/*! \file */ /* ************************************************************************ * Copyright (C) 2022-2025 Advanced Micro Devices, Inc. All rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to deal * in the Software without restriction, including without limitation the rights * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell * copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in * all copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN * THE SOFTWARE. * * ************************************************************************ */ #include "rocsparse_enum.hpp" #include "testing.hpp" #include "auto_testing_bad_arg.hpp" template void testing_bsrgeam_bad_arg(const Arguments& arg) { static const size_t safe_size = 1; T h_alpha = 0.6; T h_beta = 0.2; // Create rocsparse handle rocsparse_local_handle local_handle; // Create matrix descriptors rocsparse_local_mat_descr local_descr_A; rocsparse_local_mat_descr local_descr_B; rocsparse_local_mat_descr local_descr_C; rocsparse_handle handle = local_handle; rocsparse_direction dir = rocsparse_direction_row; rocsparse_int mb = safe_size; rocsparse_int nb = safe_size; rocsparse_int block_dim = safe_size; const T* alpha = &h_alpha; const rocsparse_mat_descr descr_A = local_descr_A; rocsparse_int nnzb_A = safe_size; const T* bsr_val_A = (const T*)0x4; const rocsparse_int* bsr_row_ptr_A = (const rocsparse_int*)0x4; const rocsparse_int* bsr_col_ind_A = (const rocsparse_int*)0x4; const T* beta = &h_beta; const rocsparse_mat_descr descr_B = local_descr_B; rocsparse_int nnzb_B = safe_size; const T* bsr_val_B = (const T*)0x4; const rocsparse_int* bsr_row_ptr_B = (const rocsparse_int*)0x4; const rocsparse_int* bsr_col_ind_B = (const rocsparse_int*)0x4; const rocsparse_mat_descr descr_C = local_descr_C; T* bsr_val_C = (T*)0x4; host_dense_vector hbsr_row_ptr_C(safe_size + 1); hbsr_row_ptr_C[0] = 0; hbsr_row_ptr_C[1] = 1; device_dense_vector dbsr_row_ptr_C(hbsr_row_ptr_C); rocsparse_int* bsr_row_ptr_C = (rocsparse_int*)dbsr_row_ptr_C; rocsparse_int* bsr_col_ind_C = (rocsparse_int*)0x4; rocsparse_int* nnzb_C = (rocsparse_int*)0x4; #define PARAMS_NNZB \ handle, dir, mb, nb, block_dim, descr_A, nnzb_A, bsr_row_ptr_A, bsr_col_ind_A, descr_B, \ nnzb_B, bsr_row_ptr_B, bsr_col_ind_B, descr_C, bsr_row_ptr_C, nnzb_C #define PARAMS \ handle, dir, mb, nb, block_dim, alpha, descr_A, nnzb_A, bsr_val_A, bsr_row_ptr_A, \ bsr_col_ind_A, beta, descr_B, nnzb_B, bsr_val_B, bsr_row_ptr_B, bsr_col_ind_B, descr_C, \ bsr_val_C, bsr_row_ptr_C, bsr_col_ind_C bad_arg_analysis(rocsparse_bsrgeam_nnzb, PARAMS_NNZB); bad_arg_analysis(rocsparse_bsrgeam, PARAMS); for(auto val : rocsparse_matrix_type_t::values) { if(val != rocsparse_matrix_type_general) { CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_A, val)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_B, val)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_C, val)); EXPECT_ROCSPARSE_STATUS(rocsparse_bsrgeam_nnzb(PARAMS_NNZB), rocsparse_status_not_implemented); EXPECT_ROCSPARSE_STATUS(rocsparse_bsrgeam(PARAMS), rocsparse_status_not_implemented); } } CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_A, rocsparse_matrix_type_general)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_B, rocsparse_matrix_type_general)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_C, rocsparse_matrix_type_general)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_A, rocsparse_matrix_type_general)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_B, rocsparse_matrix_type_general)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_type(descr_C, rocsparse_matrix_type_general)); // Check block_dim == 0 block_dim = 0; EXPECT_ROCSPARSE_STATUS(rocsparse_bsrgeam_nnzb(PARAMS_NNZB), rocsparse_status_invalid_size); EXPECT_ROCSPARSE_STATUS(rocsparse_bsrgeam(PARAMS), rocsparse_status_invalid_size); block_dim = safe_size; CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_storage_mode(descr_A, rocsparse_storage_mode_unsorted)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_storage_mode(descr_B, rocsparse_storage_mode_unsorted)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_storage_mode(descr_C, rocsparse_storage_mode_unsorted)); EXPECT_ROCSPARSE_STATUS(rocsparse_bsrgeam_nnzb(PARAMS_NNZB), rocsparse_status_requires_sorted_storage); EXPECT_ROCSPARSE_STATUS(rocsparse_bsrgeam(PARAMS), rocsparse_status_requires_sorted_storage); #undef PARAMS #undef PARAMS_NNZB } template void testing_bsrgeam(const Arguments& arg) { rocsparse_direction dir = arg.direction; rocsparse_int M = arg.M; rocsparse_int N = arg.N; rocsparse_int block_dim = arg.block_dim; rocsparse_index_base baseA = arg.baseA; rocsparse_index_base baseB = arg.baseB; rocsparse_index_base baseC = arg.baseC; static constexpr bool full_rank = false; rocsparse_matrix_factory matrix_factory(arg, arg.timing ? false : true, full_rank); rocsparse_matrix_factory_random matrix_factory_random(full_rank); rocsparse_int Mb = (M + block_dim - 1) / block_dim; rocsparse_int Nb = (N + block_dim - 1) / block_dim; host_scalar h_alpha(arg.get_alpha()); host_scalar h_beta(arg.get_beta()); // Create rocsparse handle rocsparse_local_handle handle; // Create matrix descriptor rocsparse_local_mat_descr descrA; rocsparse_local_mat_descr descrB; rocsparse_local_mat_descr descrC; // Set matrix index base CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_index_base(descrA, baseA)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_index_base(descrB, baseB)); CHECK_ROCSPARSE_ERROR(rocsparse_set_mat_index_base(descrC, baseC)); // Allocate host memory for matrices host_vector hbsr_row_ptr_A; host_vector hbsr_col_ind_A; host_vector hbsr_val_A; host_vector hbsr_row_ptr_B; host_vector hbsr_col_ind_B; host_vector hbsr_val_B; // Sample matrix rocsparse_int nnzb_A; rocsparse_int nnzb_B; rocsparse_int hnnzb_C_gold; rocsparse_int hnnzb_C_1; rocsparse_int hnnzb_C_2; // Sample A matrix_factory.init_bsr( hbsr_row_ptr_A, hbsr_col_ind_A, hbsr_val_A, dir, Mb, Nb, nnzb_A, block_dim, baseA); // Sample B matrix_factory_random.init_gebsr(hbsr_row_ptr_B, hbsr_col_ind_B, hbsr_val_B, dir, Mb, Nb, nnzb_B, block_dim, block_dim, baseB, rocsparse_matrix_type_general, rocsparse_fill_mode_lower, rocsparse_storage_mode_sorted); // Allocate device memory device_vector dbsr_row_ptr_A(Mb + 1); device_vector dbsr_col_ind_A(nnzb_A); device_vector dbsr_val_A(size_t(nnzb_A) * block_dim * block_dim); device_vector dbsr_row_ptr_B(Mb + 1); device_vector dbsr_col_ind_B(nnzb_B); device_vector dbsr_val_B(size_t(nnzb_B) * block_dim * block_dim); device_scalar d_alpha(h_alpha); device_scalar d_beta(h_beta); device_vector dbsr_row_ptr_C_1(Mb + 1); device_vector dbsr_row_ptr_C_2(Mb + 1); device_vector dnnzb_C_2(1); // Copy data from CPU to device dbsr_row_ptr_A.transfer_from(hbsr_row_ptr_A); dbsr_col_ind_A.transfer_from(hbsr_col_ind_A); dbsr_val_A.transfer_from(hbsr_val_A); dbsr_row_ptr_B.transfer_from(hbsr_row_ptr_B); dbsr_col_ind_B.transfer_from(hbsr_col_ind_B); dbsr_val_B.transfer_from(hbsr_val_B); if(arg.unit_check) { // Obtain nnzb of C // Pointer mode host CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_host)); CHECK_ROCSPARSE_ERROR(rocsparse_bsrgeam_nnzb(handle, dir, Mb, Nb, block_dim, descrA, nnzb_A, dbsr_row_ptr_A, dbsr_col_ind_A, descrB, nnzb_B, dbsr_row_ptr_B, dbsr_col_ind_B, descrC, dbsr_row_ptr_C_1, &hnnzb_C_1)); // Pointer mode device CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_device)); CHECK_ROCSPARSE_ERROR(rocsparse_bsrgeam_nnzb(handle, dir, Mb, Nb, block_dim, descrA, nnzb_A, dbsr_row_ptr_A, dbsr_col_ind_A, descrB, nnzb_B, dbsr_row_ptr_B, dbsr_col_ind_B, descrC, dbsr_row_ptr_C_2, dnnzb_C_2)); // Copy output to host host_vector hbsr_row_ptr_C_1(Mb + 1); host_vector hbsr_row_ptr_C_2(Mb + 1); hbsr_row_ptr_C_1.transfer_from(dbsr_row_ptr_C_1); hbsr_row_ptr_C_2.transfer_from(dbsr_row_ptr_C_2); CHECK_HIP_ERROR( hipMemcpy(&hnnzb_C_2, dnnzb_C_2, sizeof(rocsparse_int), hipMemcpyDeviceToHost)); // CPU bsrgemm_nnzb host_vector hbsr_row_ptr_C_gold(Mb + 1); host_bsrgeam_nnzb(dir, Mb, Nb, block_dim, h_alpha, hbsr_row_ptr_A, hbsr_col_ind_A, h_beta, hbsr_row_ptr_B, hbsr_col_ind_B, hbsr_row_ptr_C_gold, &hnnzb_C_gold, baseA, baseB, baseC); // Check nnz of C unit_check_scalar(hnnzb_C_gold, hnnzb_C_1); unit_check_scalar(hnnzb_C_gold, hnnzb_C_2); // Check row pointers of C hbsr_row_ptr_C_gold.unit_check(hbsr_row_ptr_C_1); hbsr_row_ptr_C_gold.unit_check(hbsr_row_ptr_C_2); // Allocate device memory for C device_vector dbsr_col_ind_C_1(hnnzb_C_1); device_vector dbsr_col_ind_C_2(hnnzb_C_2); device_vector dbsr_val_C_1(size_t(hnnzb_C_1) * block_dim * block_dim); device_vector dbsr_val_C_2(size_t(hnnzb_C_2) * block_dim * block_dim); // Perform matrix matrix multiplication // Pointer mode host CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_host)); CHECK_ROCSPARSE_ERROR(rocsparse_bsrgeam(handle, dir, Mb, Nb, block_dim, h_alpha, descrA, nnzb_A, dbsr_val_A, dbsr_row_ptr_A, dbsr_col_ind_A, h_beta, descrB, nnzb_B, dbsr_val_B, dbsr_row_ptr_B, dbsr_col_ind_B, descrC, dbsr_val_C_1, dbsr_row_ptr_C_1, dbsr_col_ind_C_1)); // Pointer mode device CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_device)); CHECK_ROCSPARSE_ERROR(rocsparse_bsrgeam(handle, dir, Mb, Nb, block_dim, d_alpha, descrA, nnzb_A, dbsr_val_A, dbsr_row_ptr_A, dbsr_col_ind_A, d_beta, descrB, nnzb_B, dbsr_val_B, dbsr_row_ptr_B, dbsr_col_ind_B, descrC, dbsr_val_C_2, dbsr_row_ptr_C_2, dbsr_col_ind_C_2)); // Copy output to host host_vector hbsr_col_ind_C_1(hnnzb_C_1); host_vector hbsr_col_ind_C_2(hnnzb_C_2); host_vector hbsr_val_C_1(size_t(hnnzb_C_1) * block_dim * block_dim); host_vector hbsr_val_C_2(size_t(hnnzb_C_2) * block_dim * block_dim); hbsr_col_ind_C_1.transfer_from(dbsr_col_ind_C_1); hbsr_col_ind_C_2.transfer_from(dbsr_col_ind_C_2); hbsr_val_C_1.transfer_from(dbsr_val_C_1); hbsr_val_C_2.transfer_from(dbsr_val_C_2); // CPU bsrgemm host_vector hbsr_col_ind_C_gold(hnnzb_C_gold); host_vector hbsr_val_C_gold(size_t(hnnzb_C_gold) * block_dim * block_dim); host_bsrgeam(dir, Mb, Nb, block_dim, h_alpha, hbsr_row_ptr_A, hbsr_col_ind_A, hbsr_val_A, h_beta, hbsr_row_ptr_B, hbsr_col_ind_B, hbsr_val_B, hbsr_row_ptr_C_gold, hbsr_col_ind_C_gold, hbsr_val_C_gold, baseA, baseB, baseC); // Check C if(ROCSPARSE_REPRODUCIBILITY) { rocsparse_reproducibility::save("hbsr_row_ptr_C_1", hbsr_row_ptr_C_1); rocsparse_reproducibility::save("hbsr_col_ind_C_1", hbsr_col_ind_C_1); rocsparse_reproducibility::save("hbsr_val_C_1", hbsr_val_C_1); rocsparse_reproducibility::save("hbsr_row_ptr_C_2", hbsr_row_ptr_C_2); rocsparse_reproducibility::save("hbsr_col_ind_C_2", hbsr_col_ind_C_2); rocsparse_reproducibility::save("hbsr_val_C_2", hbsr_val_C_2); } hbsr_col_ind_C_gold.unit_check(hbsr_col_ind_C_1); hbsr_col_ind_C_gold.unit_check(hbsr_col_ind_C_2); hbsr_val_C_gold.near_check(hbsr_val_C_1); hbsr_val_C_gold.near_check(hbsr_val_C_2); } if(arg.timing) { CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_host)); rocsparse_int nnzb_C; CHECK_ROCSPARSE_ERROR(rocsparse_bsrgeam_nnzb(handle, dir, Mb, Nb, block_dim, descrA, nnzb_A, dbsr_row_ptr_A, dbsr_col_ind_A, descrB, nnzb_B, dbsr_row_ptr_B, dbsr_col_ind_B, descrC, dbsr_row_ptr_C_1, &nnzb_C)); device_vector dbsr_col_ind_C(nnzb_C); device_vector dbsr_val_C(block_dim * block_dim * nnzb_C); const double gpu_analysis_time_used = rocsparse_clients::run_benchmark(arg, rocsparse_bsrgeam_nnzb, handle, dir, Mb, Nb, block_dim, descrA, nnzb_A, dbsr_row_ptr_A, dbsr_col_ind_A, descrB, nnzb_B, dbsr_row_ptr_B, dbsr_col_ind_B, descrC, dbsr_row_ptr_C_1, &nnzb_C); const double gpu_solve_time_used = rocsparse_clients::run_benchmark(arg, rocsparse_bsrgeam, handle, dir, Mb, Nb, block_dim, h_alpha, descrA, nnzb_A, dbsr_val_A, dbsr_row_ptr_A, dbsr_col_ind_A, h_beta, descrB, nnzb_B, dbsr_val_B, dbsr_row_ptr_B, dbsr_col_ind_B, descrC, dbsr_val_C, dbsr_row_ptr_C_1, dbsr_col_ind_C); double gflop_count = bsrgeam_gflop_count(block_dim, nnzb_A, nnzb_B, nnzb_C, h_alpha, h_beta); double gbyte_count = bsrgeam_gbyte_count(Mb, block_dim, nnzb_A, nnzb_B, nnzb_C, h_alpha, h_beta); double gpu_gflops = get_gpu_gflops(gpu_solve_time_used, gflop_count); double gpu_gbyte = get_gpu_gbyte(gpu_solve_time_used, gbyte_count); display_timing_info(display_key_t::Mb, Mb, display_key_t::Nb, Nb, display_key_t::bdim, block_dim, display_key_t::nnzb_A, nnzb_A, display_key_t::nnzb_B, nnzb_B, display_key_t::nnzb_C, nnzb_C, display_key_t::alpha, *h_alpha, display_key_t::beta, *h_beta, display_key_t::gflops, gpu_gflops, display_key_t::bandwidth, gpu_gbyte, display_key_t::analysis_time_ms, get_gpu_time_msec(gpu_analysis_time_used), display_key_t::time_ms, get_gpu_time_msec(gpu_solve_time_used)); } } #define INSTANTIATE(TYPE) \ template void testing_bsrgeam_bad_arg(const Arguments& arg); \ template void testing_bsrgeam(const Arguments& arg) INSTANTIATE(float); INSTANTIATE(double); INSTANTIATE(rocsparse_float_complex); INSTANTIATE(rocsparse_double_complex); void testing_bsrgeam_extra(const Arguments& arg) {}