6de00bf75f
* [WIP] Fix compilation after nd4j changes (#37) * Fix compilation. * Some tests fixed * Disable tests temporarily. * Restored test * Tests restored. * Test restored. * [WIP] perf tests (#40) * special maxpool test Signed-off-by: raver119 <raver119@gmail.com> * special maxpool test Signed-off-by: raver119 <raver119@gmail.com> * Shyrma bnorm bp (#41) Batchnorm backprop mkldnn * Add SameDiff memory reuse memory manager (array cache) (#39) * Attention op comments Signed-off-by: AlexDBlack <blacka101@gmail.com> * ArrayCacheMemoryMgr - first pass Signed-off-by: AlexDBlack <blacka101@gmail.com> * Tweak array cache for use with SameDiff identity arrays Signed-off-by: AlexDBlack <blacka101@gmail.com> * ArrayCacheMemoryMgr javadoc and properly get max memory Signed-off-by: AlexDBlack <blacka101@gmail.com> * LRU cache policy + add tests Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fixes Signed-off-by: AlexDBlack <blacka101@gmail.com> * Resize arrays internally if required for ArrayCacheMemoryMgr Signed-off-by: AlexDBlack <blacka101@gmail.com> * Test improvement Signed-off-by: AlexDBlack <blacka101@gmail.com> * Small polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * SameDiff op runtime benchmarking listener (#42) Signed-off-by: AlexDBlack <blacka101@gmail.com> * INLINE_LOOPS for windows Signed-off-by: raver119 <raver119@gmail.com> * [WIP] ThreadPool (#8) This PR removes OpenMP use in 95% of cases
87 lines
3.5 KiB
C++
Executable File
87 lines
3.5 KiB
C++
Executable File
/*******************************************************************************
|
|
* 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
|
|
******************************************************************************/
|
|
|
|
/*
|
|
*
|
|
* Created on: Dec 28, 2015
|
|
* Author: agibsonccc
|
|
*/
|
|
|
|
#ifndef INDEXREDUCE_H_
|
|
#define INDEXREDUCE_H_
|
|
#include "../helpers/shape.h"
|
|
#ifdef _OPENMP
|
|
#include <omp.h>
|
|
#endif
|
|
#include <dll.h>
|
|
#include <ops/ops.h>
|
|
#include <op_boilerplate.h>
|
|
#include <OmpLaunchHelper.h>
|
|
#include <helpers/DebugHelper.h>
|
|
|
|
#ifdef __CUDACC__
|
|
#include <cuda.h>
|
|
#include <cuda_runtime.h>
|
|
#endif
|
|
|
|
|
|
#include <helpers/TAD.h>
|
|
|
|
|
|
#include "../pairwise_util.h"
|
|
|
|
#include "legacy_ops.h"
|
|
|
|
namespace functions {
|
|
namespace indexreduce {
|
|
|
|
template<typename X, typename Z>
|
|
class IndexReduce {
|
|
public:
|
|
#ifdef __CUDACC__
|
|
|
|
static __device__ void transform(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int *dimension,int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
|
|
|
|
template<typename OpType>
|
|
static __device__ void aggregatePartials(IndexValue<X> **sPartialsRef, Nd4jLong tid, Nd4jLong numElements,void *extraParams);
|
|
|
|
|
|
template<typename OpType>
|
|
static __device__ void transform(void *dx, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
|
|
|
|
|
|
static _CUDA_H void executeIndexReduceScalar(dim3 launchDims, cudaStream_t *stream, const int op, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
|
|
|
|
static _CUDA_H void executeIndexReduce(dim3 launchDims, cudaStream_t *stream, const int op, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
|
|
#else
|
|
|
|
static Nd4jLong execScalar(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams);
|
|
|
|
static void exec(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
|
|
|
|
template<typename OpType>
|
|
static _CUDA_H Nd4jLong execScalar(void *x, Nd4jLong *xShapeInfo, void *extraParams);
|
|
|
|
template<typename OpType>
|
|
static _CUDA_H void exec(void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
|
|
#endif
|
|
};
|
|
}
|
|
}
|
|
|
|
#endif /* INDEXREDUCE_H_ */
|
|
|