【cuda学习日记】8.1 GPU加速库 --cuSPARSE

8.1.1 cuSPARSE概念

cuSPARSE是一个线性代数库，内含很多通用的稀疏线性代数函数。

稠密存储方式（dense）

把矩阵中的每个元素都存储起来，不管它是否为零。
在这里插入图片描述

坐标稀疏矩阵格式（COO）

坐标稀疏矩阵格式（COO）存储了稀疏矩阵中每个非零元素的行坐标、列坐标及其元素值。因此当通过给定的行列坐标对存储在COO格式中的稀疏矩阵进行检索的时候，如果没有匹配值，那么这个位置的值为零. 只有当矩阵中的非零元素不到1/3时，才会节约存储空间。
在这里插入图片描述

压缩稀疏行方式（CSR）和压缩稀疏列方式（CSC）compressed sparse row/col

参考： https://blog.csdn.net/m0_57102661/article/details/140112320

#include <iostream>
#include <cuda_runtime.h>
#include <cusparse.h>int main() {// 初始化稠密矩阵 (4x4)const int m = 4; // 行数const int n = 4; // 列数const int nnz = 9; // 非零元素个数double h_A[m * n] = {1, 0, 0, 2,0, 3, 4, 0,5, 0, 6, 0,0, 7, 0, 8};// 打印稠密矩阵std::cout << "Dense Matrix:" << std::endl;for (int i = 0; i < m; i++) {for (int j = 0; j < n; j++) {std::cout << h_A[i * n + j] << " ";}std::cout << std::endl;}// 初始化 cuSPARSEcusparseHandle_t handle;CHECK_CUSPARSE(cusparseCreate(&handle));// 分配设备内存并拷贝稠密矩阵到设备double *d_A;CHECK_CUDA(cudaMalloc((void**)&d_A, m * n * sizeof(double)));CHECK_CUDA(cudaMemcpy(d_A, h_A, m * n * sizeof(double), cudaMemcpyHostToDevice));// 创建稠密矩阵描述符cusparseDnMatDescr_t matA;CHECK_CUSPARSE(cusparseCreateDnMat(&matA, m, n, n, d_A, CUDA_R_64F, CUSPARSE_ORDER_ROW));// 创建稀疏矩阵描述符 (CSR 格式)cusparseSpMatDescr_t matB;int *d_csrRowPtr, *d_csrColInd;double *d_csrVal;CHECK_CUDA(cudaMalloc((void**)&d_csrRowPtr, (m + 1) * sizeof(int)));CHECK_CUDA(cudaMalloc((void**)&d_csrColInd, nnz * sizeof(int)));CHECK_CUDA(cudaMalloc((void**)&d_csrVal, nnz * sizeof(double)));CHECK_CUSPARSE(cusparseCreateCsr(&matB, m, n, nnz, d_csrRowPtr, d_csrColInd, d_csrVal,CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, CUDA_R_64F));// 分配缓冲区size_t bufferSize = 0;CHECK_CUSPARSE(cusparseDenseToSparse_bufferSize(handle, matA, matB, CUSPARSE_DENSETOSPARSE_ALG_DEFAULT, &bufferSize));void *d_buffer;CHECK_CUDA(cudaMalloc(&d_buffer, bufferSize));// 执行稠密矩阵转稀疏矩阵CHECK_CUSPARSE(cusparseDenseToSparse_analysis(handle, matA, matB, CUSPARSE_DENSETOSPARSE_ALG_DEFAULT, d_buffer));CHECK_CUSPARSE(cusparseDenseToSparse_convert(handle, matA, matB, CUSPARSE_DENSETOSPARSE_ALG_DEFAULT, d_buffer));// 将结果拷贝回主机int h_csrRowPtr[m + 1];int h_csrColInd[nnz];double h_csrVal[nnz];CHECK_CUDA(cudaMemcpy(h_csrRowPtr, d_csrRowPtr, (m + 1) * sizeof(int), cudaMemcpyDeviceToHost));CHECK_CUDA(cudaMemcpy(h_csrColInd, d_csrColInd, nnz * sizeof(int), cudaMemcpyDeviceToHost));CHECK_CUDA(cudaMemcpy(h_csrVal, d_csrVal, nnz * sizeof(double), cudaMemcpyDeviceToHost));// 打印 CSR 格式的结果std::cout << "\nCSR RowPtr:" << std::endl;for (int i = 0; i <= m; i++) {std::cout << h_csrRowPtr[i] << " ";}std::cout << std::endl;std::cout << "CSR ColInd:" << std::endl;for (int i = 0; i < nnz; i++) {std::cout << h_csrColInd[i] << " ";}std::cout << std::endl;std::cout << "CSR Val:" << std::endl;for (int i = 0; i < nnz; i++) {std::cout << h_csrVal[i] << " ";}std::cout << std::endl;// 释放设备内存CHECK_CUDA(cudaFree(d_A));CHECK_CUDA(cudaFree(d_csrRowPtr));CHECK_CUDA(cudaFree(d_csrColInd));CHECK_CUDA(cudaFree(d_csrVal));CHECK_CUDA(cudaFree(d_buffer));// 销毁描述符和句柄CHECK_CUSPARSE(cusparseDestroyDnMat(matA));CHECK_CUSPARSE(cusparseDestroySpMat(matB));CHECK_CUSPARSE(cusparseDestroy(handle));return 0;
}在common.h里添加```c
// 检查 CUDA 错误
#define CHECK_CUDA(func)                                                       \
{                                                                              \cudaError_t status = (func);                                               \if (status != cudaSuccess) {                                               \std::cerr << "CUDA Error: " << cudaGetErrorString(status) << std::endl; \exit(EXIT_FAILURE);                                                    \}                                                                          \
}// 检查 cuSPARSE 错误
#define CHECK_CUSPARSE(func)                                                   \
{                                                                              \cusparseStatus_t status = (func);                                          \if (status != CUSPARSE_STATUS_SUCCESS) {                                   \std::cerr << "cuSPARSE Error: " << status << std::endl;                \exit(EXIT_FAILURE);                                                    \}                                                                          \
}

编译： nvcc -o dense2csr cusparse.cu -lcusparse
输出：```c
Dense Matrix:
1 0 0 2 
0 3 4 0 
5 0 6 0 
0 7 0 8 CSR RowPtr:
0 2 4 6 8 
CSR ColInd:
0 3 1 2 0 2 1 3 0 
CSR Val:
1 2 3 4 5 6 7 8 0