diff options
Diffstat (limited to 'src/tuning/xaxpy.cc')
| -rw-r--r-- | src/tuning/xaxpy.cc | 133 |
1 files changed, 83 insertions, 50 deletions
diff --git a/src/tuning/xaxpy.cc b/src/tuning/xaxpy.cc index 20b5978e..cc9e81d3 100644 --- a/src/tuning/xaxpy.cc +++ b/src/tuning/xaxpy.cc @@ -7,13 +7,12 @@ // Author(s): // Cedric Nugteren <www.cedricnugteren.nl> // -// This file implements an auto-tuner to tune the Xaxpy OpenCL kernel. It uses the CLTune library. +// This file uses the CLTune auto-tuner to tune the xaxpy OpenCL kernels. // // ================================================================================================= #include <string> #include <vector> -#include <stdexcept> #include "internal/utilities.h" #include "internal/tuning.h" @@ -21,66 +20,100 @@ namespace clblast { // ================================================================================================= -// The Xaxpy auto-tuner +// See comment at top of file for a description of the class template <typename T> -void XaxpyTune(const Arguments<T> &args, - const std::vector<T> &x_vec, std::vector<T> &y_vec, - cltune::Tuner &tuner) { - - // The XaxpyFast kernel only works under certain conditions. Check here whether the condition is - // true for the reference kernel - if (!IsMultiple(args.n, 64)) { - throw std::runtime_error("The 'XaxpyFast' kernel requires 'n' to be a multiple of WGS*WPT*VW"); +class TuneXaxpy { + public: + + // The representative kernel and the source code + static std::string KernelFamily() { return "xaxpy"; } + static std::string KernelName() { return "XaxpyFast"; } + static std::string GetSources() { + return + #include "../src/kernels/common.opencl" + #include "../src/kernels/xaxpy.opencl" + ; } - // This points to the XaxpyFast kernel as found in the CLBlast library - std::string sources = - #include "../src/kernels/common.opencl" - #include "../src/kernels/xaxpy.opencl" - ; - auto id = tuner.AddKernelFromString(sources, "XaxpyFast", {args.n}, {1}); - tuner.SetReferenceFromString(sources, "XaxpyFast", {args.n}, {64}); - - // Sets the tunable parameters and their possible values - tuner.AddParameter(id, "WGS", {64, 128, 256, 512, 1024, 2048}); - tuner.AddParameter(id, "WPT", {1, 2, 4, 8}); - tuner.AddParameter(id, "VW", {1, 2, 4, 8}); - - // Tests for a specific precision - tuner.AddParameter(id, "PRECISION", {static_cast<size_t>(args.precision)}); - tuner.AddParameterReference("PRECISION", static_cast<size_t>(args.precision)); - - // Modifies the thread-sizes (local) based on the parameters - tuner.MulLocalSize(id, {"WGS"}); - tuner.DivGlobalSize(id, {"WPT"}); - tuner.DivGlobalSize(id, {"VW"}); - - // Sets the function's arguments - tuner.AddArgumentScalar(static_cast<int>(args.n)); - tuner.AddArgumentScalar(args.alpha); - tuner.AddArgumentInput(x_vec); - tuner.AddArgumentOutput(y_vec); -} + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { return {kArgN, kArgAlpha}; } -// ================================================================================================= + // Tests for valid arguments + static void TestValidArguments(const Arguments<T> &args) { + if (!IsMultiple(args.n, 64)) { + throw std::runtime_error("'XaxpyFast' requires 'n' to be a multiple of WGS*WPT*VW"); + } + } -// Main function which calls the common client code with the routine-specific function as argument. -void TunerXaxpy(int argc, char *argv[]) { - switch(GetPrecision(argc, argv)) { - case Precision::kHalf: throw std::runtime_error("Unsupported precision mode"); - case Precision::kSingle: TunerXY<float>(argc, argv, XaxpyTune<float>); break; - case Precision::kDouble: TunerXY<double>(argc, argv, XaxpyTune<double>); break; - case Precision::kComplexSingle: TunerXY<float2>(argc, argv, XaxpyTune<float2>); break; - case Precision::kComplexDouble: TunerXY<double2>(argc, argv, XaxpyTune<double2>); break; + // Sets the default values for the arguments + static size_t DefaultM() { return 1; } // N/A for this kernel + static size_t DefaultN() { return 4096*1024; } + static size_t DefaultK() { return 1; } // N/A for this kernel + static double DefaultFraction() { return 1.0; } // N/A for this kernel + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments<T> &args) { return args.n; } // N/A for this kernel + static size_t GetSizeY(const Arguments<T> &args) { return args.n; } // N/A for this kernel + static size_t GetSizeA(const Arguments<T> &) { return 1; } // N/A for this kernel + static size_t GetSizeB(const Arguments<T> &) { return 1; } // N/A for this kernel + static size_t GetSizeC(const Arguments<T> &) { return 1; } // N/A for this kernel + + // Sets the tuning parameters and their possible values + static void SetParameters(cltune::Tuner &tuner, const size_t id) { + tuner.AddParameter(id, "WGS", {64, 128, 256, 512, 1024, 2048}); + tuner.AddParameter(id, "WPT", {1, 2, 4, 8}); + tuner.AddParameter(id, "VW", {1, 2, 4, 8}); } -} + + // Sets the constraints and local memory size + static void SetConstraints(cltune::Tuner &, const size_t) { } + static void SetLocalMemorySize(cltune::Tuner &, const size_t, const Arguments<T> &) { } + + // Sets the base thread configuration + static std::vector<size_t> GlobalSize(const Arguments<T> &args) { return {args.n}; } + static std::vector<size_t> LocalSize() { return {1}; } + static std::vector<size_t> LocalSizeRef() { return {64}; } + + // Transforms the thread configuration based on the parameters + using TransformVector = std::vector<std::vector<std::string>>; + static TransformVector MulLocal() { return {{"WGS"}}; } + static TransformVector DivLocal() { return {}; } + static TransformVector MulGlobal() { return {}; } + static TransformVector DivGlobal() { return {{"WPT"},{"VW"}}; } + + // Sets the kernel's arguments + static void SetArguments(cltune::Tuner &tuner, const Arguments<T> &args, + std::vector<T> &x_vec, std::vector<T> &y_vec, + std::vector<T> &, std::vector<T> &, std::vector<T> &) { + tuner.AddArgumentScalar(static_cast<int>(args.n)); + tuner.AddArgumentScalar(args.alpha); + tuner.AddArgumentInput(x_vec); + tuner.AddArgumentOutput(y_vec); + } + + // Describes how to compute the performance metrics + static size_t GetMetric(const Arguments<T> &args) { + return 3 * args.n * GetBytes(args.precision); + } + static std::string PerformanceUnit() { return "GB/s"; } +}; // ================================================================================================= } // namespace clblast +// 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[]) { - clblast::TunerXaxpy(argc, argv); + switch(clblast::GetPrecision(argc, argv)) { + case clblast::Precision::kHalf: throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXaxpy<float>, float>(argc, argv); break; + case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXaxpy<double>, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXaxpy<float2>, float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXaxpy<double2>, double2>(argc, argv); break; + } return 0; } |