From 306bf67660da4f1adacaedf9066925240abf4ea9 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Sun, 6 Mar 2016 15:48:11 +0100 Subject: Added preliminary support for xHPR2 and xSPR2 routines --- test/routines/level2/xhpr2.h | 128 +++++++++++++++++++++++++++++++++++++++++++ test/routines/level2/xspr2.h | 128 +++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 256 insertions(+) create mode 100644 test/routines/level2/xhpr2.h create mode 100644 test/routines/level2/xspr2.h (limited to 'test/routines/level2') diff --git a/test/routines/level2/xhpr2.h b/test/routines/level2/xhpr2.h new file mode 100644 index 00000000..68fbc76d --- /dev/null +++ b/test/routines/level2/xhpr2.h @@ -0,0 +1,128 @@ + +// ================================================================================================= +// 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 Xhpr2 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_XHPR2_H_ +#define CLBLAST_TEST_ROUTINES_XHPR2_H_ + +#include +#include + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXhpr2 { + 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 {kArgN, + kArgLayout, kArgTriangle, + kArgXInc, kArgYInc, + kArgAPOffset, kArgXOffset, kArgYOffset, + kArgAlpha}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments &args) { + return args.n * 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 GetSizeAP(const Arguments &args) { + return ((args.n*(args.n+1)) / 2) + args.ap_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args) { + args.ap_size = GetSizeAP(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, const Buffers &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Hpr2(args.layout, args.triangle, + args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, + buffers.ap_mat(), args.ap_offset, + &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, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXhpr2(static_cast(args.layout), + static_cast(args.triangle), + args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, + buffers.ap_mat(), args.ap_offset, + 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, Queue &queue) { + std::vector result(args.ap_size, static_cast(0)); + buffers.ap_mat.Read(queue, args.ap_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments &args) { return args.ap_size - args.ap_offset; } + static size_t ResultID2(const Arguments &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments &args, const size_t id1, const size_t) { + return id1 + args.ap_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &args) { + return 5 * ((args.n*(args.n+1)) / 2); + } + static size_t GetBytes(const Arguments &args) { + return ((args.n*(args.n+1)) + 2 * args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XHPR2_H_ +#endif diff --git a/test/routines/level2/xspr2.h b/test/routines/level2/xspr2.h new file mode 100644 index 00000000..43d66c9e --- /dev/null +++ b/test/routines/level2/xspr2.h @@ -0,0 +1,128 @@ + +// ================================================================================================= +// 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 Xspr2 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_XSPR2_H_ +#define CLBLAST_TEST_ROUTINES_XSPR2_H_ + +#include +#include + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXspr2 { + 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 {kArgN, + kArgLayout, kArgTriangle, + kArgXInc, kArgYInc, + kArgAPOffset, kArgXOffset, kArgYOffset, + kArgAlpha}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments &args) { + return args.n * 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 GetSizeAP(const Arguments &args) { + return ((args.n*(args.n+1)) / 2) + args.ap_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args) { + args.ap_size = GetSizeAP(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, const Buffers &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Spr2(args.layout, args.triangle, + args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, + buffers.ap_mat(), args.ap_offset, + &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, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXspr2(static_cast(args.layout), + static_cast(args.triangle), + args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, + buffers.ap_mat(), args.ap_offset, + 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, Queue &queue) { + std::vector result(args.ap_size, static_cast(0)); + buffers.ap_mat.Read(queue, args.ap_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments &args) { return args.ap_size - args.ap_offset; } + static size_t ResultID2(const Arguments &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments &args, const size_t id1, const size_t) { + return id1 + args.ap_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &args) { + return 5 * ((args.n*(args.n+1)) / 2); + } + static size_t GetBytes(const Arguments &args) { + return ((args.n*(args.n+1)) + 2 * args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XSPR2_H_ +#endif -- cgit v1.2.3