使用CUDA的cusolverSpScsrlsvchol求解稀疏线性问题时出现段错误

核心错误点 #

1. 未关联cusparse句柄到cusolverSp句柄
   cusolverSp 依赖 cusparse 库的内部实现，必须显式将创建好的 cusparseHandle_t 绑定到 cusolverSpHandle_t，否则调用求解函数时会因内部空指针访问触发段错误。需要添加：

runtime_assert_cuda(cusolverSpSetCuSparseHandle(handle_cusolver_sp, handle_cusparse));

2. 矩阵类型设置不符合求解器要求
   cusolverSpScsrlsvchol 是专门针对对称正定(SPD)矩阵的Cholesky分解求解器，必须将矩阵类型设置为 CUSPARSE_MATRIX_TYPE_SYMMETRIC（或对应三角存储的LOWER/UPPER类型），而非GENERAL。修改代码：

runtime_assert_cuda(cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_SYMMETRIC));

（注：你的对角矩阵本身属于对称正定矩阵，此设置符合要求）

3. 可选优化：避免输入输出复用同一指针
   虽然函数允许将dev_rhs同时作为输入右端项和输出解向量，但从代码可读性和规避潜在风险角度，建议单独分配输出向量：

float* dev_x;
runtime_assert_cuda(cudaMalloc((void**)&dev_x, 4 * sizeof(float)));
// 调用时传入dev_x作为输出参数

修复后的核心代码片段 #

cusolverSpHandle_t handle_cusolver_sp;
cusparseHandle_t   handle_cusparse;

// 创建句柄
runtime_assert_cuda(cusolverSpCreate(&handle_cusolver_sp));
runtime_assert_cuda(cusparseCreate(&handle_cusparse));
// 关联cusparse句柄到cusolverSp（关键步骤）
runtime_assert_cuda(cusolverSpSetCuSparseHandle(handle_cusolver_sp, handle_cusparse));

runtime_assert_cuda(cudaSetDevice(0));

// CPU侧数据
float host_csr_values[4]{1,1,1,1};
int   host_csr_col_id[4]{0,1,2,3};
int   host_csr_row_pt[5]{0,1,2,3,4};
float host_rhs       [4]{0,3,7,1};
int   host_singular  [1]{0};

// GPU内存分配（新增输出向量dev_x）
float* dev_csr_values;
int  * dev_csr_col_id;
int  * dev_csr_row_pt;
float* dev_rhs;
float* dev_x;
int  * dev_singular;

runtime_assert_cuda(cudaMalloc((void**)&dev_csr_values,4 * sizeof(float)));
runtime_assert_cuda(cudaMalloc((void**)&dev_csr_col_id,4 * sizeof(int)));
runtime_assert_cuda(cudaMalloc((void**)&dev_csr_row_pt,5 * sizeof(int)));
runtime_assert_cuda(cudaMalloc((void**)&dev_rhs,4 * sizeof(float)));
runtime_assert_cuda(cudaMalloc((void**)&dev_x,4 * sizeof(float)));
runtime_assert_cuda(cudaMalloc((void**)&dev_singular,1 * sizeof(int)));

// 数据拷贝到GPU
runtime_assert_cuda(cudaMemcpy(dev_csr_values, host_csr_values, 4 * sizeof(float), cudaMemcpyHostToDevice));
runtime_assert_cuda(cudaMemcpy(dev_csr_col_id, host_csr_col_id, 4 * sizeof(int), cudaMemcpyHostToDevice));
runtime_assert_cuda(cudaMemcpy(dev_csr_row_pt, host_csr_row_pt, 5 * sizeof(int), cudaMemcpyHostToDevice));
runtime_assert_cuda(cudaMemcpy(dev_rhs, host_rhs, 4 * sizeof(float), cudaMemcpyHostToDevice));

// 创建矩阵描述符，修正类型为对称
cusparseMatDescr_t descr;
runtime_assert_cuda(cusparseCreateMatDescr(&descr));
runtime_assert_cuda(cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_SYMMETRIC));
runtime_assert_cuda(cusparseSetMatIndexBase(descr, CUSPARSE_INDEX_BASE_ZERO));

// 调用求解器
runtime_assert_cuda(cusolverSpScsrlsvchol(handle_cusolver_sp,
                                          4,
                                          4,
                                          descr,
                                          dev_csr_values,
                                          dev_csr_row_pt,
                                          dev_csr_col_id,
                                          dev_rhs,
                                          0.f,    // 容差
                                          0,      // 不启用重排序
                                          dev_x,
                                          dev_singular));

额外调试建议 #

• 确保所有CUDA API调用的返回值都被runtime_assert_cuda正确检查，避免遗漏隐性错误
• 若问题仍存在，可尝试使用compute-sanitizer的racecheck或initcheck工具，进一步排查内存初始化或竞争问题

内容的提问来源于stack exchange，提问作者Finn Eggers