cavis/libnd4j/include/helpers/cuda/ConstantShapeHelper.cu
Samuel Audet 029b84e2b7
Development updates (#9053)
* RL4J: Add generic update rule (#502)

Signed-off-by: Alexandre Boulanger <aboulang2002@yahoo.com>

* Shyrma reduce (#481)

* - start working on improving of cpu legacy code for reduce ops

Signed-off-by: Yurii <iuriish@yahoo.com>

* - further work on improving legacy loops

Signed-off-by: Yurii <iuriish@yahoo.com>

* - still working on improving reduce ops

Signed-off-by: Yurii <iuriish@yahoo.com>

* - further work on improving reduce ops

Signed-off-by: Yurii <iuriish@yahoo.com>

* - testing speed run of new reduce op

Signed-off-by: Yurii <iuriish@yahoo.com>

* - working on improvement of default loop for reduce op

Signed-off-by: Yurii <iuriish@yahoo.com>

* - update signatures of stuff which calls reduce ops

Signed-off-by: Yurii <iuriish@yahoo.com>

* - make corrections in cuda reduce kernels

Signed-off-by: Yurii <iuriish@yahoo.com>

* - change loop for default case in broadcast legacy ops

Signed-off-by: Yurii <iuriish@yahoo.com>

* - comment some shape stuff

Signed-off-by: Yurii <iuriish@yahoo.com>

* - comment unnecessary prints in RNGtests

Signed-off-by: Yurii <iuriish@yahoo.com>

* - finish to resolve conflicts after master has been merged

Signed-off-by: Yurii <iuriish@yahoo.com>

* - get rid of some compilation mistakes of cuda stuff

Signed-off-by: Yurii <iuriish@yahoo.com>

* - minor changes

Signed-off-by: Yurii <iuriish@yahoo.com>

* - further search for bug causing crash on java test

Signed-off-by: Yurii <iuriish@yahoo.com>

* - add scalar case in reduce_ ... exec stuff

Signed-off-by: Yurii <iuriish@yahoo.com>

* - minor corrections in NAtiveOps.cu

Signed-off-by: Yurii <iuriish@yahoo.com>

* - add switch to scalar case execReduceXD functions

Signed-off-by: Yurii <iuriish@yahoo.com>

* - add support for vectors old shape in ConstantShapeHelper::createShapeInfoWithNoUnitiesForReduce

Signed-off-by: Yurii <iuriish@yahoo.com>

* - correct cuda mirrorPad

Signed-off-by: Yurii <iuriish@yahoo.com>

* - add support for vectors old shape in cuda createShapeInfoWithNoUnitiesForReduce

Signed-off-by: Yurii <iuriish@yahoo.com>

Co-authored-by: raver119 <raver119@gmail.com>

* Add support for CUDA 11.0 (#492)

* Add support for CUDA 11.0

* libnd4j tweaks for CUDA 11

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

* bindings update, again?

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

* * Update versions of JavaCPP Presets for FFmpeg, OpenBLAS, and NumPy

* update API to match CUDA 8

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

* * Update version of JavaCPP Presets for CPython

* C++ updated for cuDNN 8.0

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

* one more test

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

* one more test

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

* one more test

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

* 128-bit alignment for workspaces

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

* change seed in 1 test

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

* Fix dependecy duplication in python4j-parent pom

* Fix group id for in python4j-numpy

* few tests tweaked

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

* Remove macosx-x86_64-gpu from nd4j-tests-tensorflow

* few minor tweaks for IndexReduce

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

* one test removed

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

Co-authored-by: raver119@gmail.com <raver119@gmail.com>
Co-authored-by: Serhii Shepel <9946053+sshepel@users.noreply.github.com>

* RL4J: Add SyncTrainer and AgentLearnerBuilder for a few algorithms (#504)

Signed-off-by: Alexandre Boulanger <aboulang2002@yahoo.com>

Co-authored-by: Alexandre Boulanger <44292157+aboulang2002@users.noreply.github.com>
Co-authored-by: Yurii Shyrma <iuriish@yahoo.com>
Co-authored-by: raver119 <raver119@gmail.com>
Co-authored-by: Serhii Shepel <9946053+sshepel@users.noreply.github.com>
2020-07-26 21:59:27 +09:00

225 lines
9.1 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 "../ConstantShapeHelper.h"
#include <exceptions/cuda_exception.h>
#include <array/ShapeDescriptor.h>
#include <helpers/ShapeBuilders.h>
#include <execution/AffinityManager.h>
#include <helpers/ConstantHelper.h>
#include <helpers/ShapeUtils.h>
#include <array/PrimaryPointerDeallocator.h>
#include <array/CudaPointerDeallocator.h>
namespace sd {
ConstantShapeHelper::ConstantShapeHelper() {
auto numDevices = AffinityManager::numberOfDevices();
_cache.resize(numDevices);
for (int e = 0; e < numDevices; e++) {
MAP_IMPL<ShapeDescriptor, ConstantShapeBuffer> cache;
_cache[e] = cache;
}
}
ConstantShapeHelper& ConstantShapeHelper::getInstance() {
static ConstantShapeHelper instance;
return instance;
}
ConstantShapeBuffer& ConstantShapeHelper::bufferForShapeInfo(sd::DataType dataType, char order, const std::vector<Nd4jLong> &shape) {
ShapeDescriptor descriptor(dataType, order, shape);
return bufferForShapeInfo(descriptor);
}
ConstantShapeBuffer& ConstantShapeHelper::bufferForShapeInfo(const sd::DataType dataType, const char order, const int rank, const Nd4jLong* shape) {
ShapeDescriptor descriptor(dataType, order, shape, rank);
return bufferForShapeInfo(descriptor);
}
ConstantShapeBuffer& ConstantShapeHelper::bufferForShapeInfo(const ShapeDescriptor &descriptor) {
int deviceId = AffinityManager::currentDeviceId();
std::lock_guard<std::mutex> lock(_mutex);
if (_cache[deviceId].count(descriptor) == 0) {
auto hPtr = std::make_shared<PointerWrapper>(descriptor.toShapeInfo(), std::make_shared<PrimaryPointerDeallocator>());
auto dPtr = std::make_shared<PointerWrapper>(ConstantHelper::getInstance().replicatePointer(hPtr->pointer(), shape::shapeInfoByteLength(hPtr->pointerAsT<Nd4jLong>())), std::make_shared<CudaPointerDeallocator>());
ConstantShapeBuffer buffer(hPtr, dPtr);
ShapeDescriptor descriptor1(descriptor);
_cache[deviceId][descriptor1] = buffer;
return _cache[deviceId][descriptor1];
} else {
return _cache[deviceId].at(descriptor);
}
}
ConstantShapeBuffer& ConstantShapeHelper::bufferForShapeInfo(const Nd4jLong *shapeInfo) {
ShapeDescriptor descriptor(shapeInfo);
return bufferForShapeInfo(descriptor);
}
bool ConstantShapeHelper::checkBufferExistenceForShapeInfo(ShapeDescriptor &descriptor) {
auto deviceId = AffinityManager::currentDeviceId();
std::lock_guard<std::mutex> lock(_mutex);
return _cache[deviceId].count(descriptor) != 0;
}
Nd4jLong const* ConstantShapeHelper::createShapeInfo(const sd::DataType dataType, const char order, const int rank, const Nd4jLong* shape) {
ShapeDescriptor descriptor(dataType, order, shape, rank);
return bufferForShapeInfo(descriptor).primary();
}
Nd4jLong const* ConstantShapeHelper::createShapeInfo(const sd::DataType dataType, const Nd4jLong* shapeInfo) {
return ConstantShapeHelper::createShapeInfo(dataType, shape::order(shapeInfo), shape::rank(shapeInfo), shape::shapeOf(const_cast<Nd4jLong*>(shapeInfo)));
}
Nd4jLong const* ConstantShapeHelper::emptyShapeInfo(const sd::DataType dataType) {
auto descriptor = ShapeDescriptor::emptyDescriptor(dataType);
return bufferForShapeInfo(descriptor).primary();
}
Nd4jLong const* ConstantShapeHelper::scalarShapeInfo(const sd::DataType dataType) {
auto descriptor = ShapeDescriptor::scalarDescriptor(dataType);
return bufferForShapeInfo(descriptor).primary();
}
Nd4jLong const* ConstantShapeHelper::vectorShapeInfo(const Nd4jLong length, const sd::DataType dataType) {
auto descriptor = ShapeDescriptor::vectorDescriptor(length, dataType);
return bufferForShapeInfo(descriptor).primary();
}
Nd4jLong const* ConstantShapeHelper::createShapeInfo(const sd::DataType dataType, const char order, const std::vector<Nd4jLong> &shape) {
ShapeDescriptor descriptor(dataType, order, shape);
return bufferForShapeInfo(descriptor).primary();
}
Nd4jLong const* ConstantShapeHelper::createShapeInfo(const ShapeDescriptor &descriptor) {
return bufferForShapeInfo(descriptor).primary();
}
Nd4jLong const* ConstantShapeHelper::createFromExisting(Nd4jLong *shapeInfo, bool destroyOriginal) {
ShapeDescriptor descriptor(shapeInfo);
auto result = createShapeInfo(descriptor);
if (destroyOriginal)
RELEASE(shapeInfo, nullptr);
return result;
}
Nd4jLong const* ConstantShapeHelper::createFromExisting(Nd4jLong *shapeInfo, sd::memory::Workspace *workspace) {
ShapeDescriptor descriptor(shapeInfo);
auto result = createShapeInfo(descriptor);
RELEASE(shapeInfo, workspace);
return result;
}
////////////////////////////////////////////////////////////////////////
ConstantShapeBuffer& ConstantShapeHelper::createShapeInfoWithUnitiesForBroadcast(const Nd4jLong* maxShapeInfo, const Nd4jLong* minShapeInfo, sd::memory::Workspace* workspace, const std::vector<int>& dimensions) {
Nd4jLong* newShapeInfo = nullptr;
ALLOCATE(newShapeInfo, workspace, shape::shapeInfoLength(shape::rank(maxShapeInfo)), Nd4jLong);
newShapeInfo[0] = shape::rank(maxShapeInfo);
sd::ArrayOptions::copyDataType(newShapeInfo, minShapeInfo); // type
newShapeInfo[2 * newShapeInfo[0] + 2] = shape::elementWiseStride(minShapeInfo); // ews
newShapeInfo[2 * newShapeInfo[0] + 3] = shape::order(minShapeInfo); // order
if(!dimensions.empty()) {
for(uint k = 0, j = 0, i = 0; i < shape::rank(maxShapeInfo); ++i) {
if(j < dimensions.size() && dimensions[j] == i) {
shape::shapeOf(newShapeInfo)[i] = shape::shapeOf(minShapeInfo)[k];
shape::stride(newShapeInfo)[i] = shape::stride(minShapeInfo)[k++];
++j;
}
else{
shape::shapeOf(newShapeInfo)[i] = 1;
shape::stride(newShapeInfo)[i] = 0;
if(shape::sizeAt(minShapeInfo, k) == 1 && dimensions.size() != shape::rank(minShapeInfo))
++k;
}
}
}
else{
for(int j = shape::rank(minShapeInfo) - 1, i = shape::rank(maxShapeInfo) - 1; i >=0 ; --i) {
if(j >= 0) {
shape::shapeOf(newShapeInfo)[i] = shape::shapeOf(minShapeInfo)[j];
shape::stride(newShapeInfo)[i] = shape::shapeOf(minShapeInfo)[j] == 1 ? 0 : shape::stride(minShapeInfo)[j];
--j;
}
else {
shape::shapeOf(newShapeInfo)[i] = 1;
shape::stride(newShapeInfo)[i] = 0;
}
}
}
ShapeDescriptor descriptor(newShapeInfo);
RELEASE(newShapeInfo, workspace);
return bufferForShapeInfo(descriptor);
}
////////////////////////////////////////////////////////////////////////
ConstantShapeBuffer& ConstantShapeHelper::createShapeInfoWithNoUnitiesForReduce(const Nd4jLong* inShapeInfo, const std::vector<int> &dimsWithUnities, sd::memory::Workspace* workspace) {
Nd4jLong* newShapeInfo = nullptr;
ALLOCATE(newShapeInfo, workspace, shape::shapeInfoLength(shape::rank(inShapeInfo) - dimsWithUnities.size()), Nd4jLong);
int temp;
if(dimsWithUnities.size() == 1 && shape::isCommonVector(inShapeInfo, temp) && temp == dimsWithUnities[0]) {
auto dims = ShapeUtils::evalDimsToExclude(shape::rank(inShapeInfo), {temp});
shape::excludeUnitiesFromShapeInfo(inShapeInfo, dims.data(), dims.size(), newShapeInfo);
} else {
shape::excludeUnitiesFromShapeInfo(inShapeInfo, dimsWithUnities.data(), dimsWithUnities.size(), newShapeInfo);
}
ShapeDescriptor descriptor(newShapeInfo);
RELEASE(newShapeInfo, workspace);
return bufferForShapeInfo(descriptor);
}
////////////////////////////////////////////////////////////////////////
ConstantShapeBuffer& ConstantShapeHelper::createSubArrShapeInfo(const Nd4jLong* inShapeInfo, const int* dims, const int dimsSize, sd::memory::Workspace* workspace) {
Nd4jLong* newShapeInfo = ShapeBuilders::createSubArrShapeInfo(inShapeInfo, dims, dimsSize, workspace);
ShapeDescriptor descriptor(newShapeInfo);
RELEASE(newShapeInfo, workspace);
return bufferForShapeInfo(descriptor);
}
}