/* ************************************************************************ * Copyright (C) 2022-2026 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 "testing.hpp" #include template void testing_spmm_csc_bad_arg(const Arguments& arg) { static const size_t safe_size = 100; // Create rocsparse handle rocsparse_local_handle local_handle; rocsparse_handle handle = local_handle; J m = safe_size; J n = safe_size; J k = safe_size; I nnz = safe_size; const T* alpha = (const T*)0x4; const T* beta = (const T*)0x4; void* csc_val = (void*)0x4; void* csc_col_ptr = (void*)0x4; void* csc_row_ind = (void*)0x4; void* dense_B = (void*)0x4; void* dense_C = (void*)0x4; rocsparse_operation trans_A = rocsparse_operation_none; rocsparse_operation trans_B = rocsparse_operation_none; rocsparse_index_base base = rocsparse_index_base_zero; rocsparse_order order_B = rocsparse_order_column; rocsparse_order order_C = rocsparse_order_column; rocsparse_spmm_alg alg = rocsparse_spmm_alg_default; rocsparse_spmm_stage stage = rocsparse_spmm_stage_compute; rocsparse_indextype itype = get_indextype(); rocsparse_indextype jtype = get_indextype(); rocsparse_datatype atype = get_datatype(); rocsparse_datatype btype = get_datatype(); rocsparse_datatype ctype = get_datatype(); rocsparse_datatype compute_type = get_datatype(); // SpMM structures static const bool csc_format_required = true; rocsparse_local_spmat local_mat_A(m, k, nnz, csc_col_ptr, csc_row_ind, csc_val, itype, jtype, base, atype, csc_format_required); rocsparse_local_dnmat local_mat_B(k, n, k, dense_B, btype, order_B); rocsparse_local_dnmat local_mat_C(m, n, m, dense_C, ctype, order_C); rocsparse_spmat_descr mat_A = local_mat_A; rocsparse_dnmat_descr mat_B = local_mat_B; rocsparse_dnmat_descr mat_C = local_mat_C; int nargs_to_exclude = 2; const int args_to_exclude[2] = {11, 12}; #define PARAMS \ handle, trans_A, trans_B, alpha, mat_A, mat_B, beta, mat_C, compute_type, alg, stage, \ buffer_size, temp_buffer { size_t* buffer_size = (size_t*)0x4; void* temp_buffer = (void*)0x4; select_bad_arg_analysis(rocsparse_spmm, nargs_to_exclude, args_to_exclude, PARAMS); } { size_t* buffer_size = (size_t*)0x4; void* temp_buffer = nullptr; select_bad_arg_analysis(rocsparse_spmm, nargs_to_exclude, args_to_exclude, PARAMS); } { size_t* buffer_size = nullptr; void* temp_buffer = (void*)0x4; select_bad_arg_analysis(rocsparse_spmm, nargs_to_exclude, args_to_exclude, PARAMS); } { size_t* buffer_size = nullptr; void* temp_buffer = nullptr; select_bad_arg_analysis(rocsparse_spmm, nargs_to_exclude, args_to_exclude, PARAMS); } #undef PARAMS } template void testing_spmm_csc(const Arguments& arg) { J M = arg.M; J N = arg.N; J K = arg.K; rocsparse_operation trans_A = arg.transA; rocsparse_operation trans_B = arg.transB; rocsparse_index_base base = arg.baseA; rocsparse_spmm_alg alg = arg.spmm_alg; rocsparse_order order_B = arg.orderB; rocsparse_order order_C = arg.orderC; rocsparse_int ld_multiplier_B = arg.ld_multiplier_B; rocsparse_int ld_multiplier_C = arg.ld_multiplier_C; T halpha = arg.get_alpha(); T hbeta = arg.get_beta(); // Index and data type rocsparse_indextype itype = get_indextype(); rocsparse_indextype jtype = get_indextype(); rocsparse_datatype atype = get_datatype(); rocsparse_datatype btype = get_datatype(); rocsparse_datatype ctype = get_datatype(); rocsparse_datatype ttype = get_datatype(); // Create rocsparse handle rocsparse_local_handle handle(arg); // Allocate host memory for matrix host_vector hcsc_col_ptr; host_vector hcsc_row_ind; host_vector hcsc_val; // Allocate host memory for matrix rocsparse_matrix_factory matrix_factory(arg); I nnz_A; matrix_factory.init_csc(hcsc_col_ptr, hcsc_row_ind, hcsc_val, (trans_A == rocsparse_operation_none) ? M : K, (trans_A == rocsparse_operation_none) ? K : M, nnz_A, base); // Redefine values rocsparse_init_1d_array( hcsc_val, nnz_A, arg.convert_to_int, arg.rand_gen_min, arg.rand_gen_max); // Some matrix properties J A_m = (trans_A == rocsparse_operation_none) ? M : K; J A_n = (trans_A == rocsparse_operation_none) ? K : M; J B_m = (trans_B == rocsparse_operation_none) ? K : N; J B_n = (trans_B == rocsparse_operation_none) ? N : K; J C_m = M; J C_n = N; int64_t ldb = (order_B == rocsparse_order_column) ? ((trans_B == rocsparse_operation_none) ? (int64_t(ld_multiplier_B) * K) : (int64_t(ld_multiplier_B) * N)) : ((trans_B == rocsparse_operation_none) ? (int64_t(ld_multiplier_B) * N) : (int64_t(ld_multiplier_B) * K)); int64_t ldc = (order_C == rocsparse_order_column) ? (int64_t(ld_multiplier_C) * M) : (int64_t(ld_multiplier_C) * N); int64_t nrowB = (order_B == rocsparse_order_column) ? ldb : B_m; int64_t ncolB = (order_B == rocsparse_order_column) ? B_n : ldb; int64_t nrowC = (order_C == rocsparse_order_column) ? ldc : C_m; int64_t ncolC = (order_C == rocsparse_order_column) ? C_n : ldc; int64_t nnz_B = nrowB * ncolB; int64_t nnz_C = nrowC * ncolC; // Allocate host memory for vectors host_vector hB(nnz_B); host_vector hC_1(nnz_C); host_vector hC_2(nnz_C); host_vector hC_gold(nnz_C); // Initialize data on CPU rocsparse_init_1d_array(hB, nnz_B, arg.convert_to_int, arg.rand_gen_min, arg.rand_gen_max); rocsparse_init_1d_array(hC_1, nnz_C, arg.convert_to_int, arg.rand_gen_min, arg.rand_gen_max); hC_2 = hC_1; hC_gold = hC_1; // Allocate device memory device_vector dcsc_col_ptr(A_n + 1); device_vector dcsc_row_ind(nnz_A); device_vector dcsc_val(nnz_A); device_vector dB(nnz_B); device_vector dC_1(nnz_C); device_vector dC_2(nnz_C); device_vector dalpha(1); device_vector dbeta(1); // Copy data from CPU to device CHECK_HIP_ERROR( hipMemcpy(dcsc_col_ptr, hcsc_col_ptr.data(), sizeof(I) * (A_n + 1), hipMemcpyHostToDevice)); CHECK_HIP_ERROR( hipMemcpy(dcsc_row_ind, hcsc_row_ind.data(), sizeof(J) * nnz_A, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dcsc_val, hcsc_val.data(), sizeof(A) * nnz_A, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dB, hB, sizeof(B) * nnz_B, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dC_1, hC_1, sizeof(C) * nnz_C, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dC_2, hC_2, sizeof(C) * nnz_C, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dalpha, &halpha, sizeof(T), hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dbeta, &hbeta, sizeof(T), hipMemcpyHostToDevice)); // Create descriptors static const bool csc_format_required = true; rocsparse_local_spmat mat_A(A_m, A_n, nnz_A, dcsc_col_ptr, dcsc_row_ind, dcsc_val, itype, jtype, base, atype, csc_format_required); ldb = std::max(int64_t(1), ldb); ldc = std::max(int64_t(1), ldc); rocsparse_local_dnmat mat_B(B_m, B_n, ldb, dB, btype, order_B); rocsparse_local_dnmat mat_C1(C_m, C_n, ldc, dC_1, ctype, order_C); rocsparse_local_dnmat mat_C2(C_m, C_n, ldc, dC_2, ctype, order_C); // Query SpMM buffer size_t buffer_size; CHECK_ROCSPARSE_ERROR(rocsparse_spmm(handle, trans_A, trans_B, &halpha, mat_A, mat_B, &hbeta, mat_C1, ttype, alg, rocsparse_spmm_stage_buffer_size, &buffer_size, nullptr)); // Allocate buffer void* dbuffer; CHECK_HIP_ERROR(rocsparse_hipMalloc(&dbuffer, buffer_size)); CHECK_ROCSPARSE_ERROR(rocsparse_spmm(handle, trans_A, trans_B, &halpha, mat_A, mat_B, &hbeta, mat_C1, ttype, alg, rocsparse_spmm_stage_preprocess, &buffer_size, dbuffer)); if(arg.unit_check) { // SpMM // Pointer mode host CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_host)); CHECK_ROCSPARSE_ERROR(testing::rocsparse_spmm(handle, trans_A, trans_B, &halpha, mat_A, mat_B, &hbeta, mat_C1, ttype, alg, rocsparse_spmm_stage_compute, &buffer_size, dbuffer)); if(ROCSPARSE_REPRODUCIBILITY) { rocsparse_reproducibility::save("dC_1", dC_1); } // Pointer mode device CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_device)); CHECK_ROCSPARSE_ERROR(testing::rocsparse_spmm(handle, trans_A, trans_B, dalpha, mat_A, mat_B, dbeta, mat_C2, ttype, alg, rocsparse_spmm_stage_compute, &buffer_size, dbuffer)); if(ROCSPARSE_REPRODUCIBILITY) { rocsparse_reproducibility::save("dC_2", dC_2); } // Copy output to host CHECK_HIP_ERROR(hipMemcpy(hC_1, dC_1, sizeof(C) * nnz_C, hipMemcpyDeviceToHost)); CHECK_HIP_ERROR(hipMemcpy(hC_2, dC_2, sizeof(C) * nnz_C, hipMemcpyDeviceToHost)); // CPU cscmm host_cscmm(A_m, N, A_n, trans_A, trans_B, halpha, hcsc_col_ptr, hcsc_row_ind, hcsc_val, hB, ldb, order_B, hbeta, hC_gold, ldc, order_C, base); hC_gold.near_check(hC_1, get_near_check_tol(arg)); hC_gold.near_check(hC_2, get_near_check_tol(arg)); } if(arg.timing) { CHECK_ROCSPARSE_ERROR(rocsparse_set_pointer_mode(handle, rocsparse_pointer_mode_host)); const double gpu_time_used = rocsparse_clients::run_benchmark(arg, rocsparse_spmm, handle, trans_A, trans_B, &halpha, mat_A, mat_B, &hbeta, mat_C1, ttype, alg, rocsparse_spmm_stage_compute, &buffer_size, dbuffer); double gflop_count = spmm_gflop_count(N, nnz_A, (I)C_m * (I)C_n, hbeta != static_cast(0)); double gpu_gflops = get_gpu_gflops(gpu_time_used, gflop_count); double gbyte_count = cscmm_gbyte_count( A_n, nnz_A, (I)B_m * (I)B_n, (I)C_m * (I)C_n, hbeta != static_cast(0)); double gpu_gbyte = get_gpu_gbyte(gpu_time_used, gbyte_count); display_timing_info(display_key_t::M, M, display_key_t::N, N, display_key_t::K, K, display_key_t::nnz_A, nnz_A, display_key_t::alpha, halpha, display_key_t::beta, hbeta, display_key_t::algorithm, rocsparse_spmmalg2string(alg), display_key_t::gflops, gpu_gflops, display_key_t::bandwidth, gpu_gbyte, display_key_t::time_ms, get_gpu_time_msec(gpu_time_used)); } CHECK_HIP_ERROR(rocsparse_hipFree(dbuffer)); } #define INSTANTIATE(ITYPE, JTYPE, TTYPE) \ template void testing_spmm_csc_bad_arg( \ const Arguments& arg); \ template void testing_spmm_csc(const Arguments& arg) #define INSTANTIATE_MIXED(ITYPE, JTYPE, ATYPE, XTYPE, YTYPE, TTYPE) \ template void testing_spmm_csc_bad_arg( \ const Arguments& arg); \ template void testing_spmm_csc(const Arguments& arg) INSTANTIATE(int32_t, int32_t, float); INSTANTIATE(int32_t, int32_t, double); INSTANTIATE(int32_t, int32_t, rocsparse_float_complex); INSTANTIATE(int32_t, int32_t, rocsparse_double_complex); INSTANTIATE(int64_t, int32_t, float); INSTANTIATE(int64_t, int32_t, double); INSTANTIATE(int64_t, int32_t, rocsparse_float_complex); INSTANTIATE(int64_t, int32_t, rocsparse_double_complex); INSTANTIATE(int64_t, int64_t, float); INSTANTIATE(int64_t, int64_t, double); INSTANTIATE(int64_t, int64_t, rocsparse_float_complex); INSTANTIATE(int64_t, int64_t, rocsparse_double_complex); INSTANTIATE_MIXED(int32_t, int32_t, int8_t, int8_t, int32_t, int32_t); INSTANTIATE_MIXED(int64_t, int32_t, int8_t, int8_t, int32_t, int32_t); INSTANTIATE_MIXED(int64_t, int64_t, int8_t, int8_t, int32_t, int32_t); INSTANTIATE_MIXED(int32_t, int32_t, int8_t, int8_t, float, float); INSTANTIATE_MIXED(int64_t, int32_t, int8_t, int8_t, float, float); INSTANTIATE_MIXED(int64_t, int64_t, int8_t, int8_t, float, float); INSTANTIATE_MIXED(int32_t, int32_t, _Float16, _Float16, float, float); INSTANTIATE_MIXED(int64_t, int32_t, _Float16, _Float16, float, float); INSTANTIATE_MIXED(int64_t, int64_t, _Float16, _Float16, float, float); INSTANTIATE_MIXED(int32_t, int32_t, rocsparse_bfloat16, rocsparse_bfloat16, float, float); INSTANTIATE_MIXED(int64_t, int32_t, rocsparse_bfloat16, rocsparse_bfloat16, float, float); INSTANTIATE_MIXED(int64_t, int64_t, rocsparse_bfloat16, rocsparse_bfloat16, float, float); INSTANTIATE_MIXED(int32_t, int32_t, _Float16, _Float16, _Float16, float); INSTANTIATE_MIXED(int64_t, int32_t, _Float16, _Float16, _Float16, float); INSTANTIATE_MIXED(int64_t, int64_t, _Float16, _Float16, _Float16, float); INSTANTIATE_MIXED( int32_t, int32_t, rocsparse_bfloat16, rocsparse_bfloat16, rocsparse_bfloat16, float); INSTANTIATE_MIXED( int64_t, int32_t, rocsparse_bfloat16, rocsparse_bfloat16, rocsparse_bfloat16, float); INSTANTIATE_MIXED( int64_t, int64_t, rocsparse_bfloat16, rocsparse_bfloat16, rocsparse_bfloat16, float); void testing_spmm_csc_extra(const Arguments& arg) { // Validate the CSC SpMM default-algorithm auto-selection end to end. // // This is the CSC counterpart of testing_spmm_csr_extra. A non-transposed // CSC multiply is delegated to csrmm as a transposed multiply and vice // versa, and the load-balanced kernels only apply to the non-transposed // csrmm path. For CSC that is the transposed SpMM operation, so the test // drives trans_A = transpose; the structural skew that matters is then the // longest CSC column, whose column-pointer array acts as the implicit CSR // row pointer feeding the line-nnz profile and the selection it drives. // // The kernel the default resolves to is deliberately not observable through // the public API (buffer_size is sized conservatively for the largest // auto-selectable kernel and every kernel computes the same product), so // the test runs the full buffer_size -> preprocess -> compute pipeline and // checks that the default reproduces the analytic product in both asymptotic // skew regimes and agrees with the explicit row-split and non-zero-split // kernels. // // The two synthetic matrices below sit at the asymptotic ends of the // dimensionless gate (longest-column share -> 1 vs -> 0), so the exercised // decision is independent of the device compute-unit count, and therefore // of the architecture. rocsparse_local_handle local_handle; rocsparse_handle handle = local_handle; const rocsparse_datatype ttype = get_datatype(); const rocsparse_indextype itype = get_indextype(); const rocsparse_indextype jtype = get_indextype(); const rocsparse_index_base base = rocsparse_index_base_zero; const rocsparse_order order = rocsparse_order_column; const double alpha = 1.0; const double beta = 0.0; const int32_t n = 2; // dense columns of B and C // Run the full three-stage CSC SpMM for one structure and algorithm and copy // the dense result C back to the host. Returns through an out-parameter // because the CHECK_*_ERROR macros expand to GoogleTest ASSERT_*, which may // only be used in a void-returning function. With trans_A = transpose and // every matrix and B entry equal to one (alpha = 1, beta = 0), C[i, j] is // exactly the number of non-zeros in column i, an integer represented // exactly in double. auto run_spmm = [&](const host_vector& hcol_ptr, const host_vector& hrow_ind, int32_t rows, rocsparse_spmm_alg alg, host_vector& hC) { const int32_t cols = static_cast(hcol_ptr.size()) - 1; const int64_t nnz = static_cast(hrow_ind.size()); const host_vector hval(nnz, 1.0); const host_vector hB(int64_t(rows) * n, 1.0); device_vector dcol_ptr(cols + 1); device_vector drow_ind(nnz); device_vector dval(nnz); device_vector dB(int64_t(rows) * n); device_vector dC(int64_t(cols) * n); CHECK_HIP_ERROR(hipMemcpy( dcol_ptr, hcol_ptr.data(), sizeof(int32_t) * (cols + 1), hipMemcpyHostToDevice)); CHECK_HIP_ERROR( hipMemcpy(drow_ind, hrow_ind.data(), sizeof(int32_t) * nnz, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dval, hval.data(), sizeof(double) * nnz, hipMemcpyHostToDevice)); CHECK_HIP_ERROR( hipMemcpy(dB, hB.data(), sizeof(double) * int64_t(rows) * n, hipMemcpyHostToDevice)); static const bool csc_format_required = true; rocsparse_local_spmat mat_A(rows, cols, nnz, dcol_ptr, drow_ind, dval, itype, jtype, base, ttype, csc_format_required); // op(A) = A^T is cols x rows, so op(B) = B is rows x n and C is cols x n. rocsparse_local_dnmat mat_B(rows, n, rows, dB, ttype, order); rocsparse_local_dnmat mat_C(cols, n, cols, dC, ttype, order); size_t buffer_size = 0; CHECK_ROCSPARSE_ERROR(rocsparse_spmm(handle, rocsparse_operation_transpose, rocsparse_operation_none, &alpha, mat_A, mat_B, &beta, mat_C, ttype, alg, rocsparse_spmm_stage_buffer_size, &buffer_size, nullptr)); void* dbuffer = nullptr; CHECK_HIP_ERROR(rocsparse_hipMalloc(&dbuffer, buffer_size > 0 ? buffer_size : 1)); CHECK_ROCSPARSE_ERROR(rocsparse_spmm(handle, rocsparse_operation_transpose, rocsparse_operation_none, &alpha, mat_A, mat_B, &beta, mat_C, ttype, alg, rocsparse_spmm_stage_preprocess, &buffer_size, dbuffer)); CHECK_ROCSPARSE_ERROR(rocsparse_spmm(handle, rocsparse_operation_transpose, rocsparse_operation_none, &alpha, mat_A, mat_B, &beta, mat_C, ttype, alg, rocsparse_spmm_stage_compute, &buffer_size, dbuffer)); hC.resize(int64_t(cols) * n); CHECK_HIP_ERROR( hipMemcpy(hC.data(), dC, sizeof(double) * int64_t(cols) * n, hipMemcpyDeviceToHost)); CHECK_HIP_ERROR(rocsparse_hipFree(dbuffer)); }; // For one structure, check that the default reproduces the analytic product // and matches both explicit load-balanced kernels. auto check_structure = [&](const host_vector& hcol_ptr, const host_vector& hrow_ind, int32_t rows) { const int32_t cols = static_cast(hcol_ptr.size()) - 1; // Analytic reference C[i, j] = non-zeros in column i (column-major, ld = cols). host_vector hC_ref(int64_t(cols) * n); for(int32_t j = 0; j < n; ++j) { for(int32_t i = 0; i < cols; ++i) { hC_ref[int64_t(j) * cols + i] = static_cast(hcol_ptr[i + 1] - hcol_ptr[i]); } } host_vector hC_default, hC_row, hC_nnz; run_spmm(hcol_ptr, hrow_ind, rows, rocsparse_spmm_alg_default, hC_default); run_spmm(hcol_ptr, hrow_ind, rows, rocsparse_spmm_alg_csr_row_split, hC_row); run_spmm(hcol_ptr, hrow_ind, rows, rocsparse_spmm_alg_csr_nnz_split, hC_nnz); unit_check_general(cols, n, hC_ref.data(), cols, hC_default.data(), cols); unit_check_general(cols, n, hC_ref.data(), cols, hC_row.data(), cols); unit_check_general(cols, n, hC_ref.data(), cols, hC_nnz.data(), cols); }; // Extremely column-skewed matrix: one "hub" column holds almost all // non-zeros, the rest a single entry. Longest-column share -> 1, so the gate // is free to upgrade to the non-zero-split kernel; the default must still be // correct. { const int32_t hub_nnz = 65536; const int32_t cols = 8; const int32_t rows = hub_nnz; host_vector hcol_ptr(cols + 1, 0); host_vector hrow_ind; hrow_ind.reserve(hub_nnz + (cols - 1)); for(int32_t r = 0; r < hub_nnz; ++r) { hrow_ind.push_back(r); } hcol_ptr[1] = hub_nnz; for(int32_t c = 1; c < cols; ++c) { hrow_ind.push_back(0); hcol_ptr[c + 1] = hcol_ptr[c] + 1; } check_structure(hcol_ptr, hrow_ind, rows); } // Extremely uniform matrix: every column has the same small length. // Longest-column share -> 0, so the gate keeps the row-split default; the // default must stay correct. { const int32_t col_nnz = 4; const int32_t cols = 100000; const int32_t rows = col_nnz; host_vector hcol_ptr(cols + 1, 0); host_vector hrow_ind; hrow_ind.reserve(int64_t(cols) * col_nnz); for(int32_t c = 0; c < cols; ++c) { for(int32_t r = 0; r < col_nnz; ++r) { hrow_ind.push_back(r); } hcol_ptr[c + 1] = hcol_ptr[c] + col_nnz; } check_structure(hcol_ptr, hrow_ind, rows); } }