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/level2/xgerc.hpp | 157 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 157 insertions(+) create mode 100644 test/routines/level2/xgerc.hpp (limited to 'test/routines/level2/xgerc.hpp') diff --git a/test/routines/level2/xgerc.hpp b/test/routines/level2/xgerc.hpp new file mode 100644 index 00000000..d50092cb --- /dev/null +++ b/test/routines/level2/xgerc.hpp @@ -0,0 +1,157 @@ + +// ================================================================================================= +// 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 Xgerc 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_XGERC_H_ +#define CLBLAST_TEST_ROUTINES_XGERC_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 TestXgerc { + public: + + // The BLAS level: 1, 2, or 3 + static size_t BLASLevel() { return 2; } + + // The list of arguments relevant for this routine + static std::vector GetOptions() { + return {kArgM, kArgN, + kArgLayout, + kArgALeadDim, kArgXInc, kArgYInc, + kArgAOffset, kArgXOffset, kArgYOffset, + kArgAlpha}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments &args) { + return args.m * args.x_inc + args.x_offset; + } + static size_t GetSizeY(const Arguments &args) { + return args.n * args.y_inc + args.y_offset; + } + static size_t GetSizeA(const Arguments &args) { + auto a_rotated = (args.layout == Layout::kRowMajor); + auto a_two = (a_rotated) ? args.m : args.n; + return a_two * args.a_ld + args.a_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args) { + args.a_size = GetSizeA(args); + args.x_size = GetSizeX(args); + args.y_size = GetSizeY(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments &args) { return args.n; } + static size_t DefaultLDB(const Arguments &) { return 1; } // N/A for this routine + 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 &) { return {}; } // N/A for this routine + 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 status = Gerc(args.layout, + args.m, args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, + buffers.a_mat(), args.a_offset, args.a_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 status = clblasXgerc(convertToCLBLAS(args.layout), + args.m, args.n, args.alpha, + buffers.x_vec, args.x_offset, args.x_inc, + buffers.y_vec, args.y_offset, args.y_inc, + buffers.a_mat, args.a_offset, args.a_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 x_vec_cpu(args.x_size, static_cast(0)); + std::vector y_vec_cpu(args.y_size, static_cast(0)); + buffers.a_mat.Read(queue, args.a_size, a_mat_cpu); + buffers.x_vec.Read(queue, args.x_size, x_vec_cpu); + buffers.y_vec.Read(queue, args.y_size, y_vec_cpu); + cblasXgerc(convertToCBLAS(args.layout), + args.m, args.n, args.alpha, + x_vec_cpu, args.x_offset, args.x_inc, + y_vec_cpu, args.y_offset, args.y_inc, + a_mat_cpu, args.a_offset, args.a_ld); + buffers.a_mat.Write(queue, args.a_size, a_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.a_size, static_cast(0)); + buffers.a_mat.Read(queue, args.a_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.a_ld + id2 + args.a_offset: + id2*args.a_ld + id1 + args.a_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &args) { + return 2 * args.m * args.n; + } + static size_t GetBytes(const Arguments &args) { + return (2*args.m*args.n + args.m + args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XGERC_H_ +#endif -- cgit v1.2.3