1e9ff114aa
* Added atomic tests for atomicAdd, atomicSub and atomicDiv. * Fixed atomicAdd for 16bit ints. * Fixed atomicMul for 16 floats. * Eliminated waste prints. * Fixed problems with double type on matrix inverse helepers. * Eliminated commented wrong code. * Refactored atomicMul for 16bit types. * few more minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * Fixed fake_quant_with_min_max_vars_per_channel args processing.
869 lines
42 KiB
Plaintext
869 lines
42 KiB
Plaintext
/*******************************************************************************
|
|
* 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 <ops/declarable/helpers/top_k.h>
|
|
#include <MmulHelper.h>
|
|
#include <NDArrayFactory.h>
|
|
#include <Status.h>
|
|
#include <ConstantTadHelper.h>
|
|
#include <ShapeUtils.h>
|
|
|
|
#include <cusolverDn.h>
|
|
#include <cuda_exception.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// invert the second diagonal for lower diagonal matrix
|
|
template<typename T>
|
|
static __global__ void
|
|
invertKernelLow(void *invertedBuf, Nd4jLong *invertedShape, void *inputBuf, Nd4jLong *inputShape, Nd4jLong n) {
|
|
T* inverted = reinterpret_cast<T *>(invertedBuf);
|
|
T* input = reinterpret_cast<T*>(inputBuf);
|
|
|
|
auto start = threadIdx.x + blockIdx.x * blockDim.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
|
|
for (int i = start + 1; i < n; i += step) {
|
|
Nd4jLong pos[] = {i, i - 1};
|
|
Nd4jLong posX[] = {i, i};
|
|
Nd4jLong posY[] = {i - 1, i - 1};
|
|
auto xIndex = shape::getOffset(inputShape, pos);
|
|
auto dxIndex = shape::getOffset(inputShape, posX);
|
|
auto dyIndex = shape::getOffset(inputShape, posY);
|
|
auto zIndex = shape::getOffset(invertedShape, pos);
|
|
// invert lower triangular matrix
|
|
inverted[zIndex] = -input[xIndex] / (input[dxIndex] * input[dyIndex]);
|
|
// math::atomics::nd4j_atomicAdd(&inverted[zIndex], - input[xIndex] * inverted[iIndex] / input[dIndex]);
|
|
}
|
|
}
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// invert diagonal vals to upper diagonal matrix
|
|
template<typename T>
|
|
static __global__ void
|
|
upvertKernel(void *invertedBuf, Nd4jLong *invertedShape, void *inputBuf, Nd4jLong *inputShape, Nd4jLong n) {
|
|
T *inverted = reinterpret_cast<T *>(invertedBuf);
|
|
T *input = reinterpret_cast<T *>(inputBuf);
|
|
|
|
auto start = threadIdx.x + blockIdx.x * blockDim.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
|
|
for (int i = start; i < n; i += step) {
|
|
Nd4jLong pos[] = {i, i};
|
|
auto xIndex = shape::getOffset(inputShape, pos);
|
|
auto zIndex = shape::getOffset(invertedShape, pos);
|
|
// math::atomics::nd4j_atomicDiv(&inverted[zIndex], input[xIndex]);
|
|
// invert diagonal elements
|
|
inverted[zIndex] /= input[xIndex];
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// invert upper second diagonal
|
|
template<typename T>
|
|
static __global__ void
|
|
upvertKernelUp(void *invertedBuf, Nd4jLong *invertedShape, void *inputBuf, Nd4jLong *inputShape, Nd4jLong n) {
|
|
|
|
__shared__ T* inverted;
|
|
__shared__ T* input;
|
|
if (threadIdx.x == 0) {
|
|
inverted = reinterpret_cast<T *>(invertedBuf);
|
|
input = reinterpret_cast<T *>(inputBuf);
|
|
}
|
|
__syncthreads();
|
|
|
|
auto start = threadIdx.x + blockIdx.x * blockDim.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
|
|
for (int i = start; i < n - 1; i += step) {
|
|
Nd4jLong pos[] = {i, i + 1};
|
|
Nd4jLong posX[] = {i + 1, i + 1};
|
|
auto xIndex = shape::getOffset(inputShape, pos);
|
|
auto iIndex = shape::getOffset(invertedShape, posX);
|
|
auto zIndex = shape::getOffset(invertedShape, pos);
|
|
// invert upper matrix
|
|
math::atomics::nd4j_atomicAdd(&inverted[zIndex], -input[xIndex] * inverted[iIndex]); // / input[yIndex]);
|
|
//inputMatrix->t<T>(i, i + 1) * invertedMatrix->t<T>(i + 1, i + 1) / inputMatrix->t<T>(i, i)
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
template<typename T>
|
|
static __global__ void
|
|
invertLowKernel(void *invertedBuf, Nd4jLong *invertedShape, void *inputBuf, Nd4jLong *inputShape, Nd4jLong n) {
|
|
|
|
T *inverted = reinterpret_cast<T *>(invertedBuf);
|
|
T *input = reinterpret_cast<T *>(inputBuf);
|
|
if (threadIdx.x == 0) {
|
|
inverted = reinterpret_cast<T *>(invertedBuf);
|
|
input = reinterpret_cast<T *>(inputBuf);
|
|
}
|
|
__syncthreads();
|
|
|
|
auto tid = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = gridDim.x * blockDim.x;
|
|
|
|
for (int i = tid + 2; i < n; i += step) {
|
|
for (int j = i - 2; j >= 0; --j)
|
|
for (int k = 0; k < i; k++) {
|
|
Nd4jLong posZ[] = {i, j};
|
|
Nd4jLong posY[] = {k, j};
|
|
Nd4jLong posX[] = {i, k};
|
|
Nd4jLong posD[] = {i, i};
|
|
|
|
auto xIndex = shape::getOffset(inputShape, posX);
|
|
auto yIndex = shape::getOffset(invertedShape, posY);
|
|
auto dIndex = shape::getOffset(inputShape, posD);
|
|
auto zIndex = shape::getOffset(invertedShape, posZ);
|
|
// invert non-diagonal elements
|
|
math::atomics::nd4j_atomicAdd(&inverted[zIndex], -inverted[yIndex] * input[xIndex] / input[dIndex]);
|
|
}
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// Invertion of upper triangular matrix non-diagonal elements when main and second diagonals already processed
|
|
template<typename T>
|
|
static __global__ void
|
|
invertUpKernel(void *invertedBuf, Nd4jLong *invertedShape, void *inputBuf, Nd4jLong *inputShape, Nd4jLong n) {
|
|
__shared__ T* inverted;
|
|
__shared__ T* input;
|
|
|
|
if (threadIdx.x == 0) {
|
|
inverted = reinterpret_cast<T *>(invertedBuf);;
|
|
input = reinterpret_cast<T *>(inputBuf);
|
|
}
|
|
__syncthreads();
|
|
auto tid = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
|
|
for (int i = (int)n - tid - 2; i >= 0; i -= step) {
|
|
for (int j = i + 2; j < (int)n; j++)
|
|
for (int k = i; k < (int)n; k++) {
|
|
Nd4jLong posZ[] = {i, j};
|
|
Nd4jLong posY[] = {k, j};
|
|
Nd4jLong posX[] = {i, k};
|
|
// inversion with Joardan Gauss transformation
|
|
auto xIndex = shape::getOffset(inputShape, posX);
|
|
auto yIndex = shape::getOffset(invertedShape, posY);
|
|
auto zIndex = shape::getOffset(invertedShape, posZ);
|
|
// invert upper non-diagonal elements
|
|
math::atomics::nd4j_atomicAdd(&inverted[zIndex], -inverted[yIndex] * input[xIndex]);
|
|
}
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// procedure to invert lower-triangular matrix.
|
|
// In current case lower triangular matrix has main diagonal with general values
|
|
//
|
|
template<typename T>
|
|
static void invertLowerMatrix_(LaunchContext *context, NDArray *inputMatrix, NDArray *invertedMatrix) {
|
|
int n = inputMatrix->rows();
|
|
invertedMatrix->setIdentity();
|
|
|
|
if (inputMatrix->isIdentityMatrix()) return;
|
|
|
|
auto stream = context->getCudaStream();
|
|
|
|
// invert lower matrix
|
|
// invert main diagonal
|
|
upvertKernel<T><<<1, n, 512, *stream>>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
|
|
// invert the second diagonal
|
|
invertKernelLow<T><<<1, n, 512, *stream>>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
|
|
// invert non-diagonal elements
|
|
invertLowKernel<T><<<n, n, 512, *stream>>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// caller for invert lower matrix routine
|
|
void invertLowerMatrix(LaunchContext *context, NDArray *inputMatrix, NDArray *invertedMatrix) {
|
|
NDArray::prepareSpecialUse({invertedMatrix}, {inputMatrix});
|
|
BUILD_SINGLE_SELECTOR(inputMatrix->dataType(), invertLowerMatrix_, (context, inputMatrix, invertedMatrix), FLOAT_NATIVE);
|
|
NDArray::registerSpecialUse({invertedMatrix}, {inputMatrix});
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// procedure to invert upper-triangular matrix.
|
|
// In current case upper triangular matrix has main diagonal with all ones on it.
|
|
template<typename T>
|
|
static void invertUpperMatrix_(LaunchContext *context, NDArray* inputMatrix, NDArray* invertedMatrix) {
|
|
int n = inputMatrix->rows();
|
|
invertedMatrix->setIdentity();
|
|
auto stream = context->getCudaStream();
|
|
if (inputMatrix->isIdentityMatrix()) { // the inverse for I is I
|
|
return;
|
|
}
|
|
|
|
// invert upper matrix
|
|
// invert the second diagonal
|
|
upvertKernelUp<T><<<1, n, 512, *stream >>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(),
|
|
inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
|
|
|
|
// invert other elements
|
|
invertUpKernel<T><<<n, n, 512, *stream >>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(),inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// invertion of upper triangular matrix - runner routine
|
|
void invertUpperMatrix(LaunchContext *context, NDArray *inputMatrix, NDArray *invertedMatrix) {
|
|
NDArray::prepareSpecialUse({invertedMatrix}, {inputMatrix});
|
|
BUILD_SINGLE_SELECTOR(invertedMatrix->dataType(), invertUpperMatrix_, (context, inputMatrix, invertedMatrix), FLOAT_NATIVE);
|
|
NDArray::prepareSpecialUse({invertedMatrix}, {inputMatrix});
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// determinant kernel - accumulation product of all values on the main diagonal
|
|
template<typename T>
|
|
static __global__ void determinantKernel(T *compound, T *result, Nd4jLong len) {
|
|
auto start = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
for (auto i = start; i < len; i += step) {
|
|
auto pos = i * len + i; //shape::getOffset(0, shape::shapeOf(shape), shape::stride(shape), di, 2);
|
|
// multiply all diagonal elements
|
|
math::atomics::nd4j_atomicMul(&result[0], compound[pos]);
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// determinant logarithm - accumulation sum of all logarithm values on the main diagonal. All in logarithic values
|
|
// should be positive
|
|
template<typename T>
|
|
static __global__ void determinantLogKernel(T *compound, T *result, Nd4jLong len) {
|
|
auto start = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
for (auto i = start; i < len; i += step) {
|
|
auto pos = i * len + i; //shape::getOffset(0, shape::shapeOf(shape), shape::stride(shape), di, 2);
|
|
// sum logs of all diagonal elements
|
|
math::atomics::nd4j_atomicAdd(result, math::nd4j_log<T,T>(math::nd4j_abs(compound[pos])));
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// kernel to copy matrix with given shape to compound tensor with given pos
|
|
// output - a N-D tensor buffer with rank not less than 2, input - 2D square n x n matrix with n = rowLen
|
|
template<typename T, typename F>
|
|
static __global__ void
|
|
fillMatrix(void *output, Nd4jLong *outShape, void *input, Nd4jLong *inputShape, Nd4jLong pos, Nd4jLong rowLen) {
|
|
__shared__ F *matrix;
|
|
__shared__ T *inputBuf;
|
|
__shared__ Nd4jLong inputLen;
|
|
__shared__ Nd4jLong n2;
|
|
|
|
if (threadIdx.x == 0) {
|
|
matrix = reinterpret_cast<F*>(output);
|
|
inputBuf = reinterpret_cast<T*>(input);
|
|
inputLen = shape::length(inputShape);
|
|
n2 = rowLen * rowLen;
|
|
}
|
|
__syncthreads();
|
|
|
|
auto start = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
|
|
for (int k = pos + start, j = start; j < n2; k += step, j += step) {
|
|
auto xIndex = shape::getIndexOffset(k, inputShape);
|
|
matrix[j] = (F) inputBuf[xIndex];
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// same as above, but without type conversion
|
|
template<typename T>
|
|
static __global__ void
|
|
returnMatrix(void *output, Nd4jLong *outputShape, void *input, Nd4jLong *inputShape, Nd4jLong pos, Nd4jLong rowLen) {
|
|
__shared__ T* matrix;
|
|
__shared__ T* outputBuf;
|
|
__shared__ Nd4jLong outputLen;
|
|
__shared__ Nd4jLong n2;
|
|
|
|
if (threadIdx.x == 0) {
|
|
matrix = reinterpret_cast<T *>(input);
|
|
outputBuf = reinterpret_cast<T *>(output);
|
|
outputLen = shape::length(inputShape);
|
|
n2 = rowLen * rowLen;
|
|
}
|
|
__syncthreads();
|
|
auto start = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
|
|
for (int k = pos + start, j = start; j < n2; k += step, j += step) {
|
|
auto zIndex = shape::getIndexOffset(k, outputShape);
|
|
outputBuf[zIndex] = matrix[j];
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// fill up permutaion matrix kernel. Permutation matrix filled with zeros and ones
|
|
template<typename F>
|
|
static __global__ void fillUpPermutation(void *output, Nd4jLong *shape, int *source, int rowNum) {
|
|
F *permutation = reinterpret_cast<F *>(output);
|
|
|
|
auto start = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
for (auto i = start; i < rowNum; i += step) {
|
|
int val = source[i] - 1;
|
|
Nd4jLong posF[] = {i, val};
|
|
auto pos = shape::getOffset(shape, posF);
|
|
permutation[pos] = F(1.f);
|
|
}
|
|
}
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
// LUP decomposition runner - using CUBLAS SOLVER
|
|
// if permutation is given, then using LUP decomposition, LU decomposition otherwise
|
|
// L - lower triangular, U - upper triangular, P - permutation matricies
|
|
// PA = LU
|
|
//
|
|
// input - A matrix nxn
|
|
// compound - C matrix L + U - I, or main diagonal and lower - L matrix, from the 2nd diagonal - U matrix
|
|
template<typename T>
|
|
static void lup_(LaunchContext *context, NDArray *input, NDArray *compound, NDArray *permutation) {
|
|
auto stream = context->getCudaStream();
|
|
auto n = input->rows();
|
|
cusolverDnHandle_t cusolverH = nullptr;
|
|
// create solver handle
|
|
cusolverStatus_t status = cusolverDnCreate(&cusolverH);
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("Cannot create cuSolver handle", status);
|
|
}
|
|
// set solver stream
|
|
status = cusolverDnSetStream(cusolverH, *stream);
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("Cannot set up stream for cuda solver", status);
|
|
}
|
|
int lwork = 0;
|
|
int *d_info = nullptr;
|
|
// allocate memory for permutation vector
|
|
auto err = cudaMalloc((void **) &d_info, sizeof(int));
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot allocate memory for solver info buffer", err);
|
|
}
|
|
|
|
DataType dtype = input->dataType();
|
|
switch (dtype) { // there are two implementations with cublas for LUP decomposition - double and float
|
|
|
|
case DataType::DOUBLE: {
|
|
double *d_work = nullptr;
|
|
// compute internal buffer size
|
|
double *matrix = reinterpret_cast<double *>(input->specialBuffer());
|
|
status = cusolverDnDgetrf_bufferSize(
|
|
cusolverH,
|
|
n,
|
|
n,
|
|
matrix,
|
|
n,
|
|
&lwork);
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot create cuSolver handle", status);
|
|
}
|
|
|
|
err = cudaMalloc((void **) &d_work, sizeof(float) * lwork);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot allocate memory for solver data buffer",
|
|
err);
|
|
}
|
|
|
|
if (permutation == nullptr)
|
|
status = cusolverDnDgetrf(
|
|
cusolverH,
|
|
n,
|
|
n,
|
|
matrix,
|
|
n,
|
|
d_work,
|
|
nullptr,
|
|
d_info);
|
|
else {
|
|
NDArray permutVector('c', {n}, nd4j::DataType::INT32, context);
|
|
int *permutationBuf = reinterpret_cast<int *>(permutVector.specialBuffer());
|
|
status = cusolverDnDgetrf(
|
|
cusolverH,
|
|
n,
|
|
n,
|
|
matrix,
|
|
n,
|
|
d_work,
|
|
permutationBuf,
|
|
d_info);
|
|
fillUpPermutation<double> << < n, n, 1024, *stream >> >
|
|
(permutation->specialBuffer(), permutation->specialShapeInfo(), permutationBuf, n);
|
|
permutation->tickWriteDevice();
|
|
}
|
|
err = cudaFree(d_work);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot deallocate memory for solver data buffer",
|
|
err);
|
|
}
|
|
}
|
|
break;
|
|
case DataType::FLOAT32: {
|
|
float *matrix = reinterpret_cast<float*>(input->specialBuffer());
|
|
float *d_work = nullptr;
|
|
|
|
status = cusolverDnSgetrf_bufferSize(
|
|
cusolverH,
|
|
n,
|
|
n,
|
|
matrix,
|
|
n,
|
|
&lwork);
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot create cuSolver handle", status);
|
|
}
|
|
|
|
err = cudaMalloc((void **) &d_work, sizeof(float) * lwork);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot allocate memory for solver data buffer",
|
|
err);
|
|
}
|
|
|
|
if (permutation == nullptr)
|
|
status = cusolverDnSgetrf(
|
|
cusolverH,
|
|
n,
|
|
n,
|
|
matrix,
|
|
n,
|
|
d_work,
|
|
nullptr,
|
|
d_info);
|
|
else {
|
|
NDArray permutVector('c', {n}, nd4j::DataType::INT32, context);
|
|
int *permutationBuf = reinterpret_cast<int *>(permutVector.specialBuffer());
|
|
status = cusolverDnSgetrf(
|
|
cusolverH,
|
|
n,
|
|
n,
|
|
matrix,
|
|
n,
|
|
d_work,
|
|
permutationBuf,
|
|
d_info);
|
|
fillUpPermutation<T> <<< n, n, 128, *stream >> >
|
|
(permutation->specialBuffer(), permutation->specialShapeInfo(), permutationBuf, n);
|
|
permutation->tickWriteDevice();
|
|
}
|
|
err = cudaFree(d_work);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot deallocate memory for solver data buffer",
|
|
err);
|
|
}
|
|
|
|
}
|
|
}
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot make LU decomposition", status);
|
|
}
|
|
err = cudaFree(d_info);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::lup_: Cannot deallocate memory for solver info buffer", err);
|
|
}
|
|
cusolverDnDestroy(cusolverH);
|
|
// NDArray::registerSpecialUse({input}, {input});
|
|
input->tickWriteDevice();
|
|
}
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
|
|
BUILD_SINGLE_TEMPLATE(template void lup_,(LaunchContext * context, NDArray * input, NDArray * output, NDArray * permutation), FLOAT_NATIVE);
|
|
|
|
// ------------------------------------------------------------------------------------------------------------------ //
|
|
template<typename T>
|
|
static int determinant_(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
|
|
Nd4jLong n = input->sizeAt(-1);
|
|
Nd4jLong n2 = n * n;
|
|
std::vector<int> dims();
|
|
auto packX = ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(), {input->rankOf() - 2, input->rankOf() - 1});
|
|
//auto packZ = ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), {output->rankOf() - 1});
|
|
// DataType dtype = input->dataType();
|
|
// if (dtype != DataType::DOUBLE)
|
|
// dtype = DataType::FLOAT32;
|
|
auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, DataTypeUtils::fromT<T>(), context); //, block.getWorkspace());
|
|
auto det = NDArrayFactory::create<T>(1);
|
|
auto stream = context->getCudaStream();
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
dim3 launchDims(256, 256, 1024);
|
|
output->assign(1.f);
|
|
for (int e = 0; e < output->lengthOf(); e++) {
|
|
Nd4jLong pos = e * n2;
|
|
// if (matrix.dataType() == input->dataType())
|
|
fillMatrix<T, T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
|
|
// else
|
|
// fillMatrix<T, float><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
|
|
lup_<T>(context, &matrix, nullptr, nullptr);
|
|
// else
|
|
// lup_<float>(context, &matrix, nullptr, nullptr);
|
|
auto offset = shape::getIndexOffset(e, output->shapeInfo());
|
|
auto inputBuf = reinterpret_cast<T *>(matrix.specialBuffer());
|
|
auto outputBuf = reinterpret_cast<T *>(output->specialBuffer()) + offset;
|
|
// if (matrix.dataType() == input->dataType())
|
|
determinantKernel<T> << < launchDims.x, launchDims.y, launchDims.z, *stream >> >
|
|
(inputBuf, outputBuf, n);
|
|
// else
|
|
// determinantKernel<T, float><<<launchDims.x, launchDims.y, launchDims.z, *stream >>> (inputBuf, outputBuf, n);
|
|
}
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
|
|
return Status::OK();
|
|
}
|
|
|
|
int determinant(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), return determinant_, (context, input, output), FLOAT_NATIVE);
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
}
|
|
|
|
template<typename T>
|
|
int logAbsDeterminant_(LaunchContext *context, NDArray *input, NDArray *output) {
|
|
Nd4jLong n = input->sizeAt(-1);
|
|
Nd4jLong n2 = n * n;
|
|
std::vector<int> dims();
|
|
auto packX = ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(), {input->rankOf() - 2, input->rankOf() - 1});
|
|
//auto packZ = ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), {output->rankOf() - 1});
|
|
DataType dtype = input->dataType();
|
|
if (dtype != DataType::DOUBLE)
|
|
dtype = DataType::FLOAT32;
|
|
|
|
auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, dtype, context); //, block.getWorkspace());
|
|
auto det = NDArrayFactory::create<T>(1);
|
|
auto stream = context->getCudaStream();
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
dim3 launchDims(256, 256, 1024);
|
|
output->assign(0.f);
|
|
for (int e = 0; e < output->lengthOf(); e++) {
|
|
Nd4jLong pos = e * n2;
|
|
// if (matrix.dataType() == input->dataType())
|
|
fillMatrix<T, T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
|
|
// else
|
|
// fillMatrix<T, float><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
|
|
|
|
// if (matrix.dataType() == input->dataType())
|
|
lup_<T>(context, &matrix, nullptr, nullptr);
|
|
// else
|
|
// lup_<float>(context, &matrix, nullptr, nullptr);
|
|
auto offset = shape::getIndexOffset(e, output->shapeInfo());
|
|
auto inputBuf = reinterpret_cast<T *>(matrix.specialBuffer());
|
|
auto outputBuf = reinterpret_cast<T *>(output->specialBuffer()) + offset;
|
|
// if (matrix.dataType() == input->dataType())
|
|
determinantLogKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(inputBuf, outputBuf, n);
|
|
// else
|
|
// determinantLogKernel<T, float><<<launchDims.x, launchDims.y, launchDims.z, *stream >>> (inputBuf, outputBuf, n);
|
|
}
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
|
|
return Status::OK();
|
|
|
|
return ND4J_STATUS_OK;
|
|
}
|
|
|
|
int logAbsDeterminant(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), return logAbsDeterminant_, (context, input, output), FLOAT_NATIVE);
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
}
|
|
|
|
template<typename T>
|
|
static __global__ void
|
|
fillLowerUpperKernel(void *lowerBuf, Nd4jLong *lowerShape, void *upperBuf, Nd4jLong *upperShape,
|
|
void *matrixBuf, Nd4jLong *matrixShape, Nd4jLong n) {
|
|
|
|
__shared__ T *lowerMatrix;
|
|
__shared__ T *upperMatrix;
|
|
__shared__ T *matrix;
|
|
|
|
if (threadIdx.x == 0) {
|
|
lowerMatrix = reinterpret_cast<T *>(lowerBuf);
|
|
upperMatrix = reinterpret_cast<T *>(upperBuf);
|
|
matrix = reinterpret_cast<T *>(matrixBuf);
|
|
}
|
|
__syncthreads();
|
|
|
|
for (int k = blockIdx.x; k < n; k += gridDim.x) { // and then put all values under main diagonal on to it
|
|
for (int j = threadIdx.x; j < n; j += blockDim.x) {
|
|
Nd4jLong posX[] = {k, j};
|
|
Nd4jLong posD[] = {j, j};
|
|
auto xPos = shape::getOffset(lowerShape, posX);
|
|
auto yPos = shape::getOffset(upperShape, posX);
|
|
auto iPos = shape::getOffset(matrixShape, posX);
|
|
auto dPos = shape::getOffset(matrixShape, posD);
|
|
if (k >= j)
|
|
lowerMatrix[xPos] = matrix[iPos];//(k, j);
|
|
else
|
|
upperMatrix[yPos] = matrix[iPos]; //k, j);
|
|
}
|
|
}
|
|
}
|
|
|
|
template<typename T>
|
|
static int inverse_(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
|
|
auto n = input->sizeAt(-1);
|
|
auto n2 = n * n;
|
|
auto dtype = DataTypeUtils::fromT<T>(); //input->dataType();
|
|
// if (dtype != DataType::DOUBLE)
|
|
// dtype = DataType::FLOAT32;
|
|
NDArray matrix = NDArrayFactory::create('c', {n, n}, dtype, context);
|
|
NDArray upper = NDArrayFactory::create('c', {n, n}, dtype, context);
|
|
NDArray lower = NDArrayFactory::create('c', {n, n}, dtype, context);
|
|
NDArray compound = NDArrayFactory::create('c', {n, n}, dtype, context);
|
|
NDArray permutation = NDArrayFactory::create('c', {n, n}, dtype, context);
|
|
auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(),
|
|
{input->rankOf() - 2,
|
|
input->rankOf() - 1});
|
|
auto packZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(output->getShapeInfo(),
|
|
{output->rankOf() - 2,
|
|
output->rankOf() - 1});
|
|
auto stream = context->getCudaStream();
|
|
|
|
for (auto i = 0LL; i < packX.numberOfTads(); i++) {
|
|
fillMatrix<T, T><<<1, n2, 1024, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), i * n2, n);
|
|
matrix.tickWriteDevice();
|
|
//compound.assign(matrix);
|
|
// if (matrix.dataType() == input->dataType())
|
|
lup_<T>(context, &matrix, nullptr, nullptr);
|
|
fillLowerUpperKernel<T><<<n, n, 1024, *stream>>>(lower.specialBuffer(), lower.specialShapeInfo(), upper.specialBuffer(), upper.specialShapeInfo(), matrix.specialBuffer(), matrix.specialShapeInfo(), n);
|
|
lower.tickWriteDevice();
|
|
upper.tickWriteDevice();
|
|
// lower.printIndexedBuffer("LOWER");
|
|
// upper.printIndexedBuffer("UPPER");
|
|
matrix.assign(0);
|
|
invertUpperMatrix(context, &upper, &matrix); // U^{-1}
|
|
matrix.tickWriteDevice();
|
|
// matrix.printIndexedBuffer("Upper Inverted");
|
|
compound.assign(0);
|
|
invertLowerMatrix(context, &lower, &compound); // L{-1}
|
|
compound.tickWriteDevice();
|
|
// compound.printIndexedBuffer("Lower Inverted");
|
|
// matrix.tickWriteDevice();
|
|
// compound.tickWriteDevice();
|
|
nd4j::MmulHelper::mmul(&matrix, &compound, &upper, 1.0, 0.0);
|
|
upper.tickWriteDevice();
|
|
// upper.printIndexedBuffer("Full inverted");
|
|
returnMatrix<T> <<<1, n2, 1024, *stream>>>(output->specialBuffer(), output->specialShapeInfo(), upper.specialBuffer(), upper.specialShapeInfo(), i * n2, n);
|
|
}
|
|
return Status::OK();
|
|
}
|
|
|
|
int inverse(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), return inverse_, (context, input, output), FLOAT_NATIVE);
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
}
|
|
|
|
bool checkCholeskyInput(nd4j::LaunchContext *context, NDArray const *input) {
|
|
return true;
|
|
}
|
|
|
|
template<typename F>
|
|
__global__ void fillBatchKernel(F **dArrayBatch, F *buf, Nd4jLong *offsets, Nd4jLong batchSize) {
|
|
auto start = blockIdx.x * blockDim.x + threadIdx.x;
|
|
auto step = blockDim.x * gridDim.x;
|
|
|
|
for (auto i = start; i < batchSize; i += step) {
|
|
dArrayBatch[i] = buf + offsets[i];
|
|
}
|
|
}
|
|
|
|
template<typename F>
|
|
__global__ void
|
|
adjustResultsKernel(F *dArray, Nd4jLong *shape, Nd4jLong *offsets, Nd4jLong batchSize, Nd4jLong n) {
|
|
//auto i = blockIdx.x * blockDim.x + threadIdx.x;
|
|
Nd4jLong *shapeOf = shape::shapeOf(shape);
|
|
Nd4jLong *strideOf = shape::stride(shape);
|
|
|
|
for (auto i = blockIdx.x; i < batchSize; i += gridDim.x) {
|
|
auto current = dArray + offsets[i];
|
|
for (auto r = threadIdx.x; r < n; r += blockDim.x) {
|
|
for (auto c = r + 1; c < n; c++) {
|
|
Nd4jLong posRC[] = {r, c};
|
|
auto pos = r * n + c; //shape::getOffset(0, shapeOf, strideOf, posRC, 2);
|
|
current[pos] = 0.;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
template<typename F>
|
|
int cholesky__(LaunchContext *context, NDArray *input, NDArray *output, bool inplace) {
|
|
if (!inplace)
|
|
output->assign(input);
|
|
std::unique_ptr<NDArray> tempOutput(output->dup());
|
|
cusolverDnHandle_t handle = nullptr;
|
|
auto n = input->sizeAt(-1);
|
|
auto n2 = n * n;
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
auto status = cusolverDnCreate(&handle);
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("helpers::cholesky_: Cannot create solver handle", status);
|
|
}
|
|
F **dArrayBatch = nullptr;
|
|
auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(tempOutput->getShapeInfo(),
|
|
{tempOutput->rankOf() - 2,
|
|
tempOutput->rankOf() - 1});
|
|
const Nd4jLong batchSize = packX.numberOfTads();
|
|
int *dInfoArray = nullptr;
|
|
auto err = cudaMalloc((void **) &dArrayBatch, sizeof(F *) * batchSize);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::cholesky_: Cannot allocate memory for solver batch data buffer",
|
|
err);
|
|
}
|
|
err = cudaMalloc((void **) &dInfoArray, sizeof(int) * batchSize);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::cholesky_: Cannot allocate memory for solver errors buffer", err);
|
|
}
|
|
auto stream = context->getCudaStream();
|
|
fillBatchKernel<F> << < 1, batchSize, 128, *stream >> >
|
|
(dArrayBatch, reinterpret_cast<F *>(tempOutput->specialBuffer()), packX.specialOffsets(), batchSize);
|
|
|
|
status = cusolverDnSetStream(handle, *stream);
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("helpers::cholesky_: Cannot set stream to solver handle", status);
|
|
}
|
|
const cublasFillMode_t uplo = CUBLAS_FILL_MODE_UPPER;
|
|
if (input->dataType() == DataType::DOUBLE)
|
|
status = cusolverDnDpotrfBatched(
|
|
handle,
|
|
uplo,
|
|
n,
|
|
(double **) dArrayBatch,
|
|
n,
|
|
dInfoArray,
|
|
batchSize);
|
|
else
|
|
status = cusolverDnSpotrfBatched(
|
|
handle,
|
|
uplo,
|
|
n,
|
|
(float **) dArrayBatch,
|
|
n,
|
|
dInfoArray,
|
|
batchSize);
|
|
|
|
if (CUSOLVER_STATUS_SUCCESS != status) {
|
|
throw cuda_exception::build("helpers::cholesky_: Cholesky factorization failed for batch", status);
|
|
}
|
|
adjustResultsKernel<F> << < batchSize, n2, 128, *stream >> >
|
|
(reinterpret_cast<F *>(tempOutput->specialBuffer()), packX.specialShapeInfo(), packX.specialOffsets(), batchSize, n);
|
|
|
|
err = cudaFree(dArrayBatch);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::cholesky_: Cannot deallocate memory for solver batch data buffer",
|
|
err);
|
|
}
|
|
err = cudaFree(dInfoArray);
|
|
if (err) {
|
|
throw cuda_exception::build("helpers::cholesky_: Cannot allocate memory for solver errors buffer", err);
|
|
}
|
|
|
|
if (!inplace)
|
|
output->assign(tempOutput.get());
|
|
else
|
|
input->assign(tempOutput.get());
|
|
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
return Status::OK();
|
|
}
|
|
|
|
// template <typename T>
|
|
int cholesky_(LaunchContext *context, NDArray *input, NDArray *output, bool inplace) {
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
if (input->dataType() == DataType::DOUBLE)
|
|
cholesky__<double>(context, input, output, inplace);
|
|
else if (input->dataType() == DataType::FLOAT32)
|
|
cholesky__<float>(context, input, output, inplace);
|
|
else {
|
|
std::unique_ptr<NDArray> tempOutput(
|
|
NDArrayFactory::create_('c', input->getShapeAsVector(), DataType::FLOAT32, context));
|
|
tempOutput->assign(input);
|
|
cholesky__<float>(context, tempOutput.get(), tempOutput.get(), true);
|
|
output->assign(tempOutput.get());
|
|
}
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
return Status::OK();
|
|
}
|
|
|
|
int cholesky(nd4j::LaunchContext *context, NDArray *input, NDArray *output, bool inplace) {
|
|
// BUILD_SINGLE_SELECTOR(input->dataType(), return cholesky_, (context, input, output, inplace), FLOAT_TYPES);
|
|
return cholesky_(context, input, output, inplace);
|
|
}
|
|
// BUILD_SINGLE_TEMPLATE(template int cholesky_, (LaunchContext* context, NDArray* input, NDArray* output, bool inplace), FLOAT_TYPES);
|
|
BUILD_SINGLE_TEMPLATE(template int inverse_, (nd4j::LaunchContext * context, NDArray * input, NDArray * output),
|
|
FLOAT_NATIVE);
|
|
|
|
template<typename T>
|
|
__global__ void
|
|
logDetKernel(T *inputBuf, Nd4jLong *inputShape, Nd4jLong batchNum, Nd4jLong *tadShape, Nd4jLong *tadOffsets,
|
|
T *outputBuf, Nd4jLong *outputShape) {
|
|
|
|
__shared__ int n;
|
|
if (threadIdx.x == 0) {
|
|
n = shape::sizeAt(inputShape, -1); // * shape::sizeAt(inputShape, -1);
|
|
}
|
|
__syncthreads();
|
|
|
|
T *output = outputBuf;
|
|
T *input = inputBuf;
|
|
|
|
for (auto i = blockIdx.x; i < batchNum; i += gridDim.x) {
|
|
T *current = input + tadOffsets[i];
|
|
|
|
auto zIndex = shape::getIndexOffset(i, outputShape);
|
|
for (auto e = threadIdx.x; e < n; e += blockDim.x) {
|
|
Nd4jLong diag[] = {e, e};
|
|
auto xIndex = shape::getOffset(tadShape, diag);
|
|
math::atomics::nd4j_atomicAdd(&output[zIndex],math::nd4j_log<T, T>(current[xIndex] * current[xIndex]));
|
|
}
|
|
}
|
|
}
|
|
|
|
template<typename T>
|
|
int logdetFunctor_(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
|
|
NDArray::prepareSpecialUse({output}, {input});
|
|
auto n2 = input->sizeAt(-1) * input->sizeAt(-2);
|
|
auto stream = context->getCudaStream();
|
|
NDArray tempOutput(*input);
|
|
|
|
cholesky(context, input, &tempOutput, false);
|
|
|
|
auto outputBuf = output->dataBuffer()->specialAsT<T>(); //reinterpret_cast<T*>(output->specialBuffer()); // + e * n2; // + e * n2;
|
|
auto inputBuf = tempOutput.dataBuffer()->specialAsT<T>(); //reinterpret_cast<T*>(tempOutput->specialBuffer());
|
|
output->nullify();
|
|
auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(tempOutput.getShapeInfo(),
|
|
{tempOutput.rankOf() - 2,
|
|
tempOutput.rankOf() - 1});
|
|
logDetKernel<T> <<< 128, 512, 256, *stream >>>(inputBuf, tempOutput.specialShapeInfo(),
|
|
packX.numberOfTads(), packX.specialShapeInfo(),
|
|
packX.specialOffsets(), outputBuf, output->specialShapeInfo());
|
|
output->tickWriteDevice();
|
|
NDArray::registerSpecialUse({output}, {input});
|
|
return Status::OK();
|
|
}
|
|
|
|
int logdetFunctor(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
|
|
BUILD_SINGLE_SELECTOR(output->dataType(), return logdetFunctor_, (context, input, output), FLOAT_NATIVE);
|
|
}
|
|
|
|
// BUILD_SINGLE_TEMPLATE(template int logdetFunctor_,
|
|
// (nd4j::LaunchContext * context, NDArray * input, NDArray * output), FLOAT_NATIVE);
|
|
}
|
|
}
|
|
}
|