553 lines
24 KiB
Plaintext
553 lines
24 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
|
|
******************************************************************************/
|
|
|
|
//
|
|
// Created by raver119 on 19.01.18.
|
|
// @author Yurii Shyrma (iuriish@yahoo.com)
|
|
//
|
|
|
|
#include <ops/declarable/helpers/s_t_b.h>
|
|
#include <PointersManager.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__global__ static void batchToSpaceCuda(const void* vx, const Nd4jLong* xShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint cropBottom, const uint cropLeft) {
|
|
|
|
// input [bS, H * blockSize, W * blockSize, iC]
|
|
// output [bS, H * blockSize - cropBottom - cropTop, W * blockSize - cropLeft - cropRight, iC]
|
|
|
|
// if (cropTop = cropBottom = cropRight = cropLeft = 0) shapes are the same
|
|
// else:
|
|
// oH -> [cropBottom, iH - cropTop]
|
|
// oW -> [cropLeft, iH - cropRight]
|
|
// xLen >= zLen
|
|
|
|
const auto x = reinterpret_cast<const T*>(vx);
|
|
auto z = reinterpret_cast<T*>(vz);
|
|
|
|
__shared__ int rank;
|
|
__shared__ Nd4jLong zLen, *sharedMem;
|
|
|
|
if (threadIdx.x == 0) {
|
|
extern __shared__ unsigned char shmem[];
|
|
sharedMem = reinterpret_cast<Nd4jLong*>(shmem);
|
|
|
|
rank = shape::rank(zShapeInfo);
|
|
zLen = shape::length(zShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
auto coords = sharedMem + threadIdx.x * rank;
|
|
|
|
const auto i = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
if(i >= zLen)
|
|
return;
|
|
|
|
shape::index2coords(i, zShapeInfo, coords);
|
|
|
|
const auto zOffset = shape::getOffset(zShapeInfo, coords);
|
|
|
|
coords[1] += cropBottom;
|
|
coords[2] += cropLeft;
|
|
|
|
const auto xOffset = shape::getOffset(xShapeInfo, coords);
|
|
|
|
z[zOffset] = x[xOffset];
|
|
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
static void batchToSpaceCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint cropBottom, const uint cropLeft) {
|
|
|
|
batchToSpaceCuda<T><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(vx, xShapeInfo, vz, zShapeInfo, cropBottom, cropLeft);
|
|
}
|
|
BUILD_SINGLE_TEMPLATE(template void batchToSpaceCudaLauncher, (const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint cropBottom, const uint cropLeft), LIBND4J_TYPES);
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
void batchToSpace(nd4j::LaunchContext* context, const NDArray& input, NDArray& output, const uint cropBottom, const uint cropTop, const uint cropLeft, const uint cropRight, const uint blockSize) {
|
|
|
|
// [bS*blockSize*blockSize, H/blockSize, W/blockSize, iC] is rearranged/permuted to [bS, oH, oW, iC]
|
|
// oH = H - cropTop - cropBottom
|
|
// oW = W - cropLeft - cropRight
|
|
|
|
NDArray inputRearranged0 = input.reshape(input.ordering(), {blockSize, blockSize, output.sizeAt(0), input.sizeAt(1), input.sizeAt(2), input.sizeAt(3)});
|
|
inputRearranged0.permutei({2, 3,0, 4,1, 5});
|
|
|
|
if(input.lengthOf() == output.lengthOf()) {
|
|
|
|
output.assign(inputRearranged0);
|
|
}
|
|
else {
|
|
|
|
NDArray inputRearranged1 = inputRearranged0.reshape(input.ordering(), {output.sizeAt(0), input.sizeAt(1) * blockSize, input.sizeAt(2) * blockSize, input.sizeAt(3)});
|
|
|
|
const int threadsPerBlock = MAX_NUM_THREADS / 2;
|
|
const int blocksPerGrid = (output.lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
|
|
const int sharedMem = threadsPerBlock * sizeof(Nd4jLong) * output.rankOf() + 128;
|
|
|
|
PointersManager manager(context, "batchToSpace");
|
|
|
|
NDArray::prepareSpecialUse({&output}, {&inputRearranged1});
|
|
BUILD_SINGLE_SELECTOR(input.dataType(), batchToSpaceCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), inputRearranged1.getSpecialBuffer(), inputRearranged1.getSpecialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), cropBottom, cropLeft), LIBND4J_TYPES);
|
|
NDArray::registerSpecialUse({&output}, {&inputRearranged1});
|
|
|
|
manager.synchronize();
|
|
}
|
|
}
|
|
|
|
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Y>
|
|
__global__ static void batchToSpaceNDCuda(const void* vx, const Nd4jLong* xShapeInfo,
|
|
const void* vy, const Nd4jLong* yShapeInfo,
|
|
void* vz, const Nd4jLong* zShapeInfo,
|
|
const uint numOfSpatialDims) {
|
|
|
|
// 4D example, numOfSpatialDims = 2
|
|
// input [bS, H * blockShape[0], W * blockShape[1], iC]
|
|
// output [bS, H * blockShape[0] - cropBottom - cropTop, W * blockShape[1] - cropLeft - cropRight, iC]
|
|
|
|
// if (cropTop = cropBottom = cropRight = cropLeft = 0) shapes are the same
|
|
// else:
|
|
// oH -> [cropBottom, iH - cropTop]
|
|
// oW -> [cropLeft, iH - cropRight]
|
|
// xLen >= zLen
|
|
|
|
const auto x = reinterpret_cast<const X*>(vx);
|
|
const auto y = reinterpret_cast<const Y*>(vy);
|
|
auto z = reinterpret_cast<X*>(vz);
|
|
|
|
__shared__ int rank;
|
|
__shared__ Nd4jLong zLen, *sharedMem;
|
|
|
|
if (threadIdx.x == 0) {
|
|
|
|
extern __shared__ unsigned char shmem[];
|
|
sharedMem = reinterpret_cast<Nd4jLong*>(shmem);
|
|
|
|
rank = shape::rank(zShapeInfo);
|
|
zLen = shape::length(zShapeInfo);
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
auto coords = sharedMem + threadIdx.x * rank;
|
|
|
|
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < zLen; i += gridDim.x * blockDim.x) {
|
|
|
|
shape::index2coords(i, zShapeInfo, coords);
|
|
|
|
const auto zOffset = shape::getOffset(zShapeInfo, coords);
|
|
|
|
// evaluate spatial coordinates for x
|
|
for(uint j = 1; j <= numOfSpatialDims; ++j) {
|
|
const auto yOffset = (j - 1) * yShapeInfo[3]; // yRank = 2, calculate offset manually
|
|
coords[j] += y[yOffset]; // add crop left
|
|
}
|
|
|
|
const auto xOffset = shape::getOffset(xShapeInfo, coords);
|
|
|
|
z[zOffset] = x[xOffset];
|
|
}
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename X,typename Y>
|
|
static void batchToSpaceNDCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, const void* vy, const Nd4jLong* yShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint numOfSpatialDims) {
|
|
|
|
batchToSpaceNDCuda<X,Y><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, numOfSpatialDims);
|
|
}
|
|
BUILD_DOUBLE_TEMPLATE(template void batchToSpaceNDCudaLauncher, (const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, const void* vy, const Nd4jLong* yShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint numOfSpatialDims), LIBND4J_TYPES, INTEGER_TYPES);
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
void batchToSpaceND(nd4j::LaunchContext* context, const NDArray& input, const NDArray& blockShape, const NDArray& crop, NDArray& output) {
|
|
|
|
// 4D example, numOfSpatialDims = 2 - two spatial dimensions
|
|
// [bS*blockShape[0]*blockShape[1], iH, iW, iC] is rearranged/permuted to [bS, iH*blockShape[0] - cropTop - cropBottom, iW*blockShape[1] - cropLeft - cropRight, iC]
|
|
|
|
const uint rank = input.rankOf();
|
|
const uint numOfSpatialDims = blockShape.sizeAt(0);
|
|
|
|
//*** construct reshaping std::vector for first reshape of input array ***//
|
|
|
|
std::vector<Nd4jLong> temp(numOfSpatialDims + rank);
|
|
|
|
int i;
|
|
for(i = 0; i < numOfSpatialDims; ++i)
|
|
temp[i] = blockShape.e<Nd4jLong>(i);
|
|
temp[i++] = output.sizeAt(0);
|
|
for(int j = 1; j < rank; ++i, ++j)
|
|
temp[i] = input.sizeAt(j);
|
|
|
|
NDArray inputRearranged0 = input.reshape(input.ordering(), temp);
|
|
|
|
//*** construct permuting std::vector for permutation of input array ***//
|
|
|
|
temp[0] = numOfSpatialDims;
|
|
|
|
for(i = 1; i <= numOfSpatialDims; ++i) {
|
|
temp[2*i - 1] = numOfSpatialDims + i;
|
|
temp[2*i] = i - 1;
|
|
}
|
|
for(i = 2 * numOfSpatialDims + 1; i < temp.size(); ++i)
|
|
temp[i] = i;
|
|
|
|
inputRearranged0.permutei(temp);
|
|
|
|
|
|
if(input.lengthOf() == output.lengthOf()) {
|
|
|
|
output.assign(inputRearranged0);
|
|
}
|
|
else {
|
|
//*** construct reshaping std::vector for second reshape of input array ***//
|
|
|
|
temp.resize(rank);
|
|
|
|
temp[0] = output.sizeAt(0);
|
|
|
|
for(i = 1; i < rank; ++i)
|
|
temp[i] = (i <= numOfSpatialDims) ? input.sizeAt(i) * blockShape.e<Nd4jLong>(i - 1) : input.sizeAt(i);
|
|
|
|
NDArray inputRearranged1 = inputRearranged0.reshape(input.ordering(), temp);
|
|
|
|
const int threadsPerBlock = MAX_NUM_THREADS / 4;
|
|
const int blocksPerGrid = (output.lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
|
|
const int sharedMem = threadsPerBlock * sizeof(Nd4jLong) * output.rankOf() + 128;
|
|
|
|
PointersManager manager(context, "batchToSpaceND");
|
|
|
|
NDArray::prepareSpecialUse({&output}, {&inputRearranged1, &crop});
|
|
BUILD_DOUBLE_SELECTOR(input.dataType(), crop.dataType(), batchToSpaceNDCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), inputRearranged1.getSpecialBuffer(), inputRearranged1.getSpecialShapeInfo(), crop.getSpecialBuffer(), crop.getSpecialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), numOfSpatialDims), LIBND4J_TYPES, INTEGER_TYPES);
|
|
NDArray::registerSpecialUse({&output}, {&inputRearranged1, &crop});
|
|
|
|
manager.synchronize();
|
|
}
|
|
}
|
|
|
|
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
__global__ static void spaceToBatchCuda(const void* vx, const Nd4jLong* xShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint padBottom, const uint padTop, const uint padLeft, const uint padRight) {
|
|
|
|
// input [bS, H * blockSize - padBottom - padTop, W * blockSize - padLeft - padRight, iC]
|
|
// output [bs, H * blockSize, W * blockSize, iC]
|
|
|
|
// if (padTop = padBottom = padRight = padLeft = 0) shapes are the same
|
|
// else:
|
|
// iH -> [padBottom, oH - padTop]
|
|
// iW -> [padLeft, oW - padRight]
|
|
// zLen > xLen
|
|
|
|
const auto x = reinterpret_cast<const T*>(vx);
|
|
auto z = reinterpret_cast<T*>(vz);
|
|
|
|
__shared__ int rank;
|
|
__shared__ Nd4jLong zLen, *sharedMem;
|
|
|
|
if (threadIdx.x == 0) {
|
|
extern __shared__ unsigned char shmem[];
|
|
sharedMem = reinterpret_cast<Nd4jLong*>(shmem);
|
|
|
|
rank = shape::rank(zShapeInfo);
|
|
zLen = shape::length(zShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
auto coords = sharedMem + threadIdx.x * rank;
|
|
|
|
const auto i = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
if(i >= zLen)
|
|
return;
|
|
|
|
shape::index2coords(i, zShapeInfo, coords);
|
|
|
|
const auto zOffset = shape::getOffset(zShapeInfo, coords);
|
|
|
|
if(coords[1] >= padBottom && coords[1] < zShapeInfo[2] - padTop && coords[2] >= padLeft && coords[2] < zShapeInfo[3] - padRight) {
|
|
|
|
coords[1] -= padBottom;
|
|
coords[2] -= padLeft;
|
|
|
|
const auto xOffset = shape::getOffset(xShapeInfo, coords);
|
|
|
|
z[zOffset] = x[xOffset];
|
|
}
|
|
else
|
|
z[zOffset] = 0.f;
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename T>
|
|
static void spaceToBatchCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint padBottom, const uint padTop, const uint padLeft, const uint padRight) {
|
|
|
|
spaceToBatchCuda<T><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(vx, xShapeInfo, vz, zShapeInfo, padBottom, padTop, padLeft, padRight);
|
|
}
|
|
BUILD_SINGLE_TEMPLATE(template void spaceToBatchCudaLauncher, (const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint padBottom, const uint padTop, const uint padLeft, const uint padRight), LIBND4J_TYPES);
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
void spaceToBatch(nd4j::LaunchContext* context, const NDArray& input, NDArray& output, const uint padBottom, const uint padTop, const uint padLeft, const uint padRight, const uint blockSize) {
|
|
|
|
// [bS, iH, iW, iC] is rearranged/permuted to [bS*blockSize*blockSize, (iH + padBottom + padTop)/blockSize, (iW + padLeft + padRight)/blockSize, iC]
|
|
|
|
NDArray outputRearranged0 = output.reshape(output.ordering(), {blockSize, blockSize, input.sizeAt(0), output.sizeAt(1), output.sizeAt(2), input.sizeAt(3)});
|
|
outputRearranged0.permutei({2, 3,0, 4,1, 5});
|
|
|
|
if(input.lengthOf() == output.lengthOf()) {
|
|
|
|
outputRearranged0.assign(input);
|
|
}
|
|
else {
|
|
|
|
NDArray outputRearranged1 = outputRearranged0.reshape(output.ordering(), {input.sizeAt(0), output.sizeAt(1) * blockSize, output.sizeAt(2) * blockSize, input.sizeAt(3)});
|
|
|
|
const int threadsPerBlock = MAX_NUM_THREADS / 2;
|
|
const int blocksPerGrid = (output.lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
|
|
const int sharedMem = threadsPerBlock * sizeof(Nd4jLong) * output.rankOf() + 128;
|
|
|
|
PointersManager manager(context, "spaceToBatch");
|
|
|
|
NDArray::prepareSpecialUse({&outputRearranged1}, {&input});
|
|
BUILD_SINGLE_SELECTOR(input.dataType(), spaceToBatchCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), input.getSpecialBuffer(), input.getSpecialShapeInfo(), outputRearranged1.specialBuffer(), outputRearranged1.specialShapeInfo(), padBottom, padTop, padLeft, padRight), LIBND4J_TYPES);
|
|
NDArray::registerSpecialUse({&outputRearranged1}, {&input});
|
|
|
|
manager.synchronize();
|
|
|
|
if(output.getSpecialBuffer() != outputRearranged1.getSpecialBuffer())
|
|
outputRearranged0.assign(outputRearranged1);
|
|
}
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Y>
|
|
__global__ static void spaceToBatchNDCuda(const void* vx, const Nd4jLong* xShapeInfo,
|
|
const void* vy, const Nd4jLong* yShapeInfo,
|
|
void* vz, const Nd4jLong* zShapeInfo,
|
|
const uint numOfSpatialDims) {
|
|
|
|
// x - input, y - padding, z - output
|
|
|
|
// 4D example
|
|
// input [bS, H * blockShape[0] - padBottom - padTop, W * blockShape[1] - padLeft - padRight, iC]
|
|
// output [bS, H * blockShape[0], W * blockShape[1], iC]
|
|
|
|
// if (padTop = padBottom = padRight = padLeft = 0) shapes are the same
|
|
// else:
|
|
// iH -> [padBottom, oH - padTop]
|
|
// iW -> [padLeft, oW - padRight]
|
|
// zLen > xLen
|
|
|
|
const auto x = reinterpret_cast<const X*>(vx);
|
|
const auto y = reinterpret_cast<const Y*>(vy);
|
|
auto z = reinterpret_cast<X*>(vz);
|
|
|
|
__shared__ int rank; // xRank = zRank, yRank = 2;
|
|
__shared__ Nd4jLong zLen, totalThreads, *sharedMem;
|
|
|
|
if (threadIdx.x == 0) {
|
|
|
|
extern __shared__ unsigned char shmem[];
|
|
sharedMem = reinterpret_cast<Nd4jLong*>(shmem);
|
|
|
|
rank = shape::rank(zShapeInfo);
|
|
zLen = shape::length(zShapeInfo);
|
|
totalThreads = gridDim.x * blockDim.x;
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
auto coords = sharedMem + threadIdx.x * rank;
|
|
|
|
for (int i = blockDim.x * blockIdx.x + threadIdx.x; i < zLen; i += totalThreads) {
|
|
|
|
shape::index2coords(i, zShapeInfo, coords);
|
|
|
|
const auto zOffset = shape::getOffset(zShapeInfo, coords);
|
|
|
|
bool within = true;
|
|
|
|
for(uint j = 1; j <= numOfSpatialDims; ++j) {
|
|
|
|
// yRank = 2, calculate offset manually
|
|
const auto yOffset = (j - 1) * yShapeInfo[3];
|
|
const auto padLeft = y[yOffset];
|
|
const auto padRight = y[yOffset + yShapeInfo[4]];
|
|
|
|
within &= (coords[j] >= padLeft && coords[j] < shape::shapeOf(const_cast<Nd4jLong*>(zShapeInfo))[j] - padRight);
|
|
|
|
if(!within)
|
|
break;
|
|
|
|
coords[j] -= padLeft; // get coordinates for x
|
|
}
|
|
|
|
if(within)
|
|
z[zOffset] = x[shape::getOffset(xShapeInfo, coords)];
|
|
else
|
|
z[zOffset] = 0.f;
|
|
}
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Y>
|
|
static void spaceToBatchNDCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, const void* vy, const Nd4jLong* yShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint numOfSpatialDims) {
|
|
|
|
spaceToBatchNDCuda<X,Y><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, numOfSpatialDims);
|
|
}
|
|
BUILD_DOUBLE_TEMPLATE(template void spaceToBatchNDCudaLauncher, (const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, const void* vy, const Nd4jLong* yShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const uint numOfSpatialDims), LIBND4J_TYPES, INTEGER_TYPES);
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
void spaceToBatchND(nd4j::LaunchContext* context, const NDArray& input, const NDArray& blockShape, const NDArray& padding, NDArray& output ) {
|
|
|
|
// 4D example with two spatial dimensions
|
|
// [bS, iH, iW, iC] is rearranged/permuted to [bS*blockShape[0]*blockShape[1], (iH + padBottom + padTop)/blockShape[0], (iW + padLeft + padRight)/blockShape[1], iC]
|
|
|
|
const uint rank = input.rankOf();
|
|
|
|
const uint numOfSpatialDims = blockShape.sizeAt(0);
|
|
|
|
//*** construct reshaping std::vector for first reshape of output array ***//
|
|
std::vector<Nd4jLong> temp(numOfSpatialDims + rank);
|
|
|
|
int i;
|
|
for(i = 0; i < numOfSpatialDims; ++i)
|
|
temp[i] = blockShape.e<Nd4jLong>(i);
|
|
temp[i++] = input.sizeAt(0);
|
|
for(int j = 1; j < rank; ++i, ++j)
|
|
temp[i] = output.sizeAt(j);
|
|
|
|
NDArray outputRearranged0 = output.reshape(output.ordering(), temp);
|
|
|
|
//*** construct permuting std::vector for permutation of output array ***//
|
|
|
|
temp[0] = numOfSpatialDims;
|
|
|
|
for(i = 1; i <= numOfSpatialDims; ++i) {
|
|
temp[2*i - 1] = numOfSpatialDims + i;
|
|
temp[2*i] = i - 1;
|
|
}
|
|
for(i = 2 * numOfSpatialDims + 1; i < temp.size(); ++i)
|
|
temp[i] = i;
|
|
|
|
outputRearranged0.permutei(temp);
|
|
|
|
// ****** //
|
|
|
|
if(input.lengthOf() == output.lengthOf()) {
|
|
outputRearranged0.assign(input);
|
|
}
|
|
else {
|
|
|
|
//*** construct reshaping std::vector for second reshape of output array ***//
|
|
temp.resize(rank);
|
|
|
|
temp[0] = input.sizeAt(0);
|
|
|
|
for(i = 1; i < rank; ++i)
|
|
temp[i] = (i <= numOfSpatialDims) ? output.sizeAt(i) * blockShape.e<Nd4jLong>(i - 1) : output.sizeAt(i);
|
|
|
|
NDArray outputRearranged1 = outputRearranged0.reshape(output.ordering(), temp);
|
|
|
|
const int threadsPerBlock = MAX_NUM_THREADS / 4;
|
|
const int blocksPerGrid = (output.lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
|
|
const int sharedMem = threadsPerBlock * sizeof(Nd4jLong) * output.rankOf() + 128;
|
|
|
|
PointersManager manager(context, "spaceToBatchND");
|
|
|
|
NDArray::prepareSpecialUse({&outputRearranged1}, {&input, &padding});
|
|
BUILD_DOUBLE_SELECTOR(input.dataType(), padding.dataType(), spaceToBatchNDCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), input.getSpecialBuffer(), input.getSpecialShapeInfo(), padding.getSpecialBuffer(), padding.getSpecialShapeInfo(), outputRearranged1.specialBuffer(), outputRearranged1.specialShapeInfo(), numOfSpatialDims), LIBND4J_TYPES, INTEGER_TYPES);
|
|
NDArray::registerSpecialUse({&outputRearranged1}, {&input, &padding});
|
|
|
|
manager.synchronize();
|
|
|
|
if(output.getSpecialBuffer() != outputRearranged1.getSpecialBuffer())
|
|
outputRearranged0.assign(outputRearranged1);
|
|
}
|
|
}
|
|
|
|
|
|
/*
|
|
template <int N, bool B2S>
|
|
struct SpaceToBatchHelper {
|
|
template <typename T>
|
|
static void run(T *ptrSpace, const Nd4jLong *space_shape, const Nd4jLong *space_strides, const Nd4jLong *block_shape, const Nd4jLong *pad_start, const Nd4jLong *block_offsets, T *ptrBatch, const Nd4jLong *batch_shape, const Nd4jLong *batch_strides) {
|
|
for (int batch_pos = 0; batch_pos < batch_shape[0]; ++batch_pos) {
|
|
const int space_pos = batch_pos * block_shape[0] + block_offsets[0] - pad_start[0];
|
|
if (space_pos >= 0 && space_pos < space_shape[0]) {
|
|
SpaceToBatchHelper<N - 1, B2S>::run(ptrSpace + space_pos * space_strides[0], space_shape + 1, space_strides + 1, block_shape + 1, pad_start + 1, block_offsets + 1, ptrBatch, batch_shape + 1, batch_strides + 1);
|
|
} else {
|
|
if (!B2S)
|
|
for (int i = 0; i < batch_strides[0]; i++)
|
|
ptrBatch[i] = (T) 0.f;
|
|
}
|
|
|
|
ptrBatch += batch_strides[0];
|
|
}
|
|
}
|
|
};
|
|
|
|
template <bool B2S>
|
|
struct SpaceToBatchHelper<0, B2S> {
|
|
template <typename T>
|
|
static void run(T *ptrSpace, const Nd4jLong *space_shape, const Nd4jLong *space_strides, const Nd4jLong *block_shape, const Nd4jLong *pad_start, const Nd4jLong *block_offsets, T *ptrBatch, const Nd4jLong *batch_shape, const Nd4jLong *batch_strides) {
|
|
int str = batch_strides[-1];
|
|
for (int i = 0; i < str; i++)
|
|
if (B2S)
|
|
ptrSpace[i] = ptrBatch[i];
|
|
else
|
|
ptrBatch[i] = ptrSpace[i];
|
|
}
|
|
};
|
|
|
|
template <typename T, int NUM_BLOCK_DIMS, bool B2S>
|
|
void _execute(nd4j::LaunchContext * context, void *vptrSpace, const Nd4jLong *space_shape, const Nd4jLong *space_strides, const Nd4jLong *block_shape, const Nd4jLong *pad_start, const Nd4jLong *block_offsets, void *vptrBatch, const Nd4jLong *batch_shape, const Nd4jLong *batch_strides) {
|
|
auto ptrSpace = reinterpret_cast<T *>(vptrSpace);
|
|
auto ptrBatch = reinterpret_cast<T *>(vptrBatch);
|
|
SpaceToBatchHelper<NUM_BLOCK_DIMS, B2S>::run(ptrSpace, space_shape, space_strides, block_shape, pad_start, block_offsets, ptrBatch, batch_shape, batch_strides);
|
|
};
|
|
|
|
Nd4jStatus _batchToSpace(nd4j::LaunchContext * context, int internal_block_dims, NDArray *input, NDArray *output, std::vector<Nd4jLong> &internal_input_shape, std::vector<Nd4jLong> &internal_output_shape, Nd4jLong *block_shape, Nd4jLong *crops) {
|
|
|
|
return Status::OK();
|
|
}
|
|
|
|
#define STB_DIM (0, 1),\
|
|
(1, 2),\
|
|
(2, 3),\
|
|
(3, 4)
|
|
|
|
#define STB_BOOL (0, false),\
|
|
(1, true)
|
|
|
|
BUILD_TRIPLE_TEMPLATE(template void _execute, (nd4j::LaunchContext * context, void *ptrSpace, const Nd4jLong *space_shape, const Nd4jLong *space_strides, const Nd4jLong *block_shape, const Nd4jLong *pad_start, const Nd4jLong *block_offsets, void *ptrBatch, const Nd4jLong *batch_shape, const Nd4jLong *batch_strides), LIBND4J_TYPES, STB_DIM, STB_BOOL);
|
|
|
|
#undef STB_BOOL
|
|
#undef STB_DIM
|
|
*/
|
|
|
|
}
|
|
}
|
|
} |