* 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>
212 lines
7.8 KiB
C++
212 lines
7.8 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, created on 21.09.2018
|
|
// @author raver119@gmail.com
|
|
//
|
|
|
|
|
|
#include <helpers/TAD.h>
|
|
#include<ops/declarable/helpers/ismax.h>
|
|
#include <helpers/ConstantTadHelper.h>
|
|
#include <execution/Threads.h>
|
|
|
|
namespace sd {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
template <typename X, typename Z>
|
|
static void ismax_(const NDArray* input, NDArray* output, const std::vector<int>& dimensions) {
|
|
|
|
if (input->isVector()) {
|
|
int dimensionsLength = dimensions.size();
|
|
int length = input->lengthOf();
|
|
if (!dimensions.empty() && (input->shapeOf())[dimensions[0]] == 1) {
|
|
for (int i = 0; i < length; i++)
|
|
output->p<Z>(i, 1);
|
|
}
|
|
else {
|
|
int eleStride = shape::elementWiseStride(input->shapeInfo());
|
|
if (eleStride == 1) {
|
|
int maxIdx = 0;
|
|
auto currMax = input->e<X>(0);
|
|
if (length < ELEMENT_THRESHOLD) {
|
|
|
|
for (int i = 0; i < length; i++) {
|
|
if (currMax < input->e<X>(i)) {
|
|
currMax = input->e<X>(i);
|
|
maxIdx = i;
|
|
}
|
|
output->p<Z>(i, 0);
|
|
}
|
|
}
|
|
else {
|
|
|
|
{
|
|
int maxIdxLocal = maxIdx;
|
|
auto currMaxLocal = currMax;
|
|
|
|
for (int i = 0; i < length; i++) {
|
|
if (currMaxLocal < input->e<X>(i)) {
|
|
currMaxLocal = input->e<X>(i);
|
|
maxIdxLocal = i;
|
|
}
|
|
output->p<Z>(i, 0);
|
|
}
|
|
|
|
PRAGMA_OMP_CRITICAL
|
|
{
|
|
if (currMax < currMaxLocal) {
|
|
currMax = currMaxLocal;
|
|
maxIdx = maxIdxLocal;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
output->p<Z>(maxIdx, 1);
|
|
}
|
|
else {
|
|
int maxIdx = 0;
|
|
auto currMax = input->e<X>(0);
|
|
if (length < ELEMENT_THRESHOLD) {
|
|
|
|
for (int i = 0; i < length; i++) {
|
|
if (currMax < input->e<X>(i)) {
|
|
currMax = input->e<X>(i);
|
|
maxIdx = i;
|
|
}
|
|
output->p<Z>(i, 0.f);
|
|
}
|
|
}
|
|
else {
|
|
|
|
{
|
|
int maxIdxLocal = maxIdx;
|
|
auto currMaxLocal = currMax;
|
|
for (int i = 0; i < length; i++) {
|
|
if (currMaxLocal < input->e<X>(i)) {
|
|
currMaxLocal = input->e<X>(i);
|
|
maxIdxLocal = i;
|
|
}
|
|
output->p<Z>(i, 0.f);
|
|
}
|
|
|
|
PRAGMA_OMP_CRITICAL
|
|
{
|
|
if (currMax < currMaxLocal) {
|
|
currMax = currMaxLocal;
|
|
maxIdx = maxIdxLocal;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
output->p<Z>(maxIdx, 1);
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
int dimensionsLength = dimensions.size();
|
|
//int tads = tad.numTads;
|
|
//decompose in to several sub tads after
|
|
//moving all dimensions (in sorted order)
|
|
//to the back.
|
|
//permuted version of the input shape info for setting up the tad problem
|
|
auto tadPack = sd::ConstantTadHelper::getInstance()->tadForDimensions(input->shapeInfo(), const_cast<int*>(dimensions.data()), dimensionsLength);
|
|
auto tadPackZ = sd::ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), const_cast<int*>(dimensions.data()), dimensionsLength);
|
|
|
|
|
|
auto tadShapeShapeInfo = tadPack.primaryShapeInfo();
|
|
auto tadOffsets = tadPack.primaryOffsets();
|
|
auto zOfsets = tadPackZ.platformOffsets();
|
|
|
|
int tadLength = shape::length(tadShapeShapeInfo);
|
|
int tads = tadPack.numberOfTads();
|
|
|
|
int tadsPerThread = tads / TAD_THRESHOLD;
|
|
int num_threads = sd::math::nd4j_max<int>(1, tadsPerThread);
|
|
num_threads = sd::math::nd4j_min<int>(num_threads, omp_get_max_threads());
|
|
|
|
auto tadEWS = shape::elementWiseStride(tadShapeShapeInfo);
|
|
auto zEWS = shape::elementWiseStride(tadPackZ.primaryShapeInfo());
|
|
|
|
int span = (tads / num_threads) + 8;
|
|
|
|
auto func = PRAGMA_THREADS_FOR {
|
|
for (auto r = start; r < stop; r++) {
|
|
auto rX = const_cast<NDArray*>(input)->bufferAsT<X>() + tadOffsets[r];
|
|
auto rZ = output->bufferAsT<Z>() + zOfsets[r];
|
|
|
|
auto maxValue = rX[0];
|
|
int maxIdx = 0;
|
|
if (tadEWS == 1 && zEWS == 1) {
|
|
for (int i = 0; i < tadLength; i++) {
|
|
if (rX[i] > maxValue) {
|
|
maxIdx = i;
|
|
maxValue = rX[i];
|
|
}
|
|
}
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int i = 0; i < tadLength; i++) {
|
|
rZ[i] = maxIdx == i ? (Z) 1 : (Z) 0;
|
|
}
|
|
}
|
|
else if (tadEWS > 1 && zEWS > 1) {
|
|
for (int i = 0; i < tadLength; i++) {
|
|
if (rX[i * tadEWS] > maxValue) {
|
|
maxIdx = i;
|
|
maxValue = rX[i * tadEWS];
|
|
}
|
|
}
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int i = 0; i < tadLength; i++) {
|
|
rZ[i * zEWS] = maxIdx == i ? (Z) 1 : (Z) 0;
|
|
}
|
|
} else {
|
|
for (int i = 0; i < tadLength; i++) {
|
|
auto xOffset = shape::getIndexOffset(i, tadShapeShapeInfo);
|
|
if (rX[xOffset] > maxValue) {
|
|
maxIdx = i;
|
|
maxValue = rX[xOffset];
|
|
}
|
|
}
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int i = 0; i < tadLength; i++) {
|
|
auto zOffset = shape::getIndexOffset(i, tadPackZ.primaryShapeInfo());
|
|
rZ[zOffset] = maxIdx == i ? (Z) 1 : (Z) 0;
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
samediff::Threads::parallel_tad(func, 0, tads);
|
|
}
|
|
}
|
|
|
|
|
|
void ismax(sd::LaunchContext * context, const NDArray *input, NDArray *output, const std::vector<int>& dimensions) {
|
|
BUILD_DOUBLE_SELECTOR(input->dataType(), output->dataType(), ismax_, (input, output, dimensions), LIBND4J_TYPES, LIBND4J_TYPES);
|
|
}
|
|
|
|
|
|
}
|
|
}
|
|
}
|
|
|