cavis/libnd4j/include/helpers/cuda/ConstantTadHelper.cu
raver119 7abc574eeb
Snapshot update (#8194)
* fix double consumption of rng on cpu

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

* Shyrma docs (#222)

* - documenting and profiling matrix_set_diag cuda kernel

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

* - correct formula of pnorm pooling in cuda 2d/3d kernels
- remove helper matrix_diag which duplicates work of helper matrix_set_diag

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

* cublasHandle sharing + lock

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

* cublasHandle sharing + lock

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

* Documentation from serialization/deserialization in NLP (#221)

* refactoring

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Javadocs

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Javadoc fixed

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Cleanup

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* dedicated lock for getCudaCublasHandle

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

* Small fixes (#223)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* ELU DL4J fixes (#224)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* javadoc (#225)

Signed-off-by: Robert Altena <Rob@Ra-ai.com>

* Small test compilation fix (#226)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8182 remove spark version suffix (#227)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* [WIP] Thread safety (#229)

* sync after cublas*gemm

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

* mutex for CublasHelper

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

* don't store cublasHandle in LaunchContext, it's per-device anyway

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

* some printout

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

* check for field instead

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

* pew-pew

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

* don't release ContextBuffers until device changed

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

* small tweak

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

* some logging in sgemm

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

* stream sync

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

* some more logging

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

* some more error checks

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

* one fancy test

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

* one fancy test

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

* minor AffinityManager fix

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

* cudaEvent error logging improvement

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

* ConstantHelper thread safety

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

* - minor corrections in ConstantTadHelper

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

* ConstantShapeHelper thread safety

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

* ConstantTadHelper.cu updated

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

* logging off

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

* logging off

Signed-off-by: raver119 <raver119@gmail.com>
2019-09-03 22:02:02 +03:00

114 lines
4.7 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 "../ConstantTadHelper.h"
#include <TAD.h>
#include <ConstantHelper.h>
#include <AffinityManager.h>
#include <exceptions/cuda_exception.h>
#include <execution/LaunchContext.h>
#include <ShapeUtils.h>
namespace nd4j {
ConstantTadHelper::ConstantTadHelper() {
auto numDevices = AffinityManager::numberOfDevices();
for (int e = 0; e < numDevices; e++) {
std::map<TadDescriptor, TadPack> pack;
_cache.emplace_back(pack);
}
}
ConstantTadHelper* ConstantTadHelper::getInstance() {
if (!_INSTANCE)
_INSTANCE = new ConstantTadHelper();
return _INSTANCE;
}
TadPack ConstantTadHelper::tadForDimensions(const Nd4jLong *originalShape, int dimension, const bool keepUnitiesInShape) {
return tadForDimensions(originalShape, &dimension, 1, keepUnitiesInShape);
}
TadPack ConstantTadHelper::tadForDimensions(const Nd4jLong *originalShape, const std::vector<int> &dimensions, const bool keepUnitiesInShape) {
return tadForDimensions(originalShape, const_cast<int *>(dimensions.data()), dimensions.size(), keepUnitiesInShape);
}
TadPack ConstantTadHelper::tadForDimensions(const Nd4jLong *originalShape, int* dimensions, int dimLength, const bool keepUnitiesInShape) {
TadDescriptor tadDescriptor(originalShape, dimensions, dimLength, keepUnitiesInShape);
return tadForDimensions(tadDescriptor);
}
TadPack ConstantTadHelper::tadForDimensions(ShapeDescriptor &descriptor, std::vector<int> &dimensions, const bool keepUnitiesInShape) {
TadDescriptor tadDescriptor(descriptor, dimensions, keepUnitiesInShape);
return tadForDimensions(tadDescriptor);
}
TadPack ConstantTadHelper::tadForDimensions(TadDescriptor &descriptor) {
const int deviceId = AffinityManager::currentDeviceId();
_mutex.lock();
if (_cache[deviceId].count(descriptor) == 0) {
const auto shapeInfo = descriptor.originalShape().toShapeInfo();
const int rank = shape::rank(shapeInfo);
const std::vector<int> dimsToExclude = ShapeUtils::evalDimsToExclude(rank, descriptor.axis());
const Nd4jLong numOfSubArrs = ShapeUtils::getNumOfSubArrs(shapeInfo, dimsToExclude);
const int subArrRank = (rank == dimsToExclude.size() || descriptor.areUnitiesinShape()) ? rank : rank - dimsToExclude.size();
auto sPtr = new Nd4jLong[shape::shapeInfoLength(subArrRank)];
auto oPtr = new Nd4jLong[numOfSubArrs];
if (numOfSubArrs > 0)
shape::calcSubArrShapeAndOffsets(shapeInfo, numOfSubArrs, dimsToExclude.size(), dimsToExclude.data(), sPtr, oPtr, descriptor.areUnitiesinShape());
Nd4jPointer soPtr;
auto res = cudaMalloc(reinterpret_cast<void**>(&soPtr), numOfSubArrs * sizeof(Nd4jLong));
if (res != 0)
throw cuda_exception::build("Memory allocation for tadOffsets failed", res);
res = cudaMemcpy(soPtr, oPtr, numOfSubArrs * sizeof(Nd4jLong), cudaMemcpyHostToDevice);
if (res != 0)
throw cuda_exception::build("tadOffsets copy failed", res);
auto ssPtr = ConstantHelper::getInstance()->replicatePointer(sPtr, shape::shapeInfoByteLength(subArrRank));
ConstantDataBuffer shapesBuffer(sPtr, ssPtr, shape::shapeInfoLength(subArrRank) * sizeof(Nd4jLong), DataType::INT64);
ConstantDataBuffer offsetsBuffer(oPtr, soPtr, numOfSubArrs * sizeof(Nd4jLong), DataType::INT64);
TadPack t(shapesBuffer, offsetsBuffer, numOfSubArrs);
_cache[deviceId][descriptor] = t;
TadPack r = _cache[deviceId][descriptor];
_mutex.unlock();
delete[] shapeInfo;
return r;
} else {
TadPack r = _cache[deviceId][descriptor];
_mutex.unlock();
return r;
}
}
nd4j::ConstantTadHelper* nd4j::ConstantTadHelper::_INSTANCE = 0;
}