cavis/libnd4j/include/helpers/cpu/loops/IndexReductionLoops.hpp
raver119 320924278d
Legacy API changes (#441)
* 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>
2020-05-09 08:06:14 +03:00

314 lines
14 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 Yurii Shyrma (iuriish@yahoo.com), created on 14.03.2019
//
#include <helpers/Loops.h>
using namespace simdOps;
//////////////////////////////////////////////////////////////////////////////
template <typename X, typename Z>
template <typename OpType>
void sd::IndexReductionLoops<X,Z>::loopIndexReduce(const X* x, const Nd4jLong* xShapeInfo,
Z* z, const Nd4jLong* zShapeInfo,
const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets,
X* extraParams) {
sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopTadXZ(xShapeInfo, zShapeInfo, tadShapeInfo);
if(kindOfLoop == sd::LoopKind::SMALLARR2DX)
kindOfLoop = sd::LoopKind::EWSNONZERO;
const Nd4jLong zLen = shape::length(zShapeInfo);
const Nd4jLong tadLen = shape::length(tadShapeInfo);
const uint tadEws = shape::elementWiseStride(tadShapeInfo);
const uint zEws = shape::elementWiseStride(zShapeInfo);
const Nd4jLong* tadShape = shape::shapeOf(const_cast<Nd4jLong*>(tadShapeInfo));
const Nd4jLong* tadStride = shape::stride(const_cast<Nd4jLong*>(tadShapeInfo));
switch (kindOfLoop) {
//*********************************************//
case sd::LoopKind::EWS1: {
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong j = 0; j < tadLen; j++) {
functions::indexreduce::IndexValue<X> comp(tad[j], j);
indexValue = OpType::update(indexValue, comp, extraParams);
}
z[i] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::EWSNONZERO: {
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong j = 0; j < tadLen; j++) {
functions::indexreduce::IndexValue<X> comp(tad[j * tadEws], j);
indexValue = OpType::update(indexValue, comp, extraParams);
}
z[i * zEws] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::RANK1: {
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong i0 = 0; i0 < tadLen; ++i0) {
functions::indexreduce::IndexValue<X> comp(tad[i0 * tadStride[0]], i0);
indexValue = OpType::update(indexValue, comp, extraParams);
}
z[i] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::RANK2: {
Nd4jLong newStride[2];
shape::updateStrides(2, tadShape, newStride, 'c');
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
const auto tadOffset = i0 * tadStride[0] + i1 * tadStride[1];
const auto tadIndex = i0 * newStride[0] + i1;
functions::indexreduce::IndexValue<X> comp(tad[tadOffset], tadIndex);
indexValue = OpType::update(indexValue, comp, extraParams);
}
}
z[i] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::RANK3: {
Nd4jLong newStride[3];
shape::updateStrides(3, tadShape, newStride, 'c');
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
const auto tadOffset = i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2];
const auto tadIndex = i0 * newStride[0] + i1 * newStride[1] + i2;
functions::indexreduce::IndexValue<X> comp(tad[tadOffset], tadIndex);
indexValue = OpType::update(indexValue, comp, extraParams);
}
}
}
z[i] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::RANK4: {
Nd4jLong newStride[4];
shape::updateStrides(4, tadShape, newStride, 'c');
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) {
const auto tadOffset = i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2] + i3 * tadStride[3];
const auto tadIndex = i0 * newStride[0] + i1 * newStride[1] + i2 * newStride[2] + i3;
functions::indexreduce::IndexValue<X> comp(tad[tadOffset], tadIndex);
indexValue = OpType::update(indexValue, comp, extraParams);
}
}
}
}
z[i] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::RANK5: {
Nd4jLong newStride[5];
shape::updateStrides(5, tadShape, newStride, 'c');
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) {
for (Nd4jLong i4 = 0; i4 < tadShape[4]; ++i4) {
const auto tadOffset = i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2] + i3 * tadStride[3] + i4 * tadStride[4];
const auto tadIndex = i0 * newStride[0] + i1 * newStride[1] + i2 * newStride[2] + i3 * newStride[3] + i4;
functions::indexreduce::IndexValue<X> comp(tad[tadOffset], tadIndex);
indexValue = OpType::update(indexValue, comp, extraParams);
}
}
}
}
}
z[i] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::X_EWSNONZERO: {
uint castZShapeInfo[MAX_RANK];
const bool canCastZ = sd::DataTypeUtils::castShapeInfo<uint>(zShapeInfo, castZShapeInfo);
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong j = 0; j < tadLen; j++) {
functions::indexreduce::IndexValue<X> comp(tad[j * tadEws], j);
indexValue = OpType::update(indexValue, comp, extraParams);
}
auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, canCastZ);
z[zOffset] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
case sd::LoopKind::Z_EWSNONZERO: {
uint castTadShapeInfo[MAX_RANK];
const bool canCastTad = sd::DataTypeUtils::castShapeInfo<uint>(tadShapeInfo, castTadShapeInfo);
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong j = 0; j < tadLen; j++) {
auto tadOffset = shape::indexOffset(j, tadShapeInfo, castTadShapeInfo, canCastTad);
functions::indexreduce::IndexValue<X> comp(tad[tadOffset], j);
indexValue = OpType::update(indexValue, comp, extraParams);
}
z[i * zEws] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
break;
//*********************************************//
default: {
uint castTadShapeInfo[MAX_RANK];
uint castZShapeInfo[MAX_RANK];
const bool canCastTad = sd::DataTypeUtils::castShapeInfo<uint>(tadShapeInfo, castTadShapeInfo);
const bool canCastZ = sd::DataTypeUtils::castShapeInfo<uint>(zShapeInfo, castZShapeInfo);
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
auto tad = const_cast<X *>(x) + tadOffsets[i];
auto indexValue = OpType::startingIndexValue(tad);
for (Nd4jLong j = 0; j < tadLen; j++) {
auto tadOffset = shape::indexOffset(j, tadShapeInfo, castTadShapeInfo, canCastTad);
functions::indexreduce::IndexValue<X> comp(tad[tadOffset], j);
indexValue = OpType::update(indexValue, comp, extraParams);
}
auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, canCastZ);
z[zOffset] = (Z) indexValue.index;
}
};
samediff::Threads::parallel_tad(func, 0, zLen);
}
}
}
template <typename X, typename Y>
void sd::IndexReductionLoops<X, Y>::wrapIndexReduce(const int opNum, const void* vx, const Nd4jLong* xShapeInfo, void* vz, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo, const Nd4jLong* tadOffsets, void* vextraParams) {
auto x = reinterpret_cast<const X *>(vx);
auto z = reinterpret_cast<Y *>(vz);
auto extraParams = reinterpret_cast<X *>(vextraParams);
DISPATCH_BY_OPNUM_TT(loopIndexReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams), INDEX_REDUCE_OPS);
}