// ================================================================================================= // 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 Xamax 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_XAMAX_H_ #define CLBLAST_TEST_ROUTINES_XAMAX_H_ #include "test/routines/common.hpp" namespace clblast { // ================================================================================================= // See comment at top of file for a description of the class template class TestXamax { public: // The BLAS level: 1, 2, or 3 static size_t BLASLevel() { return 1; } // The list of arguments relevant for this routine static std::vector GetOptions() { return {kArgN, kArgXInc, kArgXOffset, kArgImaxOffset}; } static std::vector BuffersIn() { return {kBufVecX, kBufScalar}; } static std::vector BuffersOut() { return {kBufScalar}; } // 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 GetSizeImax(const Arguments &args) { return (1 + args.imax_offset) * 2; // always a 4-byte integer, this is a hack for FP16 } // Describes how to set the sizes of all the buffers static void SetSizes(Arguments &args, Queue&) { args.x_size = GetSizeX(args); args.scalar_size = GetSizeImax(args); } // 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 = Amax(args.n, buffers.scalar(), args.imax_offset, buffers.x_vec(), args.x_offset, args.x_inc, &queue_plain, &event); if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } #elif CUDA_API auto status = Amax(args.n, buffers.scalar(), args.imax_offset, buffers.x_vec(), args.x_offset, args.x_inc, queue.GetContext()(), queue.GetDevice()()); cuStreamSynchronize(queue()); #endif 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 = clblasXamax(args.n, buffers.scalar, args.imax_offset, buffers.x_vec, args.x_offset, args.x_inc, 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, BuffersHost &buffers_host, Queue &) { cblasXamax(args.n, buffers_host.scalar, args.imax_offset, buffers_host.x_vec, args.x_offset, args.x_inc); return StatusCode::kSuccess; } #endif // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments &args, BuffersCUDA &buffers, Queue &) { auto status = cublasXamax(reinterpret_cast(args.cublas_handle), args.n, buffers.scalar, args.imax_offset, buffers.x_vec, args.x_offset, args.x_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } } #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.scalar_size, static_cast(0)); buffers.scalar.Read(queue, args.scalar_size, result); return result; } // Describes how to compute the indices of the result buffer static size_t ResultID1(const Arguments &) { return 1; } // N/A for this routine static size_t ResultID2(const Arguments &) { return 1; } // N/A for this routine static size_t GetResultIndex(const Arguments &args, const size_t, const size_t) { return args.imax_offset; } // Describes how to compute performance metrics static size_t GetFlops(const Arguments &args) { return args.n; } static size_t GetBytes(const Arguments &args) { return ((args.n) + 1) * sizeof(T); } }; // ================================================================================================= } // namespace clblast // CLBLAST_TEST_ROUTINES_XAMAX_H_ #endif