Updated the database to use the new TRSV and Invert tuners

CHANGELOG

@@ -4,6 +4,7 @@ Development (next version)
 - Made it possible to override the tuning parameters in the clients straight from JSON tuning files
 - Added OpenCL pre-processor to unroll loops and perform array-to-register promotions for compilers
   which don't do this themselves (ARM Mali) - greatly improves performance on these platforms
+- Added first tuners for the TRSV (block size) and TRSM (invert kernel) routines
 - Various minor fixes and enhancements
 - Added tuned parameters for various devices (see README)
 

CMakeLists.txt

@@ -195,7 +195,7 @@ endif()
 set(KERNELS copy_fast copy_pad transpose_fast transpose_pad xaxpy xdot xger
             xgemm xgemm_direct xgemv invert)
 set(DATABASES copy pad padtranspose transpose xaxpy xdot
-              xgemm xgemm_direct xgemv xgemv_fast xgemv_fast_rot xger
+              xgemm xgemm_direct xgemv xgemv_fast xgemv_fast_rot xger invert
               gemm_routine trsv_routine)
 set(ROUTINE_TUNERS xgemm xtrsv)
 set(LEVEL1_ROUTINES xswap xscal xcopy xaxpy xdot xdotu xdotc xnrm2 xasum xamax)

src/database/database.cpp

@@ -29,11 +29,11 @@
 #include "database/kernels/pad/pad.hpp"
 #include "database/kernels/transpose/transpose.hpp"
 #include "database/kernels/padtranspose/padtranspose.hpp"
+#include "database/kernels/invert/invert.hpp"
 
 #include "database/kernels/gemm_routine/gemm_routine.hpp"
+#include "database/kernels/trsv_routine/trsv_routine.hpp"
 
-#include "database/kernels/xtrsv.hpp"
-#include "database/kernels/invert.hpp"
 #include "database/apple_cpu_fallback.hpp"
 
 namespace clblast {
@@ -47,7 +47,6 @@ const std::vector<database::DatabaseEntry> Database::database = std::vector<data
   database::XgemvFastHalf, database::XgemvFastSingle, database::XgemvFastDouble, database::XgemvFastComplexSingle, database::XgemvFastComplexDouble,
   database::XgemvFastRotHalf, database::XgemvFastRotSingle, database::XgemvFastRotDouble, database::XgemvFastRotComplexSingle, database::XgemvFastRotComplexDouble,
   database::XgerHalf, database::XgerSingle, database::XgerDouble, database::XgerComplexSingle, database::XgerComplexDouble,
-  database::XtrsvHalf, database::XtrsvSingle, database::XtrsvDouble, database::XtrsvComplexSingle, database::XtrsvComplexDouble,
   database::XgemmHalf, database::XgemmSingle, database::XgemmDouble, database::XgemmComplexSingle, database::XgemmComplexDouble,
   database::XgemmDirectHalf, database::XgemmDirectSingle, database::XgemmDirectDouble, database::XgemmDirectComplexSingle, database::XgemmDirectComplexDouble,
   database::CopyHalf, database::CopySingle, database::CopyDouble, database::CopyComplexSingle, database::CopyComplexDouble,
@@ -55,7 +54,8 @@ const std::vector<database::DatabaseEntry> Database::database = std::vector<data
   database::TransposeHalf, database::TransposeSingle, database::TransposeDouble, database::TransposeComplexSingle, database::TransposeComplexDouble,
   database::PadtransposeHalf, database::PadtransposeSingle, database::PadtransposeDouble, database::PadtransposeComplexSingle, database::PadtransposeComplexDouble,
   database::InvertHalf, database::InvertSingle, database::InvertDouble, database::InvertComplexSingle, database::InvertComplexDouble,
-  database::GemmRoutineHalf, database::GemmRoutineSingle, database::GemmRoutineDouble, database::GemmRoutineComplexSingle, database::GemmRoutineComplexDouble
+  database::GemmRoutineHalf, database::GemmRoutineSingle, database::GemmRoutineDouble, database::GemmRoutineComplexSingle, database::GemmRoutineComplexDouble,
+  database::TrsvRoutineHalf, database::TrsvRoutineSingle, database::TrsvRoutineDouble, database::TrsvRoutineComplexSingle, database::TrsvRoutineComplexDouble
 };
 const std::vector<database::DatabaseEntry> Database::apple_cpu_fallback = std::vector<database::DatabaseEntry>{
   database::XaxpyApple, database::XdotApple,

src/database/kernels/invert.hpp

@@ -1,78 +0,0 @@
-
-// =================================================================================================
-// 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>
-//
-// Tuning parameters for the diagonal matrix inversion kernels
-//
-// =================================================================================================
-
-namespace clblast {
-namespace database {
-// =================================================================================================
-
-const DatabaseEntry InvertHalf = {
-  "Invert", Precision::kHalf, {"INTERNAL_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry InvertSingle = {
-  "Invert", Precision::kSingle, {"INTERNAL_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry InvertComplexSingle = {
-  "Invert", Precision::kComplexSingle, {"INTERNAL_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry InvertDouble = {
-  "Invert", Precision::kDouble, {"INTERNAL_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry InvertComplexDouble = {
-  "Invert", Precision::kComplexDouble, {"INTERNAL_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-} // namespace database
-} // namespace clblast

src/database/kernels/invert/invert.hpp

@@ -0,0 +1,14 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Invert' kernels.
+//
+// =================================================================================================
+
+#include "database/kernels/invert/invert_16.hpp"
+#include "database/kernels/invert/invert_32.hpp"
+#include "database/kernels/invert/invert_3232.hpp"
+#include "database/kernels/invert/invert_64.hpp"
+#include "database/kernels/invert/invert_6464.hpp"

src/database/kernels/invert/invert_16.hpp

@@ -0,0 +1,34 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Invert16' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry InvertHalf = {
+  "Invert", Precision::kHalf, {"INTERNAL_BLOCK_SIZE", "LOCALPAD", "TMMWGSX", "TMMWGSY"}, {
+    { // Intel GPUs
+      kDeviceTypeGPU, "Intel", {
+        { "default", {
+          { Name{"Intel(R) HD Graphics Skylake ULT GT2              "}, Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/invert/invert_32.hpp

@@ -0,0 +1,34 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Invert32' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry InvertSingle = {
+  "Invert", Precision::kSingle, {"INTERNAL_BLOCK_SIZE", "LOCALPAD", "TMMWGSX", "TMMWGSY"}, {
+    { // Intel GPUs
+      kDeviceTypeGPU, "Intel", {
+        { "default", {
+          { Name{"Intel(R) HD Graphics Skylake ULT GT2              "}, Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/invert/invert_3232.hpp

@@ -0,0 +1,34 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Invert3232' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry InvertComplexSingle = {
+  "Invert", Precision::kComplexSingle, {"INTERNAL_BLOCK_SIZE", "LOCALPAD", "TMMWGSX", "TMMWGSY"}, {
+    { // Intel GPUs
+      kDeviceTypeGPU, "Intel", {
+        { "default", {
+          { Name{"Intel(R) HD Graphics Skylake ULT GT2              "}, Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/invert/invert_64.hpp

@@ -0,0 +1,26 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Invert64' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry InvertDouble = {
+  "Invert", Precision::kDouble, {"INTERNAL_BLOCK_SIZE", "LOCALPAD", "TMMWGSX", "TMMWGSY"}, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/invert/invert_6464.hpp

@@ -0,0 +1,26 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Invert6464' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry InvertComplexDouble = {
+  "Invert", Precision::kComplexDouble, {"INTERNAL_BLOCK_SIZE", "LOCALPAD", "TMMWGSX", "TMMWGSY"}, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 16, 0, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/trsv_routine/trsv_routine.hpp

@@ -0,0 +1,14 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Trsv_Routine' kernels.
+//
+// =================================================================================================
+
+#include "database/kernels/trsv_routine/trsv_routine_16.hpp"
+#include "database/kernels/trsv_routine/trsv_routine_32.hpp"
+#include "database/kernels/trsv_routine/trsv_routine_3232.hpp"
+#include "database/kernels/trsv_routine/trsv_routine_64.hpp"
+#include "database/kernels/trsv_routine/trsv_routine_6464.hpp"

src/database/kernels/trsv_routine/trsv_routine_16.hpp

@@ -0,0 +1,26 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Trsv_Routine16' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry TrsvRoutineHalf = {
+  "TrsvRoutine", Precision::kHalf, {"TRSV_BLOCK_SIZE"}, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/trsv_routine/trsv_routine_32.hpp

@@ -0,0 +1,34 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Trsv_Routine32' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry TrsvRoutineSingle = {
+  "TrsvRoutine", Precision::kSingle, {"TRSV_BLOCK_SIZE"}, {
+    { // Intel GPUs
+      kDeviceTypeGPU, "Intel", {
+        { "default", {
+          { Name{"Intel(R) HD Graphics Skylake ULT GT2              "}, Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+          { kDeviceNameDefault                                        , Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/trsv_routine/trsv_routine_3232.hpp

@@ -0,0 +1,34 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Trsv_Routine3232' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry TrsvRoutineComplexSingle = {
+  "TrsvRoutine", Precision::kComplexSingle, {"TRSV_BLOCK_SIZE"}, {
+    { // Intel GPUs
+      kDeviceTypeGPU, "Intel", {
+        { "default", {
+          { Name{"Intel(R) HD Graphics Skylake ULT GT2              "}, Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+          { kDeviceNameDefault                                        , Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/trsv_routine/trsv_routine_64.hpp

@@ -0,0 +1,26 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Trsv_Routine64' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry TrsvRoutineDouble = {
+  "TrsvRoutine", Precision::kDouble, {"TRSV_BLOCK_SIZE"}, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/trsv_routine/trsv_routine_6464.hpp

@@ -0,0 +1,26 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. It
+// is auto-generated by the 'scripts/database/database.py' Python script.
+//
+// This file populates the database with best-found tuning parameters for the 'Trsv_Routine6464' kernels.
+//
+// =================================================================================================
+
+namespace clblast {
+namespace database {
+
+const DatabaseEntry TrsvRoutineComplexDouble = {
+  "TrsvRoutine", Precision::kComplexDouble, {"TRSV_BLOCK_SIZE"}, {
+    { // Default
+      kDeviceTypeAll, "default", {
+        { "default", {
+          { kDeviceNameDefault                                        , Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
+        } },
+      }
+    },
+  }
+};
+
+} // namespace database
+} // namespace clblast

src/database/kernels/xtrsv.hpp

@@ -1,78 +0,0 @@
-
-// =================================================================================================
-// 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 file populates the database with best-found tuning parameters for the 'Xtrsv' kernels.
-//
-// =================================================================================================
-
-namespace clblast {
-namespace database {
-// =================================================================================================
-
-const DatabaseEntry XtrsvHalf = {
-  "Xtrsv", Precision::kHalf, {"TRSV_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry XtrsvSingle = {
-  "Xtrsv", Precision::kSingle, {"TRSV_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry XtrsvComplexSingle = {
-  "Xtrsv", Precision::kComplexSingle, {"TRSV_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry XtrsvDouble = {
-  "Xtrsv", Precision::kDouble, {"TRSV_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-
-const DatabaseEntry XtrsvComplexDouble = {
-  "Xtrsv", Precision::kComplexDouble, {"TRSV_BLOCK_SIZE"}, {
-    { // Default
-      kDeviceTypeAll, "default", {
-        { "default", { { kDeviceNameDefault, Params{ 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } } } },
-      }
-    },
-  }
-};
-
-// =================================================================================================
-} // namespace database
-} // namespace clblast

src/routines/level2/xgemv.cpp

@@ -22,7 +22,7 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xgemv<T>::Xgemv(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xgemv", "XgemvFast", "XgemvFastRot", "Xtrsv"}, PrecisionValue<T>(), {}, {
+    Routine(queue, event, name, {"Xgemv", "XgemvFast", "XgemvFastRot", "TrsvRoutine"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level2/xgemv.opencl"
     #include "../../kernels/level2/xgemv_fast.opencl"
     #include "../../kernels/level2/xtrsv.opencl"

src/tuning/routines/xtrsv.cpp

@@ -59,7 +59,7 @@ void TuneXtrsv(int argc, char* argv[]) {
 
   // Values for the block size
   const auto from = size_t{8};
-  const auto to = size_t{64 + 1};
+  const auto to = size_t{32 + 1};
   const auto step = size_t{8};
 
   // OpenCL initialisation