cavis/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu

236 lines
11 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 16.04.2018
//
#include <ops/declarable/helpers/reverse.h>
#include <helpers/ShapeUtils.h>
#include <array/ResultSet.h>
#include <helpers/TAD.h>
#include <helpers/PointersManager.h>
#include <helpers/ConstantTadHelper.h>
namespace sd {
namespace ops {
namespace helpers {
template <typename T>
static __global__ void reverseTadKernel(const void* vinput, const Nd4jLong *inputShape, void* voutput, const Nd4jLong *outputShape, const Nd4jLong *inputTadShape, const Nd4jLong *inputTadOffsets, const Nd4jLong *outputTadShape, const Nd4jLong *outputTadOffsets, uint64_t limit, uint64_t numOfElemsToReverse, uint64_t numTads) {
auto input = reinterpret_cast<const T*>(vinput);
auto output = reinterpret_cast<T*>(voutput);
const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
const auto step = gridDim.x * blockDim.x;
// this means that we'll have additional cycle, to move middle element
auto div = numOfElemsToReverse / 2;
auto odd = numOfElemsToReverse % 2 != 0;
auto rlimit = odd ? limit / 2 + 1 : limit / 2;
// all threads operate in the same input/output space
for (uint64_t e = tid; e < rlimit; e += step) {
// finding out the TAD we're going to process
auto tadId = e / div;
if (tadId >= numTads)
continue;
// now finding out element within tad
auto idx = e % div;
//printf("TID: %i; numTads: %lld; tadLength: %lld; tadId: %i, idx: %lld\n", tid, numTads, numOfElemsToReverse, tadId, idx);
auto tadInput = input + inputTadOffsets[tadId];
auto tadOutput = output + outputTadOffsets[tadId];
// we're calculating offsets within input TAD
auto fOffset = shape::getIndexOffset(idx, inputTadShape);
auto lOffset = shape::getIndexOffset(numOfElemsToReverse - idx - 1, inputTadShape);
// now we're storing input values
auto v1 = tadInput[fOffset];
auto v2 = tadInput[lOffset];
// now we're calculating offsets within output TAD
auto zfOffset = shape::getIndexOffset(idx, outputTadShape);
auto zlOffset = shape::getIndexOffset(numOfElemsToReverse - idx - 1, outputTadShape);
// and saving values to output arrays
tadOutput[zfOffset] = v2;
tadOutput[zlOffset] = v1;
}
// moving odd element in blocks
if (odd && threadIdx.x == 0) {
for (uint64_t e = blockIdx.x; e < numTads; e += gridDim.x) {
auto tadInput = input + inputTadOffsets[e];
auto tadOutput = output + outputTadOffsets[e];
auto xOffset = shape::getIndexOffset(numOfElemsToReverse / 2, inputTadShape);
auto zOffset = shape::getIndexOffset(numOfElemsToReverse / 2, outputTadShape);
tadOutput[zOffset] = tadInput[xOffset];
}
}
}
template <typename T>
static __global__ void reverseArrayKernel(const void* input, const Nd4jLong *inputShape, void* output, const Nd4jLong *outputShape, Nd4jLong numOfElemsToReverse) {
const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
const auto step = gridDim.x * blockDim.x;
__shared__ int linearStatus;
__shared__ const T* inputArr;
__shared__ T* outputArr;
__shared__ char inputOrder, outputOrder;
if (threadIdx.x == 0) {
linearStatus = (shape::elementWiseStride(inputShape) == shape::elementWiseStride(outputShape)) && (inputOrder == outputOrder)? shape::elementWiseStride(inputShape):0;
char inputOrder = shape::order(inputShape);
char outputOrder = shape::order(outputShape);
inputArr = reinterpret_cast<const T*>(input);
outputArr = reinterpret_cast<T*>(output);
}
__syncthreads();
auto odd = numOfElemsToReverse % 2 != 0;
auto limit = numOfElemsToReverse / 2;
for (uint64_t e = tid; e < limit; e += step) {
// we're calculating offsets within input array
auto fOffset = shape::getIndexOffset(e, inputShape);
auto lOffset = shape::getIndexOffset(numOfElemsToReverse - e - 1, inputShape);
// now we're storing input values
auto v1 = inputArr[fOffset];
auto v2 = inputArr[lOffset];
// now we're calculating offsets within output array
auto zfOffset = shape::getIndexOffset(e, outputShape);
auto zlOffset = shape::getIndexOffset(numOfElemsToReverse - e - 1, outputShape);
// and saving values to output arrays
outputArr[zfOffset] = v2;
outputArr[zlOffset] = v1;
}
// in case of odd array we'll have to move middle value
if (odd && tid == 0) {
auto xOffset = shape::getIndexOffset(limit, inputShape);
auto zOffset = shape::getIndexOffset(limit, outputShape);
outputArr[zOffset] = inputArr[xOffset];
}
}
template<typename T>
static void reverseTad(sd::LaunchContext * context, const NDArray* input, NDArray* output, const Nd4jLong *inputTadShape, const Nd4jLong *inputTadOffsets, const Nd4jLong *outputTadShape, const Nd4jLong *outputTadOffsets, uint64_t tadLength) {
auto stream = context->getCudaStream();
reverseTadKernel<T><<<256, 512, 8192, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), output->specialBuffer(), output->specialShapeInfo(), inputTadShape, inputTadOffsets, outputTadShape, outputTadOffsets, input->lengthOf(), tadLength, input->lengthOf() / tadLength);
}
template<typename T>
static void reverseArray(sd::LaunchContext * context, const NDArray* input, NDArray* output, Nd4jLong numOfElemsToReverse) {
auto stream = context->getCudaStream();
Nd4jLong numOfReverse = numOfElemsToReverse;
if (numOfElemsToReverse == 0)
numOfReverse = input->lengthOf();
reverseArrayKernel<T><<<256, 512, 8192, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), output->specialBuffer(), output->specialShapeInfo(), numOfReverse);
}
///////////////////////////////////////////////////////////////////
template <typename T>
static void reverseSequence_(sd::LaunchContext * context, const NDArray* input, const NDArray* seqLengths, NDArray* output, int seqDim, const int batchDim){
int posOfNonUnityDim = -1;
seqLengths->syncToHost();
auto stream = context->getCudaStream();
if(input->isVector() || shape::isLikeVector(input->shapeInfo(), posOfNonUnityDim) || seqLengths->lengthOf() == 1) {
int numOfElemsToReverse = seqLengths->e<int>(0);
if((seqDim == 0 && input->sizeAt(0) == 1) || (batchDim == posOfNonUnityDim))
output->assign(input);
else
reverseArrayKernel<T><<<256, 512, 8192, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), output->specialBuffer(), output->specialShapeInfo(), numOfElemsToReverse);//helpers::reverseArray<T>(context, const_cast<NDArray*>(input), output, numOfElemsToReverse);
}
else {
if(seqDim > batchDim)
--seqDim;
std::vector<int> dimensions = ShapeUtils::evalDimsToExclude(input->rankOf(), {batchDim});
auto inSubArrsSet = input->allTensorsAlongDimension(dimensions);
auto outSubArrsSet = output->allTensorsAlongDimension(dimensions);
for(int i = 0; i < inSubArrsSet.size(); ++i) {
int numOfElemsToReverse = seqLengths->e<int>(i);
if(numOfElemsToReverse == 0 || numOfElemsToReverse == 1) {
outSubArrsSet.at(i)->assign(inSubArrsSet.at(i));
}
else {
auto inInnerSet = inSubArrsSet.at(i)->allTensorsAlongDimension({seqDim});
auto outInnerSet = outSubArrsSet.at(i)->allTensorsAlongDimension({seqDim});
for(int j = 0; j < inInnerSet.size(); ++j)
reverseArray<T>(context, inInnerSet.at(j), outInnerSet.at(j), numOfElemsToReverse);
}
}
}
}
void reverseSequence(sd::LaunchContext * context, const NDArray* input, const NDArray* seqLengths, NDArray* output, int seqDim, const int batchDim) {
NDArray::prepareSpecialUse({output}, {input, seqLengths});
// if op isn't inplace - copy original data into output array
if (output->specialBuffer() != input->specialBuffer())
output->assign(input);
BUILD_SINGLE_SELECTOR(input->dataType(), reverseSequence_, (context, input, seqLengths, output, seqDim, batchDim), LIBND4J_TYPES);
NDArray::registerSpecialUse({output}, {input, seqLengths});
}
//////////////////////////////////////////////////////////////////////////
void reverse(sd::LaunchContext * context, const NDArray* input, NDArray* output, const std::vector<int>* intArgs, bool isBackProp) {
// we need to reverse axis only if that's new op
std::vector<int> dimensions = isBackProp ? ShapeUtils::evalDimsToExclude(input->rankOf(), *intArgs) : *intArgs;
std::vector<int> axis = ShapeUtils::evalDimsToExclude(input->rankOf(), dimensions);
auto packX = sd::ConstantTadHelper::getInstance()->tadForDimensions(input->shapeInfo(), dimensions);
auto packZ = sd::ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), dimensions);
NDArray::prepareSpecialUse({output}, {input});
if (packX.numberOfTads() == 1) {
BUILD_SINGLE_SELECTOR(input->dataType(), reverseArray, (context, input, output, 0), LIBND4J_TYPES);
} else {
BUILD_SINGLE_SELECTOR(input->dataType(), reverseTad, (context, input, output, packX.platformShapeInfo(), packX.platformOffsets(), packZ.platformShapeInfo(), packZ.platformOffsets(), (uint64_t) (input->lengthOf() / packX.numberOfTads())), LIBND4J_TYPES);
}
NDArray::registerSpecialUse({output}, {input});
}
}
}
}