// ================================================================================================= // This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This // project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- // width of 100 characters per line. // // Author(s): // Cedric Nugteren // // This file implements a class with static methods to describe the Xtrsm routine. Examples of // such 'descriptions' are how to calculate the size a of buffer or how to run the routine. These // static methods are used by the correctness tester and the performance tester. // // ================================================================================================= #ifndef CLBLAST_TEST_ROUTINES_XTRSM_H_ #define CLBLAST_TEST_ROUTINES_XTRSM_H_ #include "test/routines/common.hpp" #include "test/routines/level3/xtrsm_data.hpp" namespace clblast { // ================================================================================================= // See comment at top of file for a description of the class template class TestXtrsm { public: // The BLAS level: 1, 2, or 3 static size_t BLASLevel() { return 3; } // The list of arguments relevant for this routine static std::vector GetOptions() { return {kArgM, kArgN, kArgLayout, kArgSide, kArgTriangle, kArgATransp, kArgDiagonal, kArgALeadDim, kArgBLeadDim, kArgAOffset, kArgBOffset, kArgAlpha}; } static std::vector BuffersIn() { return {kBufMatA, kBufMatB}; } static std::vector BuffersOut() { return {kBufMatB}; } // Describes how to obtain the sizes of the buffers static size_t GetSizeA(const Arguments &args) { const auto k = (args.side == Side::kLeft) ? args.m : args.n; return k * args.a_ld + args.a_offset; } static size_t GetSizeB(const Arguments &args) { const auto b_rotated = (args.layout == Layout::kRowMajor); const auto b_two = (b_rotated) ? args.m : args.n; return b_two * args.b_ld + args.b_offset; } // Describes how to set the sizes of all the buffers static void SetSizes(Arguments &args, Queue&) { args.a_size = GetSizeA(args); args.b_size = GetSizeB(args); } // Describes what the default values of the leading dimensions of the matrices are static size_t DefaultLDA(const Arguments &args) { return args.m; } static size_t DefaultLDB(const Arguments &args) { return args.n; } static size_t DefaultLDC(const Arguments &) { return 1; } // N/A for this routine // Describes which transpose options are relevant for this routine using Transposes = std::vector; static Transposes GetATransposes(const Transposes &all) { return all; } static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine // Describes how to prepare the input data static void PrepareData(const Arguments &args, Queue&, const int seed, std::vector&, std::vector&, std::vector& a_source_, std::vector& b_source_, std::vector&, std::vector&, std::vector&) { const auto k = (args.side == Side::kLeft) ? args.m : args.n; const auto b_one = (args.layout == Layout::kRowMajor) ? args.n : args.m; if (args.a_ld < k) { return; } if (args.b_ld < b_one) { return; } if (args.a_size <= 0 || args.b_size <= 0) { return; } // TODO: This is a copy of the clBLAS random matrix generation, make it work properly GenerateProperTrsmMatrices(args, seed, &a_source_[args.a_offset], &b_source_[args.b_offset]); } // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments &args, Buffers &buffers, Queue &queue) { #ifdef OPENCL_API auto queue_plain = queue(); auto event = cl_event{}; auto status = Trsm(args.layout, args.side, args.triangle, args.a_transpose, args.diagonal, args.m, args.n, args.alpha, buffers.a_mat(), args.a_offset, args.a_ld, buffers.b_mat(), args.b_offset, args.b_ld, &queue_plain, &event); if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } #elif CUDA_API auto status = Trsm(args.layout, args.side, args.triangle, args.a_transpose, args.diagonal, args.m, args.n, args.alpha, buffers.a_mat(), args.a_offset, args.a_ld, buffers.b_mat(), args.b_offset, args.b_ld, queue.GetContext()(), queue.GetDevice()()); cuStreamSynchronize(queue()); #endif return status; } // Describes how to run the clBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CLBLAS static StatusCode RunReference1(const Arguments &args, Buffers &buffers, Queue &queue) { auto queue_plain = queue(); auto event = cl_event{}; auto status = clblasXtrsm(convertToCLBLAS(args.layout), convertToCLBLAS(args.side), convertToCLBLAS(args.triangle), convertToCLBLAS(args.a_transpose), convertToCLBLAS(args.diagonal), args.m, args.n, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, buffers.b_mat, args.b_offset, args.b_ld, 1, &queue_plain, 0, nullptr, &event); clWaitForEvents(1, &event); return static_cast(status); } #endif // Describes how to run the CPU BLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CBLAS static StatusCode RunReference2(const Arguments &args, BuffersHost &buffers_host, Queue &) { cblasXtrsm(convertToCBLAS(args.layout), convertToCBLAS(args.side), convertToCBLAS(args.triangle), convertToCBLAS(args.a_transpose), convertToCBLAS(args.diagonal), args.m, args.n, args.alpha, buffers_host.a_mat, args.a_offset, args.a_ld, buffers_host.b_mat, args.b_offset, args.b_ld); return StatusCode::kSuccess; } #endif // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments &args, BuffersCUDA &buffers, Queue &) { auto status = cublasXtrsm(reinterpret_cast(args.cublas_handle), args.layout, convertToCUBLAS(args.side), convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), convertToCUBLAS(args.diagonal), args.m, args.n, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, buffers.b_mat, args.b_offset, args.b_ld); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } } #endif // Describes how to download the results of the computation (more importantly: which buffer) static std::vector DownloadResult(const Arguments &args, Buffers &buffers, Queue &queue) { std::vector result(args.b_size, static_cast(0)); buffers.b_mat.Read(queue, args.b_size, result); return result; } // Describes how to compute the indices of the result buffer static size_t ResultID1(const Arguments &args) { return args.m; } static size_t ResultID2(const Arguments &args) { return args.n; } static size_t GetResultIndex(const Arguments &args, const size_t id1, const size_t id2) { return (args.layout == Layout::kRowMajor) ? id1*args.b_ld + id2 + args.b_offset: id2*args.b_ld + id1 + args.b_offset; } // Describes how to compute performance metrics static size_t GetFlops(const Arguments &args) { auto k = (args.side == Side::kLeft) ? args.m : args.n; if((args.precision == Precision::kComplexSingle) || (args.precision == Precision::kComplexDouble)) { // complex flops return 4 * args.m * args.n * k; } else { // scalar flops return args.m * args.n * k; } } static size_t GetBytes(const Arguments &args) { auto k = (args.side == Side::kLeft) ? args.m : args.n; return (k*k + 2*args.m*args.n) * sizeof(T); } }; // ================================================================================================= } // namespace clblast // CLBLAST_TEST_ROUTINES_XTRSM_H_ #endif