diff options
Diffstat (limited to 'src/tuning/kernels/xgemv.cpp')
-rw-r--r-- | src/tuning/kernels/xgemv.cpp | 138 |
1 files changed, 7 insertions, 131 deletions
diff --git a/src/tuning/kernels/xgemv.cpp b/src/tuning/kernels/xgemv.cpp index 965fc53d..9e45d73f 100644 --- a/src/tuning/kernels/xgemv.cpp +++ b/src/tuning/kernels/xgemv.cpp @@ -7,135 +7,11 @@ // Author(s): // Cedric Nugteren <www.cedricnugteren.nl> // -// This file uses the 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 +// This file uses the auto-tuner to tune the xgemv OpenCL kernels. // // ================================================================================================= -#include <string> -#include <vector> - -#include "utilities/utilities.hpp" -#include "tuning/tuning.hpp" - -namespace clblast { -// ================================================================================================= - -// Settings for this kernel (default command-line arguments) -TunerDefaults GetTunerDefaults(const int) { - auto settings = TunerDefaults(); - settings.options = {kArgM, kArgN, kArgAlpha, kArgBeta}; - settings.default_m = 2048; - settings.default_n = 2048; - settings.default_num_runs = 4; - return settings; -} - -// Settings for this kernel (general) -template <typename T> -TunerSettings GetTunerSettings(const int V, const Arguments<T> &args) { - auto settings = TunerSettings(); - - // Identification of the kernel - settings.kernel_family = (V==1) ? "xgemv" : ((V==2) ? "xgemv_fast" : "xgemv_fast_rot"); - settings.kernel_name = (V==1) ? "Xgemv" : ((V==2) ? "XgemvFast" : "XgemvFastRot"); - settings.sources = -#include "../src/kernels/level2/xgemv.opencl" -#include "../src/kernels/level2/xgemv_fast.opencl" - ; - - // Buffer sizes - settings.size_x = args.n; - settings.size_y = args.m; - settings.size_a = args.m * args.n; - - // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5) - settings.inputs = {0, 1, 2}; - settings.outputs = {1}; - - // Sets the base thread configuration - settings.global_size = {args.m}; - settings.global_size_ref = settings.global_size; - settings.local_size = {1}; - settings.local_size_ref = {64}; - - // Transforms the thread configuration based on the parameters - settings.mul_local = {{"WGS"+std::to_string(V)}}; - settings.div_global = (V==1 || V==2) ? TransformVector{{"WPT"+std::to_string(V)}} : TransformVector{}; - - // Sets the tuning parameters and their possible values - if (V==1) { - settings.parameters = { - {"WGS"+std::to_string(V), {32, 64, 128, 256}}, - {"WPT"+std::to_string(V), {1, 2, 4}}, - }; - } - if (V==2) { - settings.parameters = { - {"WGS"+std::to_string(V), {16, 32, 64, 128, 256}}, - {"WPT"+std::to_string(V), {1, 2, 4}}, - {"VW"+std::to_string(V), {1, 2, 4, 8}}, - }; - } - if (V==3) { - settings.parameters = { - {"WGS"+std::to_string(V), {16, 32, 64, 128}}, - {"WPT"+std::to_string(V), {1, 2, 4, 8, 16, 32}}, - {"VW"+std::to_string(V), {1, 2, 4, 8}}, - }; - } - - // Describes how to compute the performance metrics - settings.metric_amount = (args.m*args.n + 2*args.m + args.n) * GetBytes(args.precision); - settings.performance_unit = "GB/s"; - - return settings; -} - -// Tests for valid arguments -template <typename T> -void TestValidArguments(const int, const Arguments<T> &) { } -std::vector<Constraint> SetConstraints(const int V) { - auto constraints = std::vector<Constraint>(); - if (V==2 || V==3) { - auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); }; - constraints.push_back({MultipleOfX, {"WPT"+std::to_string(V), "VW"+std::to_string(V)}}); - } - if (V==3) { - auto LargerOrEqual = [] (std::vector<size_t> v) { return v[0] >= v[1]; }; - constraints.push_back({LargerOrEqual, {"WGS"+std::to_string(V), "WPT"+std::to_string(V)}}); - } - return constraints; -} - -// Sets the kernel's arguments -template <typename T> -void SetArguments(const int V, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) { - auto a_rotated = (V==3) ? 1 : 0; - kernel.SetArgument(0, static_cast<int>(args.m)); - kernel.SetArgument(1, static_cast<int>(args.n)); - kernel.SetArgument(2, GetRealArg(args.alpha)); - kernel.SetArgument(3, GetRealArg(args.beta)); - kernel.SetArgument(4, a_rotated); - kernel.SetArgument(5, buffers[2]()); // 2 == A matrix - kernel.SetArgument(6, 0); - kernel.SetArgument(7, static_cast<int>(args.m)); - kernel.SetArgument(8, buffers[0]()); // 0 == X vector - kernel.SetArgument(9, 0); - kernel.SetArgument(10, 1); - kernel.SetArgument(11, buffers[1]()); // 1 == Y vector - kernel.SetArgument(12, 0); - kernel.SetArgument(13, 1); - kernel.SetArgument(14, 0); // Conjugate transpose - kernel.SetArgument(15, 0); // Additional parameter - kernel.SetArgument(16, 0); // Banded 'kl' - kernel.SetArgument(17, 0); // Banded 'ku' -} - -// ================================================================================================= -} // namespace clblast +#include "tuning/kernels/xgemv.hpp" // Shortcuts to the clblast namespace using half = clblast::half; @@ -147,11 +23,11 @@ template <int V> void StartVariation(int argc, char *argv[]) { const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv); switch(clblast::GetPrecision(command_line_args)) { - case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break; - case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break; - case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break; - case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break; - case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break; + case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, V, clblast::XgemvGetTunerDefaults, clblast::XgemvGetTunerSettings<half>, clblast::XgemvTestValidArguments<half>, clblast::XgemvSetConstraints, clblast::XgemvSetArguments<half>); break; + case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, V, clblast::XgemvGetTunerDefaults, clblast::XgemvGetTunerSettings<float>, clblast::XgemvTestValidArguments<float>, clblast::XgemvSetConstraints, clblast::XgemvSetArguments<float>); break; + case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, V, clblast::XgemvGetTunerDefaults, clblast::XgemvGetTunerSettings<double>, clblast::XgemvTestValidArguments<double>, clblast::XgemvSetConstraints, clblast::XgemvSetArguments<double>); break; + case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, V, clblast::XgemvGetTunerDefaults, clblast::XgemvGetTunerSettings<float2>, clblast::XgemvTestValidArguments<float2>, clblast::XgemvSetConstraints, clblast::XgemvSetArguments<float2>); break; + case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, V, clblast::XgemvGetTunerDefaults, clblast::XgemvGetTunerSettings<double2>, clblast::XgemvTestValidArguments<double2>, clblast::XgemvSetConstraints, clblast::XgemvSetArguments<double2>); break; } } |