NASP/cuda-samples
85
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2026-10-26 06:38:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

CUDA 13.2 samples update (#432)

Browse Source 
- Added Python samples for CUDA Python 1.0 release
- Renamed top-level `Samples` directory to `cpp` to accommodate Python samples.
This commit is contained in:
Dheemanth
2026-05-13 15:13:18 -07:00
committed by GitHub
parent a4526d5229
commit aeab82ff30
2072 changed files with 15077 additions and 1970 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/.vscode/c_cpp_properties.json vendored Normal file
View File

@@ -0,0 +1,18 @@
{
"configurations": [
{
"name": "Linux",
"includePath": [
"${workspaceFolder}/**",
"${workspaceFolder}/../../../Common"
],
"defines": [],
"compilerPath": "/usr/local/cuda/bin/nvcc",
"cStandard": "gnu17",
"cppStandard": "gnu++14",
"intelliSenseMode": "linux-gcc-x64",
"configurationProvider": "ms-vscode.makefile-tools"
}
],
"version": 4
}

7
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/.vscode/extensions.json vendored Normal file
View File

@@ -0,0 +1,7 @@
{
"recommendations": [
"nvidia.nsight-vscode-edition",
"ms-vscode.cpptools",
"ms-vscode.makefile-tools"
]
}

61
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/CMakeLists.txt Normal file
View File

@@ -0,0 +1,61 @@
cmake_minimum_required(VERSION 3.20)
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
project(cuSolverSp_LowlevelQR LANGUAGES C CXX)
find_package(CUDAToolkit REQUIRED)
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
set(CMAKE_CUDA_ARCHITECTURES 75 80 86 87 89 90 100 110 120)
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
if(ENABLE_CUDA_DEBUG)
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (may significantly affect performance on some targets)
else()
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -lineinfo") # add line information to all builds for debug tools (exclusive to -G option)
endif()
# Include directories and libraries
include_directories(../../../Common)
# Source file
# Add target for cuSolverSp_LowlevelQR
add_executable(cuSolverSp_LowlevelQR cuSolverSp_LowlevelQR.cpp mmio.c mmio_wrapper.cpp)
target_compile_options(cuSolverSp_LowlevelQR PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
target_compile_features(cuSolverSp_LowlevelQR PRIVATE cxx_std_17 cuda_std_17)
set_target_properties(cuSolverSp_LowlevelQR PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
target_include_directories(cuSolverSp_LowlevelQR PRIVATE
${CUDAToolkit_INCLUDE_DIRS}
)
target_link_libraries(cuSolverSp_LowlevelQR PRIVATE
CUDA::cudart
CUDA::cublas
CUDA::cusolver
)
# Copy data files to output directory
add_custom_command(TARGET cuSolverSp_LowlevelQR POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
${CMAKE_CURRENT_SOURCE_DIR}/lap2D_5pt_n32.mtx
${CMAKE_CURRENT_BINARY_DIR}
)
add_custom_command(TARGET cuSolverSp_LowlevelQR POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
${CMAKE_CURRENT_SOURCE_DIR}/lap2D_5pt_n100.mtx
${CMAKE_CURRENT_BINARY_DIR}
)
add_custom_command(TARGET cuSolverSp_LowlevelQR POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
${CMAKE_CURRENT_SOURCE_DIR}/lap3D_7pt_n20.mtx
${CMAKE_CURRENT_BINARY_DIR}
)
# Include installation configuration
include(${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/InstallSamples.cmake)
setup_samples_install()

39
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/README.md Normal file
View File

@@ -0,0 +1,39 @@
# cuSolverSp_LowlevelQR - cuSolverSp Lowlevel QR Solver
## Description
A CUDA Sample that demonstrates QR factorization using cuSolverSP's low level APIs.
## Key Concepts
Linear Algebra, CUSOLVER Library
## Supported SM Architectures
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows
## Supported CPU Architecture
x86_64, armv7l
## CUDA APIs involved
### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
cuGet, cuDoubleComplex, cuComplex
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
cudaMemcpy, cudaStreamDestroy, cudaFree, cudaMalloc, cudaStreamCreate
## Dependencies needed to build/run
[CUSOLVER](../../../README.md#cusolver), [CUSPARSE](../../../README.md#cusparse)
## Prerequisites
Download and install the [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
## References (for more details)

490
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/cuSolverSp_LowlevelQR.cpp Normal file
View File

@@ -0,0 +1,490 @@
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <assert.h>
#include <ctype.h>
#include <cuda_runtime.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include "cusolverSp.h"
#include "cusolverSp_LOWLEVEL_PREVIEW.h"
#include "helper_cuda.h"
#include "helper_cusolver.h"
template <typename T_ELEM>
int loadMMSparseMatrix(char *filename,
char elem_type,
bool csrFormat,
int *m,
int *n,
int *nnz,
T_ELEM **aVal,
int **aRowInd,
int **aColInd,
int extendSymMatrix);
void UsageSP(void)
{
printf("<options>\n");
printf("-h : display this help\n");
printf("-file=<filename> : filename containing a matrix in MM format\n");
printf("-device=<device_id> : <device_id> if want to run on specific GPU\n");
exit(0);
}
void parseCommandLineArguments(int argc, char *argv[], struct testOpts &opts)
{
memset(&opts, 0, sizeof(opts));
if (checkCmdLineFlag(argc, (const char **)argv, "-h")) {
UsageSP();
}
if (checkCmdLineFlag(argc, (const char **)argv, "file")) {
char *fileName = 0;
getCmdLineArgumentString(argc, (const char **)argv, "file", &fileName);
if (fileName) {
opts.sparse_mat_filename = fileName;
}
else {
printf("\nIncorrect filename passed to -file \n ");
UsageSP();
}
}
}
int main(int argc, char *argv[])
{
struct testOpts opts;
cusolverSpHandle_t cusolverSpH = NULL; // reordering, permutation and 1st LU factorization
cusparseHandle_t cusparseH = NULL; // residual evaluation
cudaStream_t stream = NULL;
cusparseMatDescr_t descrA = NULL; // A is a base-0 general matrix
csrqrInfoHost_t h_info = NULL; // opaque info structure for LU with parital pivoting
csrqrInfo_t d_info = NULL; // opaque info structure for LU with parital pivoting
int rowsA = 0; // number of rows of A
int colsA = 0; // number of columns of A
int nnzA = 0; // number of nonzeros of A
int baseA = 0; // base index in CSR format
// CSR(A) from I/O
int *h_csrRowPtrA = NULL; // <int> n+1
int *h_csrColIndA = NULL; // <int> nnzA
double *h_csrValA = NULL; // <double> nnzA
double *h_x = NULL; // <double> n, x = A \ b
double *h_b = NULL; // <double> n, b = ones(m,1)
double *h_bcopy = NULL; // <double> n, b = ones(m,1)
double *h_r = NULL; // <double> n, r = b - A*x
size_t size_internal = 0;
size_t size_chol = 0; // size of working space for csrlu
void *buffer_cpu = NULL; // working space for Cholesky
void *buffer_gpu = NULL; // working space for Cholesky
int *d_csrRowPtrA = NULL; // <int> n+1
int *d_csrColIndA = NULL; // <int> nnzA
double *d_csrValA = NULL; // <double> nnzA
double *d_x = NULL; // <double> n, x = A \ b
double *d_b = NULL; // <double> n, a copy of h_b
double *d_r = NULL; // <double> n, r = b - A*x
// the constants used in residual evaluation, r = b - A*x
const double minus_one = -1.0;
const double one = 1.0;
const double zero = 0.0;
// the constant used in cusolverSp
// singularity is -1 if A is invertible under tol
// tol determines the condition of singularity
int singularity = 0;
const double tol = 1.e-14;
double x_inf = 0.0; // |x|
double r_inf = 0.0; // |r|
double A_inf = 0.0; // |A|
parseCommandLineArguments(argc, argv, opts);
findCudaDevice(argc, (const char **)argv);
if (opts.sparse_mat_filename == NULL) {
opts.sparse_mat_filename = sdkFindFilePath("lap2D_5pt_n32.mtx", argv[0]);
if (opts.sparse_mat_filename != NULL)
printf("Using default input file [%s]\n", opts.sparse_mat_filename);
else
printf("Could not find lap2D_5pt_n32.mtx\n");
}
else {
printf("Using input file [%s]\n", opts.sparse_mat_filename);
}
printf("step 1: read matrix market format\n");
if (opts.sparse_mat_filename) {
if (loadMMSparseMatrix<double>(opts.sparse_mat_filename,
'd',
true,
&rowsA,
&colsA,
&nnzA,
&h_csrValA,
&h_csrRowPtrA,
&h_csrColIndA,
true)) {
return 1;
}
baseA = h_csrRowPtrA[0]; // baseA = {0,1}
}
else {
fprintf(stderr, "Error: input matrix is not provided\n");
return 1;
}
if (rowsA != colsA) {
fprintf(stderr, "Error: only support square matrix\n");
return 1;
}
printf("sparse matrix A is %d x %d with %d nonzeros, base=%d\n", rowsA, colsA, nnzA, baseA);
checkCudaErrors(cusolverSpCreate(&cusolverSpH));
checkCudaErrors(cusparseCreate(&cusparseH));
checkCudaErrors(cudaStreamCreate(&stream));
checkCudaErrors(cusolverSpSetStream(cusolverSpH, stream));
checkCudaErrors(cusparseSetStream(cusparseH, stream));
checkCudaErrors(cusparseCreateMatDescr(&descrA));
checkCudaErrors(cusparseSetMatType(descrA, CUSPARSE_MATRIX_TYPE_GENERAL));
if (baseA) {
checkCudaErrors(cusparseSetMatIndexBase(descrA, CUSPARSE_INDEX_BASE_ONE));
}
else {
checkCudaErrors(cusparseSetMatIndexBase(descrA, CUSPARSE_INDEX_BASE_ZERO));
}
h_x = (double *)malloc(sizeof(double) * colsA);
h_b = (double *)malloc(sizeof(double) * rowsA);
h_bcopy = (double *)malloc(sizeof(double) * rowsA);
h_r = (double *)malloc(sizeof(double) * rowsA);
assert(NULL != h_x);
assert(NULL != h_b);
assert(NULL != h_bcopy);
assert(NULL != h_r);
checkCudaErrors(cudaMalloc((void **)&d_csrRowPtrA, sizeof(int) * (rowsA + 1)));
checkCudaErrors(cudaMalloc((void **)&d_csrColIndA, sizeof(int) * nnzA));
checkCudaErrors(cudaMalloc((void **)&d_csrValA, sizeof(double) * nnzA));
checkCudaErrors(cudaMalloc((void **)&d_x, sizeof(double) * colsA));
checkCudaErrors(cudaMalloc((void **)&d_b, sizeof(double) * rowsA));
checkCudaErrors(cudaMalloc((void **)&d_r, sizeof(double) * rowsA));
for (int row = 0; row < rowsA; row++) {
h_b[row] = 1.0;
}
memcpy(h_bcopy, h_b, sizeof(double) * rowsA);
printf("step 2: create opaque info structure\n");
checkCudaErrors(cusolverSpCreateCsrqrInfoHost(&h_info));
printf("step 3: analyze qr(A) to know structure of L\n");
checkCudaErrors(
cusolverSpXcsrqrAnalysisHost(cusolverSpH, rowsA, colsA, nnzA, descrA, h_csrRowPtrA, h_csrColIndA, h_info));
printf("step 4: workspace for qr(A)\n");
checkCudaErrors(cusolverSpDcsrqrBufferInfoHost(cusolverSpH,
rowsA,
colsA,
nnzA,
descrA,
h_csrValA,
h_csrRowPtrA,
h_csrColIndA,
h_info,
&size_internal,
&size_chol));
if (buffer_cpu) {
free(buffer_cpu);
}
buffer_cpu = (void *)malloc(sizeof(char) * size_chol);
assert(NULL != buffer_cpu);
printf("step 5: compute A = L*L^T \n");
checkCudaErrors(cusolverSpDcsrqrSetupHost(
cusolverSpH, rowsA, colsA, nnzA, descrA, h_csrValA, h_csrRowPtrA, h_csrColIndA, zero, h_info));
checkCudaErrors(cusolverSpDcsrqrFactorHost(cusolverSpH, rowsA, colsA, nnzA, NULL, NULL, h_info, buffer_cpu));
printf("step 6: check if the matrix is singular \n");
checkCudaErrors(cusolverSpDcsrqrZeroPivotHost(cusolverSpH, h_info, tol, &singularity));
if (0 <= singularity) {
fprintf(stderr, "Error: A is not invertible, singularity=%d\n", singularity);
return 1;
}
printf("step 7: solve A*x = b \n");
checkCudaErrors(cusolverSpDcsrqrSolveHost(cusolverSpH, rowsA, colsA, h_b, h_x, h_info, buffer_cpu));
printf("step 8: evaluate residual r = b - A*x (result on CPU)\n");
// use GPU gemv to compute r = b - A*x
checkCudaErrors(cudaMemcpy(d_csrRowPtrA, h_csrRowPtrA, sizeof(int) * (rowsA + 1), cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_csrColIndA, h_csrColIndA, sizeof(int) * nnzA, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_csrValA, h_csrValA, sizeof(double) * nnzA, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_r, h_bcopy, sizeof(double) * rowsA, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_x, h_x, sizeof(double) * colsA, cudaMemcpyHostToDevice));
/* Wrap raw data into cuSPARSE generic API objects */
cusparseSpMatDescr_t matA = NULL;
if (baseA) {
checkCudaErrors(cusparseCreateCsr(&matA,
rowsA,
colsA,
nnzA,
d_csrRowPtrA,
d_csrColIndA,
d_csrValA,
CUSPARSE_INDEX_32I,
CUSPARSE_INDEX_32I,
CUSPARSE_INDEX_BASE_ONE,
CUDA_R_64F));
}
else {
checkCudaErrors(cusparseCreateCsr(&matA,
rowsA,
colsA,
nnzA,
d_csrRowPtrA,
d_csrColIndA,
d_csrValA,
CUSPARSE_INDEX_32I,
CUSPARSE_INDEX_32I,
CUSPARSE_INDEX_BASE_ZERO,
CUDA_R_64F));
}
cusparseDnVecDescr_t vecx = NULL;
checkCudaErrors(cusparseCreateDnVec(&vecx, colsA, d_x, CUDA_R_64F));
cusparseDnVecDescr_t vecAx = NULL;
checkCudaErrors(cusparseCreateDnVec(&vecAx, rowsA, d_r, CUDA_R_64F));
/* Allocate workspace for cuSPARSE */
size_t bufferSize = 0;
checkCudaErrors(cusparseSpMV_bufferSize(cusparseH,
CUSPARSE_OPERATION_NON_TRANSPOSE,
&minus_one,
matA,
vecx,
&one,
vecAx,
CUDA_R_64F,
CUSPARSE_SPMV_ALG_DEFAULT,
&bufferSize));
void *buffer = NULL;
checkCudaErrors(cudaMalloc(&buffer, bufferSize));
checkCudaErrors(cusparseSpMV(cusparseH,
CUSPARSE_OPERATION_NON_TRANSPOSE,
&minus_one,
matA,
vecx,
&one,
vecAx,
CUDA_R_64F,
CUSPARSE_SPMV_ALG_DEFAULT,
buffer));
checkCudaErrors(cudaMemcpy(h_r, d_r, sizeof(double) * rowsA, cudaMemcpyDeviceToHost));
x_inf = vec_norminf(colsA, h_x);
r_inf = vec_norminf(rowsA, h_r);
A_inf = csr_mat_norminf(rowsA, colsA, nnzA, descrA, h_csrValA, h_csrRowPtrA, h_csrColIndA);
printf("(CPU) |b - A*x| = %E \n", r_inf);
printf("(CPU) |A| = %E \n", A_inf);
printf("(CPU) |x| = %E \n", x_inf);
printf("(CPU) |b - A*x|/(|A|*|x|) = %E \n", r_inf / (A_inf * x_inf));
printf("step 9: create opaque info structure\n");
checkCudaErrors(cusolverSpCreateCsrqrInfo(&d_info));
checkCudaErrors(cudaMemcpy(d_csrRowPtrA, h_csrRowPtrA, sizeof(int) * (rowsA + 1), cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_csrColIndA, h_csrColIndA, sizeof(int) * nnzA, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_csrValA, h_csrValA, sizeof(double) * nnzA, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_b, h_bcopy, sizeof(double) * rowsA, cudaMemcpyHostToDevice));
printf("step 10: analyze qr(A) to know structure of L\n");
checkCudaErrors(
cusolverSpXcsrqrAnalysis(cusolverSpH, rowsA, colsA, nnzA, descrA, d_csrRowPtrA, d_csrColIndA, d_info));
printf("step 11: workspace for qr(A)\n");
checkCudaErrors(cusolverSpDcsrqrBufferInfo(cusolverSpH,
rowsA,
colsA,
nnzA,
descrA,
d_csrValA,
d_csrRowPtrA,
d_csrColIndA,
d_info,
&size_internal,
&size_chol));
printf("GPU buffer size = %lld bytes\n", (signed long long)size_chol);
if (buffer_gpu) {
checkCudaErrors(cudaFree(buffer_gpu));
}
checkCudaErrors(cudaMalloc(&buffer_gpu, sizeof(char) * size_chol));
printf("step 12: compute A = L*L^T \n");
checkCudaErrors(cusolverSpDcsrqrSetup(
cusolverSpH, rowsA, colsA, nnzA, descrA, d_csrValA, d_csrRowPtrA, d_csrColIndA, zero, d_info));
checkCudaErrors(cusolverSpDcsrqrFactor(cusolverSpH, rowsA, colsA, nnzA, NULL, NULL, d_info, buffer_gpu));
printf("step 13: check if the matrix is singular \n");
checkCudaErrors(cusolverSpDcsrqrZeroPivot(cusolverSpH, d_info, tol, &singularity));
if (0 <= singularity) {
fprintf(stderr, "Error: A is not invertible, singularity=%d\n", singularity);
return 1;
}
printf("step 14: solve A*x = b \n");
checkCudaErrors(cusolverSpDcsrqrSolve(cusolverSpH, rowsA, colsA, d_b, d_x, d_info, buffer_gpu));
checkCudaErrors(cudaMemcpy(d_r, h_bcopy, sizeof(double) * rowsA, cudaMemcpyHostToDevice));
checkCudaErrors(cusparseSpMV(cusparseH,
CUSPARSE_OPERATION_NON_TRANSPOSE,
&minus_one,
matA,
vecx,
&one,
vecAx,
CUDA_R_64F,
CUSPARSE_SPMV_ALG_DEFAULT,
buffer));
checkCudaErrors(cudaMemcpy(h_r, d_r, sizeof(double) * rowsA, cudaMemcpyDeviceToHost));
r_inf = vec_norminf(rowsA, h_r);
printf("(GPU) |b - A*x| = %E \n", r_inf);
printf("(GPU) |b - A*x|/(|A|*|x|) = %E \n", r_inf / (A_inf * x_inf));
if (cusolverSpH) {
checkCudaErrors(cusolverSpDestroy(cusolverSpH));
}
if (cusparseH) {
checkCudaErrors(cusparseDestroy(cusparseH));
}
if (stream) {
checkCudaErrors(cudaStreamDestroy(stream));
}
if (descrA) {
checkCudaErrors(cusparseDestroyMatDescr(descrA));
}
if (h_info) {
checkCudaErrors(cusolverSpDestroyCsrqrInfoHost(h_info));
}
if (d_info) {
checkCudaErrors(cusolverSpDestroyCsrqrInfo(d_info));
}
if (matA) {
checkCudaErrors(cusparseDestroySpMat(matA));
}
if (vecx) {
checkCudaErrors(cusparseDestroyDnVec(vecx));
}
if (vecAx) {
checkCudaErrors(cusparseDestroyDnVec(vecAx));
}
if (h_csrValA) {
free(h_csrValA);
}
if (h_csrRowPtrA) {
free(h_csrRowPtrA);
}
if (h_csrColIndA) {
free(h_csrColIndA);
}
if (h_x) {
free(h_x);
}
if (h_b) {
free(h_b);
}
if (h_bcopy) {
free(h_bcopy);
}
if (h_r) {
free(h_r);
}
if (buffer_cpu) {
free(buffer_cpu);
}
if (buffer_gpu) {
checkCudaErrors(cudaFree(buffer_gpu));
}
if (d_csrValA) {
checkCudaErrors(cudaFree(d_csrValA));
}
if (d_csrRowPtrA) {
checkCudaErrors(cudaFree(d_csrRowPtrA));
}
if (d_csrColIndA) {
checkCudaErrors(cudaFree(d_csrColIndA));
}
if (d_x) {
checkCudaErrors(cudaFree(d_x));
}
if (d_b) {
checkCudaErrors(cudaFree(d_b));
}
if (d_r) {
checkCudaErrors(cudaFree(d_r));
}
return 0;
}

29803
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/lap2D_5pt_n100.mtx Normal file
View File

File diff suppressed because it is too large Load Diff

3012
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/lap2D_5pt_n32.mtx Normal file
View File

File diff suppressed because it is too large Load Diff

30803
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/lap3D_7pt_n20.mtx Normal file
View File

File diff suppressed because it is too large Load Diff

477
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/mmio.c Normal file
View File

@@ -0,0 +1,477 @@
/*
* Matrix Market I/O library for ANSI C
*
* See http://math.nist.gov/MatrixMarket for details.
*
*
*/
/* avoid Windows warnings (for example: strcpy, fscanf, etc.) */
#if defined(_WIN32)
#define _CRT_SECURE_NO_WARNINGS
#endif
#include "mmio.h"
#include <ctype.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int mm_read_unsymmetric_sparse(const char *fname, int *M_, int *N_, int *nz_, double **val_, int **I_, int **J_)
{
FILE *f;
MM_typecode matcode;
int M, N, nz;
int i;
double *val;
int *I, *J;
if ((f = fopen(fname, "r")) == NULL)
return -1;
if (mm_read_banner(f, &matcode) != 0) {
printf("mm_read_unsymetric: Could not process Matrix Market banner ");
printf(" in file [%s]\n", fname);
return -1;
}
if (!(mm_is_real(matcode) && mm_is_matrix(matcode) && mm_is_sparse(matcode))) {
fprintf(stderr, "Sorry, this application does not support ");
fprintf(stderr, "Market Market type: [%s]\n", mm_typecode_to_str(matcode));
return -1;
}
/* find out size of sparse matrix: M, N, nz .... */
if (mm_read_mtx_crd_size(f, &M, &N, &nz) != 0) {
fprintf(stderr, "read_unsymmetric_sparse(): could not parse matrix size.\n");
return -1;
}
*M_ = M;
*N_ = N;
*nz_ = nz;
/* reserve memory for matrices */
I = (int *)malloc(nz * sizeof(int));
J = (int *)malloc(nz * sizeof(int));
val = (double *)malloc(nz * sizeof(double));
*val_ = val;
*I_ = I;
*J_ = J;
/* NOTE: when reading in doubles, ANSI C requires the use of the "l" */
/* specifier as in "%lg", "%lf", "%le", otherwise errors will occur */
/* (ANSI C X3.159-1989, Sec. 4.9.6.2, p. 136 lines 13-15) */
for (i = 0; i < nz; i++) {
if (fscanf(f, "%d %d %lg\n", &I[i], &J[i], &val[i]) != 3) {
return -1;
}
I[i]--; /* adjust from 1-based to 0-based */
J[i]--;
}
fclose(f);
return 0;
}
int mm_is_valid(MM_typecode matcode)
{
if (!mm_is_matrix(matcode))
return 0;
if (mm_is_dense(matcode) && mm_is_pattern(matcode))
return 0;
if (mm_is_real(matcode) && mm_is_hermitian(matcode))
return 0;
if (mm_is_pattern(matcode) && (mm_is_hermitian(matcode) || mm_is_skew(matcode)))
return 0;
return 1;
}
int mm_read_banner(FILE *f, MM_typecode *matcode)
{
char line[MM_MAX_LINE_LENGTH];
char banner[MM_MAX_TOKEN_LENGTH];
char mtx[MM_MAX_TOKEN_LENGTH];
char crd[MM_MAX_TOKEN_LENGTH];
char data_type[MM_MAX_TOKEN_LENGTH];
char storage_scheme[MM_MAX_TOKEN_LENGTH];
char *p;
mm_clear_typecode(matcode);
if (fgets(line, MM_MAX_LINE_LENGTH, f) == NULL)
return MM_PREMATURE_EOF;
if (sscanf(line, "%s %s %s %s %s", banner, mtx, crd, data_type, storage_scheme) != 5)
return MM_PREMATURE_EOF;
for (p = mtx; *p != '\0'; *p = tolower(*p), p++)
; /* convert to lower case */
for (p = crd; *p != '\0'; *p = tolower(*p), p++)
;
for (p = data_type; *p != '\0'; *p = tolower(*p), p++)
;
for (p = storage_scheme; *p != '\0'; *p = tolower(*p), p++)
;
/* check for banner */
if (strncmp(banner, MatrixMarketBanner, strlen(MatrixMarketBanner)) != 0)
return MM_NO_HEADER;
/* first field should be "mtx" */
if (strcmp(mtx, MM_MTX_STR) != 0)
return MM_UNSUPPORTED_TYPE;
mm_set_matrix(matcode);
/* second field describes whether this is a sparse matrix (in coordinate
storage) or a dense array */
if (strcmp(crd, MM_SPARSE_STR) == 0)
mm_set_sparse(matcode);
else if (strcmp(crd, MM_DENSE_STR) == 0)
mm_set_dense(matcode);
else
return MM_UNSUPPORTED_TYPE;
/* third field */
if (strcmp(data_type, MM_REAL_STR) == 0)
mm_set_real(matcode);
else if (strcmp(data_type, MM_COMPLEX_STR) == 0)
mm_set_complex(matcode);
else if (strcmp(data_type, MM_PATTERN_STR) == 0)
mm_set_pattern(matcode);
else if (strcmp(data_type, MM_INT_STR) == 0)
mm_set_integer(matcode);
else
return MM_UNSUPPORTED_TYPE;
/* fourth field */
if (strcmp(storage_scheme, MM_GENERAL_STR) == 0)
mm_set_general(matcode);
else if (strcmp(storage_scheme, MM_SYMM_STR) == 0)
mm_set_symmetric(matcode);
else if (strcmp(storage_scheme, MM_HERM_STR) == 0)
mm_set_hermitian(matcode);
else if (strcmp(storage_scheme, MM_SKEW_STR) == 0)
mm_set_skew(matcode);
else
return MM_UNSUPPORTED_TYPE;
return 0;
}
int mm_write_mtx_crd_size(FILE *f, int M, int N, int nz)
{
if (fprintf(f, "%d %d %d\n", M, N, nz) != 3)
return MM_COULD_NOT_WRITE_FILE;
else
return 0;
}
int mm_read_mtx_crd_size(FILE *f, int *M, int *N, int *nz)
{
char line[MM_MAX_LINE_LENGTH];
int num_items_read;
/* set return null parameter values, in case we exit with errors */
*M = *N = *nz = 0;
/* now continue scanning until you reach the end-of-comments */
do {
if (fgets(line, MM_MAX_LINE_LENGTH, f) == NULL)
return MM_PREMATURE_EOF;
} while (line[0] == '%');
/* line[] is either blank or has M,N, nz */
if (sscanf(line, "%d %d %d", M, N, nz) == 3)
return 0;
else
do {
num_items_read = fscanf(f, "%d %d %d", M, N, nz);
if (num_items_read == EOF)
return MM_PREMATURE_EOF;
} while (num_items_read != 3);
return 0;
}
int mm_read_mtx_array_size(FILE *f, int *M, int *N)
{
char line[MM_MAX_LINE_LENGTH];
int num_items_read;
/* set return null parameter values, in case we exit with errors */
*M = *N = 0;
/* now continue scanning until you reach the end-of-comments */
do {
if (fgets(line, MM_MAX_LINE_LENGTH, f) == NULL)
return MM_PREMATURE_EOF;
} while (line[0] == '%');
/* line[] is either blank or has M,N, nz */
if (sscanf(line, "%d %d", M, N) == 2)
return 0;
else /* we have a blank line */
do {
num_items_read = fscanf(f, "%d %d", M, N);
if (num_items_read == EOF)
return MM_PREMATURE_EOF;
} while (num_items_read != 2);
return 0;
}
int mm_write_mtx_array_size(FILE *f, int M, int N)
{
if (fprintf(f, "%d %d\n", M, N) != 2)
return MM_COULD_NOT_WRITE_FILE;
else
return 0;
}
/*-------------------------------------------------------------------------*/
/******************************************************************/
/* use when I[], J[], and val[]J, and val[] are already allocated */
/******************************************************************/
int mm_read_mtx_crd_data(FILE *f, int M, int N, int nz, int I[], int J[], double val[], MM_typecode matcode)
{
int i;
if (mm_is_complex(matcode)) {
for (i = 0; i < nz; i++)
if (fscanf(f, "%d %d %lg %lg", &I[i], &J[i], &val[2 * i], &val[2 * i + 1]) != 4)
return MM_PREMATURE_EOF;
}
else if (mm_is_real(matcode) || mm_is_integer(matcode)) {
for (i = 0; i < nz; i++) {
if (fscanf(f, "%d %d %lg\n", &I[i], &J[i], &val[i]) != 3)
return MM_PREMATURE_EOF;
}
}
else if (mm_is_pattern(matcode)) {
for (i = 0; i < nz; i++)
if (fscanf(f, "%d %d", &I[i], &J[i]) != 2)
return MM_PREMATURE_EOF;
}
else
return MM_UNSUPPORTED_TYPE;
return 0;
}
int mm_read_mtx_crd_entry(FILE *f, int *I, int *J, double *real, double *imag, MM_typecode matcode)
{
if (mm_is_complex(matcode)) {
if (fscanf(f, "%d %d %lg %lg", I, J, real, imag) != 4)
return MM_PREMATURE_EOF;
}
else if (mm_is_real(matcode) || mm_is_integer(matcode)) {
if (fscanf(f, "%d %d %lg\n", I, J, real) != 3)
return MM_PREMATURE_EOF;
}
else if (mm_is_pattern(matcode)) {
if (fscanf(f, "%d %d", I, J) != 2)
return MM_PREMATURE_EOF;
}
else
return MM_UNSUPPORTED_TYPE;
return 0;
}
/************************************************************************
mm_read_mtx_crd() fills M, N, nz, array of values, and return
type code, e.g. 'MCRS'
if matrix is complex, values[] is of size 2*nz,
(nz pairs of real/imaginary values)
************************************************************************/
int mm_read_mtx_crd(char *fname, int *M, int *N, int *nz, int **I, int **J, double **val, MM_typecode *matcode)
{
int ret_code;
FILE *f;
if (strcmp(fname, "stdin") == 0)
f = stdin;
else if ((f = fopen(fname, "r")) == NULL)
return MM_COULD_NOT_READ_FILE;
if ((ret_code = mm_read_banner(f, matcode)) != 0)
return ret_code;
if (!(mm_is_valid(*matcode) && mm_is_sparse(*matcode) && mm_is_matrix(*matcode)))
return MM_UNSUPPORTED_TYPE;
if ((ret_code = mm_read_mtx_crd_size(f, M, N, nz)) != 0)
return ret_code;
*I = (int *)malloc(*nz * sizeof(int));
*J = (int *)malloc(*nz * sizeof(int));
*val = NULL;
if (mm_is_complex(*matcode)) {
*val = (double *)malloc(*nz * 2 * sizeof(double));
ret_code = mm_read_mtx_crd_data(f, *M, *N, *nz, *I, *J, *val, *matcode);
if (ret_code != 0)
return ret_code;
}
else if (mm_is_real(*matcode) || mm_is_integer(*matcode)) {
*val = (double *)malloc(*nz * sizeof(double));
ret_code = mm_read_mtx_crd_data(f, *M, *N, *nz, *I, *J, *val, *matcode);
if (ret_code != 0)
return ret_code;
}
else if (mm_is_pattern(*matcode)) {
ret_code = mm_read_mtx_crd_data(f, *M, *N, *nz, *I, *J, *val, *matcode);
if (ret_code != 0)
return ret_code;
}
if (f != stdin)
fclose(f);
return 0;
}
int mm_write_banner(FILE *f, MM_typecode matcode)
{
char *str = mm_typecode_to_str(matcode);
int ret_code;
ret_code = fprintf(f, "%s %s\n", MatrixMarketBanner, str);
free(str);
if (ret_code != 2)
return MM_COULD_NOT_WRITE_FILE;
else
return 0;
}
int mm_write_mtx_crd(char fname[], int M, int N, int nz, int I[], int J[], double val[], MM_typecode matcode)
{
FILE *f;
int i;
if (strcmp(fname, "stdout") == 0)
f = stdout;
else if ((f = fopen(fname, "w")) == NULL)
return MM_COULD_NOT_WRITE_FILE;
/* print banner followed by typecode */
fprintf(f, "%s ", MatrixMarketBanner);
fprintf(f, "%s\n", mm_typecode_to_str(matcode));
/* print matrix sizes and nonzeros */
fprintf(f, "%d %d %d\n", M, N, nz);
/* print values */
if (mm_is_pattern(matcode))
for (i = 0; i < nz; i++)
fprintf(f, "%d %d\n", I[i], J[i]);
else if (mm_is_integer(matcode))
for (i = 0; i < nz; i++)
fprintf(f, "%d %d %d\n", I[i], J[i], (int)val[i]);
else if (mm_is_real(matcode))
for (i = 0; i < nz; i++)
fprintf(f, "%d %d %20.16g\n", I[i], J[i], val[i]);
else if (mm_is_complex(matcode))
for (i = 0; i < nz; i++)
fprintf(f, "%d %d %20.16g %20.16g\n", I[i], J[i], val[2 * i], val[2 * i + 1]);
else {
if (f != stdout)
fclose(f);
return MM_UNSUPPORTED_TYPE;
}
if (f != stdout)
fclose(f);
return 0;
}
/**
* Create a new copy of a string s. mm_strdup() is a common routine, but
* not part of ANSI C, so it is included here. Used by mm_typecode_to_str().
*
*/
static char *mm_strdup(const char *s)
{
size_t len = strlen(s);
char *s2 = (char *)malloc((len + 1) * sizeof(char));
return strcpy(s2, s);
}
char *mm_typecode_to_str(MM_typecode matcode)
{
char buffer[MM_MAX_LINE_LENGTH];
char *types[4];
// char *mm_strdup(const char *);
// int error =0;
/* check for MTX type */
if (mm_is_matrix(matcode))
types[0] = MM_MTX_STR;
else
return NULL; // error=1;
/* check for CRD or ARR matrix */
if (mm_is_sparse(matcode))
types[1] = MM_SPARSE_STR;
else if (mm_is_dense(matcode))
types[1] = MM_DENSE_STR;
else
return NULL;
/* check for element data type */
if (mm_is_real(matcode))
types[2] = MM_REAL_STR;
else if (mm_is_complex(matcode))
types[2] = MM_COMPLEX_STR;
else if (mm_is_pattern(matcode))
types[2] = MM_PATTERN_STR;
else if (mm_is_integer(matcode))
types[2] = MM_INT_STR;
else
return NULL;
/* check for symmetry type */
if (mm_is_general(matcode))
types[3] = MM_GENERAL_STR;
else if (mm_is_symmetric(matcode))
types[3] = MM_SYMM_STR;
else if (mm_is_hermitian(matcode))
types[3] = MM_HERM_STR;
else if (mm_is_skew(matcode))
types[3] = MM_SKEW_STR;
else
return NULL;
sprintf(buffer, "%s %s %s %s", types[0], types[1], types[2], types[3]);
return mm_strdup(buffer);
}

138
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/mmio.h Normal file
View File

@@ -0,0 +1,138 @@
/*
* Matrix Market I/O library for ANSI C
*
* See http://math.nist.gov/MatrixMarket for details.
*
*
*/
#ifndef MM_IO_H
#define MM_IO_H
#include <stdio.h>
#if defined(__cplusplus)
extern "C"
{
#endif /* __cplusplus */
#define MM_MAX_LINE_LENGTH 1025
#define MatrixMarketBanner "%%MatrixMarket"
#define MM_MAX_TOKEN_LENGTH 64
typedef char MM_typecode[4];
char *mm_typecode_to_str(MM_typecode matcode);
int mm_read_banner(FILE *f, MM_typecode *matcode);
int mm_read_mtx_crd_size(FILE *f, int *M, int *N, int *nz);
int mm_read_mtx_array_size(FILE *f, int *M, int *N);
int mm_write_banner(FILE *f, MM_typecode matcode);
int mm_write_mtx_crd_size(FILE *f, int M, int N, int nz);
int mm_write_mtx_array_size(FILE *f, int M, int N);
/********************* MM_typecode query fucntions ***************************/
#define mm_is_matrix(typecode) ((typecode)[0] == 'M')
#define mm_is_sparse(typecode) ((typecode)[1] == 'C')
#define mm_is_coordinate(typecode) ((typecode)[1] == 'C')
#define mm_is_dense(typecode) ((typecode)[1] == 'A')
#define mm_is_array(typecode) ((typecode)[1] == 'A')
#define mm_is_complex(typecode) ((typecode)[2] == 'C')
#define mm_is_real(typecode) ((typecode)[2] == 'R')
#define mm_is_pattern(typecode) ((typecode)[2] == 'P')
#define mm_is_integer(typecode) ((typecode)[2] == 'I')
#define mm_is_symmetric(typecode) ((typecode)[3] == 'S')
#define mm_is_general(typecode) ((typecode)[3] == 'G')
#define mm_is_skew(typecode) ((typecode)[3] == 'K')
#define mm_is_hermitian(typecode) ((typecode)[3] == 'H')
int mm_is_valid(MM_typecode matcode); /* too complex for a macro */
/********************* MM_typecode modify fucntions ***************************/
#define mm_set_matrix(typecode) ((*typecode)[0] = 'M')
#define mm_set_coordinate(typecode) ((*typecode)[1] = 'C')
#define mm_set_array(typecode) ((*typecode)[1] = 'A')
#define mm_set_dense(typecode) mm_set_array(typecode)
#define mm_set_sparse(typecode) mm_set_coordinate(typecode)
#define mm_set_complex(typecode) ((*typecode)[2] = 'C')
#define mm_set_real(typecode) ((*typecode)[2] = 'R')
#define mm_set_pattern(typecode) ((*typecode)[2] = 'P')
#define mm_set_integer(typecode) ((*typecode)[2] = 'I')
#define mm_set_symmetric(typecode) ((*typecode)[3] = 'S')
#define mm_set_general(typecode) ((*typecode)[3] = 'G')
#define mm_set_skew(typecode) ((*typecode)[3] = 'K')
#define mm_set_hermitian(typecode) ((*typecode)[3] = 'H')
#define mm_clear_typecode(typecode) ((*typecode)[0] = (*typecode)[1] = (*typecode)[2] = ' ', (*typecode)[3] = 'G')
#define mm_initialize_typecode(typecode) mm_clear_typecode(typecode)
/********************* Matrix Market error codes ***************************/
#define MM_COULD_NOT_READ_FILE 11
#define MM_PREMATURE_EOF 12
#define MM_NOT_MTX 13
#define MM_NO_HEADER 14
#define MM_UNSUPPORTED_TYPE 15
#define MM_LINE_TOO_LONG 16
#define MM_COULD_NOT_WRITE_FILE 17
/******************** Matrix Market internal definitions ********************
MM_matrix_typecode: 4-character sequence
ojbect sparse/ data storage
dense type scheme
string position: [0] [1] [2] [3]
Matrix typecode: M(atrix) C(oord) R(eal) G(eneral)
A(array) C(omplex) H(ermitian)
P(attern) S(ymmetric)
I(nteger) K(kew)
***********************************************************************/
#define MM_MTX_STR "matrix"
#define MM_ARRAY_STR "array"
#define MM_DENSE_STR "array"
#define MM_COORDINATE_STR "coordinate"
#define MM_SPARSE_STR "coordinate"
#define MM_COMPLEX_STR "complex"
#define MM_REAL_STR "real"
#define MM_INT_STR "integer"
#define MM_GENERAL_STR "general"
#define MM_SYMM_STR "symmetric"
#define MM_HERM_STR "hermitian"
#define MM_SKEW_STR "skew-symmetric"
#define MM_PATTERN_STR "pattern"
/* high level routines */
int mm_read_mtx_crd(char *fname, int *M, int *N, int *nz, int **I, int **J, double **val, MM_typecode *matcode);
int mm_write_mtx_crd(char fname[], int M, int N, int nz, int I[], int J[], double val[], MM_typecode matcode);
int mm_read_mtx_crd_data(FILE *f, int M, int N, int nz, int I[], int J[], double val[], MM_typecode matcode);
int mm_read_mtx_crd_entry(FILE *f, int *I, int *J, double *real, double *img, MM_typecode matcode);
int mm_read_unsymmetric_sparse(const char *fname, int *M_, int *N_, int *nz_, double **val_, int **I_, int **J_);
#if defined(__cplusplus)
}
#endif /* __cplusplus */
#endif

492
cpp/4_CUDA_Libraries/cuSolverSp_LowlevelQR/mmio_wrapper.cpp Normal file
View File

@@ -0,0 +1,492 @@
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <cusolverDn.h>
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include "mmio.h"
/* avoid Windows warnings (for example: strcpy, fscanf, etc.) */
#if defined(_WIN32)
#define _CRT_SECURE_NO_WARNINGS
#endif
/* various __inline__ __device__ function to initialize a T_ELEM */
template <typename T_ELEM> __inline__ T_ELEM cuGet(int);
template <> __inline__ float cuGet<float>(int x) { return float(x); }
template <> __inline__ double cuGet<double>(int x) { return double(x); }
template <> __inline__ cuComplex cuGet<cuComplex>(int x) { return (make_cuComplex(float(x), 0.0f)); }
template <> __inline__ cuDoubleComplex cuGet<cuDoubleComplex>(int x) { return (make_cuDoubleComplex(double(x), 0.0)); }
template <typename T_ELEM> __inline__ T_ELEM cuGet(int, int);
template <> __inline__ float cuGet<float>(int x, int y) { return float(x); }
template <> __inline__ double cuGet<double>(int x, int y) { return double(x); }
template <> __inline__ cuComplex cuGet<cuComplex>(int x, int y) { return make_cuComplex(float(x), float(y)); }
template <> __inline__ cuDoubleComplex cuGet<cuDoubleComplex>(int x, int y)
{
return (make_cuDoubleComplex(double(x), double(y)));
}
template <typename T_ELEM> __inline__ T_ELEM cuGet(float);
template <> __inline__ float cuGet<float>(float x) { return float(x); }
template <> __inline__ double cuGet<double>(float x) { return double(x); }
template <> __inline__ cuComplex cuGet<cuComplex>(float x) { return (make_cuComplex(float(x), 0.0f)); }
template <> __inline__ cuDoubleComplex cuGet<cuDoubleComplex>(float x)
{
return (make_cuDoubleComplex(double(x), 0.0));
}
template <typename T_ELEM> __inline__ T_ELEM cuGet(float, float);
template <> __inline__ float cuGet<float>(float x, float y) { return float(x); }
template <> __inline__ double cuGet<double>(float x, float y) { return double(x); }
template <> __inline__ cuComplex cuGet<cuComplex>(float x, float y) { return (make_cuComplex(float(x), float(y))); }
template <> __inline__ cuDoubleComplex cuGet<cuDoubleComplex>(float x, float y)
{
return (make_cuDoubleComplex(double(x), double(y)));
}
template <typename T_ELEM> __inline__ T_ELEM cuGet(double);
template <> __inline__ float cuGet<float>(double x) { return float(x); }
template <> __inline__ double cuGet<double>(double x) { return double(x); }
template <> __inline__ cuComplex cuGet<cuComplex>(double x) { return (make_cuComplex(float(x), 0.0f)); }
template <> __inline__ cuDoubleComplex cuGet<cuDoubleComplex>(double x)
{
return (make_cuDoubleComplex(double(x), 0.0));
}
template <typename T_ELEM> __inline__ T_ELEM cuGet(double, double);
template <> __inline__ float cuGet<float>(double x, double y) { return float(x); }
template <> __inline__ double cuGet<double>(double x, double y) { return double(x); }
template <> __inline__ cuComplex cuGet<cuComplex>(double x, double y) { return (make_cuComplex(float(x), float(y))); }
template <> __inline__ cuDoubleComplex cuGet<cuDoubleComplex>(double x, double y)
{
return (make_cuDoubleComplex(double(x), double(y)));
}
static void compress_index(const int *Ind, int nnz, int m, int *Ptr, int base)
{
int i;
/* initialize everything to zero */
for (i = 0; i < m + 1; i++) {
Ptr[i] = 0;
}
/* count elements in every row */
Ptr[0] = base;
for (i = 0; i < nnz; i++) {
Ptr[Ind[i] + (1 - base)]++;
}
/* add all the values */
for (i = 0; i < m; i++) {
Ptr[i + 1] += Ptr[i];
}
}
struct cooFormat
{
int i;
int j;
int p; // permutation
};
int cmp_cooFormat_csr(struct cooFormat *s, struct cooFormat *t)
{
if (s->i < t->i) {
return -1;
}
else if (s->i > t->i) {
return 1;
}
else {
return s->j - t->j;
}
}
int cmp_cooFormat_csc(struct cooFormat *s, struct cooFormat *t)
{
if (s->j < t->j) {
return -1;
}
else if (s->j > t->j) {
return 1;
}
else {
return s->i - t->i;
}
}
typedef int (*FUNPTR)(const void *, const void *);
typedef int (*FUNPTR2)(struct cooFormat *s, struct cooFormat *t);
static FUNPTR2 fptr_array[2] = {
cmp_cooFormat_csr,
cmp_cooFormat_csc,
};
static int verify_pattern(int m, int nnz, int *csrRowPtr, int *csrColInd)
{
int i, col, start, end, base_index;
int error_found = 0;
if (nnz != (csrRowPtr[m] - csrRowPtr[0])) {
fprintf(stderr,
"Error (nnz check failed): (csrRowPtr[%d]=%d - csrRowPtr[%d]=%d) "
"!= (nnz=%d)\n",
0,
csrRowPtr[0],
m,
csrRowPtr[m],
nnz);
error_found = 1;
}
base_index = csrRowPtr[0];
if ((0 != base_index) && (1 != base_index)) {
fprintf(stderr, "Error (base index check failed): base index = %d\n", base_index);
error_found = 1;
}
for (i = 0; (!error_found) && (i < m); i++) {
start = csrRowPtr[i] - base_index;
end = csrRowPtr[i + 1] - base_index;
if (start > end) {
fprintf(stderr,
"Error (corrupted row): csrRowPtr[%d] (=%d) > csrRowPtr[%d] (=%d)\n",
i,
start + base_index,
i + 1,
end + base_index);
error_found = 1;
}
for (col = start; col < end; col++) {
if (csrColInd[col] < base_index) {
fprintf(stderr, "Error (column vs. base index check failed): csrColInd[%d] < %d\n", col, base_index);
error_found = 1;
}
if ((col < (end - 1)) && (csrColInd[col] >= csrColInd[col + 1])) {
fprintf(stderr,
"Error (sorting of the column indecis check failed): "
"(csrColInd[%d]=%d) >= (csrColInd[%d]=%d)\n",
col,
csrColInd[col],
col + 1,
csrColInd[col + 1]);
error_found = 1;
}
}
}
return error_found;
}
template <typename T_ELEM>
int loadMMSparseMatrix(char *filename,
char elem_type,
bool csrFormat,
int *m,
int *n,
int *nnz,
T_ELEM **aVal,
int **aRowInd,
int **aColInd,
int extendSymMatrix)
{
MM_typecode matcode;
double *tempVal;
int *tempRowInd, *tempColInd;
double *tval;
int *trow, *tcol;
int *csrRowPtr, *cscColPtr;
int i, j, error, base, count;
struct cooFormat *work;
/* read the matrix */
error = mm_read_mtx_crd(filename, m, n, nnz, &trow, &tcol, &tval, &matcode);
if (error) {
fprintf(stderr, "!!!! can not open file: '%s'\n", filename);
return 1;
}
/* start error checking */
if (mm_is_complex(matcode) && ((elem_type != 'z') && (elem_type != 'c'))) {
fprintf(stderr, "!!!! complex matrix requires type 'z' or 'c'\n");
return 1;
}
if (mm_is_dense(matcode) || mm_is_array(matcode) || mm_is_pattern(matcode) /*|| mm_is_integer(matcode)*/) {
fprintf(stderr, "!!!! dense, array, pattern and integer matrices are not supported\n");
return 1;
}
/* if necessary symmetrize the pattern (transform from triangular to full) */
if ((extendSymMatrix) && (mm_is_symmetric(matcode) || mm_is_hermitian(matcode) || mm_is_skew(matcode))) {
// count number of non-diagonal elements
count = 0;
for (i = 0; i < (*nnz); i++) {
if (trow[i] != tcol[i]) {
count++;
}
}
// allocate space for the symmetrized matrix
tempRowInd = (int *)malloc((*nnz + count) * sizeof(int));
tempColInd = (int *)malloc((*nnz + count) * sizeof(int));
if (mm_is_real(matcode) || mm_is_integer(matcode)) {
tempVal = (double *)malloc((*nnz + count) * sizeof(double));
}
else {
tempVal = (double *)malloc(2 * (*nnz + count) * sizeof(double));
}
// copy the elements regular and transposed locations
for (j = 0, i = 0; i < (*nnz); i++) {
tempRowInd[j] = trow[i];
tempColInd[j] = tcol[i];
if (mm_is_real(matcode) || mm_is_integer(matcode)) {
tempVal[j] = tval[i];
}
else {
tempVal[2 * j] = tval[2 * i];
tempVal[2 * j + 1] = tval[2 * i + 1];
}
j++;
if (trow[i] != tcol[i]) {
tempRowInd[j] = tcol[i];
tempColInd[j] = trow[i];
if (mm_is_real(matcode) || mm_is_integer(matcode)) {
if (mm_is_skew(matcode)) {
tempVal[j] = -tval[i];
}
else {
tempVal[j] = tval[i];
}
}
else {
if (mm_is_hermitian(matcode)) {
tempVal[2 * j] = tval[2 * i];
tempVal[2 * j + 1] = -tval[2 * i + 1];
}
else {
tempVal[2 * j] = tval[2 * i];
tempVal[2 * j + 1] = tval[2 * i + 1];
}
}
j++;
}
}
(*nnz) += count;
// free temporary storage
free(trow);
free(tcol);
free(tval);
}
else {
tempRowInd = trow;
tempColInd = tcol;
tempVal = tval;
}
// life time of (trow, tcol, tval) is over.
// please use COO format (tempRowInd, tempColInd, tempVal)
// use qsort to sort COO format
work = (struct cooFormat *)malloc(sizeof(struct cooFormat) * (*nnz));
if (NULL == work) {
fprintf(stderr, "!!!! allocation error, malloc failed\n");
return 1;
}
for (i = 0; i < (*nnz); i++) {
work[i].i = tempRowInd[i];
work[i].j = tempColInd[i];
work[i].p = i; // permutation is identity
}
if (csrFormat) {
/* create row-major ordering of indices (sorted by row and within each row
* by column) */
qsort(work, *nnz, sizeof(struct cooFormat), (FUNPTR)fptr_array[0]);
}
else {
/* create column-major ordering of indices (sorted by column and within each
* column by row) */
qsort(work, *nnz, sizeof(struct cooFormat), (FUNPTR)fptr_array[1]);
}
// (tempRowInd, tempColInd) is sorted either by row-major or by col-major
for (i = 0; i < (*nnz); i++) {
tempRowInd[i] = work[i].i;
tempColInd[i] = work[i].j;
}
// setup base
// check if there is any row/col 0, if so base-0
// check if there is any row/col equal to matrix dimension m/n, if so base-1
int base0 = 0;
int base1 = 0;
for (i = 0; i < (*nnz); i++) {
const int row = tempRowInd[i];
const int col = tempColInd[i];
if ((0 == row) || (0 == col)) {
base0 = 1;
}
if ((*m == row) || (*n == col)) {
base1 = 1;
}
}
if (base0 && base1) {
printf("Error: input matrix is base-0 and base-1 \n");
return 1;
}
base = 0;
if (base1) {
base = 1;
}
/* compress the appropriate indices */
if (csrFormat) {
/* CSR format (assuming row-major format) */
csrRowPtr = (int *)malloc(((*m) + 1) * sizeof(csrRowPtr[0]));
if (!csrRowPtr)
return 1;
compress_index(tempRowInd, *nnz, *m, csrRowPtr, base);
*aRowInd = csrRowPtr;
*aColInd = (int *)malloc((*nnz) * sizeof(int));
}
else {
/* CSC format (assuming column-major format) */
cscColPtr = (int *)malloc(((*n) + 1) * sizeof(cscColPtr[0]));
if (!cscColPtr)
return 1;
compress_index(tempColInd, *nnz, *n, cscColPtr, base);
*aColInd = cscColPtr;
*aRowInd = (int *)malloc((*nnz) * sizeof(int));
}
/* transfrom the matrix values of type double into one of the cusparse library
* types */
*aVal = (T_ELEM *)malloc((*nnz) * sizeof(T_ELEM));
for (i = 0; i < (*nnz); i++) {
if (csrFormat) {
(*aColInd)[i] = tempColInd[i];
}
else {
(*aRowInd)[i] = tempRowInd[i];
}
if (mm_is_real(matcode) || mm_is_integer(matcode)) {
(*aVal)[i] = cuGet<T_ELEM>(tempVal[work[i].p]);
}
else {
(*aVal)[i] = cuGet<T_ELEM>(tempVal[2 * work[i].p], tempVal[2 * work[i].p + 1]);
}
}
/* check for corruption */
int error_found;
if (csrFormat) {
error_found = verify_pattern(*m, *nnz, *aRowInd, *aColInd);
}
else {
error_found = verify_pattern(*n, *nnz, *aColInd, *aRowInd);
}
if (error_found) {
fprintf(stderr, "!!!! verify_pattern failed\n");
return 1;
}
/* cleanup and exit */
free(work);
free(tempVal);
free(tempColInd);
free(tempRowInd);
return 0;
}
/* specific instantiation */
template int loadMMSparseMatrix<float>(char *filename,
char elem_type,
bool csrFormat,
int *m,
int *n,
int *nnz,
float **aVal,
int **aRowInd,
int **aColInd,
int extendSymMatrix);
template int loadMMSparseMatrix<double>(char *filename,
char elem_type,
bool csrFormat,
int *m,
int *n,
int *nnz,
double **aVal,
int **aRowInd,
int **aColInd,
int extendSymMatrix);
template int loadMMSparseMatrix<cuComplex>(char *filename,
char elem_type,
bool csrFormat,
int *m,
int *n,
int *nnz,
cuComplex **aVal,
int **aRowInd,
int **aColInd,
int extendSymMatrix);
template int loadMMSparseMatrix<cuDoubleComplex>(char *filename,
char elem_type,
bool csrFormat,
int *m,
int *n,
int *nnz,
cuDoubleComplex **aVal,
int **aRowInd,
int **aColInd,
int extendSymMatrix);