459 lines
22 KiB
C++
459 lines
22 KiB
C++
/*******************************************************************************
|
|
* 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 <op_boilerplate.h>
|
|
#include <loops/broadcasting.h>
|
|
#include <loops/legacy_ops.h>
|
|
#include <types/types.h>
|
|
#include <LoopKind.h>
|
|
#include <helpers/ConstantTadHelper.h>
|
|
|
|
using namespace simdOps;
|
|
|
|
namespace functions {
|
|
namespace broadcast {
|
|
|
|
template <typename X, typename Y, typename Z>
|
|
void Broadcast<X, Y, Z>::execInverse(const int opNum,
|
|
void *x,
|
|
Nd4jLong *xShapeInfo,
|
|
void *y,
|
|
Nd4jLong *yShapeInfo,
|
|
void *z,
|
|
Nd4jLong *zShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
Nd4jLong *xTadShapeInfo,
|
|
Nd4jLong *xTadOffset,
|
|
Nd4jLong *zTadShapeInfo,
|
|
Nd4jLong *zTadOffset) {
|
|
DISPATCH_BY_OPNUM_TTT(execInverse, PARAMS(x,
|
|
xShapeInfo,
|
|
y,
|
|
yShapeInfo,
|
|
z,
|
|
zShapeInfo,
|
|
dimension,
|
|
dimensionLength,
|
|
xTadShapeInfo,
|
|
xTadOffset,
|
|
zTadShapeInfo,
|
|
zTadOffset), BROADCAST_OPS);
|
|
}
|
|
|
|
template <typename X, typename Y, typename Z>
|
|
void Broadcast<X, Y, Z>::exec(const int opNum,
|
|
void *x,
|
|
Nd4jLong *xShapeInfo,
|
|
void *y,
|
|
Nd4jLong *yShapeInfo,
|
|
void *z,
|
|
Nd4jLong *zShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
Nd4jLong *xTadShapeInfo,
|
|
Nd4jLong *xTadOffset,
|
|
Nd4jLong *zTadShapeInfo,
|
|
Nd4jLong *zTadOffset) {
|
|
DISPATCH_BY_OPNUM_TTT(exec, PARAMS(x,
|
|
xShapeInfo,
|
|
y,
|
|
yShapeInfo,
|
|
z,
|
|
zShapeInfo,
|
|
dimension,
|
|
dimensionLength,
|
|
xTadShapeInfo,
|
|
xTadOffset,
|
|
zTadShapeInfo,
|
|
zTadOffset), BROADCAST_OPS);
|
|
}
|
|
|
|
template <typename X, typename Y, typename Z>
|
|
template<typename OpType>
|
|
void Broadcast<X, Y, Z>::exec(void *vx,
|
|
Nd4jLong *xShapeInfo,
|
|
void *vy,
|
|
Nd4jLong *yShapeInfo,
|
|
void *vz,
|
|
Nd4jLong *zShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
Nd4jLong *xTadShapeInfo,
|
|
Nd4jLong *xTadOffset,
|
|
Nd4jLong *zTadShapeInfo,
|
|
Nd4jLong *zTadOffset) {
|
|
|
|
auto x = reinterpret_cast<X *>(vx);
|
|
auto y = reinterpret_cast<Y *>(vy);
|
|
auto z = reinterpret_cast<Z *>(vz);
|
|
|
|
//decompose in to several sub tads after
|
|
//moving all dimensions (in sorted order)
|
|
//to the back.
|
|
//permuted version of the x shape info for setting up the tad problem
|
|
auto xTadShapeShapeInfo = xTadShapeInfo;
|
|
auto tadOffsets = xTadOffset;
|
|
|
|
if (xTadShapeInfo == nullptr || tadOffsets == nullptr) {
|
|
auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
|
|
|
|
xTadShapeShapeInfo = tadPack.primaryShapeInfo();
|
|
tadOffsets = tadPack.primaryOffsets();
|
|
}
|
|
|
|
//int *resultStride = shape::stride(xTadShapeShapeInfo);
|
|
unsigned int tadLength = shape::length(xTadShapeShapeInfo);//shape::length(xTadShapeShapeInfo);
|
|
unsigned int tads = shape::length(xShapeInfo) / tadLength;
|
|
|
|
if (zTadShapeInfo == nullptr) {
|
|
zTadShapeInfo = xTadShapeShapeInfo;
|
|
zTadOffset = tadOffsets;
|
|
}
|
|
|
|
auto lenZ = shape::length(zTadShapeInfo);
|
|
auto lenY = shape::length(yShapeInfo);
|
|
|
|
int tadsPerThread = tads / TAD_THRESHOLD;
|
|
int threads = nd4j::math::nd4j_max<int>(1, tadsPerThread);
|
|
threads = nd4j::math::nd4j_min<int>(threads, omp_get_max_threads());
|
|
|
|
auto xEws = shape::elementWiseStride(xTadShapeShapeInfo);
|
|
auto yEws = shape::elementWiseStride(yShapeInfo);
|
|
auto zEws = shape::elementWiseStride(zTadShapeInfo);
|
|
|
|
const nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXYZ(xTadShapeShapeInfo, yShapeInfo, zTadShapeInfo);
|
|
|
|
if (kindOfLoop == nd4j::LoopKind::EWS1) {
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
auto oX = x + tadOffsets[i];
|
|
auto oZ = z + zTadOffset[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++)
|
|
oZ[f] = OpType::op(oX[f], y[f]);
|
|
}
|
|
}
|
|
else if(kindOfLoop == nd4j::LoopKind::EWSNONZERO){
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
auto oX = x + tadOffsets[i];
|
|
auto oZ = z + zTadOffset[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++)
|
|
oZ[f * zEws] = OpType::op(oX[f * xEws], y[f * yEws]);
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, yShapeInfo) && shape::haveSameShapeAndStrides(xTadShapeShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oX = x + tadOffsets[i];
|
|
auto oZ = z + zTadOffset[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
|
|
oZ[offset] = OpType::op(oX[offset], y[offset]);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, yShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
|
|
oZ[zOffset] = OpType::op(oX[offset], y[offset]);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint yShapeInfoCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
|
|
auto yOffset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, lenY, canCastY);
|
|
oZ[offset] = OpType::op(oX[offset], y[yOffset]);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint yShapeInfoCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto xOffset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
|
|
auto offset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, lenY, canCastY);
|
|
oZ[offset] = OpType::op(oX[xOffset], y[offset]);
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
uint yShapeInfoCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
|
|
bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto xOffset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
|
|
auto yOffset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, lenY, canCastY);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
|
|
oZ[zOffset] = OpType::op(oX[xOffset], y[yOffset]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template <typename X, typename Y, typename Z>
|
|
template<typename OpType>
|
|
void Broadcast<X, Y, Z>::execInverse(void *vx,
|
|
Nd4jLong *xShapeInfo,
|
|
void *vy,
|
|
Nd4jLong *yShapeInfo,
|
|
void *vz,
|
|
Nd4jLong *zShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength,
|
|
Nd4jLong *yTadShapeInfo,
|
|
Nd4jLong *yTadOffset,
|
|
Nd4jLong *zTadShapeInfo,
|
|
Nd4jLong *zTadOffset) {
|
|
|
|
auto x = reinterpret_cast<X *>(vx);
|
|
auto y = reinterpret_cast<Y *>(vy);
|
|
auto z = reinterpret_cast<Z *>(vz);
|
|
|
|
//decompose in to several sub tads after
|
|
//moving all dimensions (in sorted order)
|
|
//to the back.
|
|
//permuted version of the x shape info for setting up the tad problem
|
|
auto yTadShapeShapeInfo = yTadShapeInfo;
|
|
auto tadOffsets = yTadOffset;
|
|
|
|
if (yTadShapeInfo == nullptr || tadOffsets == nullptr) {
|
|
auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(yShapeInfo, dimension, dimensionLength);
|
|
|
|
yTadShapeShapeInfo = tadPack.primaryShapeInfo();
|
|
tadOffsets = tadPack.primaryOffsets();
|
|
}
|
|
|
|
//int *resultStride = shape::stride(yTadShapeShapeInfo);
|
|
unsigned int tadLength = shape::length(yTadShapeShapeInfo);
|
|
unsigned int tads = shape::length(yShapeInfo) / tadLength;
|
|
|
|
if (zTadShapeInfo == nullptr) {
|
|
zTadShapeInfo = yTadShapeShapeInfo;
|
|
zTadOffset = tadOffsets;
|
|
}
|
|
|
|
auto lenZ = shape::length(zTadShapeInfo);
|
|
auto lenX = shape::length(xShapeInfo);
|
|
|
|
int tadsPerThread = tads / TAD_THRESHOLD;
|
|
int threads = nd4j::math::nd4j_max<int>(1, tadsPerThread);
|
|
threads = nd4j::math::nd4j_min<int>(threads, omp_get_max_threads());
|
|
|
|
auto yEws = shape::elementWiseStride(yTadShapeShapeInfo);
|
|
auto xEws = shape::elementWiseStride(xShapeInfo);
|
|
auto zEws = shape::elementWiseStride(zTadShapeInfo);
|
|
|
|
const nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXYZ(yTadShapeShapeInfo, xShapeInfo, zTadShapeInfo);
|
|
|
|
if(kindOfLoop == nd4j::LoopKind::EWS1) {
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (unsigned int i = 0; i < tads; i++) {
|
|
auto oY = y + tadOffsets[i];
|
|
auto oZ = z + zTadOffset[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++)
|
|
oZ[f] = OpType::op(x[f], oY[f]);
|
|
}
|
|
}
|
|
else if(kindOfLoop == nd4j::LoopKind::EWSNONZERO) {
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
auto oY = y + tadOffsets[i];
|
|
auto oZ = z + zTadOffset[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++)
|
|
oZ[f * zEws] = OpType::op(x[f * xEws], oY[f * yEws]);
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, xShapeInfo) && shape::haveSameShapeAndStrides(yTadShapeShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oY = x + tadOffsets[i];
|
|
auto oZ = z + zTadOffset[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
|
|
oZ[offset] = OpType::op(x[offset], oY[offset]);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, xShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
|
|
oZ[zOffset] = OpType::op(x[offset], oY[offset]);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
|
|
auto xOffset = shape::indexOffset(f, yShapeInfo, xShapeInfoCast, lenX, canCastX);
|
|
oZ[offset] = OpType::op(x[xOffset], oY[offset]);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto yOffset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
|
|
auto offset = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, lenX, canCastX);
|
|
oZ[offset] = OpType::op(x[offset], oY[yOffset]);
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
|
|
for (int i = 0; i < tads; i++) {
|
|
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int f = 0; f < tadLength; f++) {
|
|
auto xOffset = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, lenX, canCastX);
|
|
auto yOffset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
|
|
oZ[zOffset] = OpType::op(x[xOffset], oY[yOffset]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
} |