95 lines
3.8 KiB
Plaintext
95 lines
3.8 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
|
|
// @author Yurii Shyrma, created on 15.11.2018
|
|
//
|
|
|
|
#include <loops/special_kernels.h>
|
|
|
|
namespace nd4j {
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__device__ void
|
|
tearKernel(void *vx, Nd4jLong *xShapeInfo, Nd4jPointer *targets, Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
|
|
Nd4jLong *tadOffsets) {
|
|
|
|
|
|
|
|
__shared__ Nd4jLong tadLength;
|
|
__shared__ int tadEWS;
|
|
__shared__ int zEWS;
|
|
// __shared__ int tadRank;
|
|
__shared__ Nd4jLong numTads;
|
|
// __shared__ int zRank;
|
|
// __shared__ Nd4jLong *tadShape;
|
|
// __shared__ Nd4jLong *tadStride;
|
|
// __shared__ Nd4jLong *zShape;
|
|
// __shared__ Nd4jLong *zStride;
|
|
__shared__ T* x;
|
|
if (threadIdx.x == 0) {
|
|
tadLength = shape::length(tadShapeInfo);
|
|
tadEWS = shape::elementWiseStride(tadShapeInfo);
|
|
zEWS = shape::elementWiseStride(zShapeInfo);
|
|
numTads = shape::length(xShapeInfo) / tadLength;
|
|
x = static_cast<T *>(vx);
|
|
}
|
|
__syncthreads();
|
|
|
|
for (Nd4jLong r = blockIdx.x; r < numTads; r += gridDim.x) {
|
|
T *z = (T *) targets[r];
|
|
T *s = x + tadOffsets[r];
|
|
|
|
if (zEWS > 0 && tadEWS > 0) {
|
|
for (Nd4jLong i = threadIdx.x; i < tadLength; i += blockDim.x)
|
|
z[i * zEWS] = s[i * tadEWS];
|
|
} else {
|
|
|
|
for (Nd4jLong j = threadIdx.x; j < tadLength; j += blockDim.x) {
|
|
auto xOffset = shape::getIndexOffset(j, tadShapeInfo, tadLength);
|
|
auto zOffset = shape::getIndexOffset(j, zShapeInfo, tadLength);
|
|
|
|
z[zOffset] = s[xOffset];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__global__ void
|
|
execTearKernel(void *vx, Nd4jLong *xShapeInfo, Nd4jPointer *targets, Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
|
|
Nd4jLong *tadOffsets) {
|
|
|
|
tearKernel<T>(vx, xShapeInfo, targets, zShapeInfo, tadShapeInfo, tadOffsets);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__host__ void tearKernelGeneric(dim3 &launchDims, cudaStream_t *stream,
|
|
void *vx, Nd4jLong *xShapeInfo,
|
|
Nd4jPointer *targets, Nd4jLong *zShapeInfo,
|
|
Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {
|
|
|
|
execTearKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, targets, zShapeInfo, tadShapeInfo, tadOffsets);
|
|
nd4j::DebugHelper::checkErrorCode(stream, "tear(...) failed");
|
|
}
|
|
|
|
BUILD_SINGLE_TEMPLATE(template void ND4J_EXPORT tearKernelGeneric, (dim3 & launchDims, cudaStream_t * stream, void * vx, Nd4jLong * xShapeInfo, Nd4jPointer *targets, Nd4jLong * zShapeInfo, Nd4jLong * tadShapeInfo, Nd4jLong * tadOffsets), LIBND4J_TYPES);
|
|
} |