* - start working on implementation of sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - improving householder procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - further polishing householder stuff Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing hh pivoting qr procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing BiDiagonalUp procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing householder sequence class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing jacobi svd class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing class which performs Hessenberg decomposition of square matrix Signed-off-by: Yurii <iuriish@yahoo.com> * - add static method to JacobiSVD class which makes the continuous Givens rotation generation algorithm Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing auxiliary methods of Schur decomp class Signed-off-by: Yurii <iuriish@yahoo.com> * some references here and there Signed-off-by: raver119 <raver119@gmail.com> * - trying figure out difference between eigen and our Schur alg Signed-off-by: Yurii <iuriish@yahoo.com> * - testing fixing bugs in Schur decomposition op Signed-off-by: Yurii <iuriish@yahoo.com> * - start to implement class which performs calculation of eigen values and vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - add to EigenValsAndVecs method which calculates complex eigen vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in EigenValsAndVecs class Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing triangularSolver class Signed-off-by: Yurii <iuriish@yahoo.com> * Added a 2D routine for triangular systems solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored another test for triangularSolve2D. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored test for triangularSolve for vector-bar case. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * - implementation of FullPivLU class Signed-off-by: Yurii <iuriish@yahoo.com> * - fix bugs in FullPivLU::solve method Signed-off-by: Yurii <iuriish@yahoo.com> * - correct permutation vector in FullPivLU::solve Signed-off-by: Yurii <iuriish@yahoo.com> * - correct include headers Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - include sqrtm classes to cuda folder, investigate in what places synchronization doesn't work Signed-off-by: Yurii <iuriish@yahoo.com> * Added implementation for cuda triangularSolve2D and also refactored triangularSolve2D for cpu. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste implementations. Signed-off-by: shugeo <sgazeos@gmail.com> * - make offset calculation faster in t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - rename refference T& NDArray::t<> method Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on cuda sqrtm Signed-off-by: Yurii <iuriish@yahoo.com> * - provide correct synchronization to device in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - add tests for sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - correct fails which appeared while testing on jenkins Signed-off-by: Yurii <iuriish@yahoo.com> * - trying to find out mistake in svd::deflation method Signed-off-by: Yurii <iuriish@yahoo.com> * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. Signed-off-by: Yurii <iuriish@yahoo.com> * - change call semantic of r<> and t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - ged rid of ambiguity in * operator overloads for windows buikd Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 3 Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts with master Signed-off-by: Yurii <iuriish@yahoo.com> * cmakelists updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - minor fix in merge cpu helper - make use of reference getter Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
127 lines
4.6 KiB
C++
127 lines
4.6 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 20.04.2018
|
|
//
|
|
|
|
|
|
|
|
#include <ops/declarable/helpers/transforms.h>
|
|
#include <helpers/Loops.h>
|
|
#include <graph/RandomGenerator.h>
|
|
#include <numeric>
|
|
#include <helpers/ShapeUtils.h>
|
|
|
|
namespace sd {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
template <typename T>
|
|
void randomShuffle_(NDArray& input, NDArray& output, sd::graph::RandomGenerator& rng, const bool isInplace) {
|
|
|
|
// check edge cases first
|
|
int temp;
|
|
const int firstDim = input.sizeAt(0);
|
|
if(input.lengthOf() == 1 || firstDim == 1) {
|
|
|
|
if(!isInplace)
|
|
output.assign(input);
|
|
}
|
|
else if (input.isVector() || shape::isLikeVector(input.shapeInfo(), temp)) {
|
|
|
|
// apply Fisher-Yates shuffle
|
|
if(isInplace) {
|
|
//PRAGMA_OMP_PARALLEL_FOR_IF((firstDim-1) > Environment::getInstance()->tadThreshold())
|
|
for(int i = firstDim-1; i > 0; --i) {
|
|
int r = rng.relativeInt(i) % i;
|
|
if(i == r)
|
|
continue;
|
|
T t0 = input.t<T>(i);
|
|
T t1 = input.t<T>(r);
|
|
//math::nd4j_swap<T>(input(i), input(r));
|
|
input.r<T>(i) = t1;
|
|
input.r<T>(r) = t0;
|
|
}
|
|
}
|
|
else {
|
|
std::vector<int> indices(firstDim);
|
|
std::iota(indices.begin(), indices.end(), 0);
|
|
output.p<T>(Nd4jLong(0), input.e<T>(0));
|
|
|
|
// FIXME: parallelism!!
|
|
for(int i = firstDim-1; i > 0; --i) {
|
|
int r = rng.relativeInt(i) % i;
|
|
output.r<T>(i) = input.t<T>(indices[r]);
|
|
if(i == r)
|
|
continue;
|
|
|
|
output.r<T>(r) = input.t<T>(indices[i]);
|
|
math::nd4j_swap<int>(indices[i], indices[r]);
|
|
}
|
|
rng.rewindH(firstDim-1);
|
|
}
|
|
}
|
|
else {
|
|
|
|
// evaluate sub-arrays list of input array through all dimensions excluding first one
|
|
std::vector<int> dimensions = ShapeUtils::evalDimsToExclude(input.rankOf(), {0});
|
|
auto subArrsListIn = input.allTensorsAlongDimension(dimensions);
|
|
|
|
// apply Fisher-Yates shuffle
|
|
if(isInplace) {
|
|
//PRAGMA_OMP_PARALLEL_FOR_IF((firstDim-1) > Environment::getInstance()->elementwiseThreshold())
|
|
for(int i = firstDim - 1; i > 0; --i) {
|
|
int r = rng.relativeInt(i) % i;
|
|
|
|
if(i == r)
|
|
continue;
|
|
subArrsListIn.at(i)->swapUnsafe(*subArrsListIn.at(r));
|
|
}
|
|
}
|
|
else {
|
|
// evaluate sub-arrays list of output array through all dimensions excluding first one
|
|
auto subArrsListOut = output.allTensorsAlongDimension(dimensions);
|
|
std::vector<int> indices(firstDim);
|
|
std::iota(indices.begin(), indices.end(), 0);
|
|
bool isZeroShuffled = false;
|
|
//PRAGMA_OMP_PARALLEL_FOR_IF((firstDim-1) > Environment::getInstance()->tadThreshold())
|
|
for(int i = firstDim - 1; i > 0; --i) {
|
|
int r = rng.relativeInt(i) % i;
|
|
subArrsListOut.at(i)->assign(subArrsListIn.at(indices[r]));
|
|
if(r == 0)
|
|
isZeroShuffled = true;
|
|
if(i == r)
|
|
continue;
|
|
subArrsListOut.at(r)->assign(subArrsListIn.at(indices[i]));
|
|
math::nd4j_swap<int>(indices[i], indices[r]);
|
|
}
|
|
if(!isZeroShuffled)
|
|
subArrsListOut.at(0)->assign(subArrsListIn.at(0));
|
|
}
|
|
rng.rewindH(firstDim-1);
|
|
}
|
|
|
|
}
|
|
|
|
void randomShuffle(sd::LaunchContext * context, NDArray& input, NDArray& output, sd::graph::RandomGenerator& rng, const bool isInplace) {
|
|
BUILD_SINGLE_SELECTOR(input.dataType(), randomShuffle_, (input, output, rng, isInplace), LIBND4J_TYPES);
|
|
}
|
|
}
|
|
}
|
|
}
|