cavis/libnd4j/include/ops/declarable/helpers/cuda/qr.cu

181 lines
7.6 KiB
Plaintext
Raw Normal View History

Shugeo qr (#153) * Added qr op implementation. Initial version. * Fixed doc for qr op. Signed-off-by: shugeo <sgazeos@gmail.com> * Implementation of QR decomposition. CPU platform version. * Added a pair of tests for qr op testing. Signed-off-by: shugeo <sgazeos@gmail.com> * QR implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected norm using. * Properly calculated intermediate results with QR decomposition. * Another step to implement QR algorithm by householder. * Cpu implementatio for QR decomposition. The first working edition. * Corrected test to QR decomposition. * Added tad multithreading with QR implementation. * Finished cpu implementation for QR decomposition helpers. * Refactored tests and improved multithreading. * Refactored QR cpu implementation and update cuda implementation helpers. * Cuda QR helper implementation. The first working edition. * Eliminated waste prints. * Restore multithreading with cuda implementation. * Ops names corrected * Refactored qr op helpers to optimize. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste manual ticking. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored memory allocation to avoid waste memory usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored matrixMinor method both for cuda and cpu platforms. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored method of vmul to use raw buffers instead type conversion. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored temporary array of matricies. Signed-off-by: shugeo <sgazeos@gmail.com> Co-authored-by: Alexander Stoyakin <alexander.stoyakin@gmail.com> Co-authored-by: raver119 <raver119@gmail.com>
2020-01-22 11:59:36 +01:00
/*******************************************************************************
* 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 George A. Shulinok <sgazeos@gmail.com>
//
#include <ops/declarable/helpers/qr.h>
#include <NDArrayFactory.h>
#include <MmulHelper.h>
namespace nd4j {
namespace ops {
namespace helpers {
template <typename T>
static __global__ void matrixMinorKernel(T* outBuffer, Nd4jLong* outShape, T* inBuffer, Nd4jLong* inShape, Nd4jLong column, Nd4jLong rows, Nd4jLong columns) {
// auto tid = threadIdx.x + blockDim.x * blockIdx.x;
// auto step = blockDim.x * gridDim.x;
// if (threadIdx.x == 0) {
// for (auto i = tid; i < column; i += step) {
// Nd4jLong diagPos[] = {i, i};
// auto zIndex = shape::getOffset(outShape, diagPos);
// outBuffer[zIndex] = T(1.f);
// }
// }
// __syncthreads();
for (auto i = blockIdx.x; i < rows; i += gridDim.x)
for (auto j = threadIdx.x; j < columns; j += blockDim.x) {
Nd4jLong pos[] = {i,j};
auto zIndex = shape::getOffset(outShape, pos);
auto xIndex = shape::getOffset(inShape, pos);
if (i < column || j < column) {
outBuffer[zIndex] = i != j?T(0.f):T(1.f);
}
else
outBuffer[zIndex] = inBuffer[xIndex]; //m.t<T>(i,j) = in.t<T>(i,j);
}
}
template <typename T>
NDArray matrixMinor(LaunchContext* context, NDArray& in, Nd4jLong col) {
NDArray m = in.ulike();
m.setIdentity();
m({col, m.rows(), col, m.columns()}).assign(in({col, m.rows(), col, m.columns()}));
// auto stream = context->getCudaStream();
// matrixMinorKernel<T><<<128, 128, 256, *stream>>>(m.dataBuffer()->specialAsT<T>(), m.specialShapeInfo(),
// matrixMinorKernel<T><<<128, 128, 256, *stream>>>(m.dataBuffer()->specialAsT<T>(), m.specialShapeInfo(),
// reinterpret_cast<T*>(in.specialBuffer()), in.specialShapeInfo(), col, in.rows(), in.columns());
//
m.tickWriteDevice();
return m;
}
/* m = I - v v^T */
template <typename T>
static __global__ void vmulKernel(T* resBuf, Nd4jLong* resShape, T const* vBuff, Nd4jLong const* vShape, Nd4jLong n) {
for (auto i = blockIdx.x; i < n; i += gridDim.x)
for (auto j = threadIdx.x; j < n; j += blockDim.x) {
Nd4jLong posR[] = {i, j};
auto indexR = shape::getOffset(resShape, posR);
auto indexX = shape::getIndexOffset(i, vShape);
auto indexY = shape::getIndexOffset(j, vShape);
resBuf[indexR] = T(-2.f) * vBuff[indexX] * vBuff[indexY] + (i != j?T(0.f):T(1.f));
}
}
template <typename T>
NDArray vmul(LaunchContext* context, NDArray const& v, int n)
{
NDArray res('c', {n,n}, v.dataType(), context); // x = matrix_new(n, n);
auto stream = context->getCudaStream();
vmulKernel<T><<<128, 128, 128, *stream>>>(res.dataBuffer()->specialAsT<T>(), res.specialShapeInfo(),
reinterpret_cast<T const*>(v.getSpecialBuffer()), v.getSpecialShapeInfo(), n);
return res;
}
template <typename T>
static bool diagonalIsPositive(NDArray* matrix, Nd4jLong k) {
T hVal;
Nd4jLong pos[] = {k, k};
auto shift = shape::getOffset(matrix->shapeInfo(), pos);
cudaMemcpy(&hVal, matrix->specialBuffer(), sizeof(T), cudaMemcpyDeviceToHost);
return hVal > T(0.f);
}
template <typename T>
void qrSingle(LaunchContext* context, NDArray* matrix, NDArray* Q, NDArray* R, bool const fullMatricies) {
Nd4jLong M = matrix->sizeAt(0);
Nd4jLong N = matrix->sizeAt(1);
auto resQ = fullMatricies?Q->ulike():NDArrayFactory::create<T>(matrix->ordering(), {M,M}, Q->getContext());
auto resR = fullMatricies?R->ulike():matrix->ulike();
std::vector<NDArray> q(M);
NDArray z = *matrix;
NDArray e('c', {M}, DataTypeUtils::fromT<T>()); // two internal buffers and scalar for squared norm
for (auto k = 0; k < N && k < M - 1; k++) { // loop for columns, but not further then row number
e.nullify();
z = matrixMinor<T>(context, z, k); // minor computing for current column with given matrix z (initally is a input matrix)
auto currentColumn = z({0, 0, k, k + 1}); // retrieve k column from z to x buffer
auto norm = currentColumn.reduceAlongDimension(reduce::Norm2, {0});
if (diagonalIsPositive<T>(matrix, k)) //matrix->t<T>(k,k) > T(0.f)) // negate on positive matrix diagonal element
norm.applyTransform(transform::Neg, norm); // *= -1.f;//-norm.t<T>(0);
e.p(k, norm); // e - is filled by 0 vector except diagonal element (filled by 1)
e += currentColumn; // e[i] = x[i] + a * e[i] for each i from 0 to n - 1
auto normE = e.reduceAlongDimension(reduce::Norm2, {0});
e /= normE;
q[k] = vmul<T>(context, e, M);
auto qQ = z.ulike();
MmulHelper::matmul(&q[k], &z, &qQ, false, false);
z = std::move(qQ);
}
resQ.assign(q[0]); //
// MmulHelper::matmul(&q[0], matrix, &resR, false, false);
for (int i = 1; i < N && i < M - 1; i++) {
auto tempResQ = resQ;
MmulHelper::matmul(&q[i], &resQ, &tempResQ, false, false);
resQ = std::move(tempResQ);
}
MmulHelper::matmul(&resQ, matrix, &resR, false, false);
// resR *= -1.f;
resQ.transposei();
if (fullMatricies) {
Q->assign(resQ);
R->assign(resR);
}
else {
Q->assign(resQ({0, 0, 0, N}));
R->assign(resR({0, N, 0, 0}));
}
}
template <typename T>
Shugeo solve ls (#203) * lstsq op. Initial commit. Signed-off-by: shugeo <sgazeos@gmail.com> * Least squares linear problem solve op (lstsq). Cpu draft implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed shape routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Added test for lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Rectification for lstsq op implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected test to avoid numerical inconsistensy. Signed-off-by: shugeo <sgazeos@gmail.com> * Added prints for check computing. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected tests to use evalueate facility instead. Signed-off-by: shugeo <sgazeos@gmail.com> * CPU implementation of MatrixSolveLs op and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Added cuda implementation for helpers with lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored tests for lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Added processing for empty inputs. Signed-off-by: shugeo <sgazeos@gmail.com> * Merged tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored lstsq op for fast case. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed test. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed some issues with solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed lstsq op to avoid erros. Signed-off-by: shugeo <sgazeos@gmail.com> * Added kernel for giagonal factor Signed-off-by: shugeo <sgazeos@gmail.com> * lstsq wrapper and triangular_solve fixed * Added proper processing empty inputs and test. Signed-off-by: shugeo <sgazeos@gmail.com> * SequenceMask test * Build fixed * Added proper processing of empty inputs with solve op. Signed-off-by: shugeo <sgazeos@gmail.com> * Mapping added * Added check of input shapes with solve op. Signed-off-by: shugeo <sgazeos@gmail.com> * Added a couple of tests for lstsq op and minor changes with cuda helper for one.' Signed-off-by: shugeo <sgazeos@gmail.com> * Tests on * Refactored test for lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed test * Added another approach for lstsq op aka solve_ls. Signed-off-by: shugeo <sgazeos@gmail.com> * Finished cpu part for solve_ls op helpers. * Added helper for low triangular matrix inversion. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored alternate solve_ls cpu implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Removed alternate approach for solve_ls op. Added multithreading with matrix inversion. Signed-off-by: shugeo <sgazeos@gmail.com> * Assert fixed * Refactored multithreading for inverse matricies. Signed-off-by: shugeo <sgazeos@gmail.com> Co-authored-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
2020-02-28 09:37:26 +01:00
void qr_(LaunchContext* context, NDArray const* input, NDArray* outputQ, NDArray* outputR, bool const fullMatricies) {
Shugeo qr (#153) * Added qr op implementation. Initial version. * Fixed doc for qr op. Signed-off-by: shugeo <sgazeos@gmail.com> * Implementation of QR decomposition. CPU platform version. * Added a pair of tests for qr op testing. Signed-off-by: shugeo <sgazeos@gmail.com> * QR implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected norm using. * Properly calculated intermediate results with QR decomposition. * Another step to implement QR algorithm by householder. * Cpu implementatio for QR decomposition. The first working edition. * Corrected test to QR decomposition. * Added tad multithreading with QR implementation. * Finished cpu implementation for QR decomposition helpers. * Refactored tests and improved multithreading. * Refactored QR cpu implementation and update cuda implementation helpers. * Cuda QR helper implementation. The first working edition. * Eliminated waste prints. * Restore multithreading with cuda implementation. * Ops names corrected * Refactored qr op helpers to optimize. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste manual ticking. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored memory allocation to avoid waste memory usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored matrixMinor method both for cuda and cpu platforms. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored method of vmul to use raw buffers instead type conversion. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored temporary array of matricies. Signed-off-by: shugeo <sgazeos@gmail.com> Co-authored-by: Alexander Stoyakin <alexander.stoyakin@gmail.com> Co-authored-by: raver119 <raver119@gmail.com>
2020-01-22 11:59:36 +01:00
Nd4jLong lastDim = input->rankOf() - 1;
Nd4jLong preLastDim = input->rankOf() - 2;
NDArray::prepareSpecialUse({outputQ, outputR}, {input});
ResultSet listOutQ(outputQ->allTensorsAlongDimension({(int)preLastDim, (int)lastDim}));
ResultSet listOutR(outputR->allTensorsAlongDimension({(int)preLastDim, (int)lastDim}));
ResultSet listInput(input->allTensorsAlongDimension({(int)preLastDim, (int)lastDim}));
auto start = 0;
auto stop = listInput.size();
auto increment = 1;
for (auto batch = start; batch < stop; batch += increment) {
//qr here
qrSingle<T>(context, listInput.at(batch), listOutQ.at(batch), listOutR.at(batch), fullMatricies);
}
NDArray::registerSpecialUse({outputQ, outputR}, {input});
}
Shugeo solve ls (#203) * lstsq op. Initial commit. Signed-off-by: shugeo <sgazeos@gmail.com> * Least squares linear problem solve op (lstsq). Cpu draft implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed shape routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Added test for lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Rectification for lstsq op implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected test to avoid numerical inconsistensy. Signed-off-by: shugeo <sgazeos@gmail.com> * Added prints for check computing. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected tests to use evalueate facility instead. Signed-off-by: shugeo <sgazeos@gmail.com> * CPU implementation of MatrixSolveLs op and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Added cuda implementation for helpers with lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored tests for lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Added processing for empty inputs. Signed-off-by: shugeo <sgazeos@gmail.com> * Merged tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored lstsq op for fast case. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed test. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed some issues with solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed lstsq op to avoid erros. Signed-off-by: shugeo <sgazeos@gmail.com> * Added kernel for giagonal factor Signed-off-by: shugeo <sgazeos@gmail.com> * lstsq wrapper and triangular_solve fixed * Added proper processing empty inputs and test. Signed-off-by: shugeo <sgazeos@gmail.com> * SequenceMask test * Build fixed * Added proper processing of empty inputs with solve op. Signed-off-by: shugeo <sgazeos@gmail.com> * Mapping added * Added check of input shapes with solve op. Signed-off-by: shugeo <sgazeos@gmail.com> * Added a couple of tests for lstsq op and minor changes with cuda helper for one.' Signed-off-by: shugeo <sgazeos@gmail.com> * Tests on * Refactored test for lstsq op. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed test * Added another approach for lstsq op aka solve_ls. Signed-off-by: shugeo <sgazeos@gmail.com> * Finished cpu part for solve_ls op helpers. * Added helper for low triangular matrix inversion. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored alternate solve_ls cpu implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Removed alternate approach for solve_ls op. Added multithreading with matrix inversion. Signed-off-by: shugeo <sgazeos@gmail.com> * Assert fixed * Refactored multithreading for inverse matricies. Signed-off-by: shugeo <sgazeos@gmail.com> Co-authored-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
2020-02-28 09:37:26 +01:00
void qr(nd4j::LaunchContext* context, NDArray const* input, NDArray* outputQ, NDArray* outputR, bool const fullMatricies) {
Shugeo qr (#153) * Added qr op implementation. Initial version. * Fixed doc for qr op. Signed-off-by: shugeo <sgazeos@gmail.com> * Implementation of QR decomposition. CPU platform version. * Added a pair of tests for qr op testing. Signed-off-by: shugeo <sgazeos@gmail.com> * QR implementation. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected norm using. * Properly calculated intermediate results with QR decomposition. * Another step to implement QR algorithm by householder. * Cpu implementatio for QR decomposition. The first working edition. * Corrected test to QR decomposition. * Added tad multithreading with QR implementation. * Finished cpu implementation for QR decomposition helpers. * Refactored tests and improved multithreading. * Refactored QR cpu implementation and update cuda implementation helpers. * Cuda QR helper implementation. The first working edition. * Eliminated waste prints. * Restore multithreading with cuda implementation. * Ops names corrected * Refactored qr op helpers to optimize. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste manual ticking. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored memory allocation to avoid waste memory usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored matrixMinor method both for cuda and cpu platforms. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored method of vmul to use raw buffers instead type conversion. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored temporary array of matricies. Signed-off-by: shugeo <sgazeos@gmail.com> Co-authored-by: Alexander Stoyakin <alexander.stoyakin@gmail.com> Co-authored-by: raver119 <raver119@gmail.com>
2020-01-22 11:59:36 +01:00
BUILD_SINGLE_SELECTOR(input->dataType(), qr_, (context, input, outputQ, outputR, fullMatricies), FLOAT_TYPES);
}
}
}
}