* Shugeo strided slice zeros (#14) * Modified strided_slice op to properly work with empty-like shapes. * Fixed test for reduce_mean with empty-like input. * [WIP] Last merge (#15) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks * [WIP] Fixing outstanding issues for NLP (#9) * Avoid using not-inited objects * Test fixed. * Redundant method avoided for models like FastText * KMeans++ implementation * KMeans++ implementation * Disable parallel execution * KMeans++ * Tests * Dev branch merge (#16) * SameDiff: convertDataType and gradient check util improvements (#12) * GradCheck util improvements * StopGradient constructor + test * SameDiff: Add datatype conversion * Javadoc and add DataType.isNumerical() * Small fix * Fix SameDiff TF import test cases intermediate naming (workaround for bad default) * TFGraphTestAllHelper: check intermediates in execution order * Add missing debug listener * [WIP] lstmBlock fix + other changes (#13) - fixes lstmBlock issue - changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer - CheckNumerics op - fixes for ReduceBool IsInfOrNan & IsFinite * Small test fix * CheckNumerics op wrapper * Fix some issues on master (#17) * Fix DataVec test issue * Fix issue with dl4j SameDiff output layer * Dtype fix for lambda layers * #7912 BertIterator dtype fix (use float32 not global default) * [WIP] Next set of CUDA stuff (#7) New CUDA implementations and improvements * bad file * Dev branch master merge (#23) * SameDiff: convertDataType and gradient check util improvements (#12) * GradCheck util improvements * StopGradient constructor + test * SameDiff: Add datatype conversion * Javadoc and add DataType.isNumerical() * Small fix * Fix SameDiff TF import test cases intermediate naming (workaround for bad default) * TFGraphTestAllHelper: check intermediates in execution order * Add missing debug listener * [WIP] lstmBlock fix + other changes (#13) - fixes lstmBlock issue - changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer - CheckNumerics op - fixes for ReduceBool IsInfOrNan & IsFinite * Small test fix * CheckNumerics op wrapper * Compatibility of deserialization (#18) Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com> * SameDiff: add activation gradient checking support for debugging (#19) * SameDiff gradient checker: first pass on activation gradient checks * Fixes + tests for activation gradient checking * Javadoc * [WIP] Some nd4j data type corrections (#20) * Adjust data type * Set correct Data type. * Size of proper data type. * fix averaged cpu load (#22) * SameDiff ops, TF import and fixes (#24) * CheckNumerics tests + fixes + misc fixes Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fake quant Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fixes Signed-off-by: AlexDBlack <blacka101@gmail.com> * FakeQuantWithMinMaxArgs Signed-off-by: AlexDBlack <blacka101@gmail.com> * CheckNumerics fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix libnd4j ALL_INTS and ALL_FLOATS declaration (uint and bfloat types) Signed-off-by: AlexDBlack <blacka101@gmail.com> * Small fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * Javadoc Signed-off-by: AlexDBlack <blacka101@gmail.com> * Exception tweak Signed-off-by: AlexDBlack <blacka101@gmail.com> * fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix for out of scope stack allocated var use Signed-off-by: AlexDBlack <blacka101@gmail.com> * Ignores Signed-off-by: AlexDBlack <blacka101@gmail.com> * Ignore for known failing test (already logged issue) Signed-off-by: AlexDBlack <blacka101@gmail.com> * Merge upstream to fork (#25) * Add thousand-separator commas to TotalParams (#7915) * Add thousand-separator commas to TotalParams The number of parameters can be quite large, and it would help the reading of the summary printout to have the TotalParams column & values at the bottom have thousand-separator-commas in them. * Add thousand-separator commas to MultiLayerNetwork Corresponding change to MultiLayerNetwork Signed-off-by: Jxtps Jxtps <jxtps435@gmail.com> * Update contributing and issue/PR templates (#7934) Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix link to AdaDelta paper (#7942) Fix link to AdaDelta paper hosted on matthewzeiler.com Signed-off-by: Jxtps * Fixes, and ignores for known/logged failing issues (#7943) Signed-off-by: AlexDBlack <blacka101@gmail.com> * SameDiff + DL4J/SameDiff: Multiple fixes (#28) * #7919 HDF5 attribute buffer length fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7909 Arbiter constructor exception ux improvements Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7925 RNN output layer length checks Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7939 Add listener for validating inputs are not incorrectly modified Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7939 Integrate NonInplaceValidationListener into tests * #7844 DL4J SameDiff fixes for variable minibatch size * DL4J SameDiff fixes - ensure gradient for input placeholder is available Signed-off-by: AlexDBlack <blacka101@gmail.com> * Tweaks to ExternalErrorsFunction - use placeholders, make more robust * Another fix * More fixes * More SameDiff/DL4J fixes * Scope out scalar array creation in BaseScalarOp * Remove debug code Signed-off-by: AlexDBlack <blacka101@gmail.com> * [WIP] Final dev branch merge (#29) * SameDiff: convertDataType and gradient check util improvements (#12) * GradCheck util improvements * StopGradient constructor + test * SameDiff: Add datatype conversion * Javadoc and add DataType.isNumerical() * Small fix * Fix SameDiff TF import test cases intermediate naming (workaround for bad default) * TFGraphTestAllHelper: check intermediates in execution order * Add missing debug listener * [WIP] lstmBlock fix + other changes (#13) - fixes lstmBlock issue - changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer - CheckNumerics op - fixes for ReduceBool IsInfOrNan & IsFinite * Small test fix * CheckNumerics op wrapper * Compatibility of deserialization (#18) Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com> * SameDiff: add activation gradient checking support for debugging (#19) * SameDiff gradient checker: first pass on activation gradient checks * Fixes + tests for activation gradient checking * Javadoc * [WIP] Some nd4j data type corrections (#20) * Adjust data type * Set correct Data type. * Size of proper data type. * fix averaged cpu load (#22) * [WIP] Multiple dataset iterators (#27) * Splitting dataset into arbitrary number * Fixes * Multiple split of iterator * Test * Test * Some fixes * signature change * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * one more test for sequential use of DataSetIteratorSplitter Signed-off-by: raver119 <raver119@gmail.com> * Fixes * Fixes * one more test for Alexander Signed-off-by: raver119 <raver119@gmail.com> * Some fixes * Some fixes * one more test for Alexander Signed-off-by: raver119 <raver119@gmail.com> * minor test fix Signed-off-by: raver119 <raver119@gmail.com> * Some fixes * Some fixes * couple of assertions tweaked Signed-off-by: raver119 <raver119@gmail.com> * MDS splitter test :/ Signed-off-by: raver119 <raver119@gmail.com> * Minor refactoring * Multi dataset * Some fixes * More tests * Small number of test fixes/improvements (failures on CI) (#31) Signed-off-by: AlexDBlack <blacka101@gmail.com> * [WIP] More CUDA stuff (#26) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * LRN BP CUDA Signed-off-by: raver119 <raver119@gmail.com> * less memory Signed-off-by: raver119 <raver119@gmail.com> * Fixed bug with crop_and_resize op helper. * get rid of unnecessary index-calculation dunction Signed-off-by: Yurii <yurii@skymind.io> * Fixed sort with nth_element cuda-based helper. * Refactored nth_element. * Refactored nth_element op and tests. * Modified usage of dim array with sortTad routine. * Refactored main routine of helper for non_max_image_suppression op. * non_max_image_suppression op helper with cuda kernel implementation. Initial revision. * fix vol2col cuda kernel * meh Signed-off-by: raver119 <raver119@gmail.com> * topK concept Signed-off-by: raver119 <raver119@gmail.com> * unsorted topK with scanWitdh of 1 Signed-off-by: raver119 <raver119@gmail.com> * correct vol2col tests * sorted/unsorted topK Signed-off-by: raver119 <raver119@gmail.com> * implementation and fixing col2im/col2vol * Corrected usage flags with input/output with reverse op. * dup is const now Signed-off-by: raver119 <raver119@gmail.com> * percentile op Signed-off-by: raver119 <raver119@gmail.com> * group tests for mapool2d Signed-off-by: Yurii <yurii@skymind.io> * special test for george Signed-off-by: raver119 <raver119@gmail.com> * less threads for sortTad Signed-off-by: raver119 <raver119@gmail.com> * provide conv2d for cuda Signed-off-by: Yurii <yurii@skymind.io> * remove auther in sort tad kernel code Signed-off-by: Yurii <yurii@skymind.io> * provide depthwise_conv2d for cuda Signed-off-by: Yurii <yurii@skymind.io> * - max_pooling_with_argmax - null check for special use Signed-off-by: raver119 <raver119@gmail.com> * dts cuda Signed-off-by: raver119 <raver119@gmail.com> * provide sconv2d for cuda Signed-off-by: Yurii <yurii@skymind.io> * std cuda Signed-off-by: raver119 <raver119@gmail.com> * Refactored non_max_suppression op to conform TF implementation. * Improved suppression helper. * provide pooling3d for cuda Signed-off-by: Yurii <yurii@skymind.io> * minor lstm rearrangements Signed-off-by: raver119 <raver119@gmail.com> * more of minor lstm rearrangements Signed-off-by: raver119 <raver119@gmail.com> * (bi)dynamic_rnn Signed-off-by: raver119 <raver119@gmail.com> * templates init order Signed-off-by: raver119 <raver119@gmail.com> * Refactored non_max_suppression op. * Added cuda kernel for non_max_suppression. * CPU sort by key/value Signed-off-by: raver119 <raver119@gmail.com> * CPU sort TAD by key/value Signed-off-by: raver119 <raver119@gmail.com> * CPU sort TAD by key/value tests Signed-off-by: raver119 <raver119@gmail.com> * Eliminate compiler error with cuda implementation. * - repaired gradCheck in cuda - provide conv2d_bp for cuda Signed-off-by: Yurii <yurii@skymind.io> * missed signature Signed-off-by: raver119 <raver119@gmail.com> * provide depthwise_conv2d_bp for cuda Signed-off-by: Yurii <yurii@skymind.io> * Implementation of lup helper with cuda kernel. Initial commit. * further work on backprops for convolutions Signed-off-by: Yurii <yurii@skymind.io> * CUDA linear sort by key/val Signed-off-by: raver119 <raver119@gmail.com> * CUDA tad sort by key/val Signed-off-by: raver119 <raver119@gmail.com> * start providing of backprop for pooling2d/3d Signed-off-by: Yurii <yurii@skymind.io> * Added atomicAdd for bool datatype. * dynamic partition concept Signed-off-by: raver119 <raver119@gmail.com> * dynamic partition concept Signed-off-by: raver119 <raver119@gmail.com> * dynamic partition scalar CUDA Signed-off-by: raver119 <raver119@gmail.com> * important comment Signed-off-by: raver119 <raver119@gmail.com> * fix pooling2d/3d backprop helpers Signed-off-by: Yurii <yurii@skymind.io> * Added non-linear test with dynamic_partition. * Improved test for dynamic_partition. * dynamic_partition TAD concept Signed-off-by: raver119 <raver119@gmail.com> * - dynamic_partition TAD CUDA impl - dynamic_partition TAD CPU fix Signed-off-by: raver119 <raver119@gmail.com> * - rewrite cpu code for usampling2d/3d - write cuda code for usampling2d/3d Signed-off-by: Yurii <yurii@skymind.io> * dynamic_stitch CUDA vector case Signed-off-by: raver119 <raver119@gmail.com> * dynamic_stitch CUDA TAD case concept Signed-off-by: raver119 <raver119@gmail.com> * dynamic_stitch CUDA TAD case impl Signed-off-by: raver119 <raver119@gmail.com> * Added tests for dynamic_stitch 3D-4D cases. * minor tests tweaks Signed-off-by: raver119 <raver119@gmail.com> * Fixed type check for dynamic stitch. * min/max bp Signed-off-by: raver119 <raver119@gmail.com> * rewrite code for upsampling2d/3d cpu Signed-off-by: Yurii <yurii@skymind.io> * reduce min/max/norm_max bp Signed-off-by: raver119 <raver119@gmail.com> * lup implementation. Additional enhancements. * provide code for upsamling2d/3d backprop Signed-off-by: Yurii <yurii@skymind.io> * weightedCrossEntropyWithLogits Signed-off-by: raver119 <raver119@gmail.com> * Fixed template math atomicMul for 64bit ints. * Refactored dynamic_partition_bp op. * inverseBroadcast fix Signed-off-by: raver119 <raver119@gmail.com> * DynamicPartitionBP test datatype fixed. * - nd4j_atomicMul Windows fix - cpu/NDArrayLambda.hpp excluded from CUDA Signed-off-by: raver119 <raver119@gmail.com>
243 lines
13 KiB
Plaintext
243 lines
13 KiB
Plaintext
/*******************************************************************************
|
|
* 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 Yurii Shyrma, created on 25.02.2018
|
|
//
|
|
|
|
|
|
#include<ops/declarable/helpers/batchnorm.h>
|
|
#include <helpers/ShapeUtils.h>
|
|
#include <OmpLaunchHelper.h>
|
|
#include <ConstantTadHelper.h>
|
|
#include <PointersManager.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__global__ static void batchnormCuda(const void* vx, const Nd4jLong* xShapeInfo,
|
|
const void* vMean, const Nd4jLong* meanShapeInfo,
|
|
const void* vVariance, const Nd4jLong* varianceShapeInfo,
|
|
const void* vGamma, const Nd4jLong* gammaShapeInfo,
|
|
const void* vBeta, const Nd4jLong* betaShapeInfo,
|
|
void* vz, const Nd4jLong* zShapeInfo,
|
|
const Nd4jLong* xTadShapeInfo, const Nd4jLong* xTadOffsets,
|
|
const Nd4jLong* zTadShapeInfo, const Nd4jLong* zTadOffsets,
|
|
const T epsilon) {
|
|
|
|
const auto x = reinterpret_cast<const T*>(vx);
|
|
auto z = reinterpret_cast<T*>(vz);
|
|
const auto mean = reinterpret_cast<const T*>(vMean);
|
|
const auto variance = reinterpret_cast<const T*>(vVariance);
|
|
const auto gamma = reinterpret_cast<const T*>(vGamma);
|
|
const auto beta = reinterpret_cast<const T*>(vBeta);
|
|
|
|
// maxRank = xRank = zRank, minRank = meanRank = varianceRank = gammaRank = betaRank
|
|
__shared__ Nd4jLong minLen, tadLen, totalThreads;
|
|
|
|
if (threadIdx.x == 0) {
|
|
|
|
totalThreads = gridDim.x * blockDim.x;
|
|
|
|
minLen = shape::length(meanShapeInfo);
|
|
tadLen = shape::length(xShapeInfo) / minLen;
|
|
}
|
|
__syncthreads();
|
|
|
|
const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
for (uint i = tid; i < minLen; i += totalThreads) {
|
|
|
|
const auto meanOffset = shape::getIndexOffset(i, meanShapeInfo, minLen);
|
|
const auto varianceOffset = shape::getIndexOffset(i, varianceShapeInfo, minLen);
|
|
|
|
T sigmaInvGam = 1. / nd4j::math::nd4j_sqrt<T, T>(variance[varianceOffset] + epsilon);
|
|
|
|
if(gamma != nullptr)
|
|
sigmaInvGam *= gamma[shape::getIndexOffset(i, gammaShapeInfo, minLen)];
|
|
|
|
auto betaOffset = 0;
|
|
if(beta != nullptr)
|
|
betaOffset = shape::getIndexOffset(i, betaShapeInfo, minLen);
|
|
|
|
const auto xTad = x + xTadOffsets[i];
|
|
auto zTad = z + zTadOffsets[i];
|
|
|
|
for (uint j = 0; j < tadLen; ++j) {
|
|
|
|
const auto xTadOffset = shape::getIndexOffset(j, xTadShapeInfo, tadLen);
|
|
const auto zTadOffset = shape::getIndexOffset(j, zTadShapeInfo, tadLen);
|
|
|
|
zTad[zTadOffset] = (xTad[xTadOffset] - mean[meanOffset]) * sigmaInvGam;
|
|
|
|
if(beta != nullptr)
|
|
zTad[zTadOffset] += beta[betaOffset];
|
|
}
|
|
}
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__global__ static void batchnormCuda2(const void* vx, const Nd4jLong* xShapeInfo,
|
|
const void* vMean, const Nd4jLong* meanShapeInfo,
|
|
const void* vVariance, const Nd4jLong* varianceShapeInfo,
|
|
const void* vGamma, const Nd4jLong* gammaShapeInfo,
|
|
const void* vBeta, const Nd4jLong* betaShapeInfo,
|
|
void* vz, const Nd4jLong* zShapeInfo,
|
|
const int numDims, const int* dims,
|
|
const T epsilon) {
|
|
|
|
const auto x = reinterpret_cast<const T*>(vx);
|
|
auto z = reinterpret_cast<T*>(vz);
|
|
const auto mean = reinterpret_cast<const T*>(vMean);
|
|
const auto variance = reinterpret_cast<const T*>(vVariance);
|
|
const auto gamma = reinterpret_cast<const T*>(vGamma);
|
|
const auto beta = reinterpret_cast<const T*>(vBeta);
|
|
|
|
__shared__ int xRank, minRank; // xRank == zRank. minRank = meanRank = varianceRank = gammaRank = betaRank
|
|
__shared__ Nd4jLong xLen, totalThreads, *sharedMem; // xLen = zLen
|
|
|
|
|
|
if (threadIdx.x == 0) {
|
|
|
|
extern __shared__ unsigned char shmem[];
|
|
sharedMem = reinterpret_cast<Nd4jLong*>(shmem);
|
|
totalThreads = gridDim.x * blockDim.x;
|
|
|
|
xLen = shape::length(xShapeInfo);
|
|
xRank = shape::rank(xShapeInfo);
|
|
minRank = shape::rank(meanShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
auto coords = sharedMem + threadIdx.x * xRank;
|
|
const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
for (uint i = tid; i < xLen; i += totalThreads) {
|
|
|
|
shape::index2coords(xRank, shape::shapeOf(const_cast<Nd4jLong*>(xShapeInfo)), i, xLen, coords);
|
|
|
|
const auto xOffset = shape::getOffset(0, shape::shapeOf(const_cast<Nd4jLong*>(xShapeInfo)), shape::stride(const_cast<Nd4jLong*>(xShapeInfo)), coords, xRank);
|
|
const auto zOffset = shape::getOffset(0, shape::shapeOf(const_cast<Nd4jLong*>(zShapeInfo)), shape::stride(const_cast<Nd4jLong*>(zShapeInfo)), coords, xRank);
|
|
|
|
if(minRank == xRank) {
|
|
for (uint i = 0, j = 0; i < xRank; ++i) {
|
|
if(j < numDims && i != dims[j])
|
|
coords[i] = 0;
|
|
else
|
|
++j;
|
|
}
|
|
}
|
|
else // minRank = numDims = 1 in this case
|
|
coords[0] = coords[dims[0]];
|
|
|
|
const auto meanOffset = shape::getOffset(0, shape::shapeOf(const_cast<Nd4jLong*>(meanShapeInfo)), shape::stride(const_cast<Nd4jLong*>(meanShapeInfo)), coords, minRank);
|
|
const auto varianceOffset = shape::getOffset(0, shape::shapeOf(const_cast<Nd4jLong*>(varianceShapeInfo)), shape::stride(const_cast<Nd4jLong*>(varianceShapeInfo)), coords, minRank);
|
|
|
|
T sigmaInvGam = 1. / nd4j::math::nd4j_sqrt<T, T>(variance[varianceOffset] + epsilon);
|
|
|
|
if(gamma != nullptr) {
|
|
const auto gammaOffset = shape::getOffset(0, shape::shapeOf(const_cast<Nd4jLong*>(gammaShapeInfo)), shape::stride(const_cast<Nd4jLong*>(gammaShapeInfo)), coords, minRank);
|
|
sigmaInvGam *= gamma[gammaOffset];
|
|
}
|
|
|
|
z[zOffset] = (x[xOffset] - mean[meanOffset]) * sigmaInvGam;
|
|
|
|
if(beta != nullptr) {
|
|
const auto betaOffset = shape::getOffset(0, shape::shapeOf(const_cast<Nd4jLong*>(betaShapeInfo)), shape::stride(const_cast<Nd4jLong*>(betaShapeInfo)), coords, minRank);
|
|
z[zOffset] += beta[betaOffset];
|
|
}
|
|
}
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__host__ static void batchnormCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const cudaStream_t *stream,
|
|
const void* vx, const Nd4jLong* xShapeInfo,
|
|
const void* vMean, const Nd4jLong* meanShapeInfo,
|
|
const void* vVariance, const Nd4jLong* varianceShapeInfo,
|
|
const void* vGamma, const Nd4jLong* gammaShapeInfo,
|
|
const void* vBeta, const Nd4jLong* betaShapeInfo,
|
|
void* vz, const Nd4jLong* zShapeInfo,
|
|
const Nd4jLong* xTadShapeInfo, const Nd4jLong* xTadOffsets,
|
|
const Nd4jLong* zTadShapeInfo, const Nd4jLong* zTadOffsets,
|
|
const double epsilon) {
|
|
|
|
batchnormCuda<T><<<blocksPerGrid, threadsPerBlock, 1024, *stream>>>(vx, xShapeInfo, vMean, meanShapeInfo, vVariance, varianceShapeInfo, vGamma, gammaShapeInfo, vBeta, betaShapeInfo, vz, zShapeInfo, xTadShapeInfo, xTadOffsets, zTadShapeInfo, zTadOffsets, static_cast<T>(epsilon));
|
|
}
|
|
BUILD_SINGLE_TEMPLATE(template void batchnormCudaLauncher, (const int blocksPerGrid, const int threadsPerBlock, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, const void* vMean, const Nd4jLong* meanShapeInfo, const void* vVariance, const Nd4jLong* varianceShapeInfo, const void* vGamma, const Nd4jLong* gammaShapeInfo, const void* vBeta, const Nd4jLong* betaShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const Nd4jLong* xTadShapeInfo, const Nd4jLong* xTadOffsets, const Nd4jLong* zTadShapeInfo, const Nd4jLong* zTadOffsets, const double epsilon), FLOAT_TYPES);
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__host__ static void batchnormCudaLauncher2(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream,
|
|
const void* vx, const Nd4jLong* xShapeInfo,
|
|
const void* vMean, const Nd4jLong* meanShapeInfo,
|
|
const void* vVariance, const Nd4jLong* varianceShapeInfo,
|
|
const void* vGamma, const Nd4jLong* gammaShapeInfo,
|
|
const void* vBeta, const Nd4jLong* betaShapeInfo,
|
|
void* vz, const Nd4jLong* zShapeInfo,
|
|
const int numDims, const int* dims,
|
|
const double epsilon) {
|
|
|
|
batchnormCuda2<T><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(vx, xShapeInfo, vMean, meanShapeInfo, vVariance, varianceShapeInfo, vGamma, gammaShapeInfo, vBeta, betaShapeInfo, vz, zShapeInfo, numDims, dims, static_cast<T>(epsilon));
|
|
}
|
|
BUILD_SINGLE_TEMPLATE(template void batchnormCudaLauncher2, (const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, const void* vMean, const Nd4jLong* meanShapeInfo, const void* vVariance, const Nd4jLong* varianceShapeInfo, const void* vGamma, const Nd4jLong* gammaShapeInfo, const void* vBeta, const Nd4jLong* betaShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const int numDims, const int* dims, const double epsilon), FLOAT_TYPES);
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
void batchnorm(const NDArray* input, const NDArray* mean, const NDArray* variance, const NDArray* gamma, const NDArray* beta, NDArray* output, const std::vector<int>& axes, const double epsilon) {
|
|
|
|
std::vector<int> dimsToExclude = ShapeUtils::evalDimsToExclude(input->rankOf(), axes);
|
|
|
|
auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(), dimsToExclude);
|
|
auto packZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), dimsToExclude);
|
|
|
|
const int threadsPerBlock = MAX_NUM_THREADS / 2;
|
|
const int blocksPerGrid = (mean->lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
|
|
|
|
PointersManager manager(input->getContext(), "batchnorm");
|
|
|
|
NDArray::prepareSpecialUse({output}, {input, mean, variance, gamma, beta});
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), batchnormCudaLauncher, (blocksPerGrid, threadsPerBlock, input->getContext()->getCudaStream(), input->getSpecialBuffer(), input->getSpecialShapeInfo(), mean->getSpecialBuffer(), mean->getSpecialShapeInfo(), variance->getSpecialBuffer(), variance->getSpecialShapeInfo(), gamma ? gamma->getSpecialBuffer() : nullptr, gamma ? gamma->getSpecialShapeInfo() : nullptr, beta ? beta->getSpecialBuffer() : nullptr, beta ? beta->getSpecialShapeInfo() : nullptr, output->specialBuffer(), output->specialShapeInfo(), packX.platformShapeInfo(), packX.platformOffsets(), packZ.platformShapeInfo(), packZ.platformOffsets(), epsilon), FLOAT_TYPES);
|
|
NDArray::registerSpecialUse({output}, {input, mean, variance, gamma, beta});
|
|
|
|
manager.synchronize();
|
|
|
|
|
|
// const int threadsPerBlock = MAX_NUM_THREADS / 4;
|
|
// const int blocksPerGrid = (input->lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
|
|
// const int sharedMem = sizeof(Nd4jLong) * threadsPerBlock * input->rankOf() + 128;
|
|
|
|
// PointersManager manager(input->getContext(), "batchnorm");
|
|
|
|
// const int* dims = reinterpret_cast<int*>(manager.replicatePointer(axes.data(), axes.size() * sizeof(int)));
|
|
|
|
// NDArray::prepareSpecialUse({output}, {input, mean, variance, gamma, beta});
|
|
// BUILD_SINGLE_SELECTOR(input->dataType(), batchnormCudaLauncher2, (blocksPerGrid, threadsPerBlock, sharedMem, input->getContext()->getCudaStream(), input->getSpecialBuffer(), input->getSpecialShapeInfo(), mean->getSpecialBuffer(), mean->getSpecialShapeInfo(), variance->getSpecialBuffer(), variance->getSpecialShapeInfo(), gamma ? gamma->getSpecialBuffer() : nullptr, gamma ? gamma->getSpecialShapeInfo() : nullptr, beta ? beta->getSpecialBuffer() : nullptr, beta ? beta->getSpecialShapeInfo() : nullptr, output->specialBuffer(), output->specialShapeInfo(), axes.size(), dims, epsilon), FLOAT_TYPES);
|
|
// NDArray::registerSpecialUse({output}, {input, mean, variance, gamma, beta});
|
|
|
|
// manager.synchronize();
|
|
}
|
|
|
|
|
|
}
|
|
}
|
|
}
|
|
|