* initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * another initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * another initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Refactored buffer() and shapeInfo() methods usage with NDArray class. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt Graph class methods to use const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt choose op to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt where op shape method to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt lstsq op to use constant empty shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt matrix_diag_part op shape routine to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt determinant ops to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt mean_pairwssqerr_loss ops to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt shape methods for loss ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt log_loss op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt shape methods for ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt dilation2d ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted deconv2d ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted dynamicRNN op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods for ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods for lstm layer ops. Signed-off-by: shugeo <sgazeos@gmail.com> * few updates Signed-off-by: raver119@gmail.com <raver119@gmail.com> * first cuda tweak Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Adopt constant shapes for sconv2d ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt constant shapes for gru ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt constant shapes with shape methods for segment ops and so on. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted constant shapes with unsorted_segment_* ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted constant shapes with gamma op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods of reduce_stddev ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods for reduce_* ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt shape method for squeeze op. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt strided_slice shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored concat op shape method to adopt constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape method for mirror_pad op. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted split op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted tile ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Added const cast for mkldnn routines handles. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored logSoftMaxForVector_ routine to conform with proper data and shape pointer casts. Signed-off-by: shugeo <sgazeos@gmail.com> * Cosmetic changes to proper usage of constant pointers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored a couple shape comparators for strides and addBias helpers to proper use data pointers with inplace option. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored depthToSpace helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored histogram helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored im2col helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored gather and gatherND helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage on percentile helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed gather shape with helpers and range buffer usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with space to depth helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage and constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with LUP decomposition> Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored onehot_ helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pad and prefix to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactoed softmax helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed space to batch helpers to use buffers properly. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed stack and split helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with sparse to dense helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with mindistance_ helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with tile helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed constant shape usage with legacy pairwise bool ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored a couple of methods to adopt constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed broadcasting with constant shape." Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const usage with inplace reverse and constant shapes with legacy reduction. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored legacy ops with const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored sort to adopt constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected sort for constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed constant shape usage with special methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored Context to conform with constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * CUDA broadcasting headers Signed-off-by: raver119@gmail.com <raver119@gmail.com> * pairwise/indexreduce/random headers Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Refactored native ops to adopt constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * legacy reduce3/scalar headers Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Corrected pullRow signature and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected routines to proper use of constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored tests to use constant shapes properly. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored legacy ops tests to use constant shapes properly. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored buffer usage with NDArray tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed native ops tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed special concat routine. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with test. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with a test. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored TAD.h and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored calcStrides* routines to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed miscelaneous errors with constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * NativeOps const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Corrected definitions for declared functions. Signed-off-by: shugeo <sgazeos@gmail.com> * NativeOps const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * few more const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed const shapes with shape routines. Signed-off-by: shugeo <sgazeos@gmail.com> * few more const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed shape method for broadcastable case. Signed-off-by: shugeo <sgazeos@gmail.com> * few more const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * xw_plus_b BP shape fn restored Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed signatures with broadcasting. Signed-off-by: shugeo <sgazeos@gmail.com> * Repaired backprops shape methods for a set of operations. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored broadcast bool for cuda. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored methods for 3 args with const qualifier. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed a couple of kernel signatures for broadcasting. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernels signatures for const buffers and shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pairwise methods to persistent buffers and shapes usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt const to buffers and shapes with kernels. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt const to buffers and shapes with scalar kernels. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored indexreduce kernels signatures to use const buffers and shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pairwise kernels to adopt cons shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pairwise bool kernels to adopt cons shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored random special ops to conform with const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored native ops to conform with const shapes and buffers under cuda platform. Signed-off-by: shugeo <sgazeos@gmail.com> * Cosmetical changes only. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shapes and buffers error. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected start pos routine. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored methods to conform with const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored helpers to use proper methods instead. Signed-off-by: shugeo <sgazeos@gmail.com> * bunch of changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next bunch of changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next bunch of changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed execScalar declaration. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed execScalar declaration. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected const shape cases with sort and so on. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shapes for sort. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored kernel declarations to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernels declarations to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected kernel declarations to adopt const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernels declarations to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed segment helpers kernels declarations and so on to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shape usage with segment and solve helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernel declaration with adjustWeight helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed cuda implementations for constant shape helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted const shape usage with kernels. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted top_k kernels to use const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected kernels declarations to adopt const shapes with helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored NDArray definitions to adopt const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shapes with image suppression helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Slight improvement with buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored buffer usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored buffer usage with tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shape usage with definitions. Signed-off-by: shugeo <sgazeos@gmail.com> * minor updates on cpu side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Refactored const shape usage with ConstantDescritor and native ops with cuda platform. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored tear and tile kernels to adopt with const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * softmax_loop fix Signed-off-by: raver119 <raver119@gmail.com> * update missing signature Signed-off-by: raver119@gmail.com <raver119@gmail.com> * softmax again Signed-off-by: raver119@gmail.com <raver119@gmail.com> * few more missing consts Signed-off-by: raver119 <raver119@gmail.com> * new methods updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
735 lines
38 KiB
C++
735 lines
38 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 <system/op_boilerplate.h>
|
|
#include <loops/broadcasting_bool.h>
|
|
#include <loops/legacy_ops.h>
|
|
#include <types/types.h>
|
|
#include <helpers/LoopKind.h>
|
|
#include <helpers/ConstantTadHelper.h>
|
|
#include <execution/Threads.h>
|
|
|
|
using namespace simdOps;
|
|
|
|
namespace functions {
|
|
namespace broadcast {
|
|
|
|
template <typename X, typename Y>
|
|
void BroadcastBool<X, Y>::exec(const int opNum,
|
|
const void *x, const Nd4jLong *xShapeInfo,
|
|
const void *y, const Nd4jLong *yShapeInfo,
|
|
void *z, const Nd4jLong *zShapeInfo,
|
|
void *extraParams,
|
|
int *dimension, int dimensionLength,
|
|
const Nd4jLong *xTadShapeInfo, const Nd4jLong *xTadOffset,
|
|
const Nd4jLong *zTadShapeInfo, const Nd4jLong *zTadOffset,
|
|
uint64_t start, uint64_t stop) {
|
|
DISPATCH_BY_OPNUM_TT(exec, PARAMS(x,
|
|
xShapeInfo,
|
|
y,
|
|
yShapeInfo,
|
|
z,
|
|
zShapeInfo,
|
|
extraParams,
|
|
dimension,
|
|
dimensionLength,
|
|
xTadShapeInfo,
|
|
xTadOffset,
|
|
zTadShapeInfo,
|
|
zTadOffset, start, stop), BROADCAST_BOOL_OPS);
|
|
}
|
|
|
|
template <typename X, typename Y>
|
|
void BroadcastBool<X, Y>::exec(const int opNum,
|
|
const void *x, const Nd4jLong *xShapeInfo,
|
|
const void *y, const Nd4jLong *yShapeInfo,
|
|
void *z, const Nd4jLong *zShapeInfo,
|
|
void* extraParams) {
|
|
|
|
DISPATCH_BY_OPNUM_TT(exec, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, extraParams), BROADCAST_BOOL_OPS);
|
|
}
|
|
|
|
template <typename X, typename Y>
|
|
void BroadcastBool<X, Y>::execInverse(const int opNum,
|
|
const void *x, const Nd4jLong *xShapeInfo,
|
|
const void *y, const Nd4jLong *yShapeInfo,
|
|
void *z, const Nd4jLong *zShapeInfo,
|
|
void *extraParams,
|
|
int *dimension, int dimensionLength,
|
|
const Nd4jLong *xTadShapeInfo, const Nd4jLong *xTadOffset,
|
|
const Nd4jLong *zTadShapeInfo, const Nd4jLong *zTadOffset,
|
|
uint64_t start, uint64_t stop) {
|
|
DISPATCH_BY_OPNUM_TT(execInverse, PARAMS(x,
|
|
xShapeInfo,
|
|
y,
|
|
yShapeInfo,
|
|
z,
|
|
zShapeInfo,
|
|
extraParams,
|
|
dimension,
|
|
dimensionLength,
|
|
xTadShapeInfo,
|
|
xTadOffset,
|
|
zTadShapeInfo,
|
|
zTadOffset, start, stop), BROADCAST_BOOL_OPS);
|
|
}
|
|
|
|
template <typename X, typename Z>
|
|
template<typename OpType>
|
|
void BroadcastBool<X, Z>::exec(const void *vx, const Nd4jLong *xShapeInfo,
|
|
const void *vy, const Nd4jLong *yShapeInfo,
|
|
void *vz, const Nd4jLong *zShapeInfo,
|
|
void *vextraParams,
|
|
int *dimension, int dimensionLength,
|
|
const Nd4jLong *xTadShapeInfo, const Nd4jLong *xTadOffset,
|
|
const Nd4jLong *zTadShapeInfo, const Nd4jLong *zTadOffset,
|
|
uint64_t start, uint64_t stop) {
|
|
|
|
auto x = reinterpret_cast<const X *>(vx);
|
|
auto y = reinterpret_cast<const X *>(vy);
|
|
auto z = reinterpret_cast<Z *>(vz);
|
|
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
|
|
|
//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 = sd::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
|
|
|
|
xTadShapeShapeInfo = const_cast<Nd4jLong*>(tadPack.primaryShapeInfo());
|
|
tadOffsets = const_cast<Nd4jLong*>(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 = sd::math::nd4j_max<int>(1, tadsPerThread);
|
|
threads = sd::math::nd4j_min<int>(threads, sd::Environment::getInstance()->maxThreads());
|
|
|
|
auto xEws = shape::elementWiseStride(xTadShapeShapeInfo);
|
|
auto yEws = shape::elementWiseStride(yShapeInfo);
|
|
auto zEws = shape::elementWiseStride(zTadShapeInfo);
|
|
|
|
const sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopXYZ(xTadShapeShapeInfo, yShapeInfo, zTadShapeInfo);
|
|
|
|
if (kindOfLoop == sd::LoopKind::EWS1) {
|
|
for (auto i = start; i < stop; 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], extraParams);
|
|
}
|
|
}
|
|
else if(kindOfLoop == sd::LoopKind::EWSNONZERO) {
|
|
for (auto i = start; i < stop; 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], extraParams);
|
|
};
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, yShapeInfo) && shape::haveSameShapeAndStrides(xTadShapeShapeInfo, zTadShapeInfo)) {
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, canCastX);
|
|
oZ[offset] = OpType::op(oX[offset], y[offset], extraParams);
|
|
}
|
|
};
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, yShapeInfo)) {
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastZ = sd::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, canCastX);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, canCastZ);
|
|
oZ[zOffset] = OpType::op(oX[offset], y[offset], extraParams);
|
|
}
|
|
};
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, zTadShapeInfo)) {
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint yShapeInfoCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, canCastX);
|
|
auto yOffset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, canCastY);
|
|
oZ[offset] = OpType::op(oX[offset], y[yOffset], extraParams);
|
|
}
|
|
};
|
|
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yShapeInfo, zTadShapeInfo)) {
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint yShapeInfoCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto xOffset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, canCastX);
|
|
auto offset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, canCastY);
|
|
oZ[offset] = OpType::op(oX[xOffset], y[offset], extraParams);
|
|
}
|
|
};
|
|
}
|
|
else {
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
uint yShapeInfoCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
|
|
bool canCastZ = sd::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oX = x + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto xOffset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, canCastX);
|
|
auto yOffset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, canCastY);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, canCastZ);
|
|
oZ[zOffset] = OpType::op(oX[xOffset], y[yOffset], extraParams);
|
|
}
|
|
};
|
|
}
|
|
}
|
|
|
|
template <typename X, typename Z>
|
|
template<typename OpType>
|
|
void BroadcastBool<X, Z>::execInverse(const void *vx, const Nd4jLong *xShapeInfo,
|
|
const void *vy, const Nd4jLong *yShapeInfo,
|
|
void *vz, const Nd4jLong *zShapeInfo,
|
|
void *vextraParams,
|
|
int *dimension, int dimensionLength,
|
|
const Nd4jLong *yTadShapeInfo, const Nd4jLong *yTadOffset,
|
|
const Nd4jLong *zTadShapeInfo, const Nd4jLong *zTadOffset,
|
|
uint64_t start, uint64_t stop) {
|
|
|
|
auto x = reinterpret_cast<const X *>(vx);
|
|
auto y = reinterpret_cast<const X *>(vy);
|
|
auto z = reinterpret_cast<Z *>(vz);
|
|
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
|
|
|
//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 = sd::ConstantTadHelper::getInstance()->tadForDimensions(yShapeInfo, dimension, dimensionLength);
|
|
|
|
yTadShapeShapeInfo = const_cast<Nd4jLong*>(tadPack.primaryShapeInfo());
|
|
tadOffsets = const_cast<Nd4jLong*>(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 = sd::math::nd4j_max<int>(1, tadsPerThread);
|
|
threads = sd::math::nd4j_min<int>(threads, sd::Environment::getInstance()->maxThreads());
|
|
|
|
auto yEws = shape::elementWiseStride(yTadShapeShapeInfo);
|
|
auto xEws = shape::elementWiseStride(xShapeInfo);
|
|
auto zEws = shape::elementWiseStride(zTadShapeInfo);
|
|
|
|
const sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopXYZ(yTadShapeShapeInfo, xShapeInfo, zTadShapeInfo);
|
|
|
|
if (kindOfLoop == sd::LoopKind::EWS1) {
|
|
for (auto i = start; i < stop; 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], extraParams);
|
|
}
|
|
}
|
|
else if(kindOfLoop == sd::LoopKind::EWSNONZERO) {
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oY = y + tadOffsets[i];
|
|
auto oZ = z + zTadOffset[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (uint f = 0; f < tadLength; f++)
|
|
oZ[f * zEws] = OpType::op(x[f * xEws], oY[f * yEws], extraParams);
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, xShapeInfo) && shape::haveSameShapeAndStrides(yTadShapeShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oY = y + 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, canCastY);
|
|
oZ[offset] = OpType::op(x[offset], oY[offset], extraParams);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, xShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastZ = sd::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, canCastY);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, canCastZ);
|
|
oZ[zOffset] = OpType::op(x[offset], oY[offset], extraParams);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto offset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, canCastY);
|
|
auto xOffset = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, canCastX);
|
|
oZ[offset] = OpType::op(x[xOffset], oY[offset], extraParams);
|
|
}
|
|
}
|
|
}
|
|
else if(shape::haveSameShapeAndStrides(xShapeInfo, zTadShapeInfo)) {
|
|
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto yOffset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, canCastY);
|
|
auto offset = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, canCastX);
|
|
oZ[offset] = OpType::op(x[offset], oY[yOffset], extraParams);
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
uint tadShapeShapeInfoCast[MAX_RANK];
|
|
uint tadShapeInfoZCast[MAX_RANK];
|
|
bool canCastX = sd::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastY = sd::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
|
|
bool canCastZ = sd::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
|
|
|
|
for (auto i = start; i < stop; i ++) {
|
|
auto oZ = z + zTadOffset[i];
|
|
auto oY = y + tadOffsets[i];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (unsigned int f = 0; f < tadLength; f++) {
|
|
auto xOffset = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, canCastX);
|
|
auto yOffset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, canCastY);
|
|
auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, canCastZ);
|
|
oZ[zOffset] = OpType::op(x[xOffset], oY[yOffset], extraParams);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z, typename OpType>
|
|
static void execRank1(const X *x, const Nd4jLong *xShapeInfo, const X *y, const Nd4jLong *yShapeInfo, Z* z, const Nd4jLong *zShapeInfo, X* extraParams) {
|
|
|
|
uint zAxis0 = shape::sizeAt(zShapeInfo, 0);
|
|
Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, 0);
|
|
Nd4jLong yStrd0 = shape::strideAt(yShapeInfo, 0);
|
|
Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, 0);
|
|
|
|
auto func = PRAGMA_THREADS_FOR{
|
|
|
|
if(zStrd0 == 1 && xStrd0 == 1 && yStrd0 == 0) {
|
|
for (auto i0 = start; i0 < stop; ++i0)
|
|
z[i0] = OpType::op(x[i0], *y, extraParams);
|
|
}
|
|
else if(zStrd0 == 1 && xStrd0 == 0 && yStrd0 == 1) {
|
|
for (auto i0 = start; i0 < stop; ++i0)
|
|
z[i0] = OpType::op(*x, y[i0], extraParams);
|
|
}
|
|
else if(zStrd0 == 1 && xStrd0 == 1 && yStrd0 == 1) {
|
|
for (auto i0 = start; i0 < stop; ++i0)
|
|
z[i0] = OpType::op(x[i0], y[i0], extraParams);
|
|
}
|
|
else {
|
|
for (auto i0 = start; i0 < stop; ++i0)
|
|
z[i0 * zStrd0] = OpType::op(x[i0 * xStrd0], y[i0 * yStrd0], extraParams);
|
|
}
|
|
};
|
|
samediff::Threads::parallel_tad(func, 0, zAxis0);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z, typename OpType>
|
|
static void execRank2(const X *x, const Nd4jLong *xShapeInfo, const X *y, const Nd4jLong *yShapeInfo, Z* z, const Nd4jLong *zShapeInfo, X* extraParams) {
|
|
|
|
uint zAxis0 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 1);
|
|
Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 1);
|
|
Nd4jLong yStrd0 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 1);
|
|
Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 1);
|
|
|
|
uint zAxis1 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 0);
|
|
Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 0);
|
|
Nd4jLong yStrd1 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 0);
|
|
Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 0);
|
|
|
|
auto func = PRAGMA_THREADS_FOR{
|
|
|
|
for (auto i0 = start; i0 < stop; ++i0) {
|
|
|
|
auto x0 = x + i0 * xStrd0;
|
|
auto y0 = y + i0 * yStrd0;
|
|
auto z0 = z + i0 * zStrd0;
|
|
|
|
if(zStrd1 == 1 && xStrd1 == 1 && yStrd1 == 0)
|
|
for (uint i1 = 0; i1 < zAxis1; ++i1)
|
|
z0[i1] = OpType::op(x0[i1], *y0, extraParams);
|
|
else if(zStrd1 == 1 && xStrd1 == 0 && yStrd1 == 1)
|
|
for (uint i1 = 0; i1 < zAxis1; ++i1)
|
|
z0[i1] = OpType::op(*x0, y0[i1], extraParams);
|
|
else if(zStrd1 == 1 && xStrd1 == 1 && yStrd1 == 1)
|
|
for (uint i1 = 0; i1 < zAxis1; ++i1)
|
|
z0[i1] = OpType::op(x0[i1], y0[i1], extraParams);
|
|
else
|
|
for (uint i1 = 0; i1 < zAxis1; ++i1)
|
|
z0[i1 * zStrd1] = OpType::op(x0[i1 * xStrd1], y0[i1 * yStrd1], extraParams);
|
|
}
|
|
};
|
|
|
|
samediff::Threads::parallel_tad(func, 0, zAxis0);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z, typename OpType>
|
|
static void execRank3(const X *x, const Nd4jLong *xShapeInfo, const X *y, const Nd4jLong *yShapeInfo, Z* z, const Nd4jLong *zShapeInfo, X* extraParams) {
|
|
|
|
uint zAxis0 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 2);
|
|
Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 2);
|
|
Nd4jLong yStrd0 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 2);
|
|
Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 2);
|
|
|
|
uint zAxis1 = shape::sizeAt(zShapeInfo, 1);
|
|
Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, 1);
|
|
Nd4jLong yStrd1 = shape::strideAt(yShapeInfo, 1);
|
|
Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, 1);
|
|
|
|
uint zAxis2 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 0);
|
|
Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 0);
|
|
Nd4jLong yStrd2 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 0);
|
|
Nd4jLong zStrd2 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 0);
|
|
|
|
auto func = PRAGMA_THREADS_FOR_2D {
|
|
|
|
for (auto i0 = start_x; i0 < stop_x; ++i0) {
|
|
for (auto i1 = start_y; i1 < stop_y; ++i1) {
|
|
|
|
auto x1 = x + i0 * xStrd0 + i1 * xStrd1;
|
|
auto y1 = y + i0 * yStrd0 + i1 * yStrd1;
|
|
auto z1 = z + i0 * zStrd0 + i1 * zStrd1;
|
|
|
|
if(zStrd2 == 1 && xStrd2 == 1 && yStrd2 == 0)
|
|
for (uint i2 = 0; i2 < zAxis2; ++i2)
|
|
z1[i2] = OpType::op(x1[i2], *y1, extraParams);
|
|
else if(zStrd2 == 1 && xStrd2 == 0 && yStrd2 == 1)
|
|
for (uint i2 = 0; i2 < zAxis2; ++i2)
|
|
z1[i2] = OpType::op(*x1, y1[i2], extraParams);
|
|
else if(zStrd2 == 1 && xStrd2 == 1 && yStrd2 == 1)
|
|
for (uint i2 = 0; i2 < zAxis2; ++i2)
|
|
z1[i2] = OpType::op(x1[i2], y1[i2], extraParams);
|
|
else
|
|
for (uint i2 = 0; i2 < zAxis2; ++i2)
|
|
z1[i2 * zStrd2] = OpType::op(x1[i2 * xStrd2], y1[i2 * yStrd2], extraParams);
|
|
}
|
|
}
|
|
};
|
|
|
|
samediff::Threads::parallel_for(func, 0,zAxis0,1, 0,zAxis1,1);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z, typename OpType>
|
|
static void execRank4(const X *x, const Nd4jLong *xShapeInfo, const X *y, const Nd4jLong *yShapeInfo, Z* z, const Nd4jLong *zShapeInfo, X* extraParams) {
|
|
|
|
uint zAxis0 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 3);
|
|
Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 3);
|
|
Nd4jLong yStrd0 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 3);
|
|
Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 3);
|
|
|
|
uint zAxis1 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 2);
|
|
Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 2);
|
|
Nd4jLong yStrd1 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 2);
|
|
Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 2);
|
|
|
|
uint zAxis2 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 1);
|
|
Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 1);
|
|
Nd4jLong yStrd2 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 1);
|
|
Nd4jLong zStrd2 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 2 : 1);
|
|
|
|
uint zAxis3 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 0);
|
|
Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 0);
|
|
Nd4jLong yStrd3 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 0);
|
|
Nd4jLong zStrd3 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 0);
|
|
|
|
auto func = PRAGMA_THREADS_FOR_3D {
|
|
|
|
for (auto i0 = start_x; i0 < stop_x; ++i0) {
|
|
for (auto i1 = start_y; i1 < stop_y; ++i1) {
|
|
for (auto i2 = start_z; i2 < stop_z; ++i2) {
|
|
|
|
auto x2 = x + i0 * xStrd0 + i1 * xStrd1 + i2 * xStrd2;
|
|
auto y2 = y + i0 * yStrd0 + i1 * yStrd1 + i2 * yStrd2;
|
|
auto z2 = z + i0 * zStrd0 + i1 * zStrd1 + i2 * zStrd2;
|
|
|
|
if(zStrd3 == 1 && xStrd3 == 1 && yStrd3 == 0)
|
|
for (uint i3 = 0; i3 < zAxis3; ++i3)
|
|
z2[i3] = OpType::op(x2[i3], *y2, extraParams);
|
|
else if(zStrd3 == 1 && xStrd3 == 0 && yStrd3 == 1)
|
|
for (uint i3 = 0; i3 < zAxis3; ++i3)
|
|
z2[i3] = OpType::op(*x2, y2[i3], extraParams);
|
|
else if(zStrd3 == 1 && xStrd3 == 1 && yStrd3 == 1)
|
|
for (uint i3 = 0; i3 < zAxis3; ++i3)
|
|
z2[i3] = OpType::op(x2[i3], y2[i3], extraParams);
|
|
else
|
|
for (uint i3 = 0; i3 < zAxis3; ++i3)
|
|
z2[i3 * zStrd3] = OpType::op(x2[i3 * xStrd3], y2[i3 * yStrd3], extraParams);
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
samediff::Threads::parallel_for(func, 0,zAxis0,1, 0,zAxis1,1, 0,zAxis2,1);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z, typename OpType>
|
|
static void execRank5(const X *x, const Nd4jLong *xShapeInfo, const X *y, const Nd4jLong *yShapeInfo, Z* z, const Nd4jLong *zShapeInfo, X* extraParams) {
|
|
|
|
uint zAxis0 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 4);
|
|
Nd4jLong xStrd0 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 4);
|
|
Nd4jLong yStrd0 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 4);
|
|
Nd4jLong zStrd0 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 0 : 4);
|
|
|
|
uint zAxis1 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 3);
|
|
Nd4jLong xStrd1 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 3);
|
|
Nd4jLong yStrd1 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 3);
|
|
Nd4jLong zStrd1 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 1 : 3);
|
|
|
|
uint zAxis2 = shape::sizeAt(zShapeInfo, 2);
|
|
Nd4jLong xStrd2 = shape::strideAt(xShapeInfo, 2);
|
|
Nd4jLong yStrd2 = shape::strideAt(yShapeInfo, 2);
|
|
Nd4jLong zStrd2 = shape::strideAt(zShapeInfo, 2);
|
|
|
|
uint zAxis3 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 1);
|
|
Nd4jLong xStrd3 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 1);
|
|
Nd4jLong yStrd3 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 1);
|
|
Nd4jLong zStrd3 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 3 : 1);
|
|
|
|
uint zAxis4 = shape::sizeAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 4 : 0);
|
|
Nd4jLong xStrd4 = shape::strideAt(xShapeInfo, shape::order(zShapeInfo) == 'c' ? 4 : 0);
|
|
Nd4jLong yStrd4 = shape::strideAt(yShapeInfo, shape::order(zShapeInfo) == 'c' ? 4 : 0);
|
|
Nd4jLong zStrd4 = shape::strideAt(zShapeInfo, shape::order(zShapeInfo) == 'c' ? 4 : 0);
|
|
|
|
auto func = PRAGMA_THREADS_FOR_3D {
|
|
|
|
for (auto i0 = start_x; i0 < stop_x; ++i0) {
|
|
for (auto i1 = start_y; i1 < stop_y; ++i1) {
|
|
for (auto i2 = start_z; i2 < stop_z; ++i2) {
|
|
for (uint i3 = 0; i3 < zAxis3; ++i3) {
|
|
|
|
auto x3 = x + i0 * xStrd0 + i1 * xStrd1 + i2 * xStrd2 + i3 * xStrd3;
|
|
auto y3 = y + i0 * yStrd0 + i1 * yStrd1 + i2 * yStrd2 + i3 * yStrd3;
|
|
auto z3 = z + i0 * zStrd0 + i1 * zStrd1 + i2 * zStrd2 + i3 * zStrd3;
|
|
|
|
if(zStrd4 == 1 && xStrd4 == 1 && yStrd4 == 0)
|
|
for (uint i4 = 0; i4 < zAxis4; ++i4)
|
|
z3[i4] = OpType::op(x3[i4], *y3, extraParams);
|
|
else if(zStrd4 == 1 && xStrd4 == 0 && yStrd4 == 1)
|
|
for (uint i4 = 0; i4 < zAxis4; ++i4)
|
|
z3[i4] = OpType::op(*x3, y3[i4], extraParams);
|
|
else if(zStrd4 == 1 && xStrd4 == 1 && yStrd4 == 1)
|
|
for (uint i4 = 0; i4 < zAxis4; ++i4)
|
|
z3[i4] = OpType::op(x3[i4], y3[i4], extraParams);
|
|
else
|
|
for (uint i4 = 0; i4 < zAxis4; ++i4)
|
|
z3[i4 * zStrd4] = OpType::op(x3[i4 * xStrd4], y3[i4 * yStrd4], extraParams);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
samediff::Threads::parallel_for(func, 0,zAxis0,1, 0,zAxis1,1, 0,zAxis2,1);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z, typename OpType>
|
|
static void execDefault(const X *x, const Nd4jLong *xShapeInfo, const X *y, const Nd4jLong *yShapeInfo, Z* z, const Nd4jLong *zShapeInfo, X* extraParams) {
|
|
|
|
const bool xzSameOffsets = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo);
|
|
const bool yzSameOffsets = shape::haveSameShapeAndStrides(yShapeInfo, zShapeInfo);
|
|
|
|
const int rank = shape::rank(zShapeInfo); // xRank = yRank = zRank
|
|
|
|
auto func = PRAGMA_THREADS_FOR{
|
|
|
|
int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK];
|
|
|
|
for (auto i = start; i < stop; ++i) {
|
|
|
|
shape::index2coordsCPU(start, i, zShapeInfo, zCoords);
|
|
|
|
for (uint j = 0; j < rank; ++j) {
|
|
xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j];
|
|
yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j];
|
|
}
|
|
|
|
const auto zOffset = shape::getOffset(zShapeInfo, zCoords);
|
|
const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords);
|
|
const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords);
|
|
|
|
z[zOffset] = OpType::op(x[xOffset], y[yOffset], extraParams);
|
|
}
|
|
};
|
|
|
|
samediff::Threads::parallel_for(func, 0, shape::length(zShapeInfo));
|
|
}
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z>
|
|
template<typename OpType>
|
|
void BroadcastBool<X, Z>::exec(const void *vx, const Nd4jLong *xShapeInfo,
|
|
const void *vy, const Nd4jLong *yShapeInfo,
|
|
void *vz, const Nd4jLong *zShapeInfo,
|
|
void *vextraParams) {
|
|
|
|
const X* x = reinterpret_cast<const X*>(vx);
|
|
const X* y = reinterpret_cast<const X*>(vy);
|
|
Z* z = reinterpret_cast<Z*>(vz);
|
|
|
|
X* extraParams = reinterpret_cast<X*>(vextraParams);
|
|
|
|
const int rank = shape::rank(zShapeInfo); // xRank = yRank = zRank
|
|
|
|
switch (rank) {
|
|
|
|
case 1:
|
|
execRank1<X,Z, OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, extraParams);
|
|
break;
|
|
case 2:
|
|
execRank2<X,Z, OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, extraParams);
|
|
break;
|
|
case 3:
|
|
execRank3<X,Z, OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, extraParams);
|
|
break;
|
|
case 4:
|
|
execRank4<X,Z, OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, extraParams);
|
|
break;
|
|
case 5:
|
|
execRank5<X,Z, OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, extraParams);
|
|
break;
|
|
default:
|
|
execDefault<X,Z, OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, extraParams);
|
|
}
|
|
}
|
|
|
|
//BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT BroadcastBool, , LIBND4J_TYPES, BOOL_TYPES);
|
|
|
|
|
|
}
|
|
} |