEquals(reqMem + reqMem % 8, workspace.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, workspace.getPrimaryOffset()); assertEquals("Failed on iteration " + x, 10, array.sumNumber().doubleValue(), 0.01); @@ -752,13 +752,13 @@ public class BasicWorkspaceTests extends BaseNd4jTest { // checking if allocation actually happened long reqMemory = 5 * Nd4j.sizeOfDataType(DOUBLE); - assertEquals(reqMemory + reqMemory % 8, workspace.getPrimaryOffset()); + assertEquals(reqMemory + reqMemory % 16, workspace.getPrimaryOffset()); array.assign(1.0f); INDArray dup = array.dup(); - assertEquals((reqMemory + reqMemory % 8) * 2, workspace.getPrimaryOffset()); + assertEquals((reqMemory + reqMemory % 16) * 2, workspace.getPrimaryOffset()); assertEquals(5, dup.sumNumber().doubleValue(), 0.01); @@ -786,7 +786,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { // checking if allocation actually happened long reqMem = 5 * Nd4j.sizeOfDataType(array.dataType()); - assertEquals(reqMem + reqMem % 8, workspace.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, workspace.getPrimaryOffset()); try { INDArray array2 = Nd4j.create(DOUBLE, 10000000); @@ -795,11 +795,11 @@ public class BasicWorkspaceTests extends BaseNd4jTest { assertTrue(true); } - assertEquals(reqMem + reqMem % 8, workspace.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, workspace.getPrimaryOffset()); INDArray array2 = Nd4j.create(DOUBLE, new long[] {1, 5}, 'c'); - assertEquals((reqMem + reqMem % 8) * 2, workspace.getPrimaryOffset()); + assertEquals((reqMem + reqMem % 16) * 2, workspace.getPrimaryOffset()); } @Test @@ -817,7 +817,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { // checking if allocation actually happened long reqMem = 5 * Nd4j.sizeOfDataType(array.dataType()); - assertEquals(reqMem + reqMem % 8, workspace.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, workspace.getPrimaryOffset()); array.assign(1.0f); @@ -841,7 +841,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { // checking if allocation actually happened long reqMem = 5 * Nd4j.sizeOfDataType(array.dataType()); - assertEquals(reqMem + reqMem % 8, workspace.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, workspace.getPrimaryOffset()); array.assign(1.0f); @@ -870,7 +870,7 @@ public class BasicWorkspaceTests extends BaseNd4jTest { // checking if allocation actually happened long reqMem = 5 * Nd4j.sizeOfDataType(array.dataType()); - assertEquals(reqMem + reqMem % 8, workspace.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, workspace.getPrimaryOffset()); assertEquals(exp, array); diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/WorkspaceProviderTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/WorkspaceProviderTests.java index 20e7f367f..fb6975c7f 100644 --- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/WorkspaceProviderTests.java +++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/workspace/WorkspaceProviderTests.java @@ -239,15 +239,15 @@ public class WorkspaceProviderTests extends BaseNd4jTest { INDArray array3 = null; long reqMem = 5 * Nd4j.sizeOfDataType(DataType.DOUBLE); - assertEquals(reqMem + reqMem % 8, ws1.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, ws1.getPrimaryOffset()); try (Nd4jWorkspace ws2 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS2") .notifyScopeEntered()) { INDArray array2 = Nd4j.create(new double[] {1f, 2f, 3f, 4f, 5f}); reqMem = 5 * Nd4j.sizeOfDataType(DataType.DOUBLE); - assertEquals(reqMem + reqMem % 8, ws1.getPrimaryOffset()); - assertEquals(reqMem + reqMem % 8, ws2.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, ws1.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, ws2.getPrimaryOffset()); try (Nd4jWorkspace ws3 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS1") .notifyScopeBorrowed()) { @@ -256,15 +256,15 @@ public class WorkspaceProviderTests extends BaseNd4jTest { array3 = array2.unsafeDuplication(); assertTrue(ws1 == array3.data().getParentWorkspace()); - assertEquals(reqMem + reqMem % 8, ws2.getPrimaryOffset()); - assertEquals((reqMem + reqMem % 8) * 2, ws1.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, ws2.getPrimaryOffset()); + assertEquals((reqMem + reqMem % 16) * 2, ws1.getPrimaryOffset()); } log.info("Current workspace: {}", Nd4j.getMemoryManager().getCurrentWorkspace()); assertTrue(ws2 == Nd4j.getMemoryManager().getCurrentWorkspace()); - assertEquals(reqMem + reqMem % 8, ws2.getPrimaryOffset()); - assertEquals((reqMem + reqMem % 8) * 2, ws1.getPrimaryOffset()); + assertEquals(reqMem + reqMem % 16, ws2.getPrimaryOffset()); + assertEquals((reqMem + reqMem % 16) * 2, ws1.getPrimaryOffset()); assertEquals(15f, array3.sumNumber().floatValue(), 0.01f); } @@ -279,20 +279,18 @@ public class WorkspaceProviderTests extends BaseNd4jTest { public void testNestedWorkspacesOverlap1() { Nd4j.setDefaultDataTypes(DataType.FLOAT, DataType.FLOAT); Nd4j.getWorkspaceManager().setDefaultWorkspaceConfiguration(basicConfiguration); - try (Nd4jWorkspace ws1 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS1") - .notifyScopeEntered()) { + try (Nd4jWorkspace ws1 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS1").notifyScopeEntered()) { INDArray array = Nd4j.create(new float[] {1f, 2f, 3f, 4f, 5f}); long reqMem = 5 * Nd4j.sizeOfDataType(); - assertEquals(reqMem + reqMem % 8, ws1.getPrimaryOffset()); - try (Nd4jWorkspace ws2 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS2") - .notifyScopeEntered()) { + assertEquals(reqMem + (Nd4jWorkspace.alignmentBase - reqMem % Nd4jWorkspace.alignmentBase), ws1.getPrimaryOffset()); + try (Nd4jWorkspace ws2 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS2").notifyScopeEntered()) { INDArray array2 = Nd4j.create(new float[] {1f, 2f, 3f, 4f, 5f}); reqMem = 5 * Nd4j.sizeOfDataType(); - assertEquals(reqMem + reqMem % 8, ws1.getPrimaryOffset()); - assertEquals(reqMem + reqMem % 8, ws2.getPrimaryOffset()); + assertEquals(reqMem + (Nd4jWorkspace.alignmentBase - reqMem % Nd4jWorkspace.alignmentBase), ws1.getPrimaryOffset()); + assertEquals(reqMem + (Nd4jWorkspace.alignmentBase - reqMem % Nd4jWorkspace.alignmentBase), ws2.getPrimaryOffset()); try (Nd4jWorkspace ws3 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS1") .notifyScopeBorrowed()) { @@ -300,8 +298,8 @@ public class WorkspaceProviderTests extends BaseNd4jTest { INDArray array3 = Nd4j.create(new float[] {1f, 2f, 3f, 4f, 5f}); - assertEquals(reqMem + reqMem % 8, ws2.getPrimaryOffset()); - assertEquals((reqMem + reqMem % 8) * 2, ws1.getPrimaryOffset()); + assertEquals(reqMem + (Nd4jWorkspace.alignmentBase - reqMem % Nd4jWorkspace.alignmentBase), ws2.getPrimaryOffset()); + assertEquals((reqMem + (Nd4jWorkspace.alignmentBase - reqMem % Nd4jWorkspace.alignmentBase)) * 2, ws1.getPrimaryOffset()); } } } @@ -901,21 +899,18 @@ public class WorkspaceProviderTests extends BaseNd4jTest { assertEquals(100 * Nd4j.sizeOfDataType(), ws1.getPrimaryOffset()); for (int x = 1; x <= 100; x++) { - try (Nd4jWorkspace ws2 = (Nd4jWorkspace) Nd4j.getWorkspaceManager() - .getWorkspaceForCurrentThread(loopConfiguration, "WS2").notifyScopeEntered()) { + try (Nd4jWorkspace ws2 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread(loopConfiguration, "WS2").notifyScopeEntered()) { INDArray array2 = Nd4j.create(x); } Nd4jWorkspace ws2 = (Nd4jWorkspace) Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS2"); long reqMemory = x * Nd4j.sizeOfDataType(); - assertEquals(reqMemory + reqMemory % 8, ws2.getLastCycleAllocations()); + assertEquals(reqMemory + reqMemory % 16, ws2.getLastCycleAllocations()); } Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS2").initializeWorkspace(); - assertEquals(100 * Nd4j.sizeOfDataType(), - Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS2") - .getCurrentSize()); + assertEquals(100 * Nd4j.sizeOfDataType(), Nd4j.getWorkspaceManager().getWorkspaceForCurrentThread("WS2").getCurrentSize()); } assertNull(Nd4j.getMemoryManager().getCurrentWorkspace()); diff --git a/nd4j/nd4j-jdbc/nd4j-jdbc-hsql/pom.xml b/nd4j/nd4j-jdbc/nd4j-jdbc-hsql/pom.xml index b0941300a..818fc69aa 100644 --- a/nd4j/nd4j-jdbc/nd4j-jdbc-hsql/pom.xml +++ b/nd4j/nd4j-jdbc/nd4j-jdbc-hsql/pom.xml @@ -94,7 +94,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-client/pom.xml b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-client/pom.xml index 75024af0f..8938e6a4d 100644 --- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-client/pom.xml +++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-client/pom.xml @@ -120,7 +120,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/pom.xml b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/pom.xml index 07a04f80d..33ada7aa7 100644 --- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/pom.xml +++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/pom.xml @@ -119,7 +119,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/pom.xml b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/pom.xml index 0325f2d52..ec93d0be7 100644 --- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/pom.xml +++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/pom.xml @@ -113,7 +113,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml b/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml index 88af3c166..6a437341b 100644 --- a/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml +++ b/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml @@ -114,7 +114,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/nd4j/nd4j-remote/nd4j-json-client/pom.xml b/nd4j/nd4j-remote/nd4j-json-client/pom.xml index e1a661c8b..9e8b42dab 100644 --- a/nd4j/nd4j-remote/nd4j-json-client/pom.xml +++ b/nd4j/nd4j-remote/nd4j-json-client/pom.xml @@ -95,7 +95,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-remote/nd4j-json-server/pom.xml b/nd4j/nd4j-remote/nd4j-json-server/pom.xml index 004145101..ded578721 100644 --- a/nd4j/nd4j-remote/nd4j-json-server/pom.xml +++ b/nd4j/nd4j-remote/nd4j-json-server/pom.xml @@ -171,7 +171,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/nd4j/nd4j-serde/nd4j-aeron/pom.xml b/nd4j/nd4j-serde/nd4j-aeron/pom.xml index c94bf86af..9245a3998 100644 --- a/nd4j/nd4j-serde/nd4j-aeron/pom.xml +++ b/nd4j/nd4j-serde/nd4j-aeron/pom.xml @@ -109,7 +109,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-serde/nd4j-arrow/pom.xml b/nd4j/nd4j-serde/nd4j-arrow/pom.xml index 69879e965..26e59c510 100644 --- a/nd4j/nd4j-serde/nd4j-arrow/pom.xml +++ b/nd4j/nd4j-serde/nd4j-arrow/pom.xml @@ -127,7 +127,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-serde/nd4j-kryo/pom.xml b/nd4j/nd4j-serde/nd4j-kryo/pom.xml index 87ddca9f0..92a1c0315 100644 --- a/nd4j/nd4j-serde/nd4j-kryo/pom.xml +++ b/nd4j/nd4j-serde/nd4j-kryo/pom.xml @@ -187,7 +187,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} diff --git a/nd4j/nd4j-uberjar/pom.xml b/nd4j/nd4j-uberjar/pom.xml index 388e48fc5..fa6bd84d0 100644 --- a/nd4j/nd4j-uberjar/pom.xml +++ b/nd4j/nd4j-uberjar/pom.xml @@ -259,7 +259,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} @@ -291,12 +291,12 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} org.nd4j - nd4j-cuda-10.2-platform + nd4j-cuda-11.0-platform ${project.version} diff --git a/nd4s/pom.xml b/nd4s/pom.xml index f10f8cf41..878cdd02a 100644 --- a/nd4s/pom.xml +++ b/nd4s/pom.xml @@ -312,11 +312,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test diff --git a/pom.xml b/pom.xml index 184eeb11f..915409e35 100644 --- a/pom.xml +++ b/pom.xml @@ -289,24 +289,24 @@ ${javacpp.platform} - 1.5.3 - 1.5.3 - 1.5.3 + 1.5.4-SNAPSHOT + 1.5.4-SNAPSHOT + 1.5.4-SNAPSHOT - 3.7.7 + 3.7.8 ${python.version}-${javacpp-presets.version} - 1.18.2 + 1.19.0 ${numpy.version}-${javacpp-presets.version} - 0.3.9-1 + 0.3.10 2020.1 4.3.0 - 4.2.2 + 4.3 1.79.0 1.12.0 0.6.1 - 0.17.1 - 1.15.2 + 0.17.2 + 1.15.3 ${tensorflow.version}-${javacpp-presets.version} 1.18 diff --git a/python4j/pom.xml b/python4j/pom.xml index 3f1d026a5..4f672b999 100644 --- a/python4j/pom.xml +++ b/python4j/pom.xml @@ -44,7 +44,8 @@ org.slf4j slf4j-api 1.6.6 - + + ch.qos.logback logback-classic ${logback.version} @@ -66,10 +67,5 @@ jsr305 3.0.2 - - org.slf4j - slf4j-api - 1.6.6 - \ No newline at end of file diff --git a/python4j/python4j-numpy/pom.xml b/python4j/python4j-numpy/pom.xml index c631f67e3..7d2ebe7de 100644 --- a/python4j/python4j-numpy/pom.xml +++ b/python4j/python4j-numpy/pom.xml @@ -55,11 +55,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.1 + nd4j-cuda-11.0 ${nd4j.version} test diff --git a/rl4j/pom.xml b/rl4j/pom.xml index c91dd7aa2..e3cdb5ca1 100644 --- a/rl4j/pom.xml +++ b/rl4j/pom.xml @@ -112,7 +112,7 @@ ${skipBackendChoice} - test-nd4j-native,test-nd4j-cuda-10.2 + test-nd4j-native,test-nd4j-cuda-11.0 false @@ -303,11 +303,11 @@ - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${nd4j.version} test diff --git a/rl4j/rl4j-ale/pom.xml b/rl4j/rl4j-ale/pom.xml index 360c3db1d..5e97eceb0 100644 --- a/rl4j/rl4j-ale/pom.xml +++ b/rl4j/rl4j-ale/pom.xml @@ -50,7 +50,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/rl4j/rl4j-api/pom.xml b/rl4j/rl4j-api/pom.xml index 629783e15..43bf18cd8 100644 --- a/rl4j/rl4j-api/pom.xml +++ b/rl4j/rl4j-api/pom.xml @@ -45,7 +45,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/rl4j/rl4j-core/pom.xml b/rl4j/rl4j-core/pom.xml index bbb66a9e9..1cac4490d 100644 --- a/rl4j/rl4j-core/pom.xml +++ b/rl4j/rl4j-core/pom.xml @@ -135,7 +135,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/Agent.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/Agent.java index 198c2a1ca..b94ed8d61 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/Agent.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/Agent.java @@ -16,8 +16,10 @@ package org.deeplearning4j.rl4j.agent; import lombok.AccessLevel; +import lombok.Data; import lombok.Getter; import lombok.NonNull; +import lombok.experimental.SuperBuilder; import org.deeplearning4j.rl4j.agent.listener.AgentListener; import org.deeplearning4j.rl4j.agent.listener.AgentListenerList; import org.deeplearning4j.rl4j.environment.Environment; @@ -69,16 +71,20 @@ public class Agent implements IAgent { * @param environment The {@link Environment} to be used * @param transformProcess The {@link TransformProcess} to be used to transform the raw observations into usable ones. * @param policy The {@link IPolicy} to be used - * @param maxEpisodeSteps The maximum number of steps an episode can have before being interrupted. Use null to have no max. + * @param configuration The configuration for the agent * @param id A user-supplied id to identify the instance. */ - public Agent(@NonNull Environment environment, @NonNull TransformProcess transformProcess, @NonNull IPolicy policy, Integer maxEpisodeSteps, String id) { - Preconditions.checkArgument(maxEpisodeSteps == null || maxEpisodeSteps > 0, "maxEpisodeSteps must be null (no maximum) or greater than 0, got", maxEpisodeSteps); + public Agent(@NonNull Environment environment, + @NonNull TransformProcess transformProcess, + @NonNull IPolicy policy, + @NonNull Configuration configuration, + String id) { + Preconditions.checkArgument(configuration.getMaxEpisodeSteps() == null || configuration.getMaxEpisodeSteps() > 0, "Configuration: maxEpisodeSteps must be null (no maximum) or greater than 0, got", configuration.getMaxEpisodeSteps()); this.environment = environment; this.transformProcess = transformProcess; this.policy = policy; - this.maxEpisodeSteps = maxEpisodeSteps; + this.maxEpisodeSteps = configuration.getMaxEpisodeSteps(); this.id = id; listeners = buildListenerList(); @@ -126,6 +132,7 @@ public class Agent implements IAgent { } onAfterEpisode(); + listeners.notifyAfterEpisode(this); } protected void reset() { @@ -217,45 +224,13 @@ public class Agent implements IAgent { // Do Nothing } - /** - * - * @param environment - * @param transformProcess - * @param policy - * @param - * @return - */ - public static Builder builder(@NonNull Environment environment, @NonNull TransformProcess transformProcess, @NonNull IPolicy policy) { - return new Builder<>(environment, transformProcess, policy); - } - - public static class Builder { - protected final Environment environment; - protected final TransformProcess transformProcess; - protected final IPolicy policy; - protected Integer maxEpisodeSteps = null; // Default, no max - protected String id; - - public Builder(@NonNull Environment environment, @NonNull TransformProcess transformProcess, @NonNull IPolicy policy) { - this.environment = environment; - this.transformProcess = transformProcess; - this.policy = policy; - } - - public Builder maxEpisodeSteps(int maxEpisodeSteps) { - Preconditions.checkArgument(maxEpisodeSteps > 0, "maxEpisodeSteps must be greater than 0, got", maxEpisodeSteps); - this.maxEpisodeSteps = maxEpisodeSteps; - - return this; - } - - public Builder id(String id) { - this.id = id; - return this; - } - - public AGENT_TYPE build() { - return (AGENT_TYPE)new Agent(environment, transformProcess, policy, maxEpisodeSteps, id); - } + @SuperBuilder + @Data + public static class Configuration { + /** + * The maximum number of steps an episode can have before being interrupted. Use null to have no max. + */ + @lombok.Builder.Default + Integer maxEpisodeSteps = null; // Default, no max } } \ No newline at end of file diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/AgentLearner.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/AgentLearner.java index 8fd963cda..80da7ff05 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/AgentLearner.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/AgentLearner.java @@ -15,9 +15,11 @@ ******************************************************************************/ package org.deeplearning4j.rl4j.agent; +import lombok.Data; import lombok.Getter; import lombok.NonNull; -import org.deeplearning4j.rl4j.agent.learning.ILearningBehavior; +import lombok.experimental.SuperBuilder; +import org.deeplearning4j.rl4j.agent.learning.behavior.ILearningBehavior; import org.deeplearning4j.rl4j.environment.Environment; import org.deeplearning4j.rl4j.environment.StepResult; import org.deeplearning4j.rl4j.observation.Observation; @@ -41,12 +43,17 @@ public class AgentLearner extends Agent implements IAgentLearner * @param environment The {@link Environment} to be used * @param transformProcess The {@link TransformProcess} to be used to transform the raw observations into usable ones. * @param policy The {@link IPolicy} to be used - * @param maxEpisodeSteps The maximum number of steps an episode can have before being interrupted. Use null to have no max. + * @param configuration The configuration for the AgentLearner * @param id A user-supplied id to identify the instance. * @param learningBehavior The {@link ILearningBehavior} that will be used to supervise the learning. */ - public AgentLearner(Environment environment, TransformProcess transformProcess, IPolicy policy, Integer maxEpisodeSteps, String id, @NonNull ILearningBehavior learningBehavior) { - super(environment, transformProcess, policy, maxEpisodeSteps, id); + public AgentLearner(Environment environment, + TransformProcess transformProcess, + IPolicy policy, + Configuration configuration, + String id, + @NonNull ILearningBehavior learningBehavior) { + super(environment, transformProcess, policy, configuration, id); this.learningBehavior = learningBehavior; } @@ -86,30 +93,8 @@ public class AgentLearner extends Agent implements IAgentLearner ++totalStepCount; } - // FIXME: parent is still visible - public static AgentLearner.Builder> builder(Environment environment, - TransformProcess transformProcess, - IPolicy policy, - ILearningBehavior learningBehavior) { - return new AgentLearner.Builder>(environment, transformProcess, policy, learningBehavior); - } - - public static class Builder> extends Agent.Builder { - - private final ILearningBehavior learningBehavior; - - public Builder(@NonNull Environment environment, - @NonNull TransformProcess transformProcess, - @NonNull IPolicy policy, - @NonNull ILearningBehavior learningBehavior) { - super(environment, transformProcess, policy); - - this.learningBehavior = learningBehavior; - } - - @Override - public AGENT_TYPE build() { - return (AGENT_TYPE)new AgentLearner(environment, transformProcess, policy, maxEpisodeSteps, id, learningBehavior); - } + @SuperBuilder + @Data + public static class Configuration extends Agent.Configuration { } } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/ITDTargetAlgorithm.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/IUpdateAlgorithm.java similarity index 56% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/ITDTargetAlgorithm.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/IUpdateAlgorithm.java index cd7588bfe..5f07369f9 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/ITDTargetAlgorithm.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/IUpdateAlgorithm.java @@ -14,25 +14,22 @@ * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm; - -import org.deeplearning4j.rl4j.learning.sync.Transition; -import org.nd4j.linalg.dataset.api.DataSet; +package org.deeplearning4j.rl4j.agent.learning.algorithm; import java.util.List; /** - * The interface of all TD target calculation algorithms. + * The interface of all features-labels based update algorithms. * - * @param The type of actions - * - * @author Alexandre Boulanger + * @param The type of experiences + * @param The type of the result. See {@link org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels FeaturesLabels} + * and {@link org.deeplearning4j.rl4j.agent.learning.update.Gradients Gradients} */ -public interface ITDTargetAlgorithm { +public interface IUpdateAlgorithm { /** - * Compute the updated estimated Q-Values for every transition - * @param transitions The transitions from the experience replay + * Compute the labels required to update the network from the training batch + * @param trainingBatch The transitions from the experience replay * @return A DataSet where every element is the observation and the estimated Q-Values for all actions */ - DataSet compute(List> transitions); + RESULT_TYPE compute(List trainingBatch); } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearning.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearning.java new file mode 100644 index 000000000..b199ecde0 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearning.java @@ -0,0 +1,104 @@ +/******************************************************************************* + * Copyright (c) 2015-2019 Skymind, Inc. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.agent.learning.algorithm; + +import lombok.Builder; +import lombok.Data; +import lombok.NonNull; +import lombok.experimental.SuperBuilder; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; +import org.deeplearning4j.rl4j.experience.StateActionPair; +import org.deeplearning4j.rl4j.helper.INDArrayHelper; +import org.deeplearning4j.rl4j.network.CommonLabelNames; +import org.deeplearning4j.rl4j.network.IOutputNeuralNet; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.nd4j.linalg.api.ndarray.INDArray; +import org.nd4j.linalg.factory.Nd4j; + +import java.util.List; + +/** + * This the "Algorithm S2 Asynchronous n-step Q-learning" of Asynchronous Methods for Deep Reinforcement Learning + * @see https://arxiv.org/pdf/1602.01783.pdf, page 13 + */ +public class NStepQLearning implements IUpdateAlgorithm> { + + private final ITrainableNeuralNet current; + private final IOutputNeuralNet target; + private final int actionSpaceSize; + private final double gamma; + + /** + * @param current The θ' parameters (the thread-specific network) + * @param target The θ– parameters (the global target network) + * @param actionSpaceSize The numbers of possible actions that can be taken on the environment + */ + public NStepQLearning(@NonNull ITrainableNeuralNet current, + @NonNull IOutputNeuralNet target, + int actionSpaceSize, + @NonNull Configuration configuration) { + this.current = current; + this.target = target; + this.actionSpaceSize = actionSpaceSize; + this.gamma = configuration.getGamma(); + } + + @Override + public Gradients compute(List> trainingBatch) { + int size = trainingBatch.size(); + + StateActionPair stateActionPair = trainingBatch.get(size - 1); + + INDArray data = stateActionPair.getObservation().getData(); + INDArray features = INDArrayHelper.createBatchForShape(size, data.shape()); + INDArray labels = Nd4j.create(size, actionSpaceSize); + + double r; + if (stateActionPair.isTerminal()) { + r = 0; + } else { + INDArray output = target.output(data); + r = Nd4j.max(output).getDouble(0); + } + + for (int i = size - 1; i >= 0; --i) { + stateActionPair = trainingBatch.get(i); + data = stateActionPair.getObservation().getData(); + + features.putRow(i, data); + + r = stateActionPair.getReward() + gamma * r; + INDArray row = current.output(data); + row = row.putScalar(stateActionPair.getAction(), r); + labels.putRow(i, row); + } + + FeaturesLabels featuresLabels = new FeaturesLabels(features); + featuresLabels.putLabels(CommonLabelNames.QValues, labels); + return current.computeGradients(featuresLabels); + } + + @SuperBuilder + @Data + public static class Configuration { + /** + * The discount factor (default is 0.99) + */ + @Builder.Default + double gamma = 0.99; + } +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/BaseDQNAlgorithm.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/BaseDQNAlgorithm.java similarity index 70% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/BaseDQNAlgorithm.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/BaseDQNAlgorithm.java index 6cae384d5..a56880239 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/BaseDQNAlgorithm.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/BaseDQNAlgorithm.java @@ -14,8 +14,9 @@ * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm; +package org.deeplearning4j.rl4j.agent.learning.algorithm.dqn; +import lombok.NonNull; import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.nd4j.linalg.api.ndarray.INDArray; @@ -25,27 +26,24 @@ import org.nd4j.linalg.api.ndarray.INDArray; * @author Alexandre Boulanger * */ -public abstract class BaseDQNAlgorithm extends BaseTDTargetAlgorithm { +public abstract class BaseDQNAlgorithm extends BaseTransitionTDAlgorithm { private final IOutputNeuralNet targetQNetwork; /** - * In litterature, this corresponds to Q{net}(s(t+1), a) + * In literature, this corresponds to Qnet(s(t+1), a) */ protected INDArray qNetworkNextObservation; /** - * In litterature, this corresponds to Q{tnet}(s(t+1), a) + * In literature, this corresponds to Qtnet(s(t+1), a) */ protected INDArray targetQNetworkNextObservation; - protected BaseDQNAlgorithm(IOutputNeuralNet qNetwork, IOutputNeuralNet targetQNetwork, double gamma) { - super(qNetwork, gamma); - this.targetQNetwork = targetQNetwork; - } - - protected BaseDQNAlgorithm(IOutputNeuralNet qNetwork, IOutputNeuralNet targetQNetwork, double gamma, double errorClamp) { - super(qNetwork, gamma, errorClamp); + protected BaseDQNAlgorithm(IOutputNeuralNet qNetwork, + @NonNull IOutputNeuralNet targetQNetwork, + BaseTransitionTDAlgorithm.Configuration configuration) { + super(qNetwork, configuration); this.targetQNetwork = targetQNetwork; } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/BaseTDTargetAlgorithm.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/BaseTransitionTDAlgorithm.java similarity index 66% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/BaseTDTargetAlgorithm.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/BaseTransitionTDAlgorithm.java index 460567a86..7dc3c9475 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/BaseTDTargetAlgorithm.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/BaseTransitionTDAlgorithm.java @@ -14,21 +14,25 @@ * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm; +package org.deeplearning4j.rl4j.agent.learning.algorithm.dqn; +import lombok.Builder; +import lombok.Data; +import lombok.NonNull; +import lombok.experimental.SuperBuilder; +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; import org.deeplearning4j.rl4j.learning.sync.Transition; +import org.deeplearning4j.rl4j.network.CommonLabelNames; import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import java.util.List; /** - * The base of all TD target calculation algorithms that use deep learning. - * - * @author Alexandre Boulanger + * The base of all {@link Transition Transition-based} TD algorithms. */ -public abstract class BaseTDTargetAlgorithm implements ITDTargetAlgorithm { +public abstract class BaseTransitionTDAlgorithm implements IUpdateAlgorithm> { protected final IOutputNeuralNet qNetwork; protected final double gamma; @@ -39,27 +43,16 @@ public abstract class BaseTDTargetAlgorithm implements ITDTargetAlgorithmerrorClamp away from the previous value. Double.NaN will disable the clamping. + * @param configuration The {@link Configuration} to use */ - protected BaseTDTargetAlgorithm(IOutputNeuralNet qNetwork, double gamma, double errorClamp) { + protected BaseTransitionTDAlgorithm(@NonNull IOutputNeuralNet qNetwork, @NonNull Configuration configuration) { this.qNetwork = qNetwork; - this.gamma = gamma; + this.gamma = configuration.getGamma(); - this.errorClamp = errorClamp; + this.errorClamp = configuration.getErrorClamp(); isClamped = !Double.isNaN(errorClamp); } - /** - * - * @param qNetwork The Q-Network - * @param gamma The discount factor - * Note: Error clamping is disabled with this ctor - */ - protected BaseTDTargetAlgorithm(IOutputNeuralNet qNetwork, double gamma) { - this(qNetwork, gamma, Double.NaN); - } - /** * Called just before the calculation starts * @param observations A INDArray of all observations stacked on dimension 0 @@ -80,7 +73,7 @@ public abstract class BaseTDTargetAlgorithm implements ITDTargetAlgorithm> transitions) { + public FeaturesLabels compute(List> transitions) { int size = transitions.size(); @@ -103,6 +96,25 @@ public abstract class BaseTDTargetAlgorithm implements ITDTargetAlgorithmerrorClamp away from the previous value. Double.NaN will disable the clamping (default). + */ + @Builder.Default + double errorClamp = Double.NaN; } } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/DoubleDQN.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/DoubleDQN.java similarity index 77% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/DoubleDQN.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/DoubleDQN.java index caeb85fb6..f7d99276b 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/DoubleDQN.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/DoubleDQN.java @@ -14,7 +14,7 @@ * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm; +package org.deeplearning4j.rl4j.agent.learning.algorithm.dqn; import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.nd4j.linalg.api.ndarray.INDArray; @@ -29,15 +29,13 @@ public class DoubleDQN extends BaseDQNAlgorithm { private static final int ACTION_DIMENSION_IDX = 1; - // In litterature, this corresponds to: max_{a}Q(s_{t+1}, a) + // In literature, this corresponds to: maxa Q(st+1, a) private INDArray maxActionsFromQNetworkNextObservation; - public DoubleDQN(IOutputNeuralNet qNetwork, IOutputNeuralNet targetQNetwork, double gamma) { - super(qNetwork, targetQNetwork, gamma); - } - - public DoubleDQN(IOutputNeuralNet qNetwork, IOutputNeuralNet targetQNetwork, double gamma, double errorClamp) { - super(qNetwork, targetQNetwork, gamma, errorClamp); + public DoubleDQN(IOutputNeuralNet qNetwork, + IOutputNeuralNet targetQNetwork, + BaseTransitionTDAlgorithm.Configuration configuration) { + super(qNetwork, targetQNetwork, configuration); } @Override @@ -48,8 +46,8 @@ public class DoubleDQN extends BaseDQNAlgorithm { } /** - * In litterature, this corresponds to:
- * Q(s_t, a_t) = R_{t+1} + \gamma * Q_{tar}(s_{t+1}, max_{a}Q(s_{t+1}, a)) + * In literature, this corresponds to:
+ * Q(st, at) = Rt+1 + γ * Qtar(st+1, maxa Q(st+1, a)) * @param batchIdx The index in the batch of the current transition * @param reward The reward of the current transition * @param isTerminal True if it's the last transition of the "game" diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/StandardDQN.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/StandardDQN.java similarity index 79% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/StandardDQN.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/StandardDQN.java index 6cd047c74..c3a4c3be2 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/StandardDQN.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/StandardDQN.java @@ -14,7 +14,7 @@ * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm; +package org.deeplearning4j.rl4j.agent.learning.algorithm.dqn; import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.nd4j.linalg.api.ndarray.INDArray; @@ -29,16 +29,15 @@ public class StandardDQN extends BaseDQNAlgorithm { private static final int ACTION_DIMENSION_IDX = 1; - // In litterature, this corresponds to: max_{a}Q_{tar}(s_{t+1}, a) + /** + * In literature, this corresponds to: maxa Qtar(st+1, a) + */ private INDArray maxActionsFromQTargetNextObservation; - public StandardDQN(IOutputNeuralNet qNetwork, IOutputNeuralNet targetQNetwork, double gamma) { - super(qNetwork, targetQNetwork, gamma); + public StandardDQN(IOutputNeuralNet qNetwork, IOutputNeuralNet targetQNetwork, Configuration configuration) { + super(qNetwork, targetQNetwork, configuration); } - public StandardDQN(IOutputNeuralNet qNetwork, IOutputNeuralNet targetQNetwork, double gamma, double errorClamp) { - super(qNetwork, targetQNetwork, gamma, errorClamp); - } @Override protected void initComputation(INDArray observations, INDArray nextObservations) { @@ -48,8 +47,8 @@ public class StandardDQN extends BaseDQNAlgorithm { } /** - * In litterature, this corresponds to:
- * Q(s_t, a_t) = R_{t+1} + \gamma * max_{a}Q_{tar}(s_{t+1}, a) + * In literature, this corresponds to:
+ * Q(st, at) = Rt+1 + γ * maxa Qtar(st+1, a) * @param batchIdx The index in the batch of the current transition * @param reward The reward of the current transition * @param isTerminal True if it's the last transition of the "game" diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/ILearningBehavior.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/behavior/ILearningBehavior.java similarity index 94% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/ILearningBehavior.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/behavior/ILearningBehavior.java index 0187d8c3a..286655378 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/ILearningBehavior.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/behavior/ILearningBehavior.java @@ -13,7 +13,7 @@ * * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.agent.learning; +package org.deeplearning4j.rl4j.agent.learning.behavior; import org.deeplearning4j.rl4j.observation.Observation; diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/LearningBehavior.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/behavior/LearningBehavior.java similarity index 90% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/LearningBehavior.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/behavior/LearningBehavior.java index 85c7ec4ce..66709482f 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/LearningBehavior.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/behavior/LearningBehavior.java @@ -13,10 +13,11 @@ * * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.agent.learning; +package org.deeplearning4j.rl4j.agent.learning.behavior; import lombok.Builder; -import org.deeplearning4j.rl4j.agent.update.IUpdateRule; +import lombok.NonNull; +import org.deeplearning4j.rl4j.agent.learning.update.IUpdateRule; import org.deeplearning4j.rl4j.experience.ExperienceHandler; import org.deeplearning4j.rl4j.observation.Observation; @@ -30,10 +31,10 @@ import org.deeplearning4j.rl4j.observation.Observation; @Builder public class LearningBehavior implements ILearningBehavior { - @Builder.Default - private int experienceUpdateSize = 64; - + @NonNull private final ExperienceHandler experienceHandler; + + @NonNull private final IUpdateRule updateRule; @Override diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabels.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabels.java new file mode 100644 index 000000000..1d6e1249b --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabels.java @@ -0,0 +1,64 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.agent.learning.update; + +import lombok.Getter; +import org.nd4j.linalg.api.ndarray.INDArray; + +import java.util.HashMap; + +/** + * A container that holds the features and the associated labels. + */ +public class FeaturesLabels { + + @Getter + private final INDArray features; + + private final HashMap labels = new HashMap(); + + /** + * @param features + */ + public FeaturesLabels(INDArray features) { + this.features = features; + } + + /** + * @return The number of examples in features and each labels. + */ + public long getBatchSize() { + return features.shape()[0]; + } + + /** + * Add labels by name + * @param name + * @param labels + */ + public void putLabels(String name, INDArray labels) { + this.labels.put(name, labels); + } + + /** + * Get the labels associated to the name. + * @param name + * @return + */ + public INDArray getLabels(String name) { + return this.labels.get(name); + } +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/Gradients.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/Gradients.java new file mode 100644 index 000000000..e97de2042 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/Gradients.java @@ -0,0 +1,58 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.agent.learning.update; + +import lombok.Getter; +import org.deeplearning4j.nn.gradient.Gradient; + +import java.util.HashMap; + +/** + * A {@link Gradient} container used to update neural networks. + */ +public class Gradients { + + @Getter + private final long batchSize; + + private final HashMap gradients = new HashMap(); + + /** + * @param batchSize The size of the training batch used to create this instance + */ + public Gradients(long batchSize) { + this.batchSize = batchSize; + } + + /** + * Add a {@link Gradient} by name. + * @param name + * @param gradient + */ + public void putGradient(String name, Gradient gradient) { + gradients.put(name, gradient); + } + + /** + * Get a {@link Gradient} by name + * @param name + * @return + */ + public Gradient getGradient(String name) { + return gradients.get(name); + } + +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/IUpdateRule.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/IUpdateRule.java similarity index 93% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/IUpdateRule.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/IUpdateRule.java index d679cba24..99ae979b2 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/IUpdateRule.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/IUpdateRule.java @@ -13,7 +13,7 @@ * * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.agent.update; +package org.deeplearning4j.rl4j.agent.learning.update; import java.util.List; diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRule.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRule.java new file mode 100644 index 000000000..5d909cfbf --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRule.java @@ -0,0 +1,54 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.agent.learning.update; + +import lombok.Getter; +import lombok.NonNull; +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.update.updater.INeuralNetUpdater; + +import java.util.List; + +/** + * This implementation of {@link IUpdateRule} delegates the features-labels or gradients computations to + * a {@link IUpdateAlgorithm}, and the networks update to a {@link INeuralNetUpdater}. + * + * @param The type of result returned by the IUpdateAlgorithm + * @param The type of experience + */ +public class UpdateRule implements IUpdateRule { + + private final INeuralNetUpdater updater; + + private final IUpdateAlgorithm updateAlgorithm; + + @Getter + private int updateCount = 0; + + public UpdateRule(@NonNull IUpdateAlgorithm updateAlgorithm, + @NonNull INeuralNetUpdater updater) { + this.updateAlgorithm = updateAlgorithm; + this.updater = updater; + } + + @Override + public void update(List trainingBatch) { + RESULT_TYPE featuresLabels = updateAlgorithm.compute(trainingBatch); + updater.update(featuresLabels); + ++updateCount; + } + +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/NeuralNetUpdater.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/GradientsNeuralNetUpdater.java similarity index 55% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/NeuralNetUpdater.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/GradientsNeuralNetUpdater.java index 3dc778251..c657e3fa2 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/NeuralNetUpdater.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/GradientsNeuralNetUpdater.java @@ -13,15 +13,19 @@ * * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.agent.update.neuralnetupdater; +package org.deeplearning4j.rl4j.agent.learning.update.updater; +import lombok.Builder; +import lombok.Data; +import lombok.NonNull; +import lombok.experimental.SuperBuilder; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; -import org.nd4j.linalg.dataset.api.DataSet; /** * A {@link INeuralNetUpdater} that updates a neural network and sync a target network at defined intervals */ -public class NeuralNetUpdater implements INeuralNetUpdater { +public class GradientsNeuralNetUpdater implements INeuralNetUpdater { private final ITrainableNeuralNet current; private final ITrainableNeuralNet target; @@ -29,27 +33,29 @@ public class NeuralNetUpdater implements INeuralNetUpdater { private int updateCount = 0; private final int targetUpdateFrequency; + // TODO: Add async support /** * @param current The current {@link ITrainableNeuralNet network} * @param target The target {@link ITrainableNeuralNet network} - * @param targetUpdateFrequency Will synchronize the target network at every targetUpdateFrequency updates + * + * Note: Presently async is not supported */ - public NeuralNetUpdater(ITrainableNeuralNet current, - ITrainableNeuralNet target, - int targetUpdateFrequency) { + public GradientsNeuralNetUpdater(@NonNull ITrainableNeuralNet current, + @NonNull ITrainableNeuralNet target, + @NonNull Configuration configuration) { this.current = current; this.target = target; - this.targetUpdateFrequency = targetUpdateFrequency; + this.targetUpdateFrequency = configuration.getTargetUpdateFrequency(); } /** * Update the current network - * @param featuresLabels A Dataset that will be used to update the network. + * @param gradients A {@link Gradients} that will be used to update the network. */ @Override - public void update(DataSet featuresLabels) { - current.fit(featuresLabels); + public void update(Gradients gradients) { + current.applyGradients(gradients); syncTargetNetwork(); } @@ -59,4 +65,13 @@ public class NeuralNetUpdater implements INeuralNetUpdater { } } + @SuperBuilder + @Data + public static class Configuration { + /** + * Will synchronize the target network at every targetUpdateFrequency updates (default: no update) + */ + @Builder.Default + int targetUpdateFrequency = Integer.MAX_VALUE; + } } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/INeuralNetUpdater.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/INeuralNetUpdater.java similarity index 63% rename from rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/INeuralNetUpdater.java rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/INeuralNetUpdater.java index f17c4f11e..6d9fae1f8 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/INeuralNetUpdater.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/INeuralNetUpdater.java @@ -13,18 +13,17 @@ * * SPDX-License-Identifier: Apache-2.0 ******************************************************************************/ -package org.deeplearning4j.rl4j.agent.update.neuralnetupdater; - -import org.nd4j.linalg.dataset.api.DataSet; +package org.deeplearning4j.rl4j.agent.learning.update.updater; /** - * The role of INeuralNetUpdater implementations is to update a {@link org.deeplearning4j.rl4j.network.NeuralNet NeuralNet} - * from a {@link DataSet}.

+ * The role of INeuralNetUpdater implementations is to update a {@link org.deeplearning4j.rl4j.network.NeuralNet NeuralNet}.

+ * @param The type of the data needed to to update the netwok. See {@link org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels FeaturesLabels} + * and {@link org.deeplearning4j.rl4j.agent.learning.update.Gradients Gradients}. */ -public interface INeuralNetUpdater { +public interface INeuralNetUpdater { /** * Update a {@link org.deeplearning4j.rl4j.network.NeuralNet NeuralNet}. - * @param featuresLabels A Dataset that will be used to update the network. + * @param dataType */ - void update(DataSet featuresLabels); + void update(DATA_TYPE dataType); } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdater.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdater.java new file mode 100644 index 000000000..33d30f652 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdater.java @@ -0,0 +1,81 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.agent.learning.update.updater; + +import lombok.Builder; +import lombok.Data; +import lombok.NonNull; +import lombok.experimental.SuperBuilder; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.nd4j.common.base.Preconditions; + +/** + * A {@link INeuralNetUpdater} that updates a neural network and sync a target network at defined intervals + */ +public class LabelsNeuralNetUpdater implements INeuralNetUpdater { + + private final ITrainableNeuralNet current; + private final ITrainableNeuralNet target; + + private int updateCount = 0; + private final int targetUpdateFrequency; + + // TODO: Add async support + /** + * @param current The current {@link ITrainableNeuralNet network} + * @param target The target {@link ITrainableNeuralNet network} + * @param configuration The {@link Configuration} to use + * + * Note: Presently async is not supported + */ + public LabelsNeuralNetUpdater(@NonNull ITrainableNeuralNet current, + @NonNull ITrainableNeuralNet target, + @NonNull Configuration configuration) { + Preconditions.checkArgument(configuration.getTargetUpdateFrequency() > 0, "Configuration: targetUpdateFrequency must be greater than 0, got: ", configuration.getTargetUpdateFrequency()); + this.current = current; + this.target = target; + + this.targetUpdateFrequency = configuration.getTargetUpdateFrequency(); + } + + /** + * Update the current network + * @param featuresLabels A {@link FeaturesLabels} that will be used to update the network. + */ + @Override + public void update(FeaturesLabels featuresLabels) { + current.fit(featuresLabels); + syncTargetNetwork(); + } + + private void syncTargetNetwork() { + if(++updateCount % targetUpdateFrequency == 0) { + target.copy(current); + } + } + + @SuperBuilder + @Data + public static class Configuration { + /** + * Will synchronize the target network at every targetUpdateFrequency updates (default: no update) + */ + @Builder.Default + int targetUpdateFrequency = Integer.MAX_VALUE; + } + +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListener.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListener.java index f176da144..91b83c59f 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListener.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListener.java @@ -63,4 +63,12 @@ public interface AgentListener { * @return A {@link ListenerResponse}. */ AgentListener.ListenerResponse onAfterStep(Agent agent, StepResult stepResult); + + /** + * Called after the episode has ended. + * + * @param agent The agent that generated the event + * + */ + void onAfterEpisode(Agent agent); } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListenerList.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListenerList.java index 48538aeaf..e697c4c53 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListenerList.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/listener/AgentListenerList.java @@ -86,4 +86,16 @@ public class AgentListenerList { return true; } + + /** + * This method will notify all listeners that an episode has finished. + * + * @param agent The agent that generated the event. + */ + public void notifyAfterEpisode(Agent agent) { + for (AgentListener listener : listeners) { + listener.onAfterEpisode(agent); + } + } + } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/DQNNeuralNetUpdateRule.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/DQNNeuralNetUpdateRule.java deleted file mode 100644 index c359e02ce..000000000 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/DQNNeuralNetUpdateRule.java +++ /dev/null @@ -1,56 +0,0 @@ -/******************************************************************************* - * Copyright (c) 2020 Konduit K.K. - * - * This program and the accompanying materials are made available under the - * terms of the Apache License, Version 2.0 which is available at - * https://www.apache.org/licenses/LICENSE-2.0. - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT - * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the - * License for the specific language governing permissions and limitations - * under the License. - * - * SPDX-License-Identifier: Apache-2.0 - ******************************************************************************/ -package org.deeplearning4j.rl4j.agent.update; - -import lombok.Getter; -import org.deeplearning4j.rl4j.agent.update.neuralnetupdater.INeuralNetUpdater; -import org.deeplearning4j.rl4j.agent.update.neuralnetupdater.NeuralNetUpdater; -import org.deeplearning4j.rl4j.learning.sync.Transition; -import org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm.DoubleDQN; -import org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm.ITDTargetAlgorithm; -import org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm.StandardDQN; -import org.deeplearning4j.rl4j.network.dqn.IDQN; -import org.nd4j.linalg.dataset.api.DataSet; - -import java.util.List; - -// Temporary class that will be replaced with a more generic class that delegates gradient computation -// and network update to sub components. -public class DQNNeuralNetUpdateRule implements IUpdateRule> { - - private final IDQN targetQNetwork; - private final INeuralNetUpdater updater; - - private final ITDTargetAlgorithm tdTargetAlgorithm; - - @Getter - private int updateCount = 0; - - public DQNNeuralNetUpdateRule(IDQN qNetwork, int targetUpdateFrequency, boolean isDoubleDQN, double gamma, double errorClamp) { - this.targetQNetwork = qNetwork.clone(); - tdTargetAlgorithm = isDoubleDQN - ? new DoubleDQN(qNetwork, targetQNetwork, gamma, errorClamp) - : new StandardDQN(qNetwork, targetQNetwork, gamma, errorClamp); - updater = new NeuralNetUpdater(qNetwork, targetQNetwork, targetUpdateFrequency); - - } - - @Override - public void update(List> trainingBatch) { - DataSet targets = tdTargetAlgorithm.compute(trainingBatch); - updater.update(targets); - } -} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/Gradients.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/Gradients.java deleted file mode 100644 index 4307efe1e..000000000 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/agent/update/Gradients.java +++ /dev/null @@ -1,26 +0,0 @@ -/******************************************************************************* - * Copyright (c) 2020 Konduit K.K. - * - * This program and the accompanying materials are made available under the - * terms of the Apache License, Version 2.0 which is available at - * https://www.apache.org/licenses/LICENSE-2.0. - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT - * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the - * License for the specific language governing permissions and limitations - * under the License. - * - * SPDX-License-Identifier: Apache-2.0 - ******************************************************************************/ -package org.deeplearning4j.rl4j.agent.update; - -import lombok.Value; -import org.deeplearning4j.nn.gradient.Gradient; - -// Work in progress -@Value -public class Gradients { - private Gradient[] gradients; // Temporary: we'll need something better than a Gradient[] - private int batchSize; -} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilder.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilder.java new file mode 100644 index 000000000..f41f9361c --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilder.java @@ -0,0 +1,168 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.builder; + +import lombok.AccessLevel; +import lombok.Data; +import lombok.Getter; +import lombok.NonNull; +import lombok.experimental.SuperBuilder; +import org.apache.commons.lang3.NotImplementedException; +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.AgentLearner; +import org.deeplearning4j.rl4j.agent.IAgentLearner; +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.behavior.ILearningBehavior; +import org.deeplearning4j.rl4j.agent.learning.behavior.LearningBehavior; +import org.deeplearning4j.rl4j.agent.learning.update.IUpdateRule; +import org.deeplearning4j.rl4j.agent.learning.update.UpdateRule; +import org.deeplearning4j.rl4j.agent.learning.update.updater.INeuralNetUpdater; +import org.deeplearning4j.rl4j.agent.listener.AgentListener; +import org.deeplearning4j.rl4j.environment.Environment; +import org.deeplearning4j.rl4j.experience.ExperienceHandler; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.deeplearning4j.rl4j.observation.transform.TransformProcess; +import org.deeplearning4j.rl4j.policy.IPolicy; + +import java.util.List; + +/** + * A base {@link IAgentLearner} builder that should be helpful in several common scenarios.

+ * Note: Classes implementing BaseAgentLearnerBuilder should be careful not to re-use a stateful and/or non thread-safe dependency + * through several calls to build(). In doubt, use a new instance. + * @param The type of action + * @param The type of experiences + * @param The response type of {@link org.deeplearning4j.rl4j.network.IOutputNeuralNet IOutputNeuralNet}.output() + */ +public abstract class BaseAgentLearnerBuilder implements Builder> { + + private final Configuration configuration; + private final Builder> environmentBuilder; + private final Builder transformProcessBuilder; + protected final INetworksHandler networks; + + protected int createdAgentLearnerCount; + + public BaseAgentLearnerBuilder(@NonNull Configuration configuration, + @NonNull ITrainableNeuralNet neuralNet, + @NonNull Builder> environmentBuilder, + @NonNull Builder transformProcessBuilder) { + this.configuration = configuration; + this.environmentBuilder = environmentBuilder; + this.transformProcessBuilder = transformProcessBuilder; + + // TODO: Support async setups + if(configuration.isAsynchronous()) { + throw new NotImplementedException("Asynchronous BaseAgentLearnerBuilder is not yet implemented"); + } + this.networks = new SyncNetworkHandler(neuralNet); + } + + @Getter(AccessLevel.PROTECTED) + private Environment environment; + + @Getter(AccessLevel.PROTECTED) + private TransformProcess transformProcess; + + @Getter(AccessLevel.PROTECTED) + private IPolicy policy; + + @Getter(AccessLevel.PROTECTED) + private ExperienceHandler experienceHandler; + + @Getter(AccessLevel.PROTECTED) + private IUpdateAlgorithm updateAlgorithm; + + @Getter(AccessLevel.PROTECTED) + private INeuralNetUpdater neuralNetUpdater; + + @Getter(AccessLevel.PROTECTED) + private IUpdateRule updateRule; + + @Getter(AccessLevel.PROTECTED) + private ILearningBehavior learningBehavior; + + protected abstract IPolicy buildPolicy(); + protected abstract ExperienceHandler buildExperienceHandler(); + protected abstract IUpdateAlgorithm buildUpdateAlgorithm(); + protected abstract INeuralNetUpdater buildNeuralNetUpdater(); + protected IUpdateRule buildUpdateRule() { + return new UpdateRule(getUpdateAlgorithm(), getNeuralNetUpdater()); + } + protected ILearningBehavior buildLearningBehavior() { + return LearningBehavior.builder() + .experienceHandler(getExperienceHandler()) + .updateRule(getUpdateRule()) + .build(); + } + + protected void resetForNewBuild() { + environment = environmentBuilder.build(); + transformProcess = transformProcessBuilder.build(); + policy = buildPolicy(); + experienceHandler = buildExperienceHandler(); + updateAlgorithm = buildUpdateAlgorithm(); + neuralNetUpdater = buildNeuralNetUpdater(); + updateRule = buildUpdateRule(); + learningBehavior = buildLearningBehavior(); + + ++createdAgentLearnerCount; + } + + protected String getThreadId() { + return "AgentLearner-" + createdAgentLearnerCount; + } + + protected IAgentLearner buildAgentLearner() { + AgentLearner result = new AgentLearner(getEnvironment(), getTransformProcess(), getPolicy(), configuration.getAgentLearnerConfiguration(), getThreadId(), getLearningBehavior()); + if(configuration.getAgentLearnerListeners() != null) { + for (AgentListener listener : configuration.getAgentLearnerListeners()) { + result.addListener(listener); + } + } + + return result; + } + + /** + * Build a properly assembled / configured IAgentLearner. + * @return a {@link IAgentLearner} + */ + @Override + public IAgentLearner build() { + resetForNewBuild(); + return buildAgentLearner(); + } + + @SuperBuilder + @Data + public static class Configuration { + /** + * The configuration that will be used to build the {@link AgentLearner} + */ + AgentLearner.Configuration agentLearnerConfiguration; + + /** + * A list of {@link AgentListener AgentListeners} that will be added to the AgentLearner. (default = null; no listeners) + */ + List> agentLearnerListeners; + + /** + * Tell the builder that the AgentLearners will be used in an asynchronous setup + */ + boolean asynchronous; + } +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseDQNAgentLearnerBuilder.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseDQNAgentLearnerBuilder.java new file mode 100644 index 000000000..5306a5319 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/BaseDQNAgentLearnerBuilder.java @@ -0,0 +1,93 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.builder; + +import lombok.AccessLevel; +import lombok.Data; +import lombok.EqualsAndHashCode; +import lombok.Getter; +import lombok.experimental.SuperBuilder; +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.IAgentLearner; +import org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.BaseTransitionTDAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.updater.INeuralNetUpdater; +import org.deeplearning4j.rl4j.agent.learning.update.updater.LabelsNeuralNetUpdater; +import org.deeplearning4j.rl4j.environment.Environment; +import org.deeplearning4j.rl4j.environment.IActionSchema; +import org.deeplearning4j.rl4j.experience.ExperienceHandler; +import org.deeplearning4j.rl4j.experience.ReplayMemoryExperienceHandler; +import org.deeplearning4j.rl4j.learning.sync.Transition; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.deeplearning4j.rl4j.observation.transform.TransformProcess; +import org.deeplearning4j.rl4j.policy.DQNPolicy; +import org.deeplearning4j.rl4j.policy.EpsGreedy; +import org.deeplearning4j.rl4j.policy.INeuralNetPolicy; +import org.deeplearning4j.rl4j.policy.IPolicy; +import org.nd4j.linalg.api.rng.Random; + +/** + * A base {@link IAgentLearner} builder that will setup these: + *

  • a epsilon-greedy policy
    • + *
  • a replay-memory experience handler
    • + *
  • a neural net updater that expects feature-labels update data
    • + * + * Used as the base of DQN builders. + */ +public abstract class BaseDQNAgentLearnerBuilder extends BaseAgentLearnerBuilder, FeaturesLabels> { + + @Getter(AccessLevel.PROTECTED) + private final Configuration configuration; + + private final Random rnd; + + public BaseDQNAgentLearnerBuilder(Configuration configuration, + ITrainableNeuralNet neuralNet, + Builder> environmentBuilder, + Builder transformProcessBuilder, + Random rnd) { + super(configuration, neuralNet, environmentBuilder, transformProcessBuilder); + this.configuration = configuration; + this.rnd = rnd; + } + + @Override + protected IPolicy buildPolicy() { + INeuralNetPolicy greedyPolicy = new DQNPolicy(networks.getThreadCurrentNetwork()); + IActionSchema actionSchema = getEnvironment().getSchema().getActionSchema(); + return new EpsGreedy(greedyPolicy, actionSchema, configuration.getPolicyConfiguration(), rnd); + } + + @Override + protected ExperienceHandler> buildExperienceHandler() { + return new ReplayMemoryExperienceHandler(configuration.getExperienceHandlerConfiguration(), rnd); + } + + @Override + protected INeuralNetUpdater buildNeuralNetUpdater() { + return new LabelsNeuralNetUpdater(networks.getThreadCurrentNetwork(), networks.getTargetNetwork(), configuration.getNeuralNetUpdaterConfiguration()); + } + + @EqualsAndHashCode(callSuper = true) + @SuperBuilder + @Data + public static class Configuration extends BaseAgentLearnerBuilder.Configuration { + EpsGreedy.Configuration policyConfiguration; + ReplayMemoryExperienceHandler.Configuration experienceHandlerConfiguration; + LabelsNeuralNetUpdater.Configuration neuralNetUpdaterConfiguration; + BaseTransitionTDAlgorithm.Configuration updateAlgorithmConfiguration; + } +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/DoubleDQNBuilder.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/DoubleDQNBuilder.java new file mode 100644 index 000000000..752c53ca7 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/DoubleDQNBuilder.java @@ -0,0 +1,56 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.builder; + +import lombok.Data; +import lombok.EqualsAndHashCode; +import lombok.experimental.SuperBuilder; +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.IAgentLearner; +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.DoubleDQN; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.environment.Environment; +import org.deeplearning4j.rl4j.learning.sync.Transition; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.deeplearning4j.rl4j.observation.transform.TransformProcess; +import org.nd4j.linalg.api.rng.Random; + +/** + * A {@link IAgentLearner} builder that will setup a {@link DoubleDQN double-DQN} algorithm in addition to the setup done by {@link BaseDQNAgentLearnerBuilder}. + */ +public class DoubleDQNBuilder extends BaseDQNAgentLearnerBuilder { + + + public DoubleDQNBuilder(Configuration configuration, + ITrainableNeuralNet neuralNet, + Builder> environmentBuilder, + Builder transformProcessBuilder, + Random rnd) { + super(configuration, neuralNet, environmentBuilder, transformProcessBuilder, rnd); + } + + @Override + protected IUpdateAlgorithm> buildUpdateAlgorithm() { + return new DoubleDQN(networks.getThreadCurrentNetwork(), networks.getTargetNetwork(), getConfiguration().getUpdateAlgorithmConfiguration()); + } + + @EqualsAndHashCode(callSuper = true) + @SuperBuilder + @Data + public static class Configuration extends BaseDQNAgentLearnerBuilder.Configuration { + } +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/INetworksHandler.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/INetworksHandler.java new file mode 100644 index 000000000..d4c308bdb --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/INetworksHandler.java @@ -0,0 +1,43 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.builder; + +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; + +/** + * An interface that abstract what the different networks are depending on the setup (sync vs async) + */ +public interface INetworksHandler { + /** + * @return The global shared target parameters θ– + */ + ITrainableNeuralNet getTargetNetwork(); + + /** + * @return The thread-specific parameters θ' + */ + ITrainableNeuralNet getThreadCurrentNetwork(); + + /** + * @return The global shared parameters θ + */ + ITrainableNeuralNet getGlobalCurrentNetwork(); + + /** + * Perform the required changes before a new IAgentLearner is built + */ + void resetForNewBuild(); +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/NStepQLearningBuilder.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/NStepQLearningBuilder.java new file mode 100644 index 000000000..e6e7a7d11 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/NStepQLearningBuilder.java @@ -0,0 +1,96 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.builder; + +import lombok.Data; +import lombok.EqualsAndHashCode; +import lombok.experimental.SuperBuilder; +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.IAgentLearner; +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.algorithm.NStepQLearning; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; +import org.deeplearning4j.rl4j.agent.learning.update.updater.GradientsNeuralNetUpdater; +import org.deeplearning4j.rl4j.agent.learning.update.updater.INeuralNetUpdater; +import org.deeplearning4j.rl4j.environment.Environment; +import org.deeplearning4j.rl4j.environment.IActionSchema; +import org.deeplearning4j.rl4j.experience.ExperienceHandler; +import org.deeplearning4j.rl4j.experience.StateActionExperienceHandler; +import org.deeplearning4j.rl4j.experience.StateActionPair; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.deeplearning4j.rl4j.observation.transform.TransformProcess; +import org.deeplearning4j.rl4j.policy.DQNPolicy; +import org.deeplearning4j.rl4j.policy.EpsGreedy; +import org.deeplearning4j.rl4j.policy.INeuralNetPolicy; +import org.deeplearning4j.rl4j.policy.IPolicy; +import org.nd4j.linalg.api.rng.Random; + +/** + * A {@link IAgentLearner} builder that will setup a {@link NStepQLearning n-step Q-Learning} algorithm with these: + *
  • a epsilon-greedy policy
    • + *
  • a n-step state-action-reward experience handler
    • + *
  • a neural net updater that expects gradient update data
    • + *
  • a n-step Q-Learning gradient conputation algorithm
    • + */ +public class NStepQLearningBuilder extends BaseAgentLearnerBuilder, Gradients>{ + + + private final Configuration configuration; + private final Random rnd; + + public NStepQLearningBuilder(Configuration configuration, + ITrainableNeuralNet neuralNet, + Builder> environmentBuilder, + Builder transformProcessBuilder, + Random rnd) { + super(configuration, neuralNet, environmentBuilder, transformProcessBuilder); + this.configuration = configuration; + this.rnd = rnd; + } + + @Override + protected IPolicy buildPolicy() { + INeuralNetPolicy greedyPolicy = new DQNPolicy(networks.getThreadCurrentNetwork()); + IActionSchema actionSchema = getEnvironment().getSchema().getActionSchema(); + return new EpsGreedy(greedyPolicy, actionSchema, configuration.getPolicyConfiguration(), rnd); + } + + @Override + protected ExperienceHandler> buildExperienceHandler() { + return new StateActionExperienceHandler(configuration.getExperienceHandlerConfiguration()); + } + + @Override + protected IUpdateAlgorithm> buildUpdateAlgorithm() { + IActionSchema actionSchema = getEnvironment().getSchema().getActionSchema(); + return new NStepQLearning(networks.getThreadCurrentNetwork(), networks.getTargetNetwork(), actionSchema.getActionSpaceSize(), configuration.getNstepQLearningConfiguration()); + } + + @Override + protected INeuralNetUpdater buildNeuralNetUpdater() { + return new GradientsNeuralNetUpdater(networks.getThreadCurrentNetwork(), networks.getTargetNetwork(), configuration.getNeuralNetUpdaterConfiguration()); + } + + @EqualsAndHashCode(callSuper = true) + @SuperBuilder + @Data + public static class Configuration extends BaseAgentLearnerBuilder.Configuration { + EpsGreedy.Configuration policyConfiguration; + GradientsNeuralNetUpdater.Configuration neuralNetUpdaterConfiguration; + NStepQLearning.Configuration nstepQLearningConfiguration; + StateActionExperienceHandler.Configuration experienceHandlerConfiguration; + } +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/StandardDQNBuilder.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/StandardDQNBuilder.java new file mode 100644 index 000000000..462dfacbb --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/StandardDQNBuilder.java @@ -0,0 +1,57 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.builder; + +import lombok.Data; +import lombok.EqualsAndHashCode; +import lombok.experimental.SuperBuilder; +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.IAgentLearner; +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.StandardDQN; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.environment.Environment; +import org.deeplearning4j.rl4j.learning.sync.Transition; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.deeplearning4j.rl4j.observation.transform.TransformProcess; +import org.nd4j.linalg.api.rng.Random; + +/** + * A {@link IAgentLearner} builder that will setup a {@link StandardDQN standard DQN} algorithm in addition to the setup done by {@link BaseDQNAgentLearnerBuilder}. + */ +public class StandardDQNBuilder extends BaseDQNAgentLearnerBuilder { + + + public StandardDQNBuilder(Configuration configuration, + ITrainableNeuralNet neuralNet, + Builder> environmentBuilder, + Builder transformProcessBuilder, + Random rnd) { + super(configuration, neuralNet, environmentBuilder, transformProcessBuilder, rnd); + } + + @Override + protected IUpdateAlgorithm> buildUpdateAlgorithm() { + return new StandardDQN(networks.getThreadCurrentNetwork(), networks.getTargetNetwork(), getConfiguration().getUpdateAlgorithmConfiguration()); + } + + @EqualsAndHashCode(callSuper = true) + @SuperBuilder + @Data + public static class Configuration extends BaseDQNAgentLearnerBuilder.Configuration { + } +} + diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/SyncNetworkHandler.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/SyncNetworkHandler.java new file mode 100644 index 000000000..109392d23 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/builder/SyncNetworkHandler.java @@ -0,0 +1,50 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.builder; + +import lombok.Getter; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; + +/** + * A {@link INetworksHandler} implementation for synchronous setups.

    + * The target network is cloned from the input network + * The thread-current and the global-current uses the input network directly. + * Note that there is no difference between the thread-current and the global-current in a sync setup. + */ +public class SyncNetworkHandler implements INetworksHandler { + + @Getter + final ITrainableNeuralNet targetNetwork; + + @Getter + ITrainableNeuralNet threadCurrentNetwork; + + @Getter + final ITrainableNeuralNet globalCurrentNetwork; + + public SyncNetworkHandler(ITrainableNeuralNet network) { + globalCurrentNetwork = network; + targetNetwork = network.clone(); + + // In sync setup, the thread current and the global current is the same network + threadCurrentNetwork = network; + } + + @Override + public void resetForNewBuild() { + // Do Nothing + } +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IActionSchema.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IActionSchema.java index 9e6e81a7b..ce46fcee0 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IActionSchema.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IActionSchema.java @@ -19,6 +19,8 @@ import lombok.Value; // Work in progress public interface IActionSchema { + int getActionSpaceSize(); + ACTION getNoOp(); // Review: A schema should be data-only and not have behavior diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IntegerActionSchema.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IntegerActionSchema.java index cdf172da6..7a51b1f39 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IntegerActionSchema.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/environment/IntegerActionSchema.java @@ -15,13 +15,16 @@ ******************************************************************************/ package org.deeplearning4j.rl4j.environment; +import lombok.Getter; import org.nd4j.linalg.api.rng.Random; import org.nd4j.linalg.factory.Nd4j; // Work in progress public class IntegerActionSchema implements IActionSchema { - private final int numActions; + @Getter + private final int actionSpaceSize; + private final int noOpAction; private final Random rnd; @@ -30,7 +33,7 @@ public class IntegerActionSchema implements IActionSchema { } public IntegerActionSchema(int numActions, int noOpAction, Random rnd) { - this.numActions = numActions; + this.actionSpaceSize = numActions; this.noOpAction = noOpAction; this.rnd = rnd; } @@ -42,6 +45,6 @@ public class IntegerActionSchema implements IActionSchema { @Override public Integer getRandomAction() { - return rnd.nextInt(numActions); + return rnd.nextInt(actionSpaceSize); } } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandler.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandler.java index c7f7d51ae..27a47dcaf 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandler.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandler.java @@ -15,7 +15,10 @@ ******************************************************************************/ package org.deeplearning4j.rl4j.experience; +import lombok.Builder; +import lombok.Data; import lombok.EqualsAndHashCode; +import lombok.experimental.SuperBuilder; import org.deeplearning4j.rl4j.learning.sync.ExpReplay; import org.deeplearning4j.rl4j.learning.sync.IExpReplay; import org.deeplearning4j.rl4j.learning.sync.Transition; @@ -47,8 +50,8 @@ public class ReplayMemoryExperienceHandler implements ExperienceHandler(maxReplayMemorySize, batchSize, random)); + public ReplayMemoryExperienceHandler(Configuration configuration, Random random) { + this(new ExpReplay(configuration.maxReplayMemorySize, configuration.batchSize, random)); } public void addExperience(Observation observation, A action, double reward, boolean isTerminal) { @@ -91,28 +94,19 @@ public class ReplayMemoryExperienceHandler implements ExperienceHandler build() { - return new ReplayMemoryExperienceHandler(maxReplayMemorySize, batchSize, random); - } } } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandler.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandler.java index a8fae47bc..b81a5fcc0 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandler.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandler.java @@ -15,6 +15,9 @@ ******************************************************************************/ package org.deeplearning4j.rl4j.experience; +import lombok.Builder; +import lombok.Data; +import lombok.experimental.SuperBuilder; import org.deeplearning4j.rl4j.observation.Observation; import java.util.ArrayList; @@ -29,13 +32,14 @@ import java.util.List; * @author Alexandre Boulanger */ public class StateActionExperienceHandler implements ExperienceHandler> { + private static final int DEFAULT_BATCH_SIZE = 8; private final int batchSize; private boolean isFinalObservationSet; - public StateActionExperienceHandler(int batchSize) { - this.batchSize = batchSize; + public StateActionExperienceHandler(Configuration configuration) { + this.batchSize = configuration.getBatchSize(); } private List> stateActionPairs = new ArrayList<>(); @@ -79,4 +83,13 @@ public class StateActionExperienceHandler implements ExperienceHandler extends QLearning buildLearningBehavior(IDQN qNetwork, QLearningConfiguration conf, Random random) { - IUpdateRule> updateRule = new DQNNeuralNetUpdateRule(qNetwork, conf.getTargetDqnUpdateFreq(), conf.isDoubleDQN(), conf.getGamma(), conf.getErrorClamp()); - ExperienceHandler> experienceHandler = new ReplayMemoryExperienceHandler(conf.getExpRepMaxSize(), conf.getBatchSize(), random); + ITrainableNeuralNet target = qNetwork.clone(); + BaseTransitionTDAlgorithm.Configuration aglorithmConfiguration = BaseTransitionTDAlgorithm.Configuration.builder() + .gamma(conf.getGamma()) + .errorClamp(conf.getErrorClamp()) + .build(); + IUpdateAlgorithm> updateAlgorithm = conf.isDoubleDQN() + ? new DoubleDQN(qNetwork, target, aglorithmConfiguration) + : new StandardDQN(qNetwork, target, aglorithmConfiguration); + + LabelsNeuralNetUpdater.Configuration neuralNetUpdaterConfiguration = LabelsNeuralNetUpdater.Configuration.builder() + .targetUpdateFrequency(conf.getTargetDqnUpdateFreq()) + .build(); + INeuralNetUpdater updater = new LabelsNeuralNetUpdater(qNetwork, target, neuralNetUpdaterConfiguration); + IUpdateRule> updateRule = new UpdateRule>(updateAlgorithm, updater); + + ReplayMemoryExperienceHandler.Configuration experienceHandlerConfiguration = ReplayMemoryExperienceHandler.Configuration.builder() + .maxReplayMemorySize(conf.getExpRepMaxSize()) + .batchSize(conf.getBatchSize()) + .build(); + ExperienceHandler> experienceHandler = new ReplayMemoryExperienceHandler(experienceHandlerConfiguration, random); return LearningBehavior.>builder() .experienceHandler(experienceHandler) .updateRule(updateRule) - .experienceUpdateSize(conf.getBatchSize()) .build(); - } public MDP getMdp() { diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonGradientNames.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonGradientNames.java new file mode 100644 index 000000000..c3becf9a3 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonGradientNames.java @@ -0,0 +1,5 @@ +package org.deeplearning4j.rl4j.network; + +public abstract class CommonGradientNames { + public static final String QValues = "Q"; +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonLabelNames.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonLabelNames.java new file mode 100644 index 000000000..75d691238 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/CommonLabelNames.java @@ -0,0 +1,7 @@ +package org.deeplearning4j.rl4j.network; + +public abstract class CommonLabelNames { + + public static final String QValues = "Q"; + +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/IOutputNeuralNet.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/IOutputNeuralNet.java index 58e219ea0..404502d2e 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/IOutputNeuralNet.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/IOutputNeuralNet.java @@ -35,4 +35,9 @@ public interface IOutputNeuralNet { * @return The ouptut of the network */ INDArray output(INDArray batch); + + /** + * Clear the neural net of any previous state + */ + void reset(); } \ No newline at end of file diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ITrainableNeuralNet.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ITrainableNeuralNet.java index da91d7e6d..320db019b 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ITrainableNeuralNet.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ITrainableNeuralNet.java @@ -15,17 +15,31 @@ ******************************************************************************/ package org.deeplearning4j.rl4j.network; -import org.nd4j.linalg.dataset.api.DataSet; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; /** * An interface defining the trainable aspect of a {@link NeuralNet}. */ -public interface ITrainableNeuralNet { +public interface ITrainableNeuralNet extends IOutputNeuralNet { /** * Train the neural net using the supplied feature-labels * @param featuresLabels The feature-labels */ - void fit(DataSet featuresLabels); + void fit(FeaturesLabels featuresLabels); + + /** + * Use the supplied feature-labels to compute the {@link Gradients} on the neural network. + * @param updateLabels The feature-labels + * @return The computed {@link Gradients} + */ + Gradients computeGradients(FeaturesLabels updateLabels); + + /** + * Applies a {@link Gradients} to the network + * @param gradients + */ + void applyGradients(Gradients gradients); /** * Changes this instance to be a copy of the from network. diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/NeuralNet.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/NeuralNet.java index 7823ea906..8e7bbc166 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/NeuralNet.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/NeuralNet.java @@ -29,7 +29,7 @@ import java.io.OutputStream; * Factorisation between ActorCritic and DQN neural net. * Useful for AsyncLearning and Thread code. */ -public interface NeuralNet extends IOutputNeuralNet, ITrainableNeuralNet { +public interface NeuralNet extends ITrainableNeuralNet { /** * Returns the underlying MultiLayerNetwork or ComputationGraph objects. diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticCompGraph.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticCompGraph.java index 6e37fb0f3..9b50eeef1 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticCompGraph.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticCompGraph.java @@ -26,10 +26,11 @@ import org.deeplearning4j.nn.graph.ComputationGraph; import org.deeplearning4j.nn.layers.recurrent.RnnOutputLayer; import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr; import org.deeplearning4j.optimize.api.TrainingListener; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; import org.deeplearning4j.rl4j.observation.Observation; import org.deeplearning4j.util.ModelSerializer; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import java.io.IOException; import java.io.OutputStream; @@ -84,8 +85,21 @@ public class ActorCriticCompGraph implements IActorCritic } @Override - public void fit(DataSet featuresLabels) { - fit(featuresLabels.getFeatures(), new INDArray[] { featuresLabels.getLabels() }); + public void fit(FeaturesLabels featuresLabels) { + // TODO + throw new NotImplementedException("Not implemented: will be done with AgentLearner async support"); + } + + @Override + public Gradients computeGradients(FeaturesLabels updateLabels) { + // TODO + throw new NotImplementedException("Not implemented: will be done with AgentLearner async support"); + } + + @Override + public void applyGradients(Gradients gradients) { + // TODO + throw new NotImplementedException("Not implemented: will be done with AgentLearner async support"); } public void copy(ActorCriticCompGraph from) { diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticSeparate.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticSeparate.java index a086b6065..3e31f0a0b 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticSeparate.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/ac/ActorCriticSeparate.java @@ -25,10 +25,13 @@ import org.deeplearning4j.nn.layers.recurrent.RnnOutputLayer; import org.deeplearning4j.nn.multilayer.MultiLayerNetwork; import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr; import org.deeplearning4j.optimize.api.TrainingListener; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; +import org.deeplearning4j.rl4j.network.CommonGradientNames; +import org.deeplearning4j.rl4j.network.CommonLabelNames; import org.deeplearning4j.rl4j.observation.Observation; import org.deeplearning4j.util.ModelSerializer; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import java.io.IOException; import java.io.OutputStream; @@ -73,7 +76,6 @@ public class ActorCriticSeparate implements IAct } - public INDArray[] outputAll(INDArray batch) { if (recurrent) { return new INDArray[] {valueNet.rnnTimeStep(batch), policyNet.rnnTimeStep(batch)}; @@ -90,12 +92,24 @@ public class ActorCriticSeparate implements IAct } @Override - public void fit(DataSet featuresLabels) { + public void fit(FeaturesLabels featuresLabels) { // TODO: signature of fit() will change from DataSet to a class that has named labels to support network like // this one (labels for the value-network and another labels for the policy-network throw new NotImplementedException("Not implemented: will be done with AgentLearner async support"); } + @Override + public Gradients computeGradients(FeaturesLabels updateLabels) { + // TODO + throw new NotImplementedException("Not implemented: will be done with AgentLearner async support"); + } + + @Override + public void applyGradients(Gradients gradients) { + // TODO + throw new NotImplementedException("Not implemented: will be done with AgentLearner async support"); + } + public void copy(NN from) { valueNet.setParams(from.valueNet.params()); policyNet.setParams(from.policyNet.params()); diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/dqn/DQN.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/dqn/DQN.java index 260ea6aa2..27b5bcdf6 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/dqn/DQN.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/network/dqn/DQN.java @@ -22,10 +22,13 @@ import org.deeplearning4j.nn.gradient.Gradient; import org.deeplearning4j.nn.multilayer.MultiLayerNetwork; import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr; import org.deeplearning4j.optimize.api.TrainingListener; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; +import org.deeplearning4j.rl4j.network.CommonGradientNames; +import org.deeplearning4j.rl4j.network.CommonLabelNames; import org.deeplearning4j.rl4j.observation.Observation; import org.deeplearning4j.util.ModelSerializer; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import java.io.IOException; import java.io.OutputStream; @@ -81,8 +84,8 @@ public class DQN implements IDQN { } @Override - public void fit(DataSet featuresLabels) { - fit(featuresLabels.getFeatures(), featuresLabels.getLabels()); + public void fit(FeaturesLabels featuresLabels) { + fit(featuresLabels.getFeatures(), featuresLabels.getLabels(CommonLabelNames.QValues)); } @Override @@ -115,6 +118,41 @@ public class DQN implements IDQN { return gradient(input, labels[0]); } + + @Override + public Gradients computeGradients(FeaturesLabels updateLabels) { + mln.setInput(updateLabels.getFeatures()); + mln.setLabels(updateLabels.getLabels(CommonLabelNames.QValues)); + mln.computeGradientAndScore(); + Collection iterationListeners = mln.getListeners(); + if (iterationListeners != null && iterationListeners.size() > 0) { + for (TrainingListener l : iterationListeners) { + l.onGradientCalculation(mln); + } + } + Gradients result = new Gradients(updateLabels.getBatchSize()); + result.putGradient(CommonGradientNames.QValues, mln.gradient()); + return result; + } + + @Override + public void applyGradients(Gradients gradients) { + Gradient qValues = gradients.getGradient(CommonGradientNames.QValues); + + MultiLayerConfiguration mlnConf = mln.getLayerWiseConfigurations(); + int iterationCount = mlnConf.getIterationCount(); + int epochCount = mlnConf.getEpochCount(); + mln.getUpdater().update(mln, qValues, iterationCount, epochCount, (int)gradients.getBatchSize(), LayerWorkspaceMgr.noWorkspaces()); + mln.params().subi(qValues.gradient()); + Collection iterationListeners = mln.getListeners(); + if (iterationListeners != null && iterationListeners.size() > 0) { + for (TrainingListener listener : iterationListeners) { + listener.iterationDone(mln, iterationCount, epochCount); + } + } + mlnConf.setIterationCount(iterationCount + 1); + } + public void applyGradient(Gradient[] gradient, int batchSize) { MultiLayerConfiguration mlnConf = mln.getLayerWiseConfigurations(); int iterationCount = mlnConf.getIterationCount(); diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/DQNPolicy.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/DQNPolicy.java index ed591a1ff..3b27cc778 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/DQNPolicy.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/DQNPolicy.java @@ -18,6 +18,7 @@ package org.deeplearning4j.rl4j.policy; import lombok.AllArgsConstructor; import org.deeplearning4j.rl4j.learning.Learning; +import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.deeplearning4j.rl4j.network.dqn.DQN; import org.deeplearning4j.rl4j.network.dqn.IDQN; import org.deeplearning4j.rl4j.space.Encodable; @@ -37,14 +38,14 @@ import java.io.IOException; @AllArgsConstructor public class DQNPolicy extends Policy { - final private IDQN dqn; + final private IOutputNeuralNet neuralNet; public static DQNPolicy load(String path) throws IOException { return new DQNPolicy<>(DQN.load(path)); } - public IDQN getNeuralNet() { - return dqn; + public IOutputNeuralNet getNeuralNet() { + return neuralNet; } @Override @@ -53,12 +54,13 @@ public class DQNPolicy extends Policy { } public Integer nextAction(INDArray input) { - INDArray output = dqn.output(input); + INDArray output = neuralNet.output(input); return Learning.getMaxAction(output); } public void save(String filename) throws IOException { - dqn.save(filename); + // TODO: refac load & save. Code below should continue to work in the meantime because it's only called by the lecacy code and it's only using a DQN network with DQNPolicy + ((IDQN)neuralNet).save(filename); } } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/EpsGreedy.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/EpsGreedy.java index f7422be92..4e66765a1 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/EpsGreedy.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/EpsGreedy.java @@ -18,12 +18,14 @@ package org.deeplearning4j.rl4j.policy; import lombok.Builder; +import lombok.Data; import lombok.NonNull; +import lombok.experimental.SuperBuilder; import lombok.extern.slf4j.Slf4j; import org.deeplearning4j.rl4j.environment.IActionSchema; import org.deeplearning4j.rl4j.learning.IEpochTrainer; import org.deeplearning4j.rl4j.mdp.MDP; -import org.deeplearning4j.rl4j.network.NeuralNet; +import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.deeplearning4j.rl4j.observation.Observation; import org.deeplearning4j.rl4j.space.ActionSpace; import org.deeplearning4j.rl4j.space.Encodable; @@ -45,7 +47,7 @@ import org.nd4j.linalg.factory.Nd4j; public class EpsGreedy extends Policy { final private INeuralNetPolicy policy; - final private int updateStart; + final private int annealingStart; final private int epsilonNbStep; final private Random rnd; final private double minEpsilon; @@ -61,14 +63,14 @@ public class EpsGreedy extends Policy { @Deprecated public > EpsGreedy(Policy policy, MDP> mdp, - int updateStart, + int annealingStart, int epsilonNbStep, Random rnd, double minEpsilon, IEpochTrainer learning) { this.policy = policy; this.mdp = mdp; - this.updateStart = updateStart; + this.annealingStart = annealingStart; this.epsilonNbStep = epsilonNbStep; this.rnd = rnd; this.minEpsilon = minEpsilon; @@ -77,17 +79,17 @@ public class EpsGreedy extends Policy { this.actionSchema = null; } - public EpsGreedy(@NonNull Policy policy, @NonNull IActionSchema actionSchema, double minEpsilon, int updateStart, int epsilonNbStep) { - this(policy, actionSchema, minEpsilon, updateStart, epsilonNbStep, null); + public EpsGreedy(@NonNull Policy policy, @NonNull IActionSchema actionSchema, double minEpsilon, int annealingStart, int epsilonNbStep) { + this(policy, actionSchema, minEpsilon, annealingStart, epsilonNbStep, null); } @Builder - public EpsGreedy(@NonNull INeuralNetPolicy policy, @NonNull IActionSchema actionSchema, double minEpsilon, int updateStart, int epsilonNbStep, Random rnd) { + public EpsGreedy(@NonNull INeuralNetPolicy policy, @NonNull IActionSchema actionSchema, double minEpsilon, int annealingStart, int epsilonNbStep, Random rnd) { this.policy = policy; this.rnd = rnd == null ? Nd4j.getRandom() : rnd; this.minEpsilon = minEpsilon; - this.updateStart = updateStart; + this.annealingStart = annealingStart; this.epsilonNbStep = epsilonNbStep; this.actionSchema = actionSchema; @@ -95,7 +97,11 @@ public class EpsGreedy extends Policy { this.learning = null; } - public NeuralNet getNeuralNet() { + public EpsGreedy(INeuralNetPolicy policy, IActionSchema actionSchema, @NonNull Configuration configuration, Random rnd) { + this(policy, actionSchema, configuration.getMinEpsilon(), configuration.getAnnealingStart(), configuration.getEpsilonNbStep(), rnd); + } + + public IOutputNeuralNet getNeuralNet() { return policy.getNeuralNet(); } @@ -141,6 +147,16 @@ public class EpsGreedy extends Policy { public double getEpsilon() { int step = actionSchema != null ? annealingStep : learning.getStepCount(); - return Math.min(1.0, Math.max(minEpsilon, 1.0 - (step - updateStart) * 1.0 / epsilonNbStep)); + return Math.min(1.0, Math.max(minEpsilon, 1.0 - (step - annealingStart) * 1.0 / epsilonNbStep)); + } + + @SuperBuilder + @Data + public static class Configuration { + @Builder.Default + final int annealingStart = 0; + + final int epsilonNbStep; + final double minEpsilon; } } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/INeuralNetPolicy.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/INeuralNetPolicy.java index c213396c6..b3967e54f 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/INeuralNetPolicy.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/INeuralNetPolicy.java @@ -1,7 +1,7 @@ package org.deeplearning4j.rl4j.policy; -import org.deeplearning4j.rl4j.network.NeuralNet; +import org.deeplearning4j.rl4j.network.IOutputNeuralNet; public interface INeuralNetPolicy extends IPolicy { - NeuralNet getNeuralNet(); + IOutputNeuralNet getNeuralNet(); } diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/Policy.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/Policy.java index cf369e359..827162fa1 100644 --- a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/Policy.java +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/policy/Policy.java @@ -16,15 +16,16 @@ package org.deeplearning4j.rl4j.policy; +import lombok.experimental.SuperBuilder; import org.deeplearning4j.gym.StepReply; import org.deeplearning4j.rl4j.learning.HistoryProcessor; import org.deeplearning4j.rl4j.learning.IHistoryProcessor; import org.deeplearning4j.rl4j.learning.Learning; import org.deeplearning4j.rl4j.mdp.MDP; -import org.deeplearning4j.rl4j.network.NeuralNet; -import org.deeplearning4j.rl4j.space.Encodable; +import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.deeplearning4j.rl4j.observation.Observation; import org.deeplearning4j.rl4j.space.ActionSpace; +import org.deeplearning4j.rl4j.space.Encodable; import org.deeplearning4j.rl4j.util.LegacyMDPWrapper; /** @@ -36,7 +37,7 @@ import org.deeplearning4j.rl4j.util.LegacyMDPWrapper; */ public abstract class Policy implements INeuralNetPolicy { - public abstract NeuralNet getNeuralNet(); + public abstract IOutputNeuralNet getNeuralNet(); public abstract A nextAction(Observation obs); @@ -106,5 +107,4 @@ public abstract class Policy implements INeuralNetPolicy { return new Learning.InitMdp(0, observation, reward); } - } diff --git a/libnd4j/include/loops/cuda/reduce3.cu b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/ITrainer.java similarity index 71% rename from libnd4j/include/loops/cuda/reduce3.cu rename to rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/ITrainer.java index c1d63e8dd..7641d3a35 100644 --- a/libnd4j/include/loops/cuda/reduce3.cu +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/ITrainer.java @@ -1,33 +1,23 @@ -/******************************************************************************* - * Copyright (c) 2015-2018 Skymind, Inc. - * - * This program and the accompanying materials are made available under the - * terms of the Apache License, Version 2.0 which is available at - * https://www.apache.org/licenses/LICENSE-2.0. - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT - * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the - * License for the specific language governing permissions and limitations - * under the License. - * - * SPDX-License-Identifier: Apache-2.0 - ******************************************************************************/ - -// -// @author raver119@gmail.com -// - - -#include -#include -#include -#include -#include - -namespace functions { - namespace reduce3 { - - - } -} \ No newline at end of file +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.trainer; + +/** + * An interface describing the behavior of all trainers + */ +public interface ITrainer { + void train(); +} diff --git a/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/SyncTrainer.java b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/SyncTrainer.java new file mode 100644 index 000000000..d21e30e58 --- /dev/null +++ b/rl4j/rl4j-core/src/main/java/org/deeplearning4j/rl4j/trainer/SyncTrainer.java @@ -0,0 +1,64 @@ +/******************************************************************************* + * Copyright (c) 2020 Konduit K.K. + * + * This program and the accompanying materials are made available under the + * terms of the Apache License, Version 2.0 which is available at + * https://www.apache.org/licenses/LICENSE-2.0. + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the + * License for the specific language governing permissions and limitations + * under the License. + * + * SPDX-License-Identifier: Apache-2.0 + ******************************************************************************/ +package org.deeplearning4j.rl4j.trainer; + +import lombok.Getter; +import lombok.NonNull; +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.IAgentLearner; + +import java.util.function.Predicate; + +// TODO: Add listeners & events once AsyncTrainer is implemented + +/** + * A {@link ITrainer} implementation that will create a single {@link IAgentLearner} and perform the training in a + * synchronized setup, until a stopping condition is met. + * + * @param The type of the actions expected by the environment + */ +public class SyncTrainer implements ITrainer { + + private final Predicate> stoppingCondition; + + @Getter + private int episodeCount; + + @Getter + final IAgentLearner agentLearner; + + /** + * Build a SyncTrainer that will train until a stopping condition is met. + * @param agentLearnerBuilder the builder that will be used to create the agent-learner instance + * @param stoppingCondition the training will stop when this condition evaluates to true + */ + @lombok.Builder + public SyncTrainer(@NonNull Builder> agentLearnerBuilder, + @NonNull Predicate> stoppingCondition) { + this.stoppingCondition = stoppingCondition; + agentLearner = agentLearnerBuilder.build(); + } + + /** + * Perform the training + */ + public void train() { + while (!stoppingCondition.test(this)) { + agentLearner.run(); + ++episodeCount; + } + } +} \ No newline at end of file diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentLearnerTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentLearnerTest.java index e0c0685bf..04f11604b 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentLearnerTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentLearnerTest.java @@ -1,6 +1,6 @@ package org.deeplearning4j.rl4j.agent; -import org.deeplearning4j.rl4j.agent.learning.LearningBehavior; +import org.deeplearning4j.rl4j.agent.learning.behavior.LearningBehavior; import org.deeplearning4j.rl4j.environment.Environment; import org.deeplearning4j.rl4j.environment.IntegerActionSchema; import org.deeplearning4j.rl4j.environment.Schema; @@ -43,9 +43,10 @@ public class AgentLearnerTest { @Test public void when_episodeIsStarted_expect_learningBehaviorHandleEpisodeStartCalled() { // Arrange - AgentLearner sut = AgentLearner.builder(environmentMock, transformProcessMock, policyMock, learningBehaviorMock) + AgentLearner.Configuration configuration = AgentLearner.Configuration.builder() .maxEpisodeSteps(3) .build(); + AgentLearner sut = new AgentLearner(environmentMock, transformProcessMock, policyMock, configuration, null, learningBehaviorMock); Schema schema = new Schema(new IntegerActionSchema(0, -1)); when(environmentMock.reset()).thenReturn(new HashMap<>()); @@ -67,9 +68,10 @@ public class AgentLearnerTest { @Test public void when_runIsCalled_expect_experienceHandledWithLearningBehavior() { // Arrange - AgentLearner sut = AgentLearner.builder(environmentMock, transformProcessMock, policyMock, learningBehaviorMock) + AgentLearner.Configuration configuration = AgentLearner.Configuration.builder() .maxEpisodeSteps(4) .build(); + AgentLearner sut = new AgentLearner(environmentMock, transformProcessMock, policyMock, configuration, null, learningBehaviorMock); Schema schema = new Schema(new IntegerActionSchema(0, -1)); when(environmentMock.getSchema()).thenReturn(schema); @@ -127,9 +129,10 @@ public class AgentLearnerTest { @Test public void when_runIsCalledMultipleTimes_expect_totalStepCountCorrect() { // Arrange - AgentLearner sut = AgentLearner.builder(environmentMock, transformProcessMock, policyMock, learningBehaviorMock) + AgentLearner.Configuration configuration = AgentLearner.Configuration.builder() .maxEpisodeSteps(4) .build(); + AgentLearner sut = new AgentLearner(environmentMock, transformProcessMock, policyMock, configuration, null, learningBehaviorMock); Schema schema = new Schema(new IntegerActionSchema(0, -1)); when(environmentMock.getSchema()).thenReturn(schema); @@ -166,9 +169,10 @@ public class AgentLearnerTest { @Test public void when_runIsCalledMultipleTimes_expect_rewardSentToLearningBehaviorToBeCorrect() { // Arrange - AgentLearner sut = AgentLearner.builder(environmentMock, transformProcessMock, policyMock, learningBehaviorMock) + AgentLearner.Configuration configuration = AgentLearner.Configuration.builder() .maxEpisodeSteps(4) .build(); + AgentLearner sut = new AgentLearner(environmentMock, transformProcessMock, policyMock, configuration, null, learningBehaviorMock); Schema schema = new Schema(new IntegerActionSchema(0, -1)); when(environmentMock.getSchema()).thenReturn(schema); diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentTest.java index 0022e61f0..92ea9f4d7 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/AgentTest.java @@ -18,7 +18,6 @@ import java.util.HashMap; import java.util.List; import java.util.Map; -import static org.mockito.ArgumentMatchers.*; import static org.mockito.Mockito.*; @RunWith(MockitoJUnitRunner.class) @@ -34,7 +33,7 @@ public class AgentTest { @Test public void when_buildingWithNullEnvironment_expect_exception() { try { - Agent.builder(null, null, null).build(); + new Agent(null, null, null, null, null); fail("NullPointerException should have been thrown"); } catch (NullPointerException exception) { String expectedMessage = "environment is marked non-null but is null"; @@ -47,7 +46,7 @@ public class AgentTest { @Test public void when_buildingWithNullTransformProcess_expect_exception() { try { - Agent.builder(environmentMock, null, null).build(); + new Agent(environmentMock, null, null, null, null); fail("NullPointerException should have been thrown"); } catch (NullPointerException exception) { String expectedMessage = "transformProcess is marked non-null but is null"; @@ -60,7 +59,7 @@ public class AgentTest { @Test public void when_buildingWithNullPolicy_expect_exception() { try { - Agent.builder(environmentMock, transformProcessMock, null).build(); + new Agent(environmentMock, transformProcessMock, null, null, null); fail("NullPointerException should have been thrown"); } catch (NullPointerException exception) { String expectedMessage = "policy is marked non-null but is null"; @@ -70,15 +69,29 @@ public class AgentTest { } } + @Test + public void when_buildingWithNullConfiguration_expect_exception() { + try { + new Agent(environmentMock, transformProcessMock, policyMock, null, null); + fail("NullPointerException should have been thrown"); + } catch (NullPointerException exception) { + String expectedMessage = "configuration is marked non-null but is null"; + String actualMessage = exception.getMessage(); + + assertTrue(actualMessage.contains(expectedMessage)); + } + } + @Test public void when_buildingWithInvalidMaxSteps_expect_exception() { try { - Agent.builder(environmentMock, transformProcessMock, policyMock) - .maxEpisodeSteps(0) - .build(); + Agent.Configuration configuration = Agent.Configuration.builder() + .maxEpisodeSteps(0) + .build(); + new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); fail("IllegalArgumentException should have been thrown"); } catch (IllegalArgumentException exception) { - String expectedMessage = "maxEpisodeSteps must be greater than 0, got [0]"; + String expectedMessage = "Configuration: maxEpisodeSteps must be null (no maximum) or greater than 0, got [0]"; String actualMessage = exception.getMessage(); assertTrue(actualMessage.contains(expectedMessage)); @@ -88,9 +101,8 @@ public class AgentTest { @Test public void when_buildingWithId_expect_idSetInAgent() { // Arrange - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) - .id("TestAgent") - .build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, "TestAgent"); // Assert assertEquals("TestAgent", sut.getId()); @@ -107,8 +119,8 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(new Observation(Nd4j.create(new double[] { 123.0 }))); when(policyMock.nextAction(any(Observation.class))).thenReturn(1); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) - .build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); when(listenerMock.onBeforeStep(any(Agent.class), any(Observation.class), anyInt())).thenReturn(AgentListener.ListenerResponse.STOP); sut.addListener(listenerMock); @@ -135,7 +147,8 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(new Observation(Nd4j.create(new double[] { 123.0 }))); when(environmentMock.isEpisodeFinished()).thenReturn(true); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock).build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); Agent spy = Mockito.spy(sut); // Act @@ -156,7 +169,8 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(new Observation(Nd4j.create(new double[] { 123.0 }))); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock).build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); when(listenerMock.onBeforeEpisode(any(Agent.class))).thenReturn(AgentListener.ListenerResponse.STOP); sut.addListener(listenerMock); @@ -183,8 +197,8 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(new Observation(Nd4j.create(new double[] { 123.0 }))); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) - .build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); final Agent spy = Mockito.spy(sut); @@ -213,9 +227,10 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(new Observation(Nd4j.create(new double[] { 123.0 }))); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) + Agent.Configuration configuration = Agent.Configuration.builder() .maxEpisodeSteps(3) .build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); final Agent spy = Mockito.spy(sut); @@ -243,8 +258,8 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(Observation.SkippedObservation); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) - .build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); when(listenerMock.onBeforeStep(any(Agent.class), any(Observation.class), any())).thenReturn(AgentListener.ListenerResponse.STOP); sut.addListener(listenerMock); @@ -268,8 +283,8 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(Observation.SkippedObservation); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) - .build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); when(listenerMock.onBeforeStep(any(Agent.class), any(Observation.class), any())).thenReturn(AgentListener.ListenerResponse.STOP); sut.addListener(listenerMock); @@ -295,9 +310,10 @@ public class AgentTest { when(policyMock.nextAction(any(Observation.class))) .thenAnswer(invocation -> (int)((Observation)invocation.getArgument(0)).getData().getDouble(0)); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) + Agent.Configuration configuration = Agent.Configuration.builder() .maxEpisodeSteps(3) .build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); Agent spy = Mockito.spy(sut); @@ -335,7 +351,8 @@ public class AgentTest { when(transformProcessMock.transform(any(Map.class), anyInt(), anyBoolean())).thenReturn(new Observation(Nd4j.create(new double[] { 123.0 }))); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock).build(); + Agent.Configuration configuration = Agent.Configuration.builder().build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); when(listenerMock.onBeforeStep(any(Agent.class), any(Observation.class), any())).thenReturn(AgentListener.ListenerResponse.STOP); sut.addListener(listenerMock); @@ -365,9 +382,10 @@ public class AgentTest { when(policyMock.nextAction(any(Observation.class))).thenReturn(123); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) + Agent.Configuration configuration = Agent.Configuration.builder() .maxEpisodeSteps(1) .build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); // Act sut.run(); @@ -388,9 +406,10 @@ public class AgentTest { when(policyMock.nextAction(any(Observation.class))).thenReturn(123); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) + Agent.Configuration configuration = Agent.Configuration.builder() .maxEpisodeSteps(1) .build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); // Act sut.run(); @@ -413,9 +432,10 @@ public class AgentTest { when(policyMock.nextAction(any(Observation.class))).thenReturn(123); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) + Agent.Configuration configuration = Agent.Configuration.builder() .maxEpisodeSteps(1) .build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); Agent spy = Mockito.spy(sut); // Act @@ -438,9 +458,10 @@ public class AgentTest { when(policyMock.nextAction(any(Observation.class))).thenReturn(123); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) + Agent.Configuration configuration = Agent.Configuration.builder() .maxEpisodeSteps(1) .build(); + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); when(listenerMock.onAfterStep(any(Agent.class), any(StepResult.class))).thenReturn(AgentListener.ListenerResponse.STOP); sut.addListener(listenerMock); @@ -466,10 +487,11 @@ public class AgentTest { when(policyMock.nextAction(any(Observation.class))).thenReturn(123); - Agent sut = Agent.builder(environmentMock, transformProcessMock, policyMock) + Agent.Configuration configuration = Agent.Configuration.builder() .maxEpisodeSteps(1) .build(); - + Agent sut = new Agent(environmentMock, transformProcessMock, policyMock, configuration, null); + sut.addListener(listenerMock); Agent spy = Mockito.spy(sut); // Act @@ -477,5 +499,6 @@ public class AgentTest { // Assert verify(spy, times(1)).onAfterEpisode(); + verify(listenerMock, times(1)).onAfterEpisode(any()); } } diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearningTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearningTest.java new file mode 100644 index 000000000..ba81082f9 --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/NStepQLearningTest.java @@ -0,0 +1,129 @@ +package org.deeplearning4j.rl4j.agent.learning.algorithm; + +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; +import org.deeplearning4j.rl4j.experience.StateActionPair; +import org.deeplearning4j.rl4j.network.CommonLabelNames; +import org.deeplearning4j.rl4j.network.IOutputNeuralNet; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.deeplearning4j.rl4j.observation.Observation; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.mockito.ArgumentCaptor; +import org.mockito.Mock; +import org.mockito.junit.MockitoJUnitRunner; +import org.nd4j.linalg.api.ndarray.INDArray; +import org.nd4j.linalg.factory.Nd4j; + +import java.util.ArrayList; +import java.util.List; + +import static org.junit.Assert.assertEquals; +import static org.mockito.Mockito.*; + +@RunWith(MockitoJUnitRunner.class) +public class NStepQLearningTest { + + private static final int ACTION_SPACE_SIZE = 2; + + @Mock + ITrainableNeuralNet currentMock; + + @Mock + IOutputNeuralNet targetMock; + + NStepQLearning sut; + + private void setup(double gamma) { + when(currentMock.output(any(INDArray.class))).thenAnswer(invocation -> invocation.getArgument(0, INDArray.class).mul(-1.0)); + when(targetMock.output(any(INDArray.class))).thenAnswer(invocation -> invocation.getArgument(0, INDArray.class).mul(-2.0)); + + NStepQLearning.Configuration configuration = NStepQLearning.Configuration.builder() + .gamma(gamma) + .build(); + sut = new NStepQLearning(currentMock, targetMock, ACTION_SPACE_SIZE, configuration); + } + + @Test + public void when_isTerminal_expect_initRewardIs0() { + // Arrange + int action = 0; + setup(1.0); + + final Observation observation = new Observation(Nd4j.zeros(1, 2)); + List> experience = new ArrayList>() { + { + add(new StateActionPair(observation, action, 0.0, true)); + } + }; + + // Act + Gradients result = sut.compute(experience); + + // Assert + ArgumentCaptor argument = ArgumentCaptor.forClass(FeaturesLabels.class); + verify(currentMock, times(1)).computeGradients(argument.capture()); + + FeaturesLabels featuresLabels = argument.getValue(); + assertEquals(0.0, featuresLabels.getLabels(CommonLabelNames.QValues).getDouble(0), 0.000001); + } + + @Test + public void when_notTerminal_expect_initRewardWithMaxQFromTarget() { + // Arrange + int action = 0; + setup(1.0); + + final Observation observation = new Observation(Nd4j.create(new double[] { -123.0, -234.0 }).reshape(1, 2)); + List> experience = new ArrayList>() { + { + add(new StateActionPair(observation, action, 0.0, false)); + } + }; + + // Act + Gradients result = sut.compute(experience); + + // Assert + ArgumentCaptor argument = ArgumentCaptor.forClass(FeaturesLabels.class); + verify(currentMock, times(1)).computeGradients(argument.capture()); + + FeaturesLabels featuresLabels = argument.getValue(); + assertEquals(-2.0 * observation.getData().getDouble(0, 1), featuresLabels.getLabels(CommonLabelNames.QValues).getDouble(0), 0.000001); + } + + @Test + public void when_callingWithMultipleExperiences_expect_gradientsAreValid() { + // Arrange + double gamma = 0.9; + setup(gamma); + + List> experience = new ArrayList>() { + { + add(new StateActionPair(new Observation(Nd4j.create(new double[] { -1.1, -1.2 }).reshape(1, 2)), 0, 1.0, false)); + add(new StateActionPair(new Observation(Nd4j.create(new double[] { -2.1, -2.2 }).reshape(1, 2)), 1, 2.0, true)); + } + }; + + // Act + sut.compute(experience); + + // Assert + ArgumentCaptor argument = ArgumentCaptor.forClass(FeaturesLabels.class); + verify(currentMock, times(1)).computeGradients(argument.capture()); + + // input side -- should be a stack of observations + INDArray featuresValues = argument.getValue().getFeatures(); + assertEquals(-1.1, featuresValues.getDouble(0, 0), 0.00001); + assertEquals(-1.2, featuresValues.getDouble(0, 1), 0.00001); + assertEquals(-2.1, featuresValues.getDouble(1, 0), 0.00001); + assertEquals(-2.2, featuresValues.getDouble(1, 1), 0.00001); + + // target side + INDArray labels = argument.getValue().getLabels(CommonLabelNames.QValues); + assertEquals(1.0 + gamma * 2.0, labels.getDouble(0, 0), 0.00001); + assertEquals(1.2, labels.getDouble(0, 1), 0.00001); + assertEquals(2.1, labels.getDouble(1, 0), 0.00001); + assertEquals(2.0, labels.getDouble(1, 1), 0.00001); + } +} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/DoubleDQNTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/DoubleDQNTest.java similarity index 74% rename from rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/DoubleDQNTest.java rename to rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/DoubleDQNTest.java index d5a946a65..b4f0b3140 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/DoubleDQNTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/DoubleDQNTest.java @@ -1,6 +1,9 @@ -package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm; +package org.deeplearning4j.rl4j.agent.learning.algorithm.dqn; +import org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.DoubleDQN; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; import org.deeplearning4j.rl4j.learning.sync.Transition; +import org.deeplearning4j.rl4j.network.CommonLabelNames; import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.deeplearning4j.rl4j.observation.Observation; import org.junit.Before; @@ -9,7 +12,6 @@ import org.junit.runner.RunWith; import org.mockito.Mock; import org.mockito.junit.MockitoJUnitRunner; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import org.nd4j.linalg.factory.Nd4j; import java.util.ArrayList; @@ -28,6 +30,9 @@ public class DoubleDQNTest { @Mock IOutputNeuralNet targetQNetworkMock; + private final BaseTransitionTDAlgorithm.Configuration configuration = BaseTransitionTDAlgorithm.Configuration.builder() + .gamma(0.5) + .build(); @Before public void setup() { @@ -47,13 +52,13 @@ public class DoubleDQNTest { } }; - DoubleDQN sut = new DoubleDQN(qNetworkMock, targetQNetworkMock, 0.5); + org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.DoubleDQN sut = new org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.DoubleDQN(qNetworkMock, targetQNetworkMock, configuration); // Act - DataSet result = sut.compute(transitions); + FeaturesLabels result = sut.compute(transitions); // Assert - INDArray evaluatedQValues = result.getLabels(); + INDArray evaluatedQValues = result.getLabels(CommonLabelNames.QValues); assertEquals(1.0, evaluatedQValues.getDouble(0, 0), 0.0001); assertEquals(2.2, evaluatedQValues.getDouble(0, 1), 0.0001); } @@ -71,13 +76,13 @@ public class DoubleDQNTest { } }; - DoubleDQN sut = new DoubleDQN(qNetworkMock, targetQNetworkMock, 0.5); + org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.DoubleDQN sut = new org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.DoubleDQN(qNetworkMock, targetQNetworkMock, configuration); // Act - DataSet result = sut.compute(transitions); + FeaturesLabels result = sut.compute(transitions); // Assert - INDArray evaluatedQValues = result.getLabels(); + INDArray evaluatedQValues = result.getLabels(CommonLabelNames.QValues); assertEquals(1.0 + 0.5 * -22.0, evaluatedQValues.getDouble(0, 0), 0.0001); assertEquals(2.2, evaluatedQValues.getDouble(0, 1), 0.0001); } @@ -99,13 +104,13 @@ public class DoubleDQNTest { } }; - DoubleDQN sut = new DoubleDQN(qNetworkMock, targetQNetworkMock, 0.5); + org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.DoubleDQN sut = new DoubleDQN(qNetworkMock, targetQNetworkMock, configuration); // Act - DataSet result = sut.compute(transitions); + FeaturesLabels result = sut.compute(transitions); // Assert - INDArray evaluatedQValues = result.getLabels(); + INDArray evaluatedQValues = result.getLabels(CommonLabelNames.QValues); assertEquals(1.0 + 0.5 * -22.0, evaluatedQValues.getDouble(0, 0), 0.0001); assertEquals(2.2, evaluatedQValues.getDouble(0, 1), 0.0001); diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/StandardDQNTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/StandardDQNTest.java similarity index 73% rename from rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/StandardDQNTest.java rename to rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/StandardDQNTest.java index bc7812b36..588168d45 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/TDTargetAlgorithm/StandardDQNTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/algorithm/dqn/StandardDQNTest.java @@ -1,6 +1,9 @@ -package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete.TDTargetAlgorithm; +package org.deeplearning4j.rl4j.agent.learning.algorithm.dqn; +import org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.StandardDQN; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; import org.deeplearning4j.rl4j.learning.sync.Transition; +import org.deeplearning4j.rl4j.network.CommonLabelNames; import org.deeplearning4j.rl4j.network.IOutputNeuralNet; import org.deeplearning4j.rl4j.observation.Observation; import org.junit.Before; @@ -9,7 +12,6 @@ import org.junit.runner.RunWith; import org.mockito.Mock; import org.mockito.junit.MockitoJUnitRunner; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import org.nd4j.linalg.factory.Nd4j; import java.util.ArrayList; @@ -28,6 +30,9 @@ public class StandardDQNTest { @Mock IOutputNeuralNet targetQNetworkMock; + private final BaseTransitionTDAlgorithm.Configuration configuration = BaseTransitionTDAlgorithm.Configuration.builder() + .gamma(0.5) + .build(); @Before public void setup() { @@ -47,13 +52,13 @@ public class StandardDQNTest { } }; - StandardDQN sut = new StandardDQN(qNetworkMock, targetQNetworkMock, 0.5); + org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.StandardDQN sut = new org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.StandardDQN(qNetworkMock, targetQNetworkMock, configuration); // Act - DataSet result = sut.compute(transitions); + FeaturesLabels result = sut.compute(transitions); // Assert - INDArray evaluatedQValues = result.getLabels(); + INDArray evaluatedQValues = result.getLabels(CommonLabelNames.QValues); assertEquals(1.0, evaluatedQValues.getDouble(0, 0), 0.0001); assertEquals(2.2, evaluatedQValues.getDouble(0, 1), 0.0001); } @@ -69,13 +74,13 @@ public class StandardDQNTest { } }; - StandardDQN sut = new StandardDQN(qNetworkMock, targetQNetworkMock, 0.5); + org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.StandardDQN sut = new org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.StandardDQN(qNetworkMock, targetQNetworkMock, configuration); // Act - DataSet result = sut.compute(transitions); + FeaturesLabels result = sut.compute(transitions); // Assert - INDArray evaluatedQValues = result.getLabels(); + INDArray evaluatedQValues = result.getLabels(CommonLabelNames.QValues); assertEquals(1.0 + 0.5 * 22.0, evaluatedQValues.getDouble(0, 0), 0.0001); assertEquals(2.2, evaluatedQValues.getDouble(0, 1), 0.0001); } @@ -95,13 +100,13 @@ public class StandardDQNTest { } }; - StandardDQN sut = new StandardDQN(qNetworkMock, targetQNetworkMock, 0.5); + org.deeplearning4j.rl4j.agent.learning.algorithm.dqn.StandardDQN sut = new StandardDQN(qNetworkMock, targetQNetworkMock, configuration); // Act - DataSet result = sut.compute(transitions); + FeaturesLabels result = sut.compute(transitions); // Assert - INDArray evaluatedQValues = result.getLabels(); + INDArray evaluatedQValues = result.getLabels(CommonLabelNames.QValues); assertEquals((1.0 + 0.5 * 22.0), evaluatedQValues.getDouble(0, 0), 0.0001); assertEquals(2.2, evaluatedQValues.getDouble(0, 1), 0.0001); diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/LearningBehaviorTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/behavior/LearningBehaviorTest.java similarity index 93% rename from rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/LearningBehaviorTest.java rename to rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/behavior/LearningBehaviorTest.java index 1e39c63d5..07e34bfd2 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/LearningBehaviorTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/behavior/LearningBehaviorTest.java @@ -1,6 +1,7 @@ -package org.deeplearning4j.rl4j.agent.learning; +package org.deeplearning4j.rl4j.agent.learning.behavior; -import org.deeplearning4j.rl4j.agent.update.IUpdateRule; +import org.deeplearning4j.rl4j.agent.learning.behavior.LearningBehavior; +import org.deeplearning4j.rl4j.agent.learning.update.IUpdateRule; import org.deeplearning4j.rl4j.experience.ExperienceHandler; import org.deeplearning4j.rl4j.observation.Observation; import org.junit.Before; @@ -17,7 +18,6 @@ import java.util.List; import static org.junit.Assert.assertEquals; import static org.junit.Assert.assertFalse; -import static org.mockito.ArgumentMatchers.any; import static org.mockito.Mockito.*; @RunWith(MockitoJUnitRunner.class) diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabelsTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabelsTest.java new file mode 100644 index 000000000..4f9df3595 --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/FeaturesLabelsTest.java @@ -0,0 +1,38 @@ +package org.deeplearning4j.rl4j.agent.learning.update; + +import org.junit.Test; +import org.nd4j.linalg.api.ndarray.INDArray; +import org.nd4j.linalg.factory.Nd4j; + +import static org.junit.Assert.assertEquals; + +public class FeaturesLabelsTest { + + @Test + public void when_getBatchSizeIsCalled_expect_batchSizeIsReturned() { + // Arrange + INDArray features = Nd4j.create(5, 10); + FeaturesLabels sut = new FeaturesLabels(features); + + // Act + long batchSize = sut.getBatchSize(); + + // Assert + assertEquals(5, batchSize); + } + + @Test + public void when_puttingLabels_expect_getLabelReturnsLabels() { + // Arrange + INDArray features = Nd4j.create(5, 10); + INDArray labels = Nd4j.rand(2, 3); + FeaturesLabels sut = new FeaturesLabels(features); + sut.putLabels("test", labels); + + // Act + INDArray result = sut.getLabels("test"); + + // Assert + assertEquals(result, labels); + } +} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/GradientsTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/GradientsTest.java new file mode 100644 index 000000000..68eb86c0b --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/GradientsTest.java @@ -0,0 +1,40 @@ +package org.deeplearning4j.rl4j.agent.learning.update; + +import org.deeplearning4j.nn.gradient.Gradient; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.mockito.junit.MockitoJUnitRunner; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertSame; +import static org.mockito.Mockito.mock; + +@RunWith(MockitoJUnitRunner.class) +public class GradientsTest { + + @Test + public void when_getBatchSizeIsCalled_expect_batchSizeIsReturned() { + // Arrange + Gradients sut = new Gradients(5); + + // Act + long batchSize = sut.getBatchSize(); + + // Assert + assertEquals(5, batchSize); + } + + @Test + public void when_puttingLabels_expect_getLabelReturnsLabels() { + // Arrange + Gradient gradient = mock(Gradient.class); + Gradients sut = new Gradients(5); + sut.putGradient("test", gradient); + + // Act + Gradient result = sut.getGradient("test"); + + // Assert + assertSame(gradient, result); + } +} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRuleTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRuleTest.java new file mode 100644 index 000000000..af22911a0 --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/UpdateRuleTest.java @@ -0,0 +1,67 @@ +package org.deeplearning4j.rl4j.agent.learning.update; + +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.update.updater.INeuralNetUpdater; +import org.junit.Before; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.mockito.Mock; +import org.mockito.junit.MockitoJUnitRunner; + +import java.util.ArrayList; +import java.util.List; + +import static org.junit.Assert.assertEquals; +import static org.mockito.Mockito.*; + +@RunWith(MockitoJUnitRunner.class) +public class UpdateRuleTest { + + @Mock + private IUpdateAlgorithm updateAlgorithm; + + @Mock + private INeuralNetUpdater updater; + + private UpdateRule sut; + + @Before + public void init() { + sut = new UpdateRule(updateAlgorithm, updater); + } + + @Test + public void when_callingUpdate_expect_computeAndUpdateNetwork() { + // Arrange + List trainingBatch = new ArrayList() { + { + Integer.valueOf(1); + Integer.valueOf(2); + } + }; + final FeaturesLabels computeResult = new FeaturesLabels(null); + when(updateAlgorithm.compute(any())).thenReturn(computeResult); + + // Act + sut.update(trainingBatch); + + // Assert + verify(updateAlgorithm, times(1)).compute(trainingBatch); + verify(updater, times(1)).update(computeResult); + } + + @Test + public void when_callingUpdate_expect_updateCountIncremented() { + // Arrange + + // Act + sut.update(null); + int updateCountBefore = sut.getUpdateCount(); + sut.update(null); + int updateCountAfter = sut.getUpdateCount(); + + // Assert + assertEquals(updateCountBefore + 1, updateCountAfter); + } + +} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/GradientsNeuralNetUpdaterTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/GradientsNeuralNetUpdaterTest.java new file mode 100644 index 000000000..6df65782b --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/GradientsNeuralNetUpdaterTest.java @@ -0,0 +1,56 @@ +package org.deeplearning4j.rl4j.agent.learning.update.updater; + +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.mockito.Mock; +import org.mockito.junit.MockitoJUnitRunner; + +import static org.mockito.Mockito.*; + +@RunWith(MockitoJUnitRunner.class) +public class GradientsNeuralNetUpdaterTest { + + @Mock + ITrainableNeuralNet currentMock; + + @Mock + ITrainableNeuralNet targetMock; + + @Test + public void when_callingUpdate_expect_currentUpdatedAndtargetNotChanged() { + // Arrange + GradientsNeuralNetUpdater.Configuration configuration = GradientsNeuralNetUpdater.Configuration.builder() + .build(); + GradientsNeuralNetUpdater sut = new GradientsNeuralNetUpdater(currentMock, targetMock, configuration); + Gradients gradients = new Gradients(10); + + // Act + sut.update(gradients); + + // Assert + verify(currentMock, times(1)).applyGradients(gradients); + verify(targetMock, never()).applyGradients(any()); + } + + @Test + public void when_callingUpdate_expect_targetUpdatedFromCurrentAtFrequency() { + // Arrange + GradientsNeuralNetUpdater.Configuration configuration = GradientsNeuralNetUpdater.Configuration.builder() + .targetUpdateFrequency(3) + .build(); + GradientsNeuralNetUpdater sut = new GradientsNeuralNetUpdater(currentMock, targetMock, configuration); + Gradients gradients = new Gradients(10); + + // Act + sut.update(gradients); + sut.update(gradients); + sut.update(gradients); + + // Assert + verify(currentMock, never()).copy(any()); + verify(targetMock, times(1)).copy(currentMock); + } + +} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdaterTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdaterTest.java new file mode 100644 index 000000000..0dcfa3f4f --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/learning/update/updater/LabelsNeuralNetUpdaterTest.java @@ -0,0 +1,77 @@ +package org.deeplearning4j.rl4j.agent.learning.update.updater; + +import org.deeplearning4j.rl4j.agent.Agent; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.mockito.Mock; +import org.mockito.junit.MockitoJUnitRunner; + +import static org.junit.Assert.assertTrue; +import static org.junit.Assert.fail; +import static org.mockito.Mockito.*; + +@RunWith(MockitoJUnitRunner.class) +public class LabelsNeuralNetUpdaterTest { + + @Mock + ITrainableNeuralNet currentMock; + + @Mock + ITrainableNeuralNet targetMock; + + @Test + public void when_callingUpdateWithTargetUpdateFrequencyAt0_expect_Exception() { + // Arrange + LabelsNeuralNetUpdater.Configuration configuration = LabelsNeuralNetUpdater.Configuration.builder() + .targetUpdateFrequency(0) + .build(); + try { + LabelsNeuralNetUpdater sut = new LabelsNeuralNetUpdater(currentMock, targetMock, configuration); + fail("IllegalArgumentException should have been thrown"); + } catch (IllegalArgumentException exception) { + String expectedMessage = "Configuration: targetUpdateFrequency must be greater than 0, got: [0]"; + String actualMessage = exception.getMessage(); + + assertTrue(actualMessage.contains(expectedMessage)); + } + + } + + @Test + public void when_callingUpdate_expect_currentUpdatedAndTargetNotChanged() { + // Arrange + LabelsNeuralNetUpdater.Configuration configuration = LabelsNeuralNetUpdater.Configuration.builder() + .build(); + LabelsNeuralNetUpdater sut = new LabelsNeuralNetUpdater(currentMock, targetMock, configuration); + FeaturesLabels featureLabels = new FeaturesLabels(null); + + // Act + sut.update(featureLabels); + + // Assert + verify(currentMock, times(1)).fit(featureLabels); + verify(targetMock, never()).fit(any()); + } + + @Test + public void when_callingUpdate_expect_targetUpdatedFromCurrentAtFrequency() { + // Arrange + LabelsNeuralNetUpdater.Configuration configuration = LabelsNeuralNetUpdater.Configuration.builder() + .targetUpdateFrequency(3) + .build(); + LabelsNeuralNetUpdater sut = new LabelsNeuralNetUpdater(currentMock, targetMock, configuration); + FeaturesLabels featureLabels = new FeaturesLabels(null); + + // Act + sut.update(featureLabels); + sut.update(featureLabels); + sut.update(featureLabels); + + // Assert + verify(currentMock, never()).copy(any()); + verify(targetMock, times(1)).copy(currentMock); + } + +} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/NeuralNetUpdaterTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/NeuralNetUpdaterTest.java deleted file mode 100644 index 578da2d50..000000000 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/agent/update/neuralnetupdater/NeuralNetUpdaterTest.java +++ /dev/null @@ -1,51 +0,0 @@ -package org.deeplearning4j.rl4j.agent.update.neuralnetupdater; - -import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; -import org.junit.Test; -import org.junit.runner.RunWith; -import org.mockito.Mock; -import org.mockito.junit.MockitoJUnitRunner; -import org.nd4j.linalg.dataset.api.DataSet; - -import static org.mockito.Mockito.*; - -@RunWith(MockitoJUnitRunner.class) -public class NeuralNetUpdaterTest { - - @Mock - ITrainableNeuralNet currentMock; - - @Mock - ITrainableNeuralNet targetMock; - - @Test - public void when_callingUpdate_expect_currentUpdatedAndtargetNotChanged() { - // Arrange - NeuralNetUpdater sut = new NeuralNetUpdater(currentMock, targetMock, Integer.MAX_VALUE); - DataSet featureLabels = new org.nd4j.linalg.dataset.DataSet(); - - // Act - sut.update(featureLabels); - - // Assert - verify(currentMock, times(1)).fit(featureLabels); - verify(targetMock, never()).fit(any()); - } - - @Test - public void when_callingUpdate_expect_targetUpdatedFromCurrentAtFrequency() { - // Arrange - NeuralNetUpdater sut = new NeuralNetUpdater(currentMock, targetMock, 3); - DataSet featureLabels = new org.nd4j.linalg.dataset.DataSet(); - - // Act - sut.update(featureLabels); - sut.update(featureLabels); - sut.update(featureLabels); - - // Assert - verify(currentMock, never()).copy(any()); - verify(targetMock, times(1)).copy(currentMock); - } - -} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilderTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilderTest.java new file mode 100644 index 000000000..01a8908e3 --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/builder/BaseAgentLearnerBuilderTest.java @@ -0,0 +1,92 @@ +package org.deeplearning4j.rl4j.builder; + +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.AgentLearner; +import org.deeplearning4j.rl4j.agent.learning.algorithm.IUpdateAlgorithm; +import org.deeplearning4j.rl4j.agent.learning.update.updater.INeuralNetUpdater; +import org.deeplearning4j.rl4j.environment.Environment; +import org.deeplearning4j.rl4j.experience.ExperienceHandler; +import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; +import org.deeplearning4j.rl4j.observation.transform.TransformProcess; +import org.deeplearning4j.rl4j.policy.IPolicy; +import org.junit.Before; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.mockito.Mock; +import org.mockito.Mockito; +import org.mockito.junit.MockitoJUnitRunner; + +import static org.mockito.Mockito.*; + +@RunWith(MockitoJUnitRunner.class) +public class BaseAgentLearnerBuilderTest { + @Mock + BaseAgentLearnerBuilder.Configuration configuration; + + @Mock + ITrainableNeuralNet neuralNet; + + @Mock + Builder> environmentBuilder; + + @Mock + Builder transformProcessBuilder; + + @Mock + IUpdateAlgorithm updateAlgorithmMock; + + @Mock + INeuralNetUpdater neuralNetUpdaterMock; + + @Mock + ExperienceHandler experienceHandlerMock; + + @Mock + Environment environmentMock; + + @Mock + IPolicy policyMock; + + @Mock + TransformProcess transformProcessMock; + + BaseAgentLearnerBuilder sut; + + @Before + public void setup() { + sut = mock( + BaseAgentLearnerBuilder.class, + Mockito.withSettings() + .useConstructor(configuration, neuralNet, environmentBuilder, transformProcessBuilder) + .defaultAnswer(Mockito.CALLS_REAL_METHODS) + ); + + AgentLearner.Configuration agentLearnerConfiguration = AgentLearner.Configuration.builder().maxEpisodeSteps(200).build(); + + when(sut.buildUpdateAlgorithm()).thenReturn(updateAlgorithmMock); + when(sut.buildNeuralNetUpdater()).thenReturn(neuralNetUpdaterMock); + when(sut.buildExperienceHandler()).thenReturn(experienceHandlerMock); + when(environmentBuilder.build()).thenReturn(environmentMock); + when(transformProcessBuilder.build()).thenReturn(transformProcessMock); + when(sut.buildPolicy()).thenReturn(policyMock); + when(configuration.getAgentLearnerConfiguration()).thenReturn(agentLearnerConfiguration); + } + + @Test + public void when_buildingAgentLearner_expect_dependenciesAndAgentLearnerIsBuilt() { + // Arrange + + // Act + sut.build(); + + // Assert + verify(environmentBuilder, times(1)).build(); + verify(transformProcessBuilder, times(1)).build(); + verify(sut, times(1)).buildPolicy(); + verify(sut, times(1)).buildExperienceHandler(); + verify(sut, times(1)).buildUpdateAlgorithm(); + verify(sut, times(1)).buildNeuralNetUpdater(); + verify(sut, times(1)).buildAgentLearner(); + } + +} diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandlerTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandlerTest.java index 0d90e812d..44c2960d2 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandlerTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/ReplayMemoryExperienceHandlerTest.java @@ -21,6 +21,13 @@ public class ReplayMemoryExperienceHandlerTest { @Mock IExpReplay expReplayMock; + private ReplayMemoryExperienceHandler.Configuration buildConfiguration() { + return ReplayMemoryExperienceHandler.Configuration.builder() + .maxReplayMemorySize(10) + .batchSize(5) + .build(); + } + @Test public void when_addingFirstExperience_expect_notAddedToStoreBeforeNextObservationIsAdded() { // Arrange @@ -87,7 +94,7 @@ public class ReplayMemoryExperienceHandlerTest { @Test public void when_addingExperience_expect_getTrainingBatchSizeReturnSize() { // Arrange - ReplayMemoryExperienceHandler sut = new ReplayMemoryExperienceHandler(10, 5, Nd4j.getRandom()); + ReplayMemoryExperienceHandler sut = new ReplayMemoryExperienceHandler(buildConfiguration(), Nd4j.getRandom()); sut.addExperience(new Observation(Nd4j.create(new double[] { 1.0 })), 1, 1.0, false); sut.addExperience(new Observation(Nd4j.create(new double[] { 2.0 })), 2, 2.0, false); sut.setFinalObservation(new Observation(Nd4j.create(new double[] { 3.0 }))); @@ -102,7 +109,7 @@ public class ReplayMemoryExperienceHandlerTest { @Test public void when_experienceSizeIsSmallerThanBatchSize_expect_TrainingBatchIsNotReady() { // Arrange - ReplayMemoryExperienceHandler sut = new ReplayMemoryExperienceHandler(10, 5, Nd4j.getRandom()); + ReplayMemoryExperienceHandler sut = new ReplayMemoryExperienceHandler(buildConfiguration(), Nd4j.getRandom()); sut.addExperience(new Observation(Nd4j.create(new double[] { 1.0 })), 1, 1.0, false); sut.addExperience(new Observation(Nd4j.create(new double[] { 2.0 })), 2, 2.0, false); sut.setFinalObservation(new Observation(Nd4j.create(new double[] { 3.0 }))); @@ -116,7 +123,7 @@ public class ReplayMemoryExperienceHandlerTest { @Test public void when_experienceSizeIsGreaterOrEqualToBatchSize_expect_TrainingBatchIsReady() { // Arrange - ReplayMemoryExperienceHandler sut = new ReplayMemoryExperienceHandler(10, 5, Nd4j.getRandom()); + ReplayMemoryExperienceHandler sut = new ReplayMemoryExperienceHandler(buildConfiguration(), Nd4j.getRandom()); sut.addExperience(new Observation(Nd4j.create(new double[] { 1.0 })), 1, 1.0, false); sut.addExperience(new Observation(Nd4j.create(new double[] { 2.0 })), 2, 2.0, false); sut.addExperience(new Observation(Nd4j.create(new double[] { 3.0 })), 3, 3.0, false); diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandlerTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandlerTest.java index 2ce0d6659..6cdcd8620 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandlerTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/experience/StateActionExperienceHandlerTest.java @@ -10,10 +10,16 @@ import static org.junit.Assert.*; public class StateActionExperienceHandlerTest { + private StateActionExperienceHandler.Configuration buildConfiguration(int batchSize) { + return StateActionExperienceHandler.Configuration.builder() + .batchSize(batchSize) + .build(); + } + @Test public void when_addingExperience_expect_generateTrainingBatchReturnsIt() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(Integer.MAX_VALUE); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(Integer.MAX_VALUE)); sut.reset(); Observation observation = new Observation(Nd4j.zeros(1)); sut.addExperience(observation, 123, 234.0, true); @@ -32,7 +38,7 @@ public class StateActionExperienceHandlerTest { @Test public void when_addingMultipleExperiences_expect_generateTrainingBatchReturnsItInSameOrder() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(Integer.MAX_VALUE); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(Integer.MAX_VALUE)); sut.reset(); sut.addExperience(null, 1, 1.0, false); sut.addExperience(null, 2, 2.0, false); @@ -51,7 +57,7 @@ public class StateActionExperienceHandlerTest { @Test public void when_gettingExperience_expect_experienceStoreIsCleared() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(Integer.MAX_VALUE); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(Integer.MAX_VALUE)); sut.reset(); sut.addExperience(null, 1, 1.0, false); @@ -67,7 +73,7 @@ public class StateActionExperienceHandlerTest { @Test public void when_addingExperience_expect_getTrainingBatchSizeReturnSize() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(Integer.MAX_VALUE); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(Integer.MAX_VALUE)); sut.reset(); sut.addExperience(null, 1, 1.0, false); sut.addExperience(null, 2, 2.0, false); @@ -83,7 +89,7 @@ public class StateActionExperienceHandlerTest { @Test public void when_experienceIsEmpty_expect_TrainingBatchNotReady() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(5); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(5)); sut.reset(); // Act @@ -96,7 +102,7 @@ public class StateActionExperienceHandlerTest { @Test public void when_experienceSizeIsGreaterOrEqualToThanBatchSize_expect_TrainingBatchIsReady() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(5); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(5)); sut.reset(); sut.addExperience(null, 1, 1.0, false); sut.addExperience(null, 2, 2.0, false); @@ -114,7 +120,7 @@ public class StateActionExperienceHandlerTest { @Test public void when_experienceSizeIsSmallerThanBatchSizeButFinalObservationIsSet_expect_TrainingBatchIsReady() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(5); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(5)); sut.reset(); sut.addExperience(null, 1, 1.0, false); sut.addExperience(null, 2, 2.0, false); @@ -130,7 +136,7 @@ public class StateActionExperienceHandlerTest { @Test public void when_experienceSizeIsZeroAndFinalObservationIsSet_expect_TrainingBatchIsNotReady() { // Arrange - StateActionExperienceHandler sut = new StateActionExperienceHandler(5); + StateActionExperienceHandler sut = new StateActionExperienceHandler(buildConfiguration(5)); sut.reset(); sut.setFinalObservation(null); diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/QLearningDiscreteTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/QLearningDiscreteTest.java index e1424c286..90494961f 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/QLearningDiscreteTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/qlearning/discrete/QLearningDiscreteTest.java @@ -18,7 +18,7 @@ package org.deeplearning4j.rl4j.learning.sync.qlearning.discrete; import org.deeplearning4j.gym.StepReply; -import org.deeplearning4j.rl4j.agent.learning.ILearningBehavior; +import org.deeplearning4j.rl4j.agent.learning.behavior.ILearningBehavior; import org.deeplearning4j.rl4j.learning.IHistoryProcessor; import org.deeplearning4j.rl4j.learning.configuration.QLearningConfiguration; import org.deeplearning4j.rl4j.learning.sync.qlearning.QLearning; @@ -26,7 +26,6 @@ import org.deeplearning4j.rl4j.mdp.MDP; import org.deeplearning4j.rl4j.network.dqn.IDQN; import org.deeplearning4j.rl4j.space.Encodable; import org.deeplearning4j.rl4j.observation.Observation; -import org.deeplearning4j.rl4j.space.Box; import org.deeplearning4j.rl4j.space.DiscreteSpace; import org.deeplearning4j.rl4j.space.ObservationSpace; import org.junit.Before; @@ -101,6 +100,8 @@ public class QLearningDiscreteTest { when(mockQlearningConfiguration.getRewardFactor()).thenReturn(rewardFactor); when(mockQlearningConfiguration.getExpRepMaxSize()).thenReturn(maxExperienceReplay); when(mockQlearningConfiguration.getSeed()).thenReturn(123L); + when(mockQlearningConfiguration.getTargetDqnUpdateFreq()).thenReturn(1); + when(mockDQN.clone()).thenReturn(mockDQN); if(learningBehavior != null) { qLearningDiscrete = mock( diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/support/MockDQN.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/support/MockDQN.java index 7fb3f8300..6970aec21 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/support/MockDQN.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/learning/sync/support/MockDQN.java @@ -2,12 +2,12 @@ package org.deeplearning4j.rl4j.learning.sync.support; import org.deeplearning4j.nn.api.NeuralNetwork; import org.deeplearning4j.nn.gradient.Gradient; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; -import org.deeplearning4j.rl4j.network.NeuralNet; import org.deeplearning4j.rl4j.network.dqn.IDQN; import org.deeplearning4j.rl4j.observation.Observation; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import java.io.IOException; import java.io.OutputStream; @@ -69,7 +69,17 @@ public class MockDQN implements IDQN { } @Override - public void fit(DataSet featuresLabels) { + public void fit(FeaturesLabels featuresLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public Gradients computeGradients(FeaturesLabels updateLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public void applyGradients(Gradients gradients) { throw new UnsupportedOperationException(); } diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/policy/PolicyTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/policy/PolicyTest.java index b05e55a19..b5ef4865b 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/policy/PolicyTest.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/policy/PolicyTest.java @@ -23,6 +23,8 @@ import org.deeplearning4j.nn.conf.layers.OutputLayer; import org.deeplearning4j.nn.gradient.Gradient; import org.deeplearning4j.nn.graph.ComputationGraph; import org.deeplearning4j.nn.multilayer.MultiLayerNetwork; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; import org.deeplearning4j.rl4j.learning.IHistoryProcessor; import org.deeplearning4j.rl4j.learning.Learning; import org.deeplearning4j.rl4j.learning.configuration.QLearningConfiguration; @@ -36,16 +38,13 @@ import org.deeplearning4j.rl4j.util.LegacyMDPWrapper; import org.junit.Test; import org.nd4j.linalg.activations.Activation; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import org.nd4j.linalg.factory.Nd4j; import org.nd4j.linalg.lossfunctions.LossFunctions; import java.io.IOException; import java.io.OutputStream; -import static org.junit.Assert.assertEquals; -import static org.junit.Assert.assertNotNull; -import static org.junit.Assert.assertTrue; +import static org.junit.Assert.*; /** * @@ -90,7 +89,17 @@ public class PolicyTest { } @Override - public void fit(DataSet featuresLabels) { + public void fit(FeaturesLabels featuresLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public Gradients computeGradients(FeaturesLabels updateLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public void applyGradients(Gradients gradients) { throw new UnsupportedOperationException(); } diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockDQN.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockDQN.java index 1b9c63b65..442b94f4d 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockDQN.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockDQN.java @@ -2,13 +2,13 @@ package org.deeplearning4j.rl4j.support; import org.deeplearning4j.nn.api.NeuralNetwork; import org.deeplearning4j.nn.gradient.Gradient; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; -import org.deeplearning4j.rl4j.network.NeuralNet; import org.deeplearning4j.rl4j.network.dqn.IDQN; import org.deeplearning4j.rl4j.observation.Observation; -import org.nd4j.linalg.api.ndarray.INDArray; import org.nd4j.common.primitives.Pair; -import org.nd4j.linalg.dataset.api.DataSet; +import org.nd4j.linalg.api.ndarray.INDArray; import java.io.IOException; import java.io.OutputStream; @@ -66,7 +66,17 @@ public class MockDQN implements IDQN { } @Override - public void fit(DataSet featuresLabels) { + public void fit(FeaturesLabels featuresLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public Gradients computeGradients(FeaturesLabels updateLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public void applyGradients(Gradients gradients) { throw new UnsupportedOperationException(); } diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockNeuralNet.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockNeuralNet.java index fe5b3a7f9..36480c9f6 100644 --- a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockNeuralNet.java +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/support/MockNeuralNet.java @@ -2,11 +2,12 @@ package org.deeplearning4j.rl4j.support; import org.deeplearning4j.nn.api.NeuralNetwork; import org.deeplearning4j.nn.gradient.Gradient; +import org.deeplearning4j.rl4j.agent.learning.update.FeaturesLabels; +import org.deeplearning4j.rl4j.agent.learning.update.Gradients; import org.deeplearning4j.rl4j.network.ITrainableNeuralNet; import org.deeplearning4j.rl4j.network.NeuralNet; import org.deeplearning4j.rl4j.observation.Observation; import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.DataSet; import org.nd4j.linalg.factory.Nd4j; import java.io.IOException; @@ -42,7 +43,17 @@ public class MockNeuralNet implements NeuralNet { } @Override - public void fit(DataSet featuresLabels) { + public void fit(FeaturesLabels featuresLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public Gradients computeGradients(FeaturesLabels updateLabels) { + throw new UnsupportedOperationException(); + } + + @Override + public void applyGradients(Gradients gradients) { throw new UnsupportedOperationException(); } diff --git a/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/trainer/SyncTrainerTest.java b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/trainer/SyncTrainerTest.java new file mode 100644 index 000000000..fd99ba76d --- /dev/null +++ b/rl4j/rl4j-core/src/test/java/org/deeplearning4j/rl4j/trainer/SyncTrainerTest.java @@ -0,0 +1,59 @@ +package org.deeplearning4j.rl4j.trainer; + +import org.apache.commons.lang3.builder.Builder; +import org.deeplearning4j.rl4j.agent.IAgentLearner; +import org.junit.Before; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.mockito.Mock; +import org.mockito.junit.MockitoJUnitRunner; + +import java.util.function.Predicate; + +import static org.junit.Assert.assertEquals; +import static org.mockito.Mockito.*; + +@RunWith(MockitoJUnitRunner.class) +public class SyncTrainerTest { + + @Mock + IAgentLearner agentLearnerMock; + + @Mock + Builder agentLearnerBuilder; + + SyncTrainer sut; + + public void setup(Predicate stoppingCondition) { + when(agentLearnerBuilder.build()).thenReturn(agentLearnerMock); + + sut = new SyncTrainer(agentLearnerBuilder, stoppingCondition); + } + + @Test + public void when_training_expect_stoppingConditionWillStopTraining() { + // Arrange + Predicate stoppingCondition = t -> t.getEpisodeCount() >= 5; // Stop after 5 episodes + setup(stoppingCondition); + + // Act + sut.train(); + + // Assert + assertEquals(5, sut.getEpisodeCount()); + } + + @Test + public void when_training_expect_agentIsRun() { + // Arrange + Predicate stoppingCondition = t -> t.getEpisodeCount() >= 5; // Stop after 5 episodes + setup(stoppingCondition); + + // Act + sut.train(); + + // Assert + verify(agentLearnerMock, times(5)).run(); + } + +} diff --git a/rl4j/rl4j-doom/pom.xml b/rl4j/rl4j-doom/pom.xml index 76a31b4a8..31062c7e3 100644 --- a/rl4j/rl4j-doom/pom.xml +++ b/rl4j/rl4j-doom/pom.xml @@ -45,7 +45,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/rl4j/rl4j-gym/pom.xml b/rl4j/rl4j-gym/pom.xml index 76e2e39ff..a27991338 100644 --- a/rl4j/rl4j-gym/pom.xml +++ b/rl4j/rl4j-gym/pom.xml @@ -50,7 +50,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/rl4j/rl4j-malmo/pom.xml b/rl4j/rl4j-malmo/pom.xml index 3575da7a8..e9aa6f500 100644 --- a/rl4j/rl4j-malmo/pom.xml +++ b/rl4j/rl4j-malmo/pom.xml @@ -55,7 +55,7 @@ test-nd4j-native - test-nd4j-cuda-10.2 + test-nd4j-cuda-11.0 diff --git a/scalnet/pom.xml b/scalnet/pom.xml index 39408dbf5..951e2ed25 100644 --- a/scalnet/pom.xml +++ b/scalnet/pom.xml @@ -293,7 +293,7 @@ org.nd4j - nd4j-cuda-10.2 + nd4j-cuda-11.0 ${project.version} test