From 919bba3eaf0feaa83e787aa500d6f0d5169b02b5 Mon Sep 17 00:00:00 2001 From: CNugteren Date: Fri, 10 Jul 2015 07:19:59 +0200 Subject: Added the HERK routine, tester, and client --- test/performance/routines/xherk.cc | 40 ++++++++++++ test/routines/xherk.h | 121 +++++++++++++++++++++++++++++++++++++ test/wrapper_clblas.h | 28 +++++++++ 3 files changed, 189 insertions(+) create mode 100644 test/performance/routines/xherk.cc create mode 100644 test/routines/xherk.h (limited to 'test') diff --git a/test/performance/routines/xherk.cc b/test/performance/routines/xherk.cc new file mode 100644 index 00000000..ce18152e --- /dev/null +++ b/test/performance/routines/xherk.cc @@ -0,0 +1,40 @@ + +// ================================================================================================= +// 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 the Xherk command-line interface performance tester. +// +// ================================================================================================= + +#include "performance/client.h" +#include "routines/xherk.h" + +// ================================================================================================= + +// Shortcuts to the clblast namespace +using float2 = clblast::float2; +using double2 = clblast::double2; + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + switch(clblast::GetPrecision(argc, argv)) { + case clblast::Precision::kHalf: + throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kSingle: + throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kDouble: + throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kComplexSingle: + clblast::RunClient, float2, float>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient, double2, double>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/xherk.h b/test/routines/xherk.h new file mode 100644 index 00000000..780b9b52 --- /dev/null +++ b/test/routines/xherk.h @@ -0,0 +1,121 @@ + +// ================================================================================================= +// 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 Xherk 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_XHERK_H_ +#define CLBLAST_TEST_ROUTINES_XHERK_H_ + +#include +#include + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXherk { + public: + + // The list of arguments relevant for this routine + static std::vector GetOptions() { + return {kArgN, kArgK, + kArgLayout, kArgTriangle, kArgATransp, + kArgALeadDim, kArgCLeadDim, + kArgAOffset, 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 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.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 &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments &args) { return args.n; } + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments &args, const Buffers &buffers, + CommandQueue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Herk(args.layout, args.triangle, args.a_transpose, + args.n, args.k, args.alpha, + buffers.a_mat(), args.a_offset, args.a_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) + static StatusCode RunReference(const Arguments &args, const Buffers &buffers, + CommandQueue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXherk(static_cast(args.layout), + static_cast(args.triangle), + static_cast(args.a_transpose), + args.n, args.k, args.alpha, + buffers.a_mat(), args.a_offset, args.a_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); + } + + // Describes how to download the results of the computation (more importantly: which buffer) + static std::vector DownloadResult(const Arguments &args, Buffers &buffers, + CommandQueue &queue) { + std::vector result(args.c_size, static_cast(0)); + buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), 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 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_XHERK_H_ +#endif diff --git a/test/wrapper_clblas.h b/test/wrapper_clblas.h index a03ade5e..360adfc8 100644 --- a/test/wrapper_clblas.h +++ b/test/wrapper_clblas.h @@ -325,6 +325,34 @@ clblasStatus clblasXsyrk( num_queues, queues, num_wait_events, wait_events, events); } +// This calls {clblasCherk, clblasZherk} with the arguments forwarded. +clblasStatus clblasXherk( + clblasOrder layout, clblasUplo triangle, clblasTranspose a_transpose, + size_t n, size_t k, float alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, float beta, + const cl_mem c_mat, size_t c_offset, size_t c_ld, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + return clblasCherk(layout, triangle, a_transpose, + n, k, alpha, + a_mat, a_offset, a_ld, beta, + c_mat, c_offset, c_ld, + num_queues, queues, num_wait_events, wait_events, events); +} +clblasStatus clblasXherk( + clblasOrder layout, clblasUplo triangle, clblasTranspose a_transpose, + size_t n, size_t k, double alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, double beta, + const cl_mem c_mat, size_t c_offset, size_t c_ld, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + return clblasZherk(layout, triangle, a_transpose, + n, k, alpha, + a_mat, a_offset, a_ld, beta, + c_mat, c_offset, c_ld, + num_queues, queues, num_wait_events, wait_events, events); +} + // This calls {clblasSsyr2k, clblasDsyr2k, clblasCsyr2k, clblasZsyr2k} with the arguments forwarded. clblasStatus clblasXsyr2k( clblasOrder layout, clblasUplo triangle, clblasTranspose ab_transpose, -- cgit v1.2.3