Split the three variations of the GEMV kernel for maximal tuning freedom

2 files changed, 71 insertions, 45 deletions

diff --git a/src/tuning/tuning.cc b/src/tuning/tuning.cc
index d617af88..2dcb11d5 100644
--- a/src/tuning/tuning.cc
+++ b/src/tuning/tuning.cc
@@ -75,7 +75,8 @@ template void TunerXY<double2>(int, char**, const Tuner2<double2>&);
 // Function to get command-line argument, set-up the input buffers, configure the tuner, and collect
 // the results. Used for matrix-vector-vector routines.
 template <typename T>
-void TunerAXY(int argc, char* argv[], const Tuner3<T> &tune_function) {
+void TunerAXY(int argc, char* argv[], const size_t num_variations,
+              const Tuner3V<T> &tune_function) {
 
   // Sets the parameters and platform/device for which to tune (command-line options)
   auto help = std::string{"* Options given/available:\n"};
@@ -83,11 +84,10 @@ void TunerAXY(int argc, char* argv[], const Tuner3<T> &tune_function) {
   args.platform_id = GetArgument(argc, argv, help, kArgPlatform, size_t{0});
   args.device_id   = GetArgument(argc, argv, help, kArgDevice, size_t{0});
   args.precision   = GetArgument(argc, argv, help, kArgPrecision, Precision::kSingle);
-  args.m           = GetArgument(argc, argv, help, kArgM, size_t{1024});
-  args.n           = GetArgument(argc, argv, help, kArgN, size_t{1024});
+  args.m           = GetArgument(argc, argv, help, kArgM, size_t{2048});
+  args.n           = GetArgument(argc, argv, help, kArgN, size_t{2048});
   args.alpha       = GetArgument(argc, argv, help, kArgAlpha, GetScalar<T>());
   args.beta        = GetArgument(argc, argv, help, kArgBeta, GetScalar<T>());
-  args.layout      = GetArgument(argc, argv, help, kArgLayout, Layout::kColMajor);
   fprintf(stdout, "%s\n", help.c_str());
 
   // Creates input buffers with random data
@@ -98,36 +98,40 @@ void TunerAXY(int argc, char* argv[], const Tuner3<T> &tune_function) {
   PopulateVector(x_vec);
   PopulateVector(y_vec);
 
-  // Initializes the tuner for the chosen device
-  cltune::Tuner tuner(args.platform_id, args.device_id);
+  // Loop over the different variations of the kernel
+  for (auto variation=size_t{1}; variation<=num_variations; ++variation) {
 
-  // Use full-search to explore all parameter combinations.
-  tuner.UseFullSearch();
+    // Initializes the tuner for the chosen device
+    cltune::Tuner tuner(args.platform_id, args.device_id);
 
-  // Configures the tuning parameters (kernel specific)
-  tune_function(args, a_mat, x_vec, y_vec, tuner);
+    // Use full-search to explore all parameter combinations.
+    tuner.UseFullSearch();
 
-  // Starts the tuning process
-  tuner.Tune();
+    // Configures the tuning parameters (kernel specific)
+    tune_function(args, variation, a_mat, x_vec, y_vec, tuner);
 
-  // Prints the results to screen
-  auto time_ms = tuner.PrintToScreen();
-  tuner.PrintFormatted();
+    // Starts the tuning process
+    tuner.Tune();
 
-  // Also prints the performance of the best-case in terms of GB/s and GFLOPS
-  const auto mega_bytes = ((args.m*args.n + 2*args.m + args.n)*GetBytes(args.precision)) * 1.0e-6;
-  const auto mega_flops = (2*args.m*args.n) * 1.0e-6;
-  if (time_ms != 0.0) {
-    printf("[ -------> ] %.1lf ms or %.1lf GB/s or %.1lf GFLOPS\n",
-           time_ms, mega_bytes/time_ms, mega_flops/time_ms);
+    // Prints the results to screen
+    auto time_ms = tuner.PrintToScreen();
+    tuner.PrintFormatted();
+
+    // Also prints the performance of the best-case in terms of GB/s and GFLOPS
+    const auto mega_bytes = ((args.m*args.n + 2*args.m + args.n)*GetBytes(args.precision)) * 1.0e-6;
+    const auto mega_flops = (2*args.m*args.n) * 1.0e-6;
+    if (time_ms != 0.0) {
+      printf("[ -------> ] %.1lf ms or %.1lf GB/s or %.1lf GFLOPS\n",
+             time_ms, mega_bytes/time_ms, mega_flops/time_ms);
+    }
   }
 }
 
 // Compiles the above function
-template void TunerAXY<float>(int, char**, const Tuner3<float>&);
-template void TunerAXY<double>(int, char**, const Tuner3<double>&);
-template void TunerAXY<float2>(int, char**, const Tuner3<float2>&);
-template void TunerAXY<double2>(int, char**, const Tuner3<double2>&);
+template void TunerAXY<float>(int, char**, const size_t, const Tuner3V<float>&);
+template void TunerAXY<double>(int, char**, const size_t, const Tuner3V<double>&);
+template void TunerAXY<float2>(int, char**, const size_t, const Tuner3V<float2>&);
+template void TunerAXY<double2>(int, char**, const size_t, const Tuner3V<double2>&);
 
 // =================================================================================================
 
diff --git a/src/tuning/xgemv.cc b/src/tuning/xgemv.cc
index e2d54729..dccd250c 100644
--- a/src/tuning/xgemv.cc
+++ b/src/tuning/xgemv.cc
@@ -8,6 +8,10 @@
 //   Cedric Nugteren <www.cedricnugteren.nl>
 //
 // This file implements an auto-tuner to tune the Xgemv OpenCL kernel. It uses the CLTune library.
+// Three variations of the kernel are tuned:
+// 1: The full version of the kernel
+// 2: The fast version for non-transposed matrices
+// 3: The fast version for transposed matrices
 //
 // =================================================================================================
 
@@ -23,43 +27,60 @@ namespace clblast {
 
 // The Xgemv auto-tuner
 template <typename T>
-void XgemvTune(const Arguments<T> &args,
+void XgemvTune(const Arguments<T> &args, const size_t variation,
                const std::vector<T> &a_mat, const std::vector<T> &x_vec, std::vector<T> &y_vec,
                cltune::Tuner &tuner) {
 
+  // Sets the kernel name and the layout argument
+  auto kernel_name = (variation == 1) ? "Xgemv" : ((variation == 2) ? "XgemvFast" : "XgemvFastRot");
+  auto a_rotated = (variation == 3) ? 1 : 0;
+
   // This points to the Xgemv kernel as found in the CLBlast library
   std::string common_source =
   #include "../src/kernels/common.opencl"
   std::string kernel_source =
   #include "../src/kernels/xgemv.opencl"
   auto sources = common_source + kernel_source;
-  auto id = tuner.AddKernelFromString(sources, "XgemvFast", {args.m}, {1});
+  auto id = tuner.AddKernelFromString(sources, kernel_name, {args.m}, {1});
   tuner.SetReferenceFromString(sources, "Xgemv", {args.m}, {64});
 
-  // Sets the tunable parameters and their possible values
-  tuner.AddParameter(id, "WGS", {64, 128, 256, 512, 1024, 1536, 2048});
-  tuner.AddParameter(id, "WPT", {1, 2, 4, 8});
-  tuner.AddParameter(id, "VW", {1, 2, 4, 8});
+  // Helper for the constraints
+  auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
+
+  // Sets the tunable parameters, their possible values, the adjusted thread sizes, and constraints
+  if (variation == 1) {
+    tuner.AddParameter(id, "WGS1", {64, 128, 256, 512, 1024, 1536, 2048});
+    tuner.AddParameter(id, "WPT1", {1, 2, 4, 8});
+    tuner.MulLocalSize(id, {"WGS1"});
+    tuner.DivGlobalSize(id, {"WPT1"});
+  }
+  else if (variation == 2) {
+    tuner.AddParameter(id, "WGS2", {64, 128, 256, 512, 1024, 1536, 2048});
+    tuner.AddParameter(id, "WPT2", {1, 2, 4, 8});
+    tuner.AddParameter(id, "VW2", {1, 2, 4, 8});
+    tuner.MulLocalSize(id, {"WGS2"});
+    tuner.DivGlobalSize(id, {"WPT2"});
+    tuner.AddConstraint(id, MultipleOfX, {"WPT2", "VW2"});
+  }
+  else if (variation == 3) {
+    tuner.AddParameter(id, "WGS3", {64, 128, 256, 512, 1024, 1536, 2048});
+    tuner.AddParameter(id, "WPT3", {1, 2, 4, 8});
+    tuner.AddParameter(id, "VW3", {1, 2, 4, 8});
+    tuner.MulLocalSize(id, {"WGS3"});
+    tuner.DivGlobalSize(id, {"WPT3"});
+    tuner.AddConstraint(id, MultipleOfX, {"WGS3", "VW3"});
+  }
 
   // Tests for a specific precision
   tuner.AddParameter(id, "PRECISION", {static_cast<size_t>(args.precision)});
   tuner.AddParameterReference("PRECISION", static_cast<size_t>(args.precision));
 
-  // Sets the constraints
-  auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
-  tuner.AddConstraint(id, MultipleOfX, {"WGS", "VW"});
-  tuner.AddConstraint(id, MultipleOfX, {"WPT", "VW"});
-
-  // Modifies the thread-sizes (local) based on the parameters
-  tuner.MulLocalSize(id, {"WGS"});
-  tuner.DivGlobalSize(id, {"WPT"});
-
   // Sets the function's arguments
   tuner.AddArgumentScalar(static_cast<int>(args.m));
   tuner.AddArgumentScalar(static_cast<int>(args.n));
   tuner.AddArgumentScalar(args.alpha);
   tuner.AddArgumentScalar(args.beta);
-  tuner.AddArgumentScalar(static_cast<int>(args.layout));
+  tuner.AddArgumentScalar(static_cast<int>(a_rotated));
   tuner.AddArgumentInput(a_mat);
   tuner.AddArgumentScalar(0);
   tuner.AddArgumentScalar(static_cast<int>(args.m));
@@ -75,12 +96,13 @@ void XgemvTune(const Arguments<T> &args,
 
 // Main function which calls the common client code with the routine-specific function as argument.
 void TunerXgemv(int argc, char *argv[]) {
+  auto num_variations = size_t{3};
   switch(GetPrecision(argc, argv)) {
     case Precision::kHalf: throw std::runtime_error("Unsupported precision mode");
-    case Precision::kSingle: TunerAXY<float>(argc, argv, XgemvTune<float>); break;
-    case Precision::kDouble: TunerAXY<double>(argc, argv, XgemvTune<double>); break;
-    case Precision::kComplexSingle: TunerAXY<float2>(argc, argv, XgemvTune<float2>); break;
-    case Precision::kComplexDouble: TunerAXY<double2>(argc, argv, XgemvTune<double2>); break;
+    case Precision::kSingle: TunerAXY<float>(argc, argv, num_variations, XgemvTune<float>); break;
+    case Precision::kDouble: TunerAXY<double>(argc, argv, num_variations, XgemvTune<double>); break;
+    case Precision::kComplexSingle: TunerAXY<float2>(argc, argv, num_variations, XgemvTune<float2>); break;
+    case Precision::kComplexDouble: TunerAXY<double2>(argc, argv, num_variations, XgemvTune<double2>); break;
   }
 }