/* ************************************************************************ * Copyright (C) 2021-2026 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. * * ************************************************************************ */ #pragma once #ifndef TESTING_SPMM_BELL_HPP #define TESTING_SPMM_BELL_HPP #include "display.hpp" #include "flops.hpp" #include "gbyte.hpp" #include "hipsparse.hpp" #include "hipsparse_arguments.hpp" #include "hipsparse_graph.hpp" #include "hipsparse_test_unique_ptr.hpp" #include "unit.hpp" #include "utility.hpp" #include #include #include using namespace hipsparse; using namespace hipsparse_test; template void testing_spmm_bell_bad_arg(const Arguments& argus) { #if(!defined(CUDART_VERSION)) int32_t n = 100; int32_t m = 100; int32_t k = 100; int32_t ellBlockSize = 2; int32_t ell_cols = 10; int32_t safe_size = 100; float alpha = 0.6; float beta = 0.2; hipsparseOperation_t transA = HIPSPARSE_OPERATION_NON_TRANSPOSE; hipsparseOperation_t transB = HIPSPARSE_OPERATION_NON_TRANSPOSE; hipsparseOrder_t order = HIPSPARSE_ORDER_COL; hipsparseIndexBase_t idxBase = HIPSPARSE_INDEX_BASE_ZERO; hipsparseIndexType_t idxType = HIPSPARSE_INDEX_32I; hipDataType dataType = HIP_R_32F; hipsparseSpMMAlg_t alg = HIPSPARSE_SPMM_BLOCKED_ELL_ALG1; std::unique_ptr unique_ptr_handle(new handle_struct); hipsparseHandle_t handle = unique_ptr_handle->handle; auto dind_managed = hipsparse_unique_ptr{device_malloc(sizeof(int32_t) * safe_size), device_free}; auto dval_managed = hipsparse_unique_ptr{device_malloc(sizeof(float) * safe_size), device_free}; auto dB_managed = hipsparse_unique_ptr{device_malloc(sizeof(float) * safe_size), device_free}; auto dC_managed = hipsparse_unique_ptr{device_malloc(sizeof(float) * safe_size), device_free}; auto dbuf_managed = hipsparse_unique_ptr{device_malloc(sizeof(char) * safe_size), device_free}; int32_t* dind = (int32_t*)dind_managed.get(); float* dval = (float*)dval_managed.get(); float* dB = (float*)dB_managed.get(); float* dC = (float*)dC_managed.get(); void* dbuf = (void*)dbuf_managed.get(); // SpMM structures hipsparseSpMatDescr_t matA; hipsparseDnMatDescr_t matB, matC; size_t bsize; // Create SpMM structures verify_hipsparse_status_success( hipsparseCreateBlockedEll( &matA, m, k, ellBlockSize, ell_cols, dind, dval, idxType, idxBase, dataType), "success"); verify_hipsparse_status_success(hipsparseCreateDnMat(&matB, k, n, k, dB, dataType, order), "success"); verify_hipsparse_status_success(hipsparseCreateDnMat(&matC, m, n, m, dC, dataType, order), "success"); // SpMM buffer verify_hipsparse_status_invalid_handle(hipsparseSpMM_bufferSize( nullptr, transA, transB, &alpha, matA, matB, &beta, matC, dataType, alg, &bsize)); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_bufferSize( handle, transA, transB, nullptr, matA, matB, &beta, matC, dataType, alg, &bsize), "Error: alpha is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_bufferSize( handle, transA, transB, &alpha, nullptr, matB, &beta, matC, dataType, alg, &bsize), "Error: matA is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_bufferSize( handle, transA, transB, &alpha, matA, nullptr, &beta, matC, dataType, alg, &bsize), "Error: matB is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_bufferSize( handle, transA, transB, &alpha, matA, matB, nullptr, matC, dataType, alg, &bsize), "Error: beta is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_bufferSize( handle, transA, transB, &alpha, matA, matB, &beta, nullptr, dataType, alg, &bsize), "Error: matC is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_bufferSize( handle, transA, transB, &alpha, matA, matB, &beta, matC, dataType, alg, nullptr), "Error: bsize is nullptr"); // SpMM_preprocess verify_hipsparse_status_invalid_handle(hipsparseSpMM_preprocess( nullptr, transA, transB, &alpha, matA, matB, &beta, matC, dataType, alg, dbuf)); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_preprocess( handle, transA, transB, nullptr, matA, matB, &beta, matC, dataType, alg, dbuf), "Error: alpha is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_preprocess( handle, transA, transB, &alpha, nullptr, matB, &beta, matC, dataType, alg, dbuf), "Error: matA is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_preprocess( handle, transA, transB, &alpha, matA, nullptr, &beta, matC, dataType, alg, dbuf), "Error: matB is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_preprocess( handle, transA, transB, &alpha, matA, matB, nullptr, matC, dataType, alg, dbuf), "Error: beta is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_preprocess( handle, transA, transB, &alpha, matA, matB, &beta, nullptr, dataType, alg, dbuf), "Error: matC is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM_preprocess( handle, transA, transB, &alpha, matA, matB, &beta, nullptr, dataType, alg, nullptr), "Error: dbuf is nullptr"); // SpMM verify_hipsparse_status_invalid_handle(hipsparseSpMM( nullptr, transA, transB, &alpha, matA, matB, &beta, matC, dataType, alg, dbuf)); verify_hipsparse_status_invalid_pointer( hipsparseSpMM( handle, transA, transB, nullptr, matA, matB, &beta, matC, dataType, alg, dbuf), "Error: alpha is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM( handle, transA, transB, &alpha, nullptr, matB, &beta, matC, dataType, alg, dbuf), "Error: matA is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM( handle, transA, transB, &alpha, matA, nullptr, &beta, matC, dataType, alg, dbuf), "Error: matB is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM( handle, transA, transB, &alpha, matA, matB, nullptr, matC, dataType, alg, dbuf), "Error: beta is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM( handle, transA, transB, &alpha, matA, matB, &beta, nullptr, dataType, alg, dbuf), "Error: matC is nullptr"); verify_hipsparse_status_invalid_pointer( hipsparseSpMM( handle, transA, transB, &alpha, matA, matB, &beta, nullptr, dataType, alg, nullptr), "Error: dbuf is nullptr"); // Destruct verify_hipsparse_status_success(hipsparseDestroySpMat(matA), "success"); verify_hipsparse_status_success(hipsparseDestroyDnMat(matB), "success"); verify_hipsparse_status_success(hipsparseDestroyDnMat(matC), "success"); #endif } template void testing_spmm_bell(Arguments argus) { #if(!defined(CUDART_VERSION) || CUDART_VERSION >= 11021) I m = argus.M; I n = argus.N; I k = argus.K; I blockDim = argus.block_dim; T h_alpha = argus.get_alpha(); T h_beta = argus.get_beta(); hipsparseOperation_t transA = argus.transA; hipsparseOperation_t transB = argus.transB; hipsparseOrder_t orderB = argus.orderB; hipsparseOrder_t orderC = argus.orderC; hipsparseIndexBase_t idxBase = argus.baseA; hipsparseSpMMAlg_t alg = argus.spmm_alg; std::string filename = argus.filename; I mb = (m + blockDim - 1) / blockDim; I kb = (k + blockDim - 1) / blockDim; I ellBlockSize = blockDim; // Index and data type hipsparseIndexType_t typeI = getIndexType(); hipDataType typeT = getDataType(); // hipSPARSE handle hipsparseLocalHandle_t handle(argus); // Generate the Blocked-ELL matrix. srand(12345ULL); I ellCols = 0; std::vector hbellColInd; std::vector hbellVal; CHECK_GENERATE_MATRIX_ERROR( generate_bell_matrix(filename, (transA == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? mb : kb, (transA == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? kb : mb, ellCols, blockDim, hbellColInd, hbellVal, idxBase)); m = mb * blockDim; k = kb * blockDim; // Some matrix properties I A_mb = (transA == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? mb : kb; I A_nb = (transA == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? kb : mb; I B_m = (transB == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? k : n; I B_n = (transB == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? n : k; I C_m = m; I C_n = n; int64_t ldb = (orderB == HIPSPARSE_ORDER_COL) ? ((transB == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? k : n) : ((transB == HIPSPARSE_OPERATION_NON_TRANSPOSE) ? n : k); int64_t ldc = (orderC == HIPSPARSE_ORDER_COL) ? m : n; ldb = std::max(int64_t(1), ldb); ldc = std::max(int64_t(1), ldc); int64_t nrowB = (orderB == HIPSPARSE_ORDER_COL) ? ldb : B_m; int64_t ncolB = (orderB == HIPSPARSE_ORDER_COL) ? B_n : ldb; int64_t nrowC = (orderC == HIPSPARSE_ORDER_COL) ? ldc : C_m; int64_t ncolC = (orderC == HIPSPARSE_ORDER_COL) ? C_n : ldc; // ell_val has (A_mb * blockDim) * ellCols elements in row-major order. int64_t nnz_A = A_mb * blockDim * ellCols; int64_t nnz_B = nrowB * ncolB; int64_t nnz_C = nrowC * ncolC; // Allocate host memory for dense matrices. std::vector hB(nnz_B); std::vector hC_1(nnz_C); std::vector hC_2(nnz_C); std::vector hC_gold(nnz_C); hipsparseInit(hB, nnz_B, 1); hipsparseInit(hC_1, nnz_C, 1); hC_2 = hC_1; hC_gold = hC_1; // allocate memory on device auto dbell_col_ind_managed = hipsparse_unique_ptr{ device_malloc(sizeof(I) * A_mb * (ellCols / ellBlockSize)), device_free}; auto dbell_val_managed = hipsparse_unique_ptr{device_malloc(sizeof(T) * nnz_A), device_free}; auto dB_managed = hipsparse_unique_ptr{device_malloc(sizeof(T) * nnz_B), device_free}; auto dC_1_managed = hipsparse_unique_ptr{device_malloc(sizeof(T) * nnz_C), device_free}; auto dC_2_managed = hipsparse_unique_ptr{device_malloc(sizeof(T) * nnz_C), device_free}; auto d_alpha_managed = hipsparse_unique_ptr{device_malloc(sizeof(T)), device_free}; auto d_beta_managed = hipsparse_unique_ptr{device_malloc(sizeof(T)), device_free}; I* dbell_col_ind = (I*)dbell_col_ind_managed.get(); T* dbell_val = (T*)dbell_val_managed.get(); T* dB = (T*)dB_managed.get(); T* dC_1 = (T*)dC_1_managed.get(); T* dC_2 = (T*)dC_2_managed.get(); T* d_alpha = (T*)d_alpha_managed.get(); T* d_beta = (T*)d_beta_managed.get(); // Copy data from CPU to device. CHECK_HIP_ERROR(hipMemcpy(dbell_col_ind, hbellColInd.data(), sizeof(I) * A_mb * (ellCols / ellBlockSize), hipMemcpyHostToDevice)); CHECK_HIP_ERROR( hipMemcpy(dbell_val, hbellVal.data(), sizeof(T) * nnz_A, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dB, hB.data(), sizeof(T) * nnz_B, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dC_1, hC_1.data(), sizeof(T) * nnz_C, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(dC_2, hC_2.data(), sizeof(T) * nnz_C, hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(d_alpha, &h_alpha, sizeof(T), hipMemcpyHostToDevice)); CHECK_HIP_ERROR(hipMemcpy(d_beta, &h_beta, sizeof(T), hipMemcpyHostToDevice)); // Create matrices. hipsparseSpMatDescr_t matA; CHECK_HIPSPARSE_ERROR(hipsparseCreateBlockedEll( &matA, m, k, ellBlockSize, ellCols, dbell_col_ind, dbell_val, typeI, idxBase, typeT)); // Create dense matrices. hipsparseDnMatDescr_t matB, matC1, matC2; CHECK_HIPSPARSE_ERROR(hipsparseCreateDnMat(&matB, B_m, B_n, ldb, dB, typeT, orderB)); CHECK_HIPSPARSE_ERROR(hipsparseCreateDnMat(&matC1, C_m, C_n, ldc, dC_1, typeT, orderC)); CHECK_HIPSPARSE_ERROR(hipsparseCreateDnMat(&matC2, C_m, C_n, ldc, dC_2, typeT, orderC)); // Query SpMM buffer size. size_t bufferSize; CHECK_HIPSPARSE_ERROR(hipsparseSpMM_bufferSize( handle, transA, transB, &h_alpha, matA, matB, &h_beta, matC1, typeT, alg, &bufferSize)); void* dbuffer; CHECK_HIP_ERROR(hipMalloc(&dbuffer, bufferSize)); if(argus.call_preprocess) { #if(!defined(CUDART_VERSION) || CUDART_VERSION >= 11021) // Host pointer mode. CHECK_HIPSPARSE_ERROR(hipsparseSetPointerMode(handle, HIPSPARSE_POINTER_MODE_HOST)); CHECK_HIPSPARSE_ERROR(hipsparseSpMM_preprocess( handle, transA, transB, &h_alpha, matA, matB, &h_beta, matC1, typeT, alg, dbuffer)); #endif #if(!defined(CUDART_VERSION)) // Device pointer mode. CHECK_HIPSPARSE_ERROR(hipsparseSetPointerMode(handle, HIPSPARSE_POINTER_MODE_DEVICE)); CHECK_HIPSPARSE_ERROR(hipsparseSpMM_preprocess( handle, transA, transB, d_alpha, matA, matB, d_beta, matC2, typeT, alg, dbuffer)); #endif } if(argus.unit_check) { // Host pointer mode. CHECK_HIPSPARSE_ERROR(hipsparseSetPointerMode(handle, HIPSPARSE_POINTER_MODE_HOST)); CHECK_HIPSPARSE_ERROR(hipsparseSpMM( handle, transA, transB, &h_alpha, matA, matB, &h_beta, matC1, typeT, alg, dbuffer)); #if(!defined(CUDART_VERSION)) // Device pointer mode. CHECK_HIPSPARSE_ERROR(hipsparseSetPointerMode(handle, HIPSPARSE_POINTER_MODE_DEVICE)); CHECK_HIPSPARSE_ERROR(hipsparseSpMM( handle, transA, transB, d_alpha, matA, matB, d_beta, matC2, typeT, alg, dbuffer)); #endif host_bellmm(A_mb, n, A_nb, ellCols, ellBlockSize, transA, transB, h_alpha, hbellColInd, hbellVal, hB, ldb, orderB, h_beta, hC_gold, ldc, orderC, idxBase); CHECK_HIP_ERROR(hipMemcpy(hC_1.data(), dC_1, sizeof(T) * nnz_C, hipMemcpyDeviceToHost)); unit_check_near(1, nnz_C, 1, hC_gold.data(), hC_1.data()); #if(!defined(CUDART_VERSION)) CHECK_HIP_ERROR(hipMemcpy(hC_2.data(), dC_2, sizeof(T) * nnz_C, hipMemcpyDeviceToHost)); unit_check_near(1, nnz_C, 1, hC_gold.data(), hC_2.data()); #endif } if(argus.timing) { int number_cold_calls = 2; int number_hot_calls = argus.iters; CHECK_HIPSPARSE_ERROR(hipsparseSetPointerMode(handle, HIPSPARSE_POINTER_MODE_HOST)); // Warm up for(int iter = 0; iter < number_cold_calls; ++iter) { CHECK_HIPSPARSE_ERROR(hipsparseSpMM( handle, transA, transB, &h_alpha, matA, matB, &h_beta, matC1, typeT, alg, dbuffer)); } double gpu_time_used = get_time_us(); // Performance run for(int iter = 0; iter < number_hot_calls; ++iter) { CHECK_HIPSPARSE_ERROR(hipsparseSpMM( handle, transA, transB, &h_alpha, matA, matB, &h_beta, matC1, typeT, alg, dbuffer)); } gpu_time_used = (get_time_us() - gpu_time_used) / number_hot_calls; double gflop_count = bellmm_gflop_count(static_cast(A_mb), static_cast(n), static_cast(ellCols), static_cast(blockDim), static_cast(C_m * C_n), h_beta != make_DataType(0)); double gbyte_count = bellmm_gbyte_count(static_cast(A_mb), static_cast(ellCols), static_cast(blockDim), static_cast(B_m * B_n), static_cast(C_m * C_n), h_beta != make_DataType(0)); double gpu_gflops = get_gpu_gflops(gpu_time_used, gflop_count); 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::block_dim, blockDim, display_key_t::transA, transA, display_key_t::transB, transB, display_key_t::alpha, h_alpha, display_key_t::beta, h_beta, display_key_t::algorithm, hipsparse_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(hipFree(dbuffer)); CHECK_HIPSPARSE_ERROR(hipsparseDestroySpMat(matA)); CHECK_HIPSPARSE_ERROR(hipsparseDestroyDnMat(matB)); CHECK_HIPSPARSE_ERROR(hipsparseDestroyDnMat(matC1)); CHECK_HIPSPARSE_ERROR(hipsparseDestroyDnMat(matC2)); #endif } #endif // TESTING_SPMM_BELL_HPP