From ef5008f5e46c4fe6d3728beff1d3277d02aae099 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Wed, 31 Jan 2018 20:41:02 +0100 Subject: Created the API and stubs for the HAD (hadamard-product) routines --- test/routines/levelx/xhad.hpp | 154 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 154 insertions(+) create mode 100644 test/routines/levelx/xhad.hpp (limited to 'test/routines/levelx') diff --git a/test/routines/levelx/xhad.hpp b/test/routines/levelx/xhad.hpp new file mode 100644 index 00000000..fc47a7d6 --- /dev/null +++ b/test/routines/levelx/xhad.hpp @@ -0,0 +1,154 @@ + +// ================================================================================================= +// 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 Xhad 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_XHAD_H_ +#define CLBLAST_TEST_ROUTINES_XHAD_H_ + +#include "test/routines/common.hpp" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXhad { +public: + + // The BLAS level: 4 for the extra routines + static size_t BLASLevel() { return 4; } + + // The list of arguments relevant for this routine + static std::vector GetOptions() { + return {kArgN, + kArgXInc, kArgYInc, + kArgXOffset, kArgYOffset, + kArgAlpha}; + } + static std::vector BuffersIn() { return {kBufVecX, kBufVecY, kBufMatC}; } + static std::vector BuffersOut() { return {kBufMatC}; } + + // 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 GetSizeC(const Arguments &args) { // used for 'vector z' + return args.n; // * args.z_inc + args.z_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args, Queue&) { + args.x_size = GetSizeX(args); + args.y_size = GetSizeY(args); + args.c_size = GetSizeC(args); // used for 'vector z' + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments &) { return 1; } // N/A for this routine + 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 prepare the input data + static void PrepareData(const Arguments&, Queue&, const int, std::vector&, + std::vector&, std::vector&, std::vector&, std::vector&, + std::vector&, std::vector&) {} // N/A for this routine + + // 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 = Had(args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, args.beta, + buffers.c_mat(), 0, 1, // used for 'vector z' + &queue_plain, &event); + if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } +#elif CUDA_API + auto status = Had(args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, args.beta, + buffers.c_mat(), 0, 1, // used for 'vector z' + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); +#endif + return status; + } + + // Describes how to run a naive version of the routine (for correctness/performance comparison). + // Note that a proper clBLAS or CPU BLAS comparison is not available for non-BLAS routines. + static StatusCode RunReference1(const Arguments &args, Buffers &buffers, Queue &queue) { + auto buffers_host = BuffersHost(); + DeviceToHost(args, buffers, buffers_host, queue, BuffersIn()); + const auto status = RunReference(args, buffers_host); + HostToDevice(args, buffers, buffers_host, queue, BuffersOut()); + return status; + } + + static StatusCode RunReference2(const Arguments &args, BuffersHost &buffers_host, Queue&) { + return RunReference(args, buffers_host); + } + static StatusCode RunReference3(const Arguments &, BuffersCUDA &, Queue &) { + return StatusCode::kUnknownError; + } + + // 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 &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments &args, const size_t id1, const size_t) { + return id1; // * args.z_inc + args.z_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &args) { + return 4 * args.n; + } + static size_t GetBytes(const Arguments &args) { + return (4 * args.n) * sizeof(T); + } +}; + +// ================================================================================================= + +template +StatusCode RunReference(const Arguments &args, BuffersHost &buffers_host) { + for (auto index = size_t{0}; index < args.n; ++index) { + const auto x = buffers_host.x_vec[index * args.x_inc + args.x_offset]; + const auto y = buffers_host.y_vec[index * args.y_inc + args.y_offset]; + const auto z = buffers_host.c_mat[index]; // * args.z_inc + args.z_offset]; + buffers_host.c_mat[index] = x * y * args.alpha + z * args.beta; + } + return StatusCode::kSuccess; +} + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XHAD_H_ +#endif -- cgit v1.2.3