raver119 c969b724bb [WIP] more CUDA stuff (#57)
* initial commit

Signed-off-by: raver119 <raver119@gmail.com>

* Added gradcheck test for dynamic_partition_bp op.

* - implementation of dilation op (cpu and cuda)

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed broadcast_dynamic_shape 1D case and tests.

* Fixed usage of default integer arguments.

* Fixed dynamic_partition_bp op and tests.

* Eliminated test with grad check for dynamic_partition_bp op.

* start working on cuda svd - porting available corresponding api from cuSOLVER library

Signed-off-by: Yurii <yurii@skymind.io>

* provide prelu_bp

Signed-off-by: Yurii <yurii@skymind.io>

* - provide gruCell_bp (old version ??)

Signed-off-by: Yurii <yurii@skymind.io>

* - polishing cumsum_bp and cumprod_bp tests

Signed-off-by: Yurii <yurii@skymind.io>

* provide sparseSoftmaxCrossEntropyWithLogits and sparseSoftmaxCrossEntropyWithLogits_grad

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed atomicMul with float input/output

* implementation of cuda kernel for triu_bp operation

Signed-off-by: Yurii <yurii@skymind.io>

* Refactored lup helper to add parrallel computing.

* cusolver libraries

Signed-off-by: raver119 <raver119@gmail.com>

* uncomment cuSolver APIs in svd.cu

Signed-off-by: Yurii <yurii@skymind.io>

* cusolver var

Signed-off-by: raver119 <raver119@gmail.com>

* - further work on cuSolver svd

Signed-off-by: Yurii <yurii@skymind.io>

* Implement usage of cuda solver to LUP decomposition.

* - correct naames in lup functions

Signed-off-by: Yurii <yurii@skymind.io>

* correct svdQR cuda

Signed-off-by: Yurii <yurii@skymind.io>

* - provide transpositions of input matrices in case of c order in svdCudaQR

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed implementation issues with LUP usign cuda solver.

* Implementation of matrix_determinant helper with cuda kernels. Working revision.

* Implemented log_matrix_determinant helper with cuda kernels.

* - implementation of batched cuda svd

Signed-off-by: Yurii <yurii@skymind.io>

* Refactored cholesky helper and implementation of cuda solver cholesky batch.

* - implementation of cuda kernel for tile bp

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of cholesky and logdet with cuda kernels.

* - implementation of cuda kernel for sru_bidirectional

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed cholesky helper.

* Cholesky op helper implementation. Working double-based cublas implementation.

* bad import excluded

Signed-off-by: raver119 <raver119@gmail.com>

* Finished with cuda implementation of cholesky helper and tests.

* - implementation of cuda kernel for sru_bidirectional_backprop operation

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of matrix_inverse op helper with cuda kernels. The first revision.

* - start working on gruCell_bp

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of matrix_inverse helper.

* - further work on new gruCell_bp

Signed-off-by: Yurii <yurii@skymind.io>

* cuBLAS related fixes

Signed-off-by: raver119 <raver119@gmail.com>

* calculateOutputShapes() now passes device buffers as well

Signed-off-by: raver119 <raver119@gmail.com>

* special concat/average/accumulate init host pointers now

Signed-off-by: raver119 <raver119@gmail.com>

* few more tweaks

Signed-off-by: raver119 <raver119@gmail.com>

* additional CudaDataBufferFactory signatures certain for data types

Signed-off-by: raver119 <raver119@gmail.com>

* cuSolver host buffer

Signed-off-by: raver119 <raver119@gmail.com>

* buffer to buffer memcpy host ptr allocation

Signed-off-by: raver119 <raver119@gmail.com>
2019-07-20 23:05:21 +10:00

400 lines
15 KiB
C++

/*******************************************************************************
* 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 (iuriish@yahoo.com), created on 19.04.2018
// @author raver119@gmail.com
//
#include <ops/declarable/helpers/activations.h>
#include <ShapeUtils.h>
#include <numeric>
#include <ConstantTadHelper.h>
namespace nd4j {
namespace ops {
namespace helpers {
template <typename T>
static void softMaxForVector_(void *input, Nd4jLong *inShapeInfo, void *output, Nd4jLong *outShapeInfo) {
T* inBuff = reinterpret_cast<T *>(input);
T* outBuff = reinterpret_cast<T *>(output);
T max = -DataTypeUtils::max<T>();
T sum = 0.;
int inEWS = shape::elementWiseStride(inShapeInfo);
int outEWS = shape::elementWiseStride(outShapeInfo);
int length = shape::length(inShapeInfo);
if (inEWS >= 1 && outEWS >= 1) {
if (inEWS == 1 && outEWS == 1) {
PRAGMA_OMP_SIMD_MAX(max)
for (int i = 0; i < length; i++)
max = nd4j::math::nd4j_max<T>(max, inBuff[i]);
PRAGMA_OMP_SIMD_SUM(sum)
for (int i = 0; i < length; i++) {
outBuff[i] = nd4j::math::nd4j_exp<T, T>(inBuff[i] - max);
sum += outBuff[i];
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++)
outBuff[i] /= sum;
}
else {
PRAGMA_OMP_SIMD_MAX(max)
for (int i = 0; i < length; i++)
max = nd4j::math::nd4j_max<T>(max, inBuff[i * inEWS]);
PRAGMA_OMP_SIMD_SUM(sum)
for (int i = 0; i < length; i++) {
T r = nd4j::math::nd4j_exp<T, T>(inBuff[i * inEWS] - max);
outBuff[i * outEWS] = r;
sum += r;
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++)
outBuff[i * outEWS] /= sum;
}
}
}
///////////////////////////////////////////////////////////////////
template <typename T>
void static _softMaxDerivForVector(nd4j::LaunchContext * context, const void *input, const Nd4jLong *inShapeInfo, void *output) {
const T* inBuff = reinterpret_cast<const T *>(input);
T* outBuff = reinterpret_cast<T *>(output);
T max = -DataTypeUtils::max<T>();
T sum = 0.;
int length = shape::length(inShapeInfo);
PRAGMA_OMP_SIMD_ARGS(reduction(OMP_MAXT:max))
for (int i = 0; i < length; i++) {
const Nd4jLong offset = shape::getIndexOffset(i, inShapeInfo, length);
max = nd4j::math::nd4j_max<T>(max, inBuff[offset]);
}
PRAGMA_OMP_PARALLEL_FOR_SIMD_ARGS(reduction(OMP_SUMT:sum))
for (int i = 0; i < length; i++) {
const Nd4jLong offset = shape::getIndexOffset(i, inShapeInfo, length);
outBuff[offset] = nd4j::math::nd4j_exp<T, T>(inBuff[offset] - max);
sum += outBuff[offset];
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++) {
const Nd4jLong offset = shape::getIndexOffset(i, inShapeInfo, length);
outBuff[offset] /= sum;
outBuff[offset] *= (1.f - outBuff[offset]); // derivative
}
}
///////////////////////////////////////////////////////////////////
void softmaxDerivative(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
const int rank = input.rankOf();
int temp;
if(shape::isCommonVector(input.getShapeInfo(), temp)) {
BUILD_SINGLE_SELECTOR(input.dataType(), _softMaxDerivForVector, (context, input.getBuffer(), input.getShapeInfo(), output.buffer()), FLOAT_TYPES);
}
else {
auto maxAlongDim = const_cast<NDArray&>(input).reduceAlongDims(reduce::Max, {dimension}, true);
(input - maxAlongDim).applyTransform(transform::Exp, &output); // output contains exponents temporarily
auto sumAlongDim = output.reduceAlongDims(reduce::Sum, {dimension}, true);
output /= sumAlongDim;
output *= (1.f - output); // derivative
}
}
///////////////////////////////////////////////////////////////////
void softMaxForVector(nd4j::LaunchContext * context, const NDArray& input, NDArray& output) {
if(!input.isVector() || !output.isVector())
throw std::runtime_error("ops::helpers::softMaxForVector function: input and output arrays must be vectors !");
auto xType = input.dataType();
BUILD_SINGLE_SELECTOR(xType, softMaxForVector_, (input.getBuffer(), input.getShapeInfo(), output.buffer(), output.shapeInfo()), FLOAT_TYPES);
}
///////////////////////////////////////////////////////////////////
template <typename T>
void logSoftMaxForVector_(void *input, Nd4jLong *inShapeInfo, void *output, Nd4jLong *outShapeInfo) {
auto inBuff = reinterpret_cast<T *>(input);
auto outBuff = reinterpret_cast<T *>(output);
T max = -DataTypeUtils::max<T>();
T sum = 0;
auto inEWS = shape::elementWiseStride(inShapeInfo);
auto length = shape::length(inShapeInfo);
if (inEWS == 1) {
PRAGMA_OMP_SIMD_MAX(max)
for (int i = 0; i < length; i++)
max = nd4j::math::nd4j_max<T>(max, inBuff[i]);
PRAGMA_OMP_SIMD_SUM(sum)
for (int i = 0; i < length; i++) {
outBuff[i] = nd4j::math::nd4j_exp<T,T>(inBuff[i] - max);
sum += outBuff[i];
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++) {
outBuff[i] /= sum;
outBuff[i] = nd4j::math::nd4j_log<T,T>(outBuff[i]);
}
}
else if (inEWS > 1) {
PRAGMA_OMP_SIMD_MAX(max)
for (int i = 0; i < length; i++)
max = nd4j::math::nd4j_max<T>(max, inBuff[i * inEWS]);
PRAGMA_OMP_SIMD_SUM(sum)
for (int i = 0; i < length; i++) {
outBuff[i * inEWS] = nd4j::math::nd4j_exp<T,T>(inBuff[i * inEWS] - max);
sum += outBuff[i * inEWS];
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++) {
outBuff[i * inEWS] /= sum;
outBuff[i * inEWS] = nd4j::math::nd4j_log<T, T>(outBuff[i * inEWS]);
}
}
}
///////////////////////////////////////////////////////////////////
void logSoftMaxForVector(nd4j::LaunchContext * context, const NDArray& input, NDArray& output) {
if(!input.isVector() || !output.isVector())
throw std::runtime_error("ops::helpers::logSoftMaxForVector function input and output arrays must be vectors !");
auto xType = input.dataType();
BUILD_SINGLE_SELECTOR(xType, logSoftMaxForVector_, (input.getBuffer(), input.getShapeInfo(), output.buffer(), output.shapeInfo()), FLOAT_TYPES);
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void softmax_(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
const int rank = input.rankOf();
if(input.isVector()) {
if(rank == 1 || input.sizeAt(dimension) != 1)
softMaxForVector_<T>(input.getBuffer(), input.getShapeInfo(), output.buffer(), output.getShapeInfo());
else
output = 1.;
}
else if(input.isSameShapeStrict(&output)) {
TadPack tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input.getShapeInfo(), {dimension});
Nd4jLong* tadShapeInfo = tadPack.primaryShapeInfo();
Nd4jLong* tadOffsets = tadPack.primaryOffsets();
const uint numOfSubArrs = tadPack.numberOfTads();
const uint tadLen = shape::length(tadShapeInfo);
if(shape::elementWiseStride(tadShapeInfo) == 1){
PRAGMA_OMP_PARALLEL_FOR_SIMD
for (uint i = 0; i < numOfSubArrs; ++i) {
T* inBuff = input.bufferAsT<T>() + tadOffsets[i];
T* outBuff = output.bufferAsT<T>() + tadOffsets[i];
T max = -DataTypeUtils::max<T>();
T sum = 0;
for(uint j = 0; j < tadLen; ++j)
max = nd4j::math::nd4j_max<T>(max, inBuff[j]);
for (uint j = 0; j < tadLen; ++j) {
T temp = nd4j::math::nd4j_exp<T,T>(inBuff[j] - max);
outBuff[j] = temp;
sum += temp;
}
for (uint j = 0; j < tadLen; ++j)
outBuff[j] /= sum;
}
}
else {
uint inShapeInfoCast[MAX_RANK];
bool canCast = nd4j::DataTypeUtils::castShapeInfo(tadShapeInfo, inShapeInfoCast);
auto offsets = new Nd4jLong[tadLen];
shape::calcOffsets(tadShapeInfo, offsets);
PRAGMA_OMP_PARALLEL_FOR_SIMD
for (uint i = 0; i < numOfSubArrs; ++i) {
T* inBuff = input.bufferAsT<T>() + tadOffsets[i];
T* outBuff = output.bufferAsT<T>() + tadOffsets[i];
T max = -DataTypeUtils::max<T>();
T sum = 0.f;
for(uint j = 0; j < tadLen; ++j)
max = nd4j::math::nd4j_max<T>(max, inBuff[offsets[j]]);
for (uint j = 0; j < tadLen; ++j) {
T temp = nd4j::math::nd4j_exp<T,T>(inBuff[offsets[j]] - max);
outBuff[offsets[j]] = temp;
sum += temp;
}
for (uint j = 0; j < tadLen; ++j)
outBuff[offsets[j]] /= sum;
}
delete []offsets;
}
}
else {
NDArray max = input.reduceAlongDims(nd4j::reduce::Max, {dimension}, true);
input.applyTrueBroadcast(nd4j::BroadcastOpsTuple::Subtract(), &max, &output, false);
output.applyTransform(nd4j::transform::Exp);
NDArray sum = output.reduceAlongDims(nd4j::reduce::Sum, {dimension}, true);
output /= sum;
}
}
///////////////////////////////////////////////////////////////////
void softmax(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
BUILD_SINGLE_SELECTOR(input.dataType(), softmax_, (context, input, output, dimension), FLOAT_TYPES);
}
//////////////////////////////////////////////////////////////////////////
void prelu(nd4j::LaunchContext * context, const NDArray& input, const NDArray& alpha, NDArray& output) {
const Nd4jLong inputLen = input.lengthOf();
const Nd4jLong* inputShapeInfo = input.getShapeInfo();
const Nd4jLong* alphaShapeInfo = alpha.getShapeInfo();
PRAGMA_OMP_PARALLEL_FOR_IF(inputLen > Environment::getInstance()->elementwiseThreshold())
for(Nd4jLong i = 0; i < inputLen; ++i) {
// FIXME: double!
double x = input.e<double>(i);
if(x < 0.0) {
// FIXME: double
output.p(i, (x * alpha.e<double>(shape::subArrayIndex(i, inputShapeInfo, alphaShapeInfo))));
} else
output.p(i, x);
}
}
//////////////////////////////////////////////////////////////////////////
void preluBP(nd4j::LaunchContext * context, const NDArray& input, const NDArray& alpha, const NDArray& dLdO, NDArray& dLdI, NDArray& dLdA) {
const Nd4jLong inputLen = input.lengthOf();
const Nd4jLong* inputShapeInfo = input.getShapeInfo();
const Nd4jLong* alphaShapeInfo = alpha.getShapeInfo();
dLdA.assign(0.0f);
for(Nd4jLong i = 0; i < inputLen; ++i) {
// FIXME: double
double x = input.e<double>(i);
double grO = dLdO.e<double>(i);
if(x < 0.0) {
Nd4jLong alphaInd = shape::subArrayIndex(i, inputShapeInfo, alphaShapeInfo);
dLdI.p(i, grO * alpha.e<double>(alphaInd));
double prevVal = dLdA.e<double>(alphaInd);
prevVal += (grO * x);
dLdA.p(alphaInd, prevVal);
}
else
dLdI.p(i, grO);
}
}
bool checkAlphaShapeLen(std::vector<Nd4jLong> const& expectedShape, Nd4jLong shapeLen) {
Nd4jLong expectedAlphaLen = std::accumulate(expectedShape.cbegin(), expectedShape.cend(), 1, std::multiplies<Nd4jLong>());
return expectedAlphaLen == shapeLen;
}
template <typename T>
static void thresholdRelu_(NDArray const& input, double threshold, NDArray& output) {
auto routine = LAMBDA_T(_x, threshold) {
return _x > (T)threshold? _x: (T)0.f;
};
const_cast<NDArray&>(input).applyLambda<T>(routine, &output);
}
void thresholdRelu(nd4j::LaunchContext * context, NDArray const& input, double threshold, NDArray& output) {
BUILD_SINGLE_SELECTOR(input.dataType(), thresholdRelu_, (input, threshold, output), FLOAT_TYPES);
}
template <typename T>
static void thresholdReluDerivative_(nd4j::LaunchContext * context, NDArray* input, double theta, NDArray* dLdO, NDArray* output) {
auto derivative = LAMBDA_TT(_x, grO, theta) {if (_x > theta) return grO; else return static_cast<T>(0); };
input->applyPairwiseLambda<T>(dLdO, derivative, output);
}
void thresholdReluDerivative(nd4j::LaunchContext * context, NDArray* input, double threshold, NDArray* dLdO, NDArray* output) {
BUILD_SINGLE_SELECTOR(input->dataType(), thresholdReluDerivative_, (context, input, threshold, dLdO, output), FLOAT_TYPES);
}
///////////////////////////////////////////////////////////////////
void logSoftmax(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
const int rank = input.rankOf();
if(input.isVector()) {
if(rank == 1 || input.sizeAt(dimension) != 1) {
BUILD_SINGLE_SELECTOR(input.dataType(), logSoftMaxForVector_, (input.getBuffer(), input.getShapeInfo(), output.buffer(), output.shapeInfo()), FLOAT_TYPES);
}
else
output = 0.;
}
else {
auto maxAlongDim = const_cast<NDArray&>(input).reduceAlongDims(reduce::Max, {dimension}, true);
(input - maxAlongDim).applyTransform(transform::Exp, &output); // output contains exponents temporarily
auto sumAlongDim = output.reduceAlongDims(reduce::Sum, {dimension}, true);
output /= sumAlongDim;
output.applyTransform(transform::Log);
}
}
BUILD_SINGLE_TEMPLATE(template void thresholdReluDerivative_, (nd4j::LaunchContext * context, NDArray* input, double threshold, NDArray* dLdO, NDArray* output), FLOAT_TYPES);
BUILD_SINGLE_TEMPLATE(template void softmax_, (nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension), FLOAT_TYPES);
BUILD_SINGLE_TEMPLATE(template void logSoftMaxForVector_, (void *input, Nd4jLong *inShapeInfo, void *output, Nd4jLong *outShapeInfo), FLOAT_TYPES);
BUILD_SINGLE_TEMPLATE(template void _softMaxDerivForVector, (nd4j::LaunchContext * context, const void *input, const Nd4jLong *inShapeInfo, void *output), FLOAT_TYPES);
}
}
}