Added a kernel selection database to select between the direct and indirect GEMM kernels

src/database/database.cpp

@@ -26,6 +26,7 @@
 #include "database/kernels/pad.hpp"
 #include "database/kernels/transpose.hpp"
 #include "database/kernels/padtranspose.hpp"
+#include "database/kernel_selection.hpp"
 
 namespace clblast {
 // =================================================================================================
@@ -43,7 +44,8 @@ const std::vector<Database::DatabaseEntry> Database::database = {
   CopyHalf, CopySingle, CopyDouble, CopyComplexSingle, CopyComplexDouble,
   PadHalf, PadSingle, PadDouble, PadComplexSingle, PadComplexDouble,
   TransposeHalf, TransposeSingle, TransposeDouble, TransposeComplexSingle, TransposeComplexDouble,
-  PadtransposeHalf, PadtransposeSingle, PadtransposeDouble, PadtransposeComplexSingle, PadtransposeComplexDouble
+  PadtransposeHalf, PadtransposeSingle, PadtransposeDouble, PadtransposeComplexSingle, PadtransposeComplexDouble,
+  KernelSelectionHalf, KernelSelectionSingle, KernelSelectionDouble, KernelSelectionComplexSingle, KernelSelectionComplexDouble
 };
 
 // =================================================================================================

src/database/database.hpp

@@ -80,6 +80,7 @@ class Database {
   static const DatabaseEntry PadHalf, PadSingle, PadDouble, PadComplexSingle, PadComplexDouble;
   static const DatabaseEntry TransposeHalf, TransposeSingle, TransposeDouble, TransposeComplexSingle, TransposeComplexDouble;
   static const DatabaseEntry PadtransposeHalf, PadtransposeSingle, PadtransposeDouble, PadtransposeComplexSingle, PadtransposeComplexDouble;
+  static const DatabaseEntry KernelSelectionHalf, KernelSelectionSingle, KernelSelectionDouble, KernelSelectionComplexSingle, KernelSelectionComplexDouble;
   static const std::vector<DatabaseEntry> database;
 
   // The constructor with a user-provided database overlay (potentially an empty vector)

src/database/kernel_selection.hpp

@@ -0,0 +1,79 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This determines when to switch between the direct (for small sizes) and in-direct GEMM kernel
+// with pre/post-processing kernels (for larger sizes). These can be set in a similar way as for the
+// regular kernel tuning parameters: they can be specific for a certain vendor or device or can use
+// some common default values.
+//
+// =================================================================================================
+
+namespace clblast {
+// =================================================================================================
+
+const Database::DatabaseEntry Database::KernelSelectionHalf = {
+  "KernelSelection", Precision::kHalf, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default",                                         { {"XGEMM_MIN_INDIRECT_SIZE",512*512*512} } },
+      }
+    },
+  }
+};
+
+// =================================================================================================
+
+const Database::DatabaseEntry Database::KernelSelectionSingle = {
+  "KernelSelection", Precision::kSingle, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default",                                         { {"XGEMM_MIN_INDIRECT_SIZE",512*512*512} } },
+      }
+    },
+  }
+};
+
+// =================================================================================================
+
+const Database::DatabaseEntry Database::KernelSelectionComplexSingle = {
+  "KernelSelection", Precision::kComplexSingle, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default",                                         { {"XGEMM_MIN_INDIRECT_SIZE",512*512*512} } },
+      }
+    },
+  }
+};
+
+// =================================================================================================
+
+const Database::DatabaseEntry Database::KernelSelectionDouble = {
+  "KernelSelection", Precision::kDouble, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default",                                         { {"XGEMM_MIN_INDIRECT_SIZE",512*512*512} } },
+      }
+    },
+  }
+};
+
+// =================================================================================================
+
+const Database::DatabaseEntry Database::KernelSelectionComplexDouble = {
+  "KernelSelection", Precision::kComplexDouble, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default",                                         { {"XGEMM_MIN_INDIRECT_SIZE",512*512*512} } },
+      }
+    },
+  }
+};
+
+// =================================================================================================
+} // namespace clblast

src/routines/level3/xgemm.cpp

@@ -22,7 +22,8 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xgemm<T>::Xgemm(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm", "XgemmDirect"},
+    Routine(queue, event, name,
+            {"Copy","Pad","Transpose","Padtranspose","Xgemm","XgemmDirect","KernelSelection"},
             PrecisionValue<T>()) {
   source_string_ =
     #include "../../kernels/level3/level3.opencl"
@@ -102,15 +103,15 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
   status = TestMatrixC(c_one, c_two, c_buffer, c_offset, c_ld);
   if (ErrorIn(status)) { return status; }
 
-  // Optionally runs the direct version of GEMM. TODO: Set this based on the arguments
-  const auto do_gemm_direct = true; // for now, for testing
-  if (do_gemm_direct) {
+  // Selects which version of GEMM to run
+  const auto do_gemm_direct = (m * n * k < db_["XGEMM_MIN_INDIRECT_SIZE"]);
+  if (do_gemm_direct) { // for small sizes (single kernel)
     return GemmDirect(m, n, k, alpha,
                       a_buffer, a_offset, a_ld, b_buffer, b_offset, b_ld, beta,
                       c_buffer, c_offset, c_ld,
                       a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate);
   }
-  else {
+  else { // for larger sizes (pre/post-processing plus a very fast kernel)
     return GemmIndirect(m, n, k, alpha,
                         a_buffer, a_offset, a_ld, b_buffer, b_offset, b_ld, beta,
                         c_buffer, c_offset, c_ld,