583 lines
27 KiB
Plaintext
583 lines
27 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 raver119@gmail.com
|
|
//
|
|
|
|
#include <op_boilerplate.h>
|
|
#include <loops/reduce_float.h>
|
|
#include <loops/legacy_ops.h>
|
|
#include <helpers/DebugHelper.h>
|
|
|
|
|
|
template <typename X, typename Z, typename OpClass>
|
|
__device__ void reduceSimpleGeneric(
|
|
void *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
void *extraParams,
|
|
void *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {
|
|
|
|
__shared__ UnifiedSharedMemory *manager;
|
|
|
|
if (threadIdx.x == 0) {
|
|
c __shared__ unsigned char shmem[];
|
|
manager = new(shmem) UnifiedSharedMemory((int *) shmem);
|
|
manager->init(sizeof(UnifiedSharedMemory), 0, sizeof(functions::reduce::ReduceFunction<T>), sizeof(shape::TAD), shape::rank(xShapeInfo));
|
|
}
|
|
__syncthreads();
|
|
|
|
|
|
functions::reduce::ReduceFunction<T>::template transformCudaXD<OpClass>(
|
|
dx,
|
|
xShapeInfo,
|
|
extraParams,
|
|
result,
|
|
resultShapeInfo,
|
|
dimension,
|
|
dimensionLength,
|
|
reductionBuffer,
|
|
manager,
|
|
tadOnlyShapeInfo,
|
|
tadOffsets);
|
|
}
|
|
|
|
template <typename T, typename OpClass>
|
|
__device__ void reduceSimpleGeneric1D(
|
|
T *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
T *extraParams,
|
|
T *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
T *reductionBuffer,
|
|
Nd4jLong *tadOnlyShapeInfo,
|
|
Nd4jLong *tadOffsets) {
|
|
|
|
functions::reduce::ReduceFunction<T>::template transformCuda1D<OpClass>(
|
|
dx,
|
|
xShapeInfo,
|
|
extraParams,
|
|
result,
|
|
resultShapeInfo,
|
|
dimension,
|
|
dimensionLength,
|
|
reductionBuffer, nullptr, tadOnlyShapeInfo, tadOffsets);
|
|
}
|
|
|
|
template <typename T, typename OpClass>
|
|
__device__ void reduceSimpleGeneric3D(
|
|
T *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
T *extraParams,
|
|
T *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
T *reductionBuffer,
|
|
Nd4jLong *tadOnlyShapeInfo,
|
|
Nd4jLong *tadOffsets) {
|
|
|
|
functions::reduce::ReduceFunction<T>::template transformCuda3D<OpClass>(
|
|
dx,
|
|
xShapeInfo,
|
|
extraParams,
|
|
result,
|
|
resultShapeInfo,
|
|
dimension,
|
|
dimensionLength,
|
|
reductionBuffer,
|
|
nullptr,
|
|
tadOnlyShapeInfo,
|
|
tadOffsets);
|
|
}
|
|
|
|
template <typename T, typename OpClass>
|
|
__device__ void reduceScalarGeneric(
|
|
T *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
T *extraParams,
|
|
T *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
T *reductionBuffer, Nd4jLong *tadOnlyShapeInfo) {
|
|
|
|
__shared__ UnifiedSharedMemory *manager;
|
|
|
|
if (threadIdx.x == 0) {
|
|
extern __shared__ unsigned char shmem[];
|
|
manager = new(shmem) UnifiedSharedMemory((int *) shmem);
|
|
manager->init(sizeof(UnifiedSharedMemory), 0, sizeof(functions::reduce::ReduceFunction<T>), sizeof(shape::TAD), 0);
|
|
}
|
|
__syncthreads();
|
|
|
|
functions::reduce::ReduceFunction<T>::template execScalarCuda<OpClass>(
|
|
dx,
|
|
xShapeInfo,
|
|
extraParams,
|
|
result,
|
|
resultShapeInfo,
|
|
reductionBuffer,
|
|
manager,
|
|
tadOnlyShapeInfo);
|
|
};
|
|
|
|
#ifndef __CLION_IDE__
|
|
// reduceScalar
|
|
DISPATCH_KERNEL_SIMPLE(reduceScalarSimple_, reduceScalarGeneric, float, INPUT(float *x, Nd4jLong *xShapeInfo, float *extraParams, float *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, float *reductionBuffer, Nd4jLong *tadOnlyShapeInfo), PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceScalarSimple_, reduceScalarGeneric, double, INPUT(double *x, Nd4jLong *xShapeInfo, double *extraParams, double *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, double *reductionBuffer, Nd4jLong *tadOnlyShapeInfo), PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceScalarSimple_, reduceScalarGeneric, float16, INPUT(float16 *x, Nd4jLong *xShapeInfo, float16 *extraParams, float16 *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, float16 *reductionBuffer, Nd4jLong *tadOnlyShapeInfo), PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_OPS))
|
|
|
|
// reduce1D
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGeneric1D_, reduceSimpleGeneric1D, float, INPUT(float *x, Nd4jLong *xShape, float *extraParams, float *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGeneric1D_, reduceSimpleGeneric1D, double, INPUT(double *x, Nd4jLong *xShape, double *extraParams, double *z, Nd4jLong *zShape, int *dimension, int dimensionLength, double *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGeneric1D_, reduceSimpleGeneric1D, float16, INPUT(float16 *x, Nd4jLong *xShape, float16 *extraParams, float16 *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float16 *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
|
|
// reduce3D
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGeneric3D_, reduceSimpleGeneric3D, float, INPUT(float *x, Nd4jLong *xShape, float *extraParams, float *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGeneric3D_, reduceSimpleGeneric3D, double, INPUT(double *x, Nd4jLong *xShape, double *extraParams, double *z, Nd4jLong *zShape, int *dimension, int dimensionLength, double *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGeneric3D_, reduceSimpleGeneric3D, float16, INPUT(float16 *x, Nd4jLong *xShape, float16 *extraParams, float16 *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float16 *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
|
|
// reduceXD
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGenericXD_, reduceSimpleGeneric, float, INPUT(float *x, Nd4jLong *xShape, float *extraParams, float *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGenericXD_, reduceSimpleGeneric, double, INPUT(double *x, Nd4jLong *xShape, double *extraParams, double *z, Nd4jLong *zShape, int *dimension, int dimensionLength, double *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
DISPATCH_KERNEL_SIMPLE(reduceSimpleGenericXD_, reduceSimpleGeneric, float16, INPUT(float16 *x, Nd4jLong *xShape, float16 *extraParams, float16 *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float16 *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets), PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
|
|
#endif
|
|
|
|
namespace functions {
|
|
namespace reduce {
|
|
|
|
template <>
|
|
_CUDA_H void ReduceFunction<float>::execReduceScalar(dim3 launchDims, cudaStream_t *stream, int opNum, float *x, Nd4jLong *xShapeInfo, float *extraParams, float *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, float *reductionBuffer, Nd4jLong *tadOnlyShapeInfo) {
|
|
|
|
DISPATCH_SIMPLE(reduceScalarSimple, float, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, nullptr, 1, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_OPS))
|
|
|
|
nd4j::DebugHelper::checkErrorCode(stream, "execReduceScalarFloat(...) failed");
|
|
}
|
|
|
|
template <>
|
|
_CUDA_H void ReduceFunction<float16>::execReduceScalar(dim3 launchDims, cudaStream_t *stream, int opNum, float16 *x, Nd4jLong *xShapeInfo, float16 *extraParams, float16 *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, float16 *reductionBuffer, Nd4jLong *tadOnlyShapeInfo) {
|
|
|
|
DISPATCH_SIMPLE(reduceScalarSimple, float16, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, nullptr, 1, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_OPS))
|
|
|
|
nd4j::DebugHelper::checkErrorCode(stream, "execReduceScalarHalf(...) failed");
|
|
}
|
|
|
|
template <>
|
|
_CUDA_H void ReduceFunction<double>::execReduceScalar(dim3 launchDims, cudaStream_t *stream, int opNum, double *x, Nd4jLong *xShapeInfo, double *extraParams, double *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, double *reductionBuffer, Nd4jLong *tadOnlyShapeInfo) {
|
|
|
|
DISPATCH_SIMPLE(reduceScalarSimple, double, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, nullptr, 1, reductionBuffer, tadOnlyShapeInfo), OPS_A(REDUCE_OPS))
|
|
|
|
nd4j::DebugHelper::checkErrorCode(stream, "execReduceScalarDouble(...) failed");
|
|
}
|
|
|
|
template <>
|
|
_CUDA_H void ReduceFunction<float>::execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, float *x, Nd4jLong *xShape, float *extraParams, float *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {
|
|
if (rank == 1) {
|
|
DISPATCH_SIMPLE(reduceSimpleGeneric1D, float, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
} else if (rank <= 3) {
|
|
DISPATCH_SIMPLE(reduceSimpleGeneric3D, float, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
} else {
|
|
DISPATCH_SIMPLE(reduceSimpleGenericXD, float, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
}
|
|
|
|
DEBUG_KERNEL(stream, opNum);
|
|
}
|
|
|
|
template <>
|
|
_CUDA_H void ReduceFunction<float16>::execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, float16 *x, Nd4jLong *xShape, float16 *extraParams, float16 *z, Nd4jLong *zShape, int *dimension, int dimensionLength, float16 *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {
|
|
if (rank == 1) {
|
|
DISPATCH_SIMPLE(reduceSimpleGeneric1D, float16, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
} else if (rank <= 3) {
|
|
DISPATCH_SIMPLE(reduceSimpleGeneric3D, float16, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
} else {
|
|
DISPATCH_SIMPLE(reduceSimpleGenericXD, float16, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
}
|
|
|
|
DEBUG_KERNEL(stream, opNum);
|
|
}
|
|
|
|
template <>
|
|
_CUDA_H void ReduceFunction<double>::execReduceXD(dim3 launchDims, cudaStream_t *stream, int opNum, int rank, double *x, Nd4jLong *xShape, double *extraParams, double *z, Nd4jLong *zShape, int *dimension, int dimensionLength, double *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {
|
|
|
|
if (rank == 1) {
|
|
DISPATCH_SIMPLE(reduceSimpleGeneric1D, double, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
} else if (rank <= 3) {
|
|
DISPATCH_SIMPLE(reduceSimpleGeneric3D, double, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
} else {
|
|
DISPATCH_SIMPLE(reduceSimpleGenericXD, double, PARAMS(x, xShape, extraParams, z, zShape, dimension, dimensionLength, reductionPointer, tadShapeInfo, tadOffsets), OPS_A(REDUCE_OPS))
|
|
}
|
|
|
|
DEBUG_KERNEL(stream, opNum);
|
|
}
|
|
|
|
template <typename T>
|
|
__device__ void initializeShared(T *extraParams, T **sPartials, int sMemSize) {
|
|
int sPartialsLength = sMemSize / sizeof(T);
|
|
T *sPartialsDeref = (T *) *sPartials;
|
|
for (int i = 0; i < sPartialsLength; i++) {
|
|
sPartialsDeref[i] = extraParams[0];
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
template <typename OpType>
|
|
__device__ void ReduceFunction<T>::transformCuda1D(T *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
T *extraParams,
|
|
T *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
T *reductionBuffer, UnifiedSharedMemory *manager, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {
|
|
|
|
if (OpType::requiresSpecialAccumulation) {
|
|
OpType::execSpecialCuda(dx, xShapeInfo, extraParams, result, resultShapeInfo, dimension, dimensionLength, reductionBuffer, manager, tadOnlyShapeInfo, tadOffsets);
|
|
return;
|
|
}
|
|
|
|
//shared memory space for storing intermediate results
|
|
__shared__ T *sPartials;// = (T *)manager->getSharedReductionBuffer();
|
|
__shared__ int tadLength;
|
|
__shared__ int tadEWS;
|
|
__shared__ int numTads;
|
|
if (threadIdx.x == 0) {
|
|
extern __shared__ unsigned char shmem[];
|
|
sPartials = (T *) shmem;
|
|
tadLength = shape::length(tadOnlyShapeInfo);//shape::tadLength(xShapeInfo, dimension, dimensionLength);
|
|
tadEWS = shape::elementWiseStride(tadOnlyShapeInfo);
|
|
numTads = shape::length(xShapeInfo) / tadLength;
|
|
}
|
|
__syncthreads();
|
|
|
|
for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
|
|
Nd4jLong tadOffsetForBlock = tadOffsets[r];
|
|
T *rX = dx + tadOffsetForBlock;
|
|
|
|
sPartials[threadIdx.x] = OpType::startingValue(rX);
|
|
|
|
if (tadEWS >= 1) {
|
|
for (int i = threadIdx.x; i < tadLength; i += blockDim.x) {
|
|
sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(rX[i * tadEWS], extraParams), extraParams);
|
|
}
|
|
} else {
|
|
__shared__ int tadRank;
|
|
__shared__ Nd4jLong *tadShape;
|
|
__shared__ Nd4jLong *tadStride;
|
|
Nd4jLong xCoord[MAX_RANK];
|
|
if (threadIdx.x == 0) {
|
|
tadRank = shape::rank(tadOnlyShapeInfo);
|
|
tadShape = shape::shapeOf(tadOnlyShapeInfo);
|
|
tadStride = shape::stride(tadOnlyShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
for (int i = threadIdx.x; i < tadLength; i += blockDim.x) {
|
|
shape::ind2subC(tadRank, tadShape, i, tadLength, xCoord);
|
|
auto xOffset = shape::getOffset(tadOffsetForBlock, tadShape, tadStride, xCoord, tadRank);
|
|
|
|
sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(dx[xOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
|
|
|
|
__syncthreads();
|
|
|
|
// aggregate. do NOT reduce for elements > tadLength
|
|
aggregatePartials<OpType>(sPartials, threadIdx.x, nd4j::math::nd4j_min<int>(blockDim.x, tadLength), extraParams);
|
|
|
|
|
|
__syncthreads();
|
|
if (threadIdx.x == 0) {
|
|
result[r] = OpType::postProcess(sPartials[threadIdx.x], tadLength, extraParams);
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
template <typename T>
|
|
template <typename OpType>
|
|
__device__ void ReduceFunction<T>::execScalarCuda(
|
|
T *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
T *extraParams,
|
|
T *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
T *reductionBuffer,
|
|
UnifiedSharedMemory *manager,
|
|
Nd4jLong *tadOnlyShapeInfo) {
|
|
int elementWiseStride = shape::elementWiseStride(xShapeInfo);
|
|
|
|
auto n = shape::length(xShapeInfo);
|
|
|
|
auto tid = blockDim.x * blockIdx.x + threadIdx.x;
|
|
|
|
//shared memory space for storing intermediate results
|
|
T *sPartials = (T *)manager->getSharedReductionBuffer();
|
|
|
|
sPartials[threadIdx.x] = OpType::startingValue(dx);
|
|
|
|
if (elementWiseStride >= 1) {
|
|
for (int i = tid; i < n; i += (blockDim.x * gridDim.x)) {
|
|
sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(dx[i * elementWiseStride], extraParams), extraParams);
|
|
}
|
|
}
|
|
else {
|
|
__shared__ int rank;
|
|
__shared__ Nd4jLong *xShape;
|
|
__shared__ Nd4jLong *xStride;
|
|
if (threadIdx.x == 0) {
|
|
rank = shape::rank(xShapeInfo);
|
|
xShape = shape::shapeOf(xShapeInfo);
|
|
xStride = shape::stride(xShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
Nd4jLong ind2sub[MAX_RANK];
|
|
|
|
for (int i = tid; i < n; i += blockDim.x * gridDim.x) {
|
|
shape::ind2subC(rank, xShape, i, n, ind2sub);
|
|
|
|
auto offset = shape::getOffset(0, xShape, xStride, ind2sub, rank);
|
|
sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(dx[offset], extraParams), extraParams);
|
|
}
|
|
}
|
|
|
|
__syncthreads();
|
|
aggregatePartials<OpType>(sPartials, threadIdx.x, nd4j::math::nd4j_min<int>(blockDim.x, n), extraParams);
|
|
|
|
|
|
__syncthreads();
|
|
|
|
if (gridDim.x > 1) {
|
|
unsigned int *tc = (unsigned int *)reductionBuffer;
|
|
__shared__ bool amLast;
|
|
tid = threadIdx.x;
|
|
if (threadIdx.x == 0) {
|
|
reductionBuffer[blockIdx.x] = sPartials[0];//this->postProcess(sPartials[0],n,extraParams);
|
|
}
|
|
__threadfence();
|
|
__syncthreads();
|
|
|
|
if (threadIdx.x == 0) {
|
|
unsigned int ticket = atomicInc(&tc[16384], gridDim.x);
|
|
amLast = (ticket == gridDim.x - 1);
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
if (amLast) {
|
|
tc[16384] = 0;
|
|
|
|
sPartials[threadIdx.x] = OpType::startingValue(dx);
|
|
|
|
for (int i = threadIdx.x; i < gridDim.x; i += blockDim.x) {
|
|
sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], reductionBuffer[i], extraParams);
|
|
}
|
|
__syncthreads();
|
|
|
|
|
|
|
|
aggregatePartials<OpType>(sPartials, threadIdx.x, nd4j::math::nd4j_min<int>(gridDim.x, blockDim.x), extraParams);
|
|
|
|
__syncthreads();
|
|
if (threadIdx.x == 0) {
|
|
result[0] = OpType::postProcess(sPartials[0], n, extraParams);
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
if (threadIdx.x == 0) {
|
|
unsigned int *tc = (unsigned *)reductionBuffer;
|
|
tc[16384] = 0;
|
|
result[0] = OpType::postProcess(sPartials[0], n, extraParams);
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
template <typename OpType>
|
|
__device__ void ReduceFunction<T>::transformCuda3D(
|
|
T *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
T *extraParams,
|
|
T *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
T *reductionBuffer, UnifiedSharedMemory *manager, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {
|
|
|
|
if (OpType::requiresSpecialAccumulation) {
|
|
OpType::execSpecialCuda(dx, xShapeInfo, extraParams, result, resultShapeInfo, dimension, dimensionLength, reductionBuffer, manager, tadOnlyShapeInfo, tadOffsets);
|
|
return;
|
|
}
|
|
|
|
//shared memory space for storing intermediate results
|
|
__shared__ T *sPartials; // = (T *)manager->getSharedReductionBuffer();
|
|
|
|
__shared__ int tadLength;
|
|
__shared__ int tadRank;
|
|
__shared__ int numTads;
|
|
__shared__ Nd4jLong *tadShape;
|
|
__shared__ Nd4jLong *tadStride;
|
|
if (threadIdx.x == 0) {
|
|
extern __shared__ unsigned char shmem[];
|
|
sPartials = (T *) shmem;
|
|
tadLength = shape::length(tadOnlyShapeInfo);//shape::tadLength(xShapeInfo, dimension, dimensionLength);
|
|
tadRank = shape::rank(tadOnlyShapeInfo);
|
|
numTads = shape::length(xShapeInfo) / tadLength;
|
|
|
|
tadShape = shape::shapeOf(tadOnlyShapeInfo);
|
|
tadStride = shape::stride(tadOnlyShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
Nd4jLong xCoord[3];
|
|
|
|
for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
|
|
Nd4jLong tadOffsetForBlock = tadOffsets[r];
|
|
|
|
sPartials[threadIdx.x] = OpType::startingValue(dx + tadOffsetForBlock);
|
|
|
|
for (int i = threadIdx.x; i < tadLength; i += blockDim.x) {
|
|
shape::ind2subC(tadRank, tadShape, i, tadLength, xCoord);
|
|
auto xOffset = shape::getOffset(tadOffsetForBlock, tadShape, tadStride, xCoord, tadRank);
|
|
|
|
sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(dx[xOffset], extraParams), extraParams);
|
|
}
|
|
__syncthreads();
|
|
|
|
// aggregate. do NOT reduce for elements > tadLength
|
|
aggregatePartials<OpType>(sPartials, threadIdx.x, nd4j::math::nd4j_min<int>(blockDim.x, tadLength), extraParams);
|
|
|
|
__syncthreads();
|
|
if (threadIdx.x == 0)
|
|
result[r] = OpType::postProcess(sPartials[threadIdx.x], tadLength, extraParams);
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
template <typename OpType>
|
|
__device__ void ReduceFunction<T>::transformCudaXD(
|
|
T *dx,
|
|
Nd4jLong *xShapeInfo,
|
|
T *extraParams,
|
|
T *result,
|
|
Nd4jLong *resultShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
T *reductionBuffer,
|
|
UnifiedSharedMemory *manager,
|
|
Nd4jLong *tadOnlyShapeInfo,
|
|
Nd4jLong *tadOffsets) {
|
|
|
|
if (OpType::requiresSpecialAccumulation) {
|
|
OpType::execSpecialCuda(dx, xShapeInfo, extraParams, result, resultShapeInfo, dimension, dimensionLength, reductionBuffer, manager, tadOnlyShapeInfo, tadOffsets);
|
|
return;
|
|
}
|
|
|
|
//shared memory space for storing intermediate results
|
|
__shared__ T *sPartials;
|
|
|
|
// __shared__ shape::TAD *tad;
|
|
__shared__ int tadLength;
|
|
__shared__ int tadRank;
|
|
__shared__ int numTads;
|
|
__shared__ Nd4jLong *tadShape;
|
|
__shared__ Nd4jLong *tadStride;
|
|
if (threadIdx.x == 0) {
|
|
extern __shared__ unsigned char shmem[];
|
|
sPartials = (T *) shmem;
|
|
tadLength = shape::length(tadOnlyShapeInfo);//shape::tadLength(xShapeInfo, dimension, dimensionLength);
|
|
tadRank = shape::rank(tadOnlyShapeInfo);
|
|
numTads = shape::length(xShapeInfo) / tadLength;
|
|
|
|
tadShape = shape::shapeOf(tadOnlyShapeInfo);
|
|
tadStride = shape::stride(tadOnlyShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
Nd4jLong xCoord[MAX_RANK];
|
|
|
|
for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
|
|
Nd4jLong tadOffsetForBlock = tadOffsets[r];
|
|
|
|
sPartials[threadIdx.x] = OpType::startingValue(dx + tadOffsetForBlock);
|
|
|
|
for (int i = threadIdx.x; i < tadLength; i += blockDim.x) {
|
|
shape::ind2subC(tadRank, tadShape, i, tadLength, xCoord);
|
|
auto xOffset = shape::getOffset(tadOffsetForBlock, tadShape, tadStride, xCoord, tadRank);
|
|
|
|
sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], OpType::op(dx[xOffset], extraParams), extraParams);
|
|
}
|
|
__syncthreads();
|
|
|
|
// aggregate. do NOT reduce for elements > tadLength
|
|
aggregatePartials<OpType>(sPartials, threadIdx.x, nd4j::math::nd4j_min<int>(blockDim.x, tadLength), extraParams);
|
|
|
|
|
|
__syncthreads();
|
|
if (threadIdx.x == 0)
|
|
result[r] = OpType::postProcess(sPartials[threadIdx.x], tadLength, extraParams);
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
template <typename OpType>
|
|
__device__ void ReduceFunction<T>::aggregatePartials(T *sPartials, Nd4jLong tid, Nd4jLong numItems, T *extraParams) {
|
|
// 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.
|
|
Nd4jLong floorPow2 = numItems;
|
|
|
|
if (floorPow2 & (floorPow2 - 1)) {
|
|
while (floorPow2 & (floorPow2 - 1)) {
|
|
floorPow2 &= floorPow2 - 1;
|
|
}
|
|
if (tid >= floorPow2) {
|
|
sPartials[tid - floorPow2] = OpType::update(sPartials[tid - floorPow2], sPartials[tid], extraParams);
|
|
}
|
|
|
|
__syncthreads();
|
|
}
|
|
|
|
|
|
for (Nd4jLong activeThreads = floorPow2 >> 1; activeThreads; activeThreads >>= 1) {
|
|
if (tid < activeThreads && tid + activeThreads < numItems) {
|
|
sPartials[tid] = OpType::update(sPartials[tid], sPartials[tid + activeThreads], extraParams);
|
|
}
|
|
__syncthreads();
|
|
}
|
|
}
|
|
|
|
#ifndef __CLION_IDE__
|
|
BUILD_CALL_1(template __device__ void ReduceFunction<float>::execScalarCuda, float, (float*, Nd4jLong*, float*, float*, Nd4jLong*, float*, UnifiedSharedMemory *, Nd4jLong*), REDUCE_OPS)
|
|
BUILD_CALL_1(template __device__ void ReduceFunction<float16>::execScalarCuda, float16, (float16*, Nd4jLong*, float16*, float16*, Nd4jLong*, float16*, UnifiedSharedMemory *, Nd4jLong*), REDUCE_OPS)
|
|
BUILD_CALL_1(template __device__ void ReduceFunction<double>::execScalarCuda, double, (double*, Nd4jLong*, double*, double*, Nd4jLong*, double*, UnifiedSharedMemory *, Nd4jLong*), REDUCE_OPS)
|
|
|
|
BUILD_CALL_1(template __device__ void ReduceFunction<float>::aggregatePartials, float, (float*, Nd4jLong, Nd4jLong, float*), REDUCE_OPS)
|
|
BUILD_CALL_1(template __device__ void ReduceFunction<float16>::aggregatePartials, float16, (float16*, Nd4jLong, Nd4jLong, float16*), REDUCE_OPS)
|
|
BUILD_CALL_1(template __device__ void ReduceFunction<double>::aggregatePartials, double, (double*, Nd4jLong, Nd4jLong, double*), REDUCE_OPS)
|
|
#endif
|
|
}
|
|
} |