cavis/libnd4j/include/helpers/cuda/ConstantTadHelper.cu
raver119 763a225c6a [WIP] More of CUDA operations (#69)
* initial commit

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

* - gruCell_bp further

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

* - further work on gruCell_bp

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

* Inverse matrix cublas implementation. Partial working revision.

* Separation of segment ops helpers. Max separation.

* Separated segment_min ops.

* Separation of segment_mean/sum/prod/sqrtN ops heleprs.

* Fixed diagonal processing with LUP decomposition.

* Modified inversion approach using current state of LU decomposition.

* Implementation of matrix_inverse op with cuda kernels. Working revision.

* Implemented sequence_mask cuda helper. Eliminated waste printf with matrix_inverse implementation. Added proper tests.

* - further work on gruCell_bp (ff/cuda)

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

* comment one test for gruCell_bp

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

* - provide cuda static_rnn

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

* Refactored random_shuffle op to use new random generator.

* Refactored random_shuffle op helper.

* Fixed debug tests with random ops tests.

* Implement random_shuffle op cuda kernel helper and tests.

* - provide cuda scatter_update

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

* Implementation of random_shuffle for linear case with cuda kernels and tests.

* Implemented random_shuffle with cuda kernels. Final revision.

* - finally gruCell_bp is completed

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

* Dropout op cuda helper implementation.

* Implemented dropout_bp cuda helper.

* Implemented alpha_dropout_bp with cuda kernel helpers.

* Refactored helper.

* Implementation of suppresion helper with cuda kernels.

* - provide cpu code fot hsvToRgb, rgbToHsv, adjustHue

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

* Using sort by value method.

* Implementation of image.non_max_suppression op cuda-based helper.

* - correcting and testing adjust_hue, adjust_saturation cpu/cuda code

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

* Added cuda device prefixes to declarations.

* Implementation of hashcode op with cuda helper. Initital revision.

* rnn cu impl removed

Signed-off-by: raver119 <raver119@gmail.com>
2019-07-20 23:20:41 +10:00

113 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 <exceptions/cuda_exception.h>
#include <execution/LaunchContext.h>
#include <ShapeUtils.h>
namespace nd4j {
ConstantTadHelper::ConstantTadHelper() {
auto numDevices = ConstantHelper::getNumberOfDevices();
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 = ConstantHelper::getCurrentDevice();
_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;
}