From f726fbdc9fef937fbe32222f0e66aac8d7e2678c Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Sat, 18 Jun 2016 20:20:13 +0200 Subject: Moved all headers into the source tree, changed headers to .hpp extension --- test/routines/level3/xher2k.hpp | 166 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 166 insertions(+) create mode 100644 test/routines/level3/xher2k.hpp (limited to 'test/routines/level3/xher2k.hpp') diff --git a/test/routines/level3/xher2k.hpp b/test/routines/level3/xher2k.hpp new file mode 100644 index 00000000..fd20bbb5 --- /dev/null +++ b/test/routines/level3/xher2k.hpp @@ -0,0 +1,166 @@ + +// ================================================================================================= +// 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 Xher2k 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_XHER2K_H_ +#define CLBLAST_TEST_ROUTINES_XHER2K_H_ + +#include +#include + +#ifdef CLBLAST_REF_CLBLAS + #include "test/wrapper_clblas.hpp" +#endif +#ifdef CLBLAST_REF_CBLAS + #include "test/wrapper_cblas.hpp" +#endif + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXher2k { + 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 {kArgN, kArgK, + kArgLayout, kArgTriangle, kArgATransp, + kArgALeadDim, kArgBLeadDim, kArgCLeadDim, + kArgAOffset, kArgBOffset, kArgCOffset, + kArgAlpha, kArgBeta}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeA(const Arguments &args) { + auto a_rotated = (args.layout == Layout::kColMajor && args.a_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.a_transpose == Transpose::kNo); + auto a_two = (a_rotated) ? args.n : args.k; + return a_two * args.a_ld + args.a_offset; + } + static size_t GetSizeB(const Arguments &args) { + auto b_rotated = (args.layout == Layout::kColMajor && args.a_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.a_transpose == Transpose::kNo); + auto b_two = (b_rotated) ? args.n : args.k; + return b_two * args.b_ld + args.b_offset; + } + static size_t GetSizeC(const Arguments &args) { + return args.n * args.c_ld + args.c_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args) { + args.a_size = GetSizeA(args); + args.b_size = GetSizeB(args); + args.c_size = GetSizeC(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments &args) { return args.k; } + static size_t DefaultLDB(const Arguments &args) { return args.k; } + static size_t DefaultLDC(const Arguments &args) { return args.n; } + + // Describes which transpose options are relevant for this routine + using Transposes = std::vector; + static Transposes GetATransposes(const Transposes &) { return {Transpose::kNo, Transpose::kConjugate}; } + static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments &args, Buffers &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto alpha2 = T{args.alpha, args.alpha}; + auto status = Her2k(args.layout, args.triangle, args.a_transpose, + args.n, args.k, alpha2, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.b_mat(), args.b_offset, args.b_ld, args.beta, + buffers.c_mat(), args.c_offset, args.c_ld, + &queue_plain, &event); + clWaitForEvents(1, &event); + 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 alpha2 = T{args.alpha, args.alpha}; + auto status = clblasXher2k(convertToCLBLAS(args.layout), + convertToCLBLAS(args.triangle), + convertToCLBLAS(args.a_transpose), + args.n, args.k, alpha2, + buffers.a_mat, args.a_offset, args.a_ld, + buffers.b_mat, args.b_offset, args.b_ld, args.beta, + buffers.c_mat, args.c_offset, args.c_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, Buffers &buffers, Queue &queue) { + std::vector a_mat_cpu(args.a_size, static_cast(0)); + std::vector b_mat_cpu(args.b_size, static_cast(0)); + std::vector c_mat_cpu(args.c_size, static_cast(0)); + buffers.a_mat.Read(queue, args.a_size, a_mat_cpu); + buffers.b_mat.Read(queue, args.b_size, b_mat_cpu); + buffers.c_mat.Read(queue, args.c_size, c_mat_cpu); + auto alpha2 = T{args.alpha, args.alpha}; + cblasXher2k(convertToCBLAS(args.layout), + convertToCBLAS(args.triangle), + convertToCBLAS(args.a_transpose), + args.n, args.k, alpha2, + a_mat_cpu, args.a_offset, args.a_ld, + b_mat_cpu, args.b_offset, args.b_ld, args.beta, + c_mat_cpu, args.c_offset, args.c_ld); + buffers.c_mat.Write(queue, args.c_size, c_mat_cpu); + return StatusCode::kSuccess; + } + #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.c_size, static_cast(0)); + buffers.c_mat.Read(queue, args.c_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments &args) { return args.n; } + 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 id1*args.c_ld + id2 + args.c_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &args) { + return 2 * args.n * args.n * args.k; + } + static size_t GetBytes(const Arguments &args) { + return (args.n*args.k + args.n*args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XHER2K_H_ +#endif -- cgit v1.2.3