diff options
Diffstat (limited to 'src/tuning/kernels/xger.cpp')
-rw-r--r-- | src/tuning/kernels/xger.cpp | 112 |
1 files changed, 48 insertions, 64 deletions
diff --git a/src/tuning/kernels/xger.cpp b/src/tuning/kernels/xger.cpp index 14a98761..c2eb1d31 100644 --- a/src/tuning/kernels/xger.cpp +++ b/src/tuning/kernels/xger.cpp @@ -25,69 +25,64 @@ template <typename T> class TuneXger { public: - // The representative kernel and the source code - static std::string KernelFamily() { return "xger"; } - static std::string KernelName() { return "Xger"; } - static std::string GetSources() { - return - #include "../src/kernels/common.opencl" - #include "../src/kernels/level2/level2.opencl" - #include "../src/kernels/level2/xger.opencl" - ; + // Settings for this kernel (default command-line arguments) + static TunerDefaults GetTunerDefaults() { + auto settings = TunerDefaults(); + settings.options = {kArgM, kArgN, kArgAlpha}; + settings.default_m = 1024; + settings.default_n = 1024; + return settings; } - // The list of arguments relevant for this routine - static std::vector<std::string> GetOptions() { return {kArgN, kArgM, kArgAlpha}; } + // Settings for this kernel (general) + static TunerSettings GetTunerSettings(const Arguments<T> &args) { + auto settings = TunerSettings(); + + // Identification of the kernel + settings.kernel_family = "xger"; + settings.kernel_name = "Xger"; + settings.sources = +#include "../src/kernels/common.opencl" +#include "../src/kernels/level2/level2.opencl" +#include "../src/kernels/level2/xger.opencl" + ; + + // Buffer sizes + settings.size_x = args.m; + settings.size_y = args.n; + settings.size_a = args.m * args.n; + + // Sets the base thread configuration + settings.global_size = {args.m, args.n}; + settings.global_size_ref = settings.global_size; + settings.local_size = {1, 1}; + settings.local_size_ref = {8, 8}; + + // Transforms the thread configuration based on the parameters + settings.mul_local = {{"WGS1", "WGS2"}}; + settings.div_global = {{"WPT", "WPT"}}; + + // Sets the tuning parameters and their possible values + settings.parameters = { + {"WGS1", {4, 8, 16, 32, 64, 128, 256, 512}}, + {"WGS2", {1, 2, 4, 8, 16, 32, 64, 128, 256}}, + {"WPT", {1, 2, 4}}, + }; + + // Describes how to compute the performance metrics + settings.metric_amount = (2*args.m*args.n + args.m + args.n) * GetBytes(args.precision); + settings.performance_unit = "GB/s"; + + return settings; + } // 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 size_t DefaultBatchCount() { return 1; } // N/A for this kernel - static double DefaultFraction() { return 1.0; } // N/A for this kernel - static size_t DefaultNumRuns() { return 10; } // run every kernel this many times for averaging - static size_t DefaultSwarmSizePSO() { return 8; } // N/A for this kernel - static double DefaultInfluenceGlobalPSO(){ return 0.1; }// N/A for this kernel - static double DefaultInfluenceLocalPSO(){ return 0.3; } // N/A for this kernel - static double DefaultInfluenceRandomPSO(){ return 0.6; }// N/A for this kernel - static size_t DefaultHeuristic(){ return static_cast<size_t> (cltune::SearchMethod::FullSearch);} - static double DefaultMaxTempAnn(){ 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.m; } - static size_t GetSizeY(const Arguments<T> &args) { return args.n; } - 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 - static size_t GetSizeTemp(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, "WGS1", {4, 8, 16, 32, 64, 128, 256, 512}); - tuner.AddParameter(id, "WGS2", {1, 2, 4, 8, 16, 32, 64, 128, 256}); - tuner.AddParameter(id, "WPT", {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> GlobalSizeRef(const Arguments<T> &args) { return GlobalSize(args); } - 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 {{"WGS1", "WGS2"}}; } - static TransformVector DivLocal() { return {}; } - static TransformVector MulGlobal() { return {}; } - static TransformVector DivGlobal() { return {{"WPT", "WPT"}}; } - // 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, @@ -107,17 +102,6 @@ class TuneXger { tuner.AddArgumentScalar(static_cast<int>(args.m)); // a_ld tuner.AddArgumentScalar(0); // a_is_rowmajor } - - // Describes how to compute the performance metrics - static size_t GetMetric(const Arguments<T> &args) { - return (2*args.m*args.n + args.m + args.n) * GetBytes(args.precision); - } - static std::string PerformanceUnit() { return "GB/s"; } - - // Returns which Heuristic to run - static size_t GetHeuristic(const Arguments<T> &args){ - return static_cast<size_t> (cltune::SearchMethod::FullSearch); - } }; // ================================================================================================= |