/* ************************************************************************ * Copyright (C) 2021-2024 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 cop- * ies 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 IM- * PLIED, 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 CONNE- * CTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. * * * ************************************************************************ */ #pragma once #include "clientcommon.hpp" template void potrs_checkBadArgs(const hipsolverHandle_t handle, const hipsolverFillMode_t uplo, const int n, const int nrhs, T dA, const int lda, const int stA, T dB, const int ldb, const int stB, V dWork, const int lwork, U dInfo, const int bc) { // handle EXPECT_ROCBLAS_STATUS( hipsolver_potrs( API, nullptr, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, dWork, lwork, dInfo, bc), HIPSOLVER_STATUS_NOT_INITIALIZED); // values EXPECT_ROCBLAS_STATUS(hipsolver_potrs(API, handle, hipsolverFillMode_t(-1), n, nrhs, dA, lda, stA, dB, ldb, stB, dWork, lwork, dInfo, bc), HIPSOLVER_STATUS_INVALID_ENUM); #if defined(__HIP_PLATFORM_HCC__) || defined(__HIP_PLATFORM_AMD__) // pointers EXPECT_ROCBLAS_STATUS(hipsolver_potrs(API, handle, uplo, n, nrhs, (T) nullptr, lda, stA, dB, ldb, stB, dWork, lwork, dInfo, bc), HIPSOLVER_STATUS_INVALID_VALUE); EXPECT_ROCBLAS_STATUS(hipsolver_potrs(API, handle, uplo, n, nrhs, dA, lda, stA, (T) nullptr, ldb, stB, dWork, lwork, dInfo, bc), HIPSOLVER_STATUS_INVALID_VALUE); EXPECT_ROCBLAS_STATUS( hipsolver_potrs( API, handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, dWork, lwork, (U) nullptr, bc), HIPSOLVER_STATUS_INVALID_VALUE); #endif } template void testing_potrs_bad_arg() { // safe arguments hipsolver_local_handle handle; int n = 1; int nrhs = 1; int lda = 1; int ldb = 1; int stA = 1; int stB = 1; int bc = 1; hipsolverFillMode_t uplo = HIPSOLVER_FILL_MODE_UPPER; if(BATCHED) { // memory allocations device_batch_vector dA(1, 1, 1); device_batch_vector dB(1, 1, 1); device_strided_batch_vector dInfo(1, 1, 1, 1); CHECK_HIP_ERROR(dA.memcheck()); CHECK_HIP_ERROR(dB.memcheck()); CHECK_HIP_ERROR(dInfo.memcheck()); int size_W; hipsolver_potrs_bufferSize( API, handle, uplo, n, nrhs, dA.data(), lda, dB.data(), ldb, &size_W, bc); device_strided_batch_vector dWork(size_W, 1, size_W, 1); if(size_W) CHECK_HIP_ERROR(dWork.memcheck()); // check bad arguments potrs_checkBadArgs(handle, uplo, n, nrhs, dA.data(), lda, stA, dB.data(), ldb, stB, dWork.data(), size_W, dInfo.data(), bc); } else { // memory allocations device_strided_batch_vector dA(1, 1, 1, 1); device_strided_batch_vector dB(1, 1, 1, 1); device_strided_batch_vector dInfo(1, 1, 1, 1); CHECK_HIP_ERROR(dA.memcheck()); CHECK_HIP_ERROR(dB.memcheck()); CHECK_HIP_ERROR(dInfo.memcheck()); int size_W; hipsolver_potrs_bufferSize( API, handle, uplo, n, nrhs, dA.data(), lda, dB.data(), ldb, &size_W, bc); device_strided_batch_vector dWork(size_W, 1, size_W, 1); if(size_W) CHECK_HIP_ERROR(dWork.memcheck()); // check bad arguments potrs_checkBadArgs(handle, uplo, n, nrhs, dA.data(), lda, stA, dB.data(), ldb, stB, dWork.data(), size_W, dInfo.data(), bc); } } template void potrs_initData(const hipsolverHandle_t handle, const hipsolverFillMode_t uplo, const int n, const int nrhs, Td& dA, const int lda, const int stA, Td& dB, const int ldb, const int stB, const int bc, Th& hA, Th& hB) { if(CPU) { rocblas_init(hA, true); rocblas_init(hB, true); int info; for(int b = 0; b < bc; ++b) { // scale to ensure positive definiteness for(int i = 0; i < n; i++) hA[b][i + i * lda] = hA[b][i + i * lda] * conj(hA[b][i + i * lda]) * 400; // do the Cholesky factorization of matrix A w/ the reference LAPACK routine cpu_potrf(uplo, n, hA[b], lda, &info); } } if(GPU) { // now copy matrices to the GPU CHECK_HIP_ERROR(dA.transfer_from(hA)); CHECK_HIP_ERROR(dB.transfer_from(hB)); } } template void potrs_getError(const hipsolverHandle_t handle, const hipsolverFillMode_t uplo, const int n, const int nrhs, Td& dA, const int lda, const int stA, Td& dB, const int ldb, const int stB, Vd& dWork, const int lwork, Ud& dInfo, const int bc, Th& hA, Th& hB, Th& hBRes, Uh& hInfo, Uh& hInfoRes, double* max_err) { // input data initialization potrs_initData(handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, bc, hA, hB); // execute computations // GPU lapack CHECK_ROCBLAS_ERROR(hipsolver_potrs(API, handle, uplo, n, nrhs, dA.data(), lda, stA, dB.data(), ldb, stB, dWork.data(), lwork, dInfo.data(), bc)); CHECK_HIP_ERROR(hBRes.transfer_from(dB)); CHECK_HIP_ERROR(hInfoRes.transfer_from(dInfo)); // CPU lapack for(int b = 0; b < bc; ++b) { cpu_potrs(uplo, n, nrhs, hA[b], lda, hB[b], ldb, hInfo[b]); } // error is ||hB - hBRes|| / ||hB|| // (THIS DOES NOT ACCOUNT FOR NUMERICAL REPRODUCIBILITY ISSUES. // IT MIGHT BE REVISITED IN THE FUTURE) // using vector-induced infinity norm double err; *max_err = 0; for(int b = 0; b < bc; ++b) { err = norm_error('I', n, nrhs, ldb, hB[b], hBRes[b]); *max_err = err > *max_err ? err : *max_err; } // check info err = 0; for(int b = 0; b < bc; ++b) { EXPECT_EQ(hInfo[b][0], hInfoRes[b][0]) << "where b = " << b; if(hInfo[b][0] != hInfoRes[b][0]) err++; } *max_err += err; } template void potrs_getPerfData(const hipsolverHandle_t handle, const hipsolverFillMode_t uplo, const int n, const int nrhs, Td& dA, const int lda, const int stA, Td& dB, const int ldb, const int stB, Vd& dWork, const int lwork, Ud& dInfo, const int bc, Th& hA, Th& hB, Uh& hInfo, double* gpu_time_used, double* cpu_time_used, const int hot_calls, const bool perf) { if(!perf) { potrs_initData( handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, bc, hA, hB); // cpu-lapack performance (only if not in perf mode) *cpu_time_used = get_time_us_no_sync(); for(int b = 0; b < bc; ++b) { cpu_potrs(uplo, n, nrhs, hA[b], lda, hB[b], ldb, hInfo[b]); } *cpu_time_used = get_time_us_no_sync() - *cpu_time_used; } potrs_initData(handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, bc, hA, hB); // cold calls for(int iter = 0; iter < 2; iter++) { potrs_initData( handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, bc, hA, hB); CHECK_ROCBLAS_ERROR(hipsolver_potrs(API, handle, uplo, n, nrhs, dA.data(), lda, stA, dB.data(), ldb, stB, dWork.data(), lwork, dInfo.data(), bc)); } // gpu-lapack performance hipStream_t stream; CHECK_ROCBLAS_ERROR(hipsolverGetStream(handle, &stream)); double start; for(int iter = 0; iter < hot_calls; iter++) { potrs_initData( handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, bc, hA, hB); start = get_time_us_sync(stream); hipsolver_potrs(API, handle, uplo, n, nrhs, dA.data(), lda, stA, dB.data(), ldb, stB, dWork.data(), lwork, dInfo.data(), bc); *gpu_time_used += get_time_us_sync(stream) - start; } *gpu_time_used /= hot_calls; } template void testing_potrs(Arguments& argus) { // get arguments hipsolver_local_handle handle; char uploC = argus.get("uplo"); int n = argus.get("n"); int nrhs = argus.get("nrhs", n); int lda = argus.get("lda", n); int ldb = argus.get("ldb", n); int stA = argus.get("strideA", lda * n); int stB = argus.get("strideB", ldb * nrhs); hipsolverFillMode_t uplo = char2hipsolver_fill(uploC); int bc = argus.batch_count; int hot_calls = argus.iters; int stBRes = (argus.unit_check || argus.norm_check) ? stB : 0; // check non-supported values if(uplo != HIPSOLVER_FILL_MODE_UPPER && uplo != HIPSOLVER_FILL_MODE_LOWER) { if(BATCHED) { EXPECT_ROCBLAS_STATUS(hipsolver_potrs(API, handle, uplo, n, nrhs, (T**)nullptr, lda, stA, (T**)nullptr, ldb, stB, (T*)nullptr, 0, (int*)nullptr, bc), HIPSOLVER_STATUS_INVALID_VALUE); } else { EXPECT_ROCBLAS_STATUS(hipsolver_potrs(API, handle, uplo, n, nrhs, (T*)nullptr, lda, stA, (T*)nullptr, ldb, stB, (T*)nullptr, 0, (int*)nullptr, bc), HIPSOLVER_STATUS_INVALID_VALUE); } if(argus.timing) rocsolver_bench_inform(inform_invalid_args); return; } // determine sizes size_t size_A = size_t(lda) * n; size_t size_B = size_t(ldb) * nrhs; double max_error = 0, gpu_time_used = 0, cpu_time_used = 0; size_t size_BRes = (argus.unit_check || argus.norm_check) ? size_B : 0; // check invalid sizes bool invalid_size = (n < 0 || nrhs < 0 || lda < n || ldb < n || bc < 0); if(invalid_size) { if(BATCHED) { EXPECT_ROCBLAS_STATUS(hipsolver_potrs(API, handle, uplo, n, nrhs, (T**)nullptr, lda, stA, (T**)nullptr, ldb, stB, (T*)nullptr, 0, (int*)nullptr, bc), HIPSOLVER_STATUS_INVALID_VALUE); } else { EXPECT_ROCBLAS_STATUS(hipsolver_potrs(API, handle, uplo, n, nrhs, (T*)nullptr, lda, stA, (T*)nullptr, ldb, stB, (T*)nullptr, 0, (int*)nullptr, bc), HIPSOLVER_STATUS_INVALID_VALUE); } if(argus.timing) rocsolver_bench_inform(inform_invalid_size); return; } // memory size query is necessary int size_W; if(BATCHED) hipsolver_potrs_bufferSize( API, handle, uplo, n, nrhs, (T**)nullptr, lda, (T**)nullptr, ldb, &size_W, bc); else hipsolver_potrs_bufferSize( API, handle, uplo, n, nrhs, (T*)nullptr, lda, (T*)nullptr, ldb, &size_W, bc); if(argus.mem_query) { rocsolver_bench_inform(inform_mem_query, size_W); return; } if(BATCHED) { // memory allocations host_batch_vector hA(size_A, 1, bc); host_batch_vector hB(size_B, 1, bc); host_batch_vector hBRes(size_BRes, 1, bc); host_strided_batch_vector hInfo(1, 1, 1, bc); host_strided_batch_vector hInfoRes(1, 1, 1, bc); device_batch_vector dA(size_A, 1, bc); device_batch_vector dB(size_B, 1, bc); device_strided_batch_vector dInfo(1, 1, 1, bc); device_strided_batch_vector dWork(size_W, 1, size_W, 1); // size_W accounts for bc if(size_A) CHECK_HIP_ERROR(dA.memcheck()); if(size_B) CHECK_HIP_ERROR(dB.memcheck()); CHECK_HIP_ERROR(dInfo.memcheck()); if(size_W) CHECK_HIP_ERROR(dWork.memcheck()); // check computations if(argus.unit_check || argus.norm_check) potrs_getError(handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, dWork, size_W, dInfo, bc, hA, hB, hBRes, hInfo, hInfoRes, &max_error); // collect performance data if(argus.timing) potrs_getPerfData(handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, dWork, size_W, dInfo, bc, hA, hB, hInfo, &gpu_time_used, &cpu_time_used, hot_calls, argus.perf); } else { // memory allocations host_strided_batch_vector hA(size_A, 1, stA, bc); host_strided_batch_vector hB(size_B, 1, stB, bc); host_strided_batch_vector hBRes(size_BRes, 1, stBRes, bc); host_strided_batch_vector hInfo(1, 1, 1, bc); host_strided_batch_vector hInfoRes(1, 1, 1, bc); device_strided_batch_vector dA(size_A, 1, stA, bc); device_strided_batch_vector dB(size_B, 1, stB, bc); device_strided_batch_vector dInfo(1, 1, 1, bc); device_strided_batch_vector dWork(size_W, 1, size_W, 1); // size_W accounts for bc if(size_A) CHECK_HIP_ERROR(dA.memcheck()); if(size_B) CHECK_HIP_ERROR(dB.memcheck()); CHECK_HIP_ERROR(dInfo.memcheck()); if(size_W) CHECK_HIP_ERROR(dWork.memcheck()); // check computations if(argus.unit_check || argus.norm_check) potrs_getError(handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, dWork, size_W, dInfo, bc, hA, hB, hBRes, hInfo, hInfoRes, &max_error); // collect performance data if(argus.timing) potrs_getPerfData(handle, uplo, n, nrhs, dA, lda, stA, dB, ldb, stB, dWork, size_W, dInfo, bc, hA, hB, hInfo, &gpu_time_used, &cpu_time_used, hot_calls, argus.perf); } // validate results for rocsolver-test // using m * machine_precision as tolerance if(argus.unit_check) ROCSOLVER_TEST_CHECK(T, max_error, n); // output results for rocsolver-bench if(argus.timing) { if(!argus.perf) { std::cerr << "\n============================================\n"; std::cerr << "Arguments:\n"; std::cerr << "============================================\n"; if(BATCHED) { rocsolver_bench_output("uplo", "n", "nrhs", "lda", "ldb", "batch_c"); rocsolver_bench_output(uploC, n, nrhs, lda, ldb, bc); } else if(STRIDED) { rocsolver_bench_output( "uplo", "n", "nrhs", "lda", "ldb", "strideA", "strideB", "batch_c"); rocsolver_bench_output(uploC, n, nrhs, lda, ldb, stA, stB, bc); } else { rocsolver_bench_output("uplo", "n", "nrhs", "lda", "ldb"); rocsolver_bench_output(uploC, n, nrhs, lda, ldb); } std::cerr << "\n============================================\n"; std::cerr << "Results:\n"; std::cerr << "============================================\n"; if(argus.norm_check) { rocsolver_bench_output("cpu_time", "gpu_time", "error"); rocsolver_bench_output(cpu_time_used, gpu_time_used, max_error); } else { rocsolver_bench_output("cpu_time", "gpu_time"); rocsolver_bench_output(cpu_time_used, gpu_time_used); } std::cerr << std::endl; } else { if(argus.norm_check) rocsolver_bench_output(gpu_time_used, max_error); else rocsolver_bench_output(gpu_time_used); } } // ensure all arguments were consumed argus.validate_consumed(); }