diff options
author | CNugteren <web@cedricnugteren.nl> | 2015-08-09 15:50:41 +0200 |
---|---|---|
committer | CNugteren <web@cedricnugteren.nl> | 2015-08-09 15:50:41 +0200 |
commit | dbdb58c6002cbd693f246f1e93919cc32ad4055a (patch) | |
tree | dd2b7b51c9d5baebb3f9ea660c9ba96e7aaef281 /src/tuning/xgemv.cc | |
parent | e4aa4519c2e2e41d3f8551d4fc8b775e7bf0d168 (diff) |
Refactored the tuners, added JSON output
Diffstat (limited to 'src/tuning/xgemv.cc')
-rw-r--r-- | src/tuning/xgemv.cc | 182 |
1 files changed, 104 insertions, 78 deletions
diff --git a/src/tuning/xgemv.cc b/src/tuning/xgemv.cc index a9d88e4b..e22b5103 100644 --- a/src/tuning/xgemv.cc +++ b/src/tuning/xgemv.cc @@ -7,8 +7,7 @@ // Author(s): // Cedric Nugteren <www.cedricnugteren.nl> // -// This file implements an auto-tuner to tune the Xgemv OpenCL kernel. It uses the CLTune library. -// Three variations of the kernel are tuned: +// This file uses the CLTune auto-tuner to tune the xgemv OpenCL kernels. Three variants are tuned: // 1: The full version of the kernel // 2: The fast version for non-transposed matrices // 3: The fast version for transposed matrices @@ -17,7 +16,6 @@ #include <string> #include <vector> -#include <stdexcept> #include "internal/utilities.h" #include "internal/tuning.h" @@ -25,93 +23,121 @@ namespace clblast { // ================================================================================================= -// The Xgemv auto-tuner -template <typename T> -void XgemvTune(const Arguments<T> &args, const size_t variation, - const std::vector<T> &a_mat, const std::vector<T> &x_vec, std::vector<T> &y_vec, - cltune::Tuner &tuner) { - - // Sets the kernel name and the layout argument - auto kernel_name = (variation == 1) ? "Xgemv" : ((variation == 2) ? "XgemvFast" : "XgemvFastRot"); - auto a_rotated = (variation == 3) ? 1 : 0; - - // This points to the Xgemv kernel as found in the CLBlast library - std::string sources = - #include "../src/kernels/common.opencl" - #include "../src/kernels/xgemv.opencl" - ; - auto id = tuner.AddKernelFromString(sources, kernel_name, {args.m}, {1}); - tuner.SetReferenceFromString(sources, "Xgemv", {args.m}, {64}); - - // Helper for the constraints - auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); }; - - // Sets the tunable parameters, their possible values, the adjusted thread sizes, and constraints - if (variation == 1) { - tuner.AddParameter(id, "WGS1", {64, 128, 256, 512, 1024, 1536, 2048}); - tuner.AddParameter(id, "WPT1", {1, 2, 4, 8}); - tuner.MulLocalSize(id, {"WGS1"}); - tuner.DivGlobalSize(id, {"WPT1"}); +// See comment at top of file for a description of the class +template <typename T, int V> +class TuneXgemv { + public: + + // The representative kernel and the source code + static std::string KernelFamily() { return "xgemv_"+std::to_string(V); } + static std::string KernelName() { return (V==1) ? "Xgemv" : ((V==2) ? "XgemvFast" : "XgemvFastRot"); } + static std::string GetSources() { + return + #include "../src/kernels/common.opencl" + #include "../src/kernels/xgemv.opencl" + ; } - else if (variation == 2) { - tuner.AddParameter(id, "WGS2", {64, 128, 256, 512, 1024, 1536, 2048}); - tuner.AddParameter(id, "WPT2", {1, 2, 4, 8}); - tuner.AddParameter(id, "VW2", {1, 2, 4, 8}); - tuner.MulLocalSize(id, {"WGS2"}); - tuner.DivGlobalSize(id, {"WPT2"}); - tuner.AddConstraint(id, MultipleOfX, {"WPT2", "VW2"}); + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { return {kArgM, kArgN, kArgAlpha, kArgBeta}; } + + // Tests for valid arguments + static void TestValidArguments(const Arguments<T> &) { } + + // Sets the default values for the arguments + static size_t DefaultM() { return 2048; } + static size_t DefaultN() { return 2048; } + 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; } + static size_t GetSizeY(const Arguments<T> &args) { return args.m; } + static size_t GetSizeA(const Arguments<T> &args) { return args.m * args.n; } + 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"+std::to_string(V), {64, 128, 256, 512, 1024, 1536, 2048}); + tuner.AddParameter(id, "WPT"+std::to_string(V), {1, 2, 4, 8}); + if (V==2 || V==3) { tuner.AddParameter(id, "VW"+std::to_string(V), {1, 2, 4, 8}); } + } + + // Sets the constraints and local memory size + static void SetConstraints(cltune::Tuner &tuner, const size_t id) { + auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); }; + if (V==2 || V==3) { + tuner.AddConstraint(id, MultipleOfX, {"WPT"+std::to_string(V), "VW"+std::to_string(V)}); + } } - else if (variation == 3) { - tuner.AddParameter(id, "WGS3", {64, 128, 256, 512, 1024, 1536, 2048}); - tuner.AddParameter(id, "WPT3", {1, 2, 4, 8}); - tuner.AddParameter(id, "VW3", {1, 2, 4, 8}); - tuner.MulLocalSize(id, {"WGS3"}); - tuner.DivGlobalSize(id, {"WPT3"}); - tuner.AddConstraint(id, MultipleOfX, {"WGS3", "VW3"}); + 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.m}; } + 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"+std::to_string(V)}}; } + static TransformVector DivLocal() { return {}; } + static TransformVector MulGlobal() { return {}; } + static TransformVector DivGlobal() { return {{"WPT"+std::to_string(V)}}; } + + // 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> &a_mat, std::vector<T> &, std::vector<T> &) { + auto a_rotated = (V==3) ? 1 : 0; + tuner.AddArgumentScalar(static_cast<int>(args.m)); + tuner.AddArgumentScalar(static_cast<int>(args.n)); + tuner.AddArgumentScalar(args.alpha); + tuner.AddArgumentScalar(args.beta); + tuner.AddArgumentScalar(static_cast<int>(a_rotated)); + tuner.AddArgumentInput(a_mat); + tuner.AddArgumentScalar(0); + tuner.AddArgumentScalar(static_cast<int>(args.m)); + tuner.AddArgumentInput(x_vec); + tuner.AddArgumentScalar(0); + tuner.AddArgumentScalar(1); + tuner.AddArgumentOutput(y_vec); + tuner.AddArgumentScalar(0); + tuner.AddArgumentScalar(1); + tuner.AddArgumentScalar(0); // Conjugate transpose } - // Tests for a specific precision - tuner.AddParameter(id, "PRECISION", {static_cast<size_t>(args.precision)}); - tuner.AddParameterReference("PRECISION", static_cast<size_t>(args.precision)); - - // Sets the function's arguments - tuner.AddArgumentScalar(static_cast<int>(args.m)); - tuner.AddArgumentScalar(static_cast<int>(args.n)); - tuner.AddArgumentScalar(args.alpha); - tuner.AddArgumentScalar(args.beta); - tuner.AddArgumentScalar(static_cast<int>(a_rotated)); - tuner.AddArgumentInput(a_mat); - tuner.AddArgumentScalar(0); - tuner.AddArgumentScalar(static_cast<int>(args.m)); - tuner.AddArgumentInput(x_vec); - tuner.AddArgumentScalar(0); - tuner.AddArgumentScalar(1); - tuner.AddArgumentOutput(y_vec); - tuner.AddArgumentScalar(0); - tuner.AddArgumentScalar(1); - tuner.AddArgumentScalar(0); // Conjugate transpose -} + // Describes how to compute the performance metrics + static size_t GetMetric(const Arguments<T> &args) { + return (args.m*args.n + 2*args.m + args.n) * GetBytes(args.precision); + } + static std::string PerformanceUnit() { return "GB/s"; } +}; // ================================================================================================= +} // namespace clblast -// Main function which calls the common client code with the routine-specific function as argument. -void TunerXgemv(int argc, char *argv[]) { - auto num_variations = size_t{3}; - switch(GetPrecision(argc, argv)) { - case Precision::kHalf: throw std::runtime_error("Unsupported precision mode"); - case Precision::kSingle: TunerAXY<float>(argc, argv, num_variations, XgemvTune<float>); break; - case Precision::kDouble: TunerAXY<double>(argc, argv, num_variations, XgemvTune<double>); break; - case Precision::kComplexSingle: TunerAXY<float2>(argc, argv, num_variations, XgemvTune<float2>); break; - case Precision::kComplexDouble: TunerAXY<double2>(argc, argv, num_variations, XgemvTune<double2>); break; +// Shortcuts to the clblast namespace +using float2 = clblast::float2; +using double2 = clblast::double2; + +// Function to tune a specific variation V (not within the clblast namespace) +template <int V> +void StartVariation(int argc, char *argv[]) { + switch(clblast::GetPrecision(argc, argv)) { + case clblast::Precision::kHalf: throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXgemv<float,V>, float>(argc, argv); break; + case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXgemv<double,V>, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXgemv<float2,V>, float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXgemv<double2,V>, double2>(argc, argv); break; } } -// ================================================================================================= -} // namespace clblast - // Main function (not within the clblast namespace) int main(int argc, char *argv[]) { - clblast::TunerXgemv(argc, argv); + StartVariation<1>(argc, argv); + StartVariation<2>(argc, argv); + StartVariation<3>(argc, argv); return 0; } |