summaryrefslogtreecommitdiff
path: root/src/tuning/kernels/xaxpy.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/tuning/kernels/xaxpy.cpp')
-rw-r--r--src/tuning/kernels/xaxpy.cpp83
1 files changed, 1 insertions, 82 deletions
diff --git a/src/tuning/kernels/xaxpy.cpp b/src/tuning/kernels/xaxpy.cpp
index dd44018c..a290b6c1 100644
--- a/src/tuning/kernels/xaxpy.cpp
+++ b/src/tuning/kernels/xaxpy.cpp
@@ -11,88 +11,7 @@
//
// =================================================================================================
-#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, kArgAlpha};
- settings.default_n = 4096*1024;
- return settings;
-}
-
-// Settings for this kernel (general)
-template <typename T>
-TunerSettings GetTunerSettings(const int, const Arguments<T> &args) {
- auto settings = TunerSettings();
-
- // Identification of the kernel
- settings.kernel_family = "xaxpy";
- settings.kernel_name = "XaxpyFastest";
- settings.sources =
-#include "../src/kernels/level1/level1.opencl"
-#include "../src/kernels/level1/xaxpy.opencl"
- ;
-
- // Buffer sizes
- settings.size_x = args.n;
- settings.size_y = args.n;
-
- // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
- settings.inputs = {0, 1};
- settings.outputs = {1};
-
- // Sets the base thread configuration
- settings.global_size = {args.n};
- 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"}};
- settings.div_global = {{"WPT"},{"VW"}};
-
- // Sets the tuning parameters and their possible values
- settings.parameters = {
- {"WGS", {64, 128, 256, 512, 1024, 2048}},
- {"WPT", {1, 2, 4, 8}},
- {"VW", {1, 2, 4, 8}},
- };
-
- // Describes how to compute the performance metrics
- settings.metric_amount = 3 * 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> &args) {
- if (!IsMultiple(args.n, 64)) {
- throw std::runtime_error("'XaxpyFastest' requires 'n' to be a multiple of WGS*WPT*VW");
- }
-}
-std::vector<Constraint> SetConstraints(const int) { return {}; }
-
-// Sets the kernel's arguments
-template <typename T>
-void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
- kernel.SetArgument(0, static_cast<int>(args.n));
- kernel.SetArgument(1, GetRealArg(args.alpha));
- kernel.SetArgument(2, buffers[0]()); // 0 == X vector
- kernel.SetArgument(3, buffers[1]()); // 1 == Y vector
-}
-
-// =================================================================================================
-} // namespace clblast
+#include "tuning/kernels/xaxpy.hpp"
// Shortcuts to the clblast namespace
using half = clblast::half;