diff options
Diffstat (limited to 'src/tuning/kernels/xdot.cpp')
-rw-r--r-- | src/tuning/kernels/xdot.cpp | 89 |
1 files changed, 1 insertions, 88 deletions
diff --git a/src/tuning/kernels/xdot.cpp b/src/tuning/kernels/xdot.cpp index 635d012a..10126392 100644 --- a/src/tuning/kernels/xdot.cpp +++ b/src/tuning/kernels/xdot.cpp @@ -12,94 +12,7 @@ // // ================================================================================================= -#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 = {kArgN}; - settings.default_n = 2*1024*1024; - 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 = "xdot_"+std::to_string(V); - settings.kernel_name = (V==1) ? "Xdot" : "XdotEpilogue"; - settings.sources = -#include "../src/kernels/level1/xdot.opencl" - ; - - // Buffer sizes - settings.size_x = args.n; - settings.size_y = args.n; - settings.size_temp = args.n; // Worst case - - // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5) - settings.inputs = {0, 1, 5}; - settings.outputs = {}; // no output checking - - // Sets the base thread configuration - settings.global_size = (V==1) ? std::vector<size_t>{2*64} : std::vector<size_t>{1}; - settings.global_size_ref = (V==1) ? std::vector<size_t>{2*64*64} : std::vector<size_t>{64}; - settings.local_size = {1}; - settings.local_size_ref = {64}; - - // Transforms the thread configuration based on the parameters - settings.mul_local = (V==1) ? TransformVector{{"WGS1"}} : TransformVector{{"WGS2"}}; - settings.mul_global = (V==1) ? TransformVector{{"WGS1"}} : TransformVector{{"WGS2"}}; - - // Sets the tuning parameters and their possible values - settings.parameters = { - {"WGS"+std::to_string(V), {32, 64, 128, 256, 512, 1024}}, - }; - - // Describes how to compute the performance metrics - settings.metric_amount = (V==1) ? (2*args.n + 1) * GetBytes(args.precision) : 1 * GetBytes(args.precision); - settings.performance_unit = (V==1) ? "GB/s" : "N/A"; - - return settings; -} - -// Tests for valid arguments -template <typename T> -void TestValidArguments(const int, const Arguments<T> &) { } -std::vector<Constraint> SetConstraints(const int) { return {}; } - -// Sets the kernel's arguments -template <typename T> -void SetArguments(const int V, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) { - if (V == 1) { - kernel.SetArgument(0, static_cast<int>(args.n)); - kernel.SetArgument(1, buffers[0]()); // 0 == X vector - kernel.SetArgument(2, 0); - kernel.SetArgument(3, 1); - kernel.SetArgument(4, buffers[1]()); // 1 == Y vector - kernel.SetArgument(5, 0); - kernel.SetArgument(6, 1); - kernel.SetArgument(7, buffers[5]()); // 5 == temp; no output checking - size varies - kernel.SetArgument(8, static_cast<int>(false)); - } - else { - kernel.SetArgument(0, buffers[5]()); // 5 == temp - kernel.SetArgument(1, buffers[0]()); // 0 == X vector; no output checking - size varies - kernel.SetArgument(2, 0); - } -} - -// ================================================================================================= -} // namespace clblast +#include "tuning/kernels/xdot.hpp" // Shortcuts to the clblast namespace using half = clblast::half; |