diff options
Diffstat (limited to 'src/tuning/pad.cc')
-rw-r--r-- | src/tuning/pad.cc | 124 |
1 files changed, 79 insertions, 45 deletions
diff --git a/src/tuning/pad.cc b/src/tuning/pad.cc index 584415c7..2ce566fb 100644 --- a/src/tuning/pad.cc +++ b/src/tuning/pad.cc @@ -7,13 +7,12 @@ // Author(s): // Cedric Nugteren <www.cedricnugteren.nl> // -// This file implements an auto-tuner to tune the pad-copy OpenCL kernels. It uses CLTune. +// This file uses the CLTune auto-tuner to tune the pad OpenCL kernels. // // ================================================================================================= #include <string> #include <vector> -#include <stdexcept> #include "internal/utilities.h" #include "internal/tuning.h" @@ -21,37 +20,68 @@ namespace clblast { // ================================================================================================= -// The pad auto-tuner +// See comment at top of file for a description of the class template <typename T> -void PadTune(const Arguments<T> &args, - const std::vector<T> &a_mat, std::vector<T> &b_mat, - cltune::Tuner &tuner) { - - // This points to the PadMatrix kernel as found in the CLBlast library. This is just one - // example of a pad kernel. However, all pad-kernels use the same tuning parameters, so one has - // to be chosen as a representative. - std::string sources = - #include "../src/kernels/common.opencl" - #include "../src/kernels/pad.opencl" - ; - auto id = tuner.AddKernelFromString(sources, "PadMatrix", {args.m, args.n}, {1, 1}); - tuner.SetReferenceFromString(sources, "PadMatrix", {args.m, args.n}, {8, 8}); - - // Sets the tunable parameters and their possible values - tuner.AddParameter(id, "PAD_DIMX", {8, 16, 32}); - tuner.AddParameter(id, "PAD_DIMY", {8, 16, 32}); - tuner.AddParameter(id, "PAD_WPTX", {1, 2, 4}); - tuner.AddParameter(id, "PAD_WPTY", {1, 2, 4}); - - // 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 (both global and local) based on the parameters - tuner.MulLocalSize(id, {"PAD_DIMX", "PAD_DIMY"}); - tuner.DivGlobalSize(id, {"PAD_WPTX", "PAD_WPTY"}); - - // Sets the function's arguments +class TunePad { + public: + + // The representative kernel and the source code + static std::string KernelFamily() { return "pad"; } + static std::string KernelName() { return "PadMatrix"; } + static std::string GetSources() { + return + #include "../src/kernels/common.opencl" + #include "../src/kernels/pad.opencl" + ; + } + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { return {kArgM, kArgN}; } + + // Tests for valid arguments + static void TestValidArguments(const Arguments<T> &) { } + + // Sets the default values for the arguments + static size_t DefaultM() { return 1024; } + static size_t DefaultN() { return 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> &) { return 1; } // N/A for this kernel + static size_t GetSizeY(const Arguments<T> &) { return 1; } // N/A for this kernel + static size_t GetSizeA(const Arguments<T> &args) { return args.m * args.n; } + static size_t GetSizeB(const Arguments<T> &args) { return args.m * args.n; } + 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, "PAD_DIMX", {8, 16, 32}); + tuner.AddParameter(id, "PAD_DIMY", {8, 16, 32}); + tuner.AddParameter(id, "PAD_WPTX", {1, 2, 4}); + tuner.AddParameter(id, "PAD_WPTY", {1, 2, 4}); + } + + // 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.m, args.n}; } + static std::vector<size_t> LocalSize() { return {1, 1}; } + static std::vector<size_t> LocalSizeRef() { return {8, 8}; } + + // Transforms the thread configuration based on the parameters + using TransformVector = std::vector<std::vector<std::string>>; + static TransformVector MulLocal() { return {{"PAD_DIMX", "PAD_DIMY"}}; } + static TransformVector DivLocal() { return {}; } + static TransformVector MulGlobal() { return {}; } + static TransformVector DivGlobal() { return {{"PAD_WPTX", "PAD_WPTY"}}; } + + // Sets the kernel's arguments + static void SetArguments(cltune::Tuner &tuner, const Arguments<T> &args, + std::vector<T> &, std::vector<T> &, + std::vector<T> &a_mat, std::vector<T> &b_mat, std::vector<T> &) { tuner.AddArgumentScalar(static_cast<int>(args.m)); tuner.AddArgumentScalar(static_cast<int>(args.n)); tuner.AddArgumentScalar(static_cast<int>(args.m)); @@ -63,27 +93,31 @@ void PadTune(const Arguments<T> &args, tuner.AddArgumentScalar(0); tuner.AddArgumentOutput(b_mat); tuner.AddArgumentScalar(0); -} - -// ================================================================================================= + } -// Main function which calls the common client code with the routine-specific function as argument. -void TunerPad(int argc, char *argv[]) { - switch(GetPrecision(argc, argv)) { - case Precision::kHalf: throw std::runtime_error("Unsupported precision mode"); - case Precision::kSingle: TunerAB<float>(argc, argv, PadTune<float>); break; - case Precision::kDouble: TunerAB<double>(argc, argv, PadTune<double>); break; - case Precision::kComplexSingle: TunerAB<float2>(argc, argv, PadTune<float2>); break; - case Precision::kComplexDouble: TunerAB<double2>(argc, argv, PadTune<double2>); break; + // Describes how to compute the performance metrics + static size_t GetMetric(const Arguments<T> &args) { + return 2 * args.m * 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::TunerPad(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::TunePad<float>, float>(argc, argv); break; + case clblast::Precision::kDouble: clblast::Tuner<clblast::TunePad<double>, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TunePad<float2>, float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TunePad<double2>, double2>(argc, argv); break; + } return 0; } |