summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2016-01-30 11:46:45 +0100
committerCedric Nugteren <web@cedricnugteren.nl>2016-01-30 11:46:45 +0100
commitdeb58709c8af78905d65238ddc1d3f2a5df97500 (patch)
treee5e0c9e26db8717e4051efd6d48ddafb02697c79
parente4e3663e614b94ca81e92672e14463f22b975242 (diff)
parent276e772a2c672ad868ba57e06d55e4991c793207 (diff)
Merge branch 'tuning_database' into development
is merge is necessary,
Diffstat
-rw-r--r--include/internal/database.h11
-rw-r--r--include/internal/database/copy.h106
-rw-r--r--include/internal/database/pad.h106
-rw-r--r--include/internal/database/padtranspose.h114
-rw-r--r--include/internal/database/transpose.h110
-rw-r--r--include/internal/database/xaxpy.h107
-rw-r--r--include/internal/database/xdot.h91
-rw-r--r--include/internal/database/xgemm.h111
-rw-r--r--include/internal/database/xgemv.h101
-rw-r--r--include/internal/tuning.h8
-rw-r--r--scripts/database/database.dbbin0 -> 2094065 bytes
-rw-r--r--scripts/database/database.py (renamed from src/database.py)89
-rw-r--r--src/database.cc8
-rw-r--r--src/routines/level3/xgemm.cc2
-rw-r--r--src/routines/level3/xher2k.cc2
-rw-r--r--src/routines/level3/xherk.cc2
-rw-r--r--src/routines/level3/xsyr2k.cc2
-rw-r--r--src/routines/level3/xsyrk.cc2
-rw-r--r--src/tuning/xgemv.cc8
19 files changed, 398 insertions, 582 deletions
diff --git a/include/internal/database.h b/include/internal/database.h
index 1ac0e646..9107f978 100644
--- a/include/internal/database.h
+++ b/include/internal/database.h
@@ -56,15 +56,8 @@ class Database {
static constexpr auto kDeviceTypeAll = "default";
// The OpenCL device vendors
- static constexpr auto kDeviceVendorNVIDIA = "NVIDIA Corporation";
- static constexpr auto kDeviceVendorAMD = "Advanced Micro Devices, Inc.";
- static constexpr auto kDeviceVendorIntel = "Intel";
static constexpr auto kDeviceVendorAll = "default";
- // The OpenCL device names
- static constexpr auto kDefaultDevice = "default";
-
-
// The database consists of separate database entries, stored together in a vector
static const DatabaseEntry XaxpySingle, XaxpyDouble, XaxpyComplexSingle, XaxpyComplexDouble;
static const DatabaseEntry XdotSingle, XdotDouble, XdotComplexSingle, XdotComplexDouble;
@@ -72,8 +65,8 @@ class Database {
static const DatabaseEntry XgemmSingle, XgemmDouble, XgemmComplexSingle, XgemmComplexDouble;
static const DatabaseEntry CopySingle, CopyDouble, CopyComplexSingle, CopyComplexDouble;
static const DatabaseEntry PadSingle, PadDouble, PadComplexSingle, PadComplexDouble;
- static const DatabaseEntry TraSingle, TraDouble, TraComplexSingle, TraComplexDouble;
- static const DatabaseEntry PadTraSingle, PadTraDouble, PadTraComplexSingle, PadTraComplexDouble;
+ static const DatabaseEntry TransposeSingle, TransposeDouble, TransposeComplexSingle, TransposeComplexDouble;
+ static const DatabaseEntry PadtransposeSingle, PadtransposeDouble, PadtransposeComplexSingle, PadtransposeComplexDouble;
static const std::vector<DatabaseEntry> database;
// The constructor
diff --git a/include/internal/database/copy.h b/include/internal/database/copy.h
index 541a352b..3bd85fa6 100644
--- a/include/internal/database/copy.h
+++ b/include/internal/database/copy.h
@@ -5,9 +5,9 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the Copy kernels.
+// This file populates the database with best-found tuning parameters for the 'Copy' kernels.
//
// =================================================================================================
@@ -16,26 +16,21 @@ namespace clblast {
const Database::DatabaseEntry Database::CopySingle = {
"Copy", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",2} } },
- { "Tesla K20m", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_WPT",2}, {"COPY_VW",4} } },
- { "Tesla K40m", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_WPT",4}, {"COPY_VW",4} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_WPT",4}, {"COPY_VW",2} } },
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } },
+ { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",4} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",2} } },
+ { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",2} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } },
}
},
}
@@ -43,27 +38,23 @@ const Database::DatabaseEntry Database::CopySingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::CopyDouble = {
- "Copy", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
- { "Tesla K20m", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",2} } },
- { "Tesla K40m", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",2} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_WPT",2}, {"COPY_VW",4} } },
+const Database::DatabaseEntry Database::CopyComplexSingle = {
+ "Copy", Precision::kComplexSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } },
+ { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } },
+ { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } },
}
},
}
@@ -71,28 +62,17 @@ const Database::DatabaseEntry Database::CopyDouble = {
// =================================================================================================
-const Database::DatabaseEntry Database::CopyComplexSingle = {
- "Copy", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_WPT",1}, {"COPY_VW",1} } },
- { "Tesla K20m", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_WPT",2}, {"COPY_VW",1} } },
- { "Tesla K40m", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
+const Database::DatabaseEntry Database::CopyDouble = {
+ "Copy", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } },
+ { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } },
}
},
}
@@ -102,25 +82,15 @@ const Database::DatabaseEntry Database::CopyComplexSingle = {
const Database::DatabaseEntry Database::CopyComplexDouble = {
"Copy", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
- { "Tesla K20m", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_WPT",1}, {"COPY_VW",1} } },
- { "Tesla K40m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_WPT",4}, {"COPY_VW",2} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } },
+ { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_WPT",1}, {"COPY_VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } },
}
},
}
diff --git a/include/internal/database/pad.h b/include/internal/database/pad.h
index 4a599648..d833a934 100644
--- a/include/internal/database/pad.h
+++ b/include/internal/database/pad.h
@@ -5,9 +5,9 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the Pad kernels.
+// This file populates the database with best-found tuning parameters for the 'Pad' kernels.
//
// =================================================================================================
@@ -16,26 +16,21 @@ namespace clblast {
const Database::DatabaseEntry Database::PadSingle = {
"Pad", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",4} } },
- { "Tesla K20m", { {"PAD_DIMX",16}, {"PAD_DIMY",32}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } },
- { "Tesla K40m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } },
+ { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+ { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
}
},
}
@@ -43,27 +38,23 @@ const Database::DatabaseEntry Database::PadSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::PadDouble = {
- "Pad", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- { "Tesla K20m", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- { "Tesla K40m", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
+const Database::DatabaseEntry Database::PadComplexSingle = {
+ "Pad", Precision::kComplexSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",4} } },
+ { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",4} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+ { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
}
},
}
@@ -71,28 +62,17 @@ const Database::DatabaseEntry Database::PadDouble = {
// =================================================================================================
-const Database::DatabaseEntry Database::PadComplexSingle = {
- "Pad", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- { "Tesla K20m", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
- { "Tesla K40m", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+const Database::DatabaseEntry Database::PadDouble = {
+ "Pad", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
+ { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } },
}
},
}
@@ -102,25 +82,15 @@ const Database::DatabaseEntry Database::PadComplexSingle = {
const Database::DatabaseEntry Database::PadComplexDouble = {
"Pad", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- { "Tesla K20m", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- { "Tesla K40m", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PAD_DIMX",8}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+ { "default", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } },
}
},
}
diff --git a/include/internal/database/padtranspose.h b/include/internal/database/padtranspose.h
index 53226c1d..dacc693f 100644
--- a/include/internal/database/padtranspose.h
+++ b/include/internal/database/padtranspose.h
@@ -5,37 +5,32 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the PadTranspose kernels.
+// This file populates the database with best-found tuning parameters for the 'Padtranspose' kernels.
//
// =================================================================================================
namespace clblast {
// =================================================================================================
-const Database::DatabaseEntry Database::PadTraSingle = {
- "PadTranspose", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PADTRA_TILE",16}, {"PADTRA_WPT",2}, {"PADTRA_PAD",1} } },
- { "Tesla K20m", { {"PADTRA_TILE",16}, {"PADTRA_WPT",2}, {"PADTRA_PAD",1} } },
- { "Tesla K40m", { {"PADTRA_TILE",32}, {"PADTRA_WPT",2}, {"PADTRA_PAD",1} } },
+const Database::DatabaseEntry Database::PadtransposeSingle = {
+ "Padtranspose", Precision::kSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } },
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } },
}
},
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PADTRA_TILE",16}, {"PADTRA_WPT",4}, {"PADTRA_PAD",0} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"PADTRA_TILE",16}, {"PADTRA_WPT",2}, {"PADTRA_PAD",0} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",2} } },
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",2} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } },
}
},
}
@@ -43,27 +38,23 @@ const Database::DatabaseEntry Database::PadTraSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::PadTraDouble = {
- "PadTranspose", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",1} } },
- { "Tesla K20m", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",1} } },
- { "Tesla K40m", { {"PADTRA_TILE",16}, {"PADTRA_WPT",2}, {"PADTRA_PAD",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PADTRA_TILE",8}, {"PADTRA_WPT",4}, {"PADTRA_PAD",0} } },
+const Database::DatabaseEntry Database::PadtransposeComplexSingle = {
+ "Padtranspose", Precision::kComplexSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } },
+ { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
}
},
}
@@ -71,28 +62,17 @@ const Database::DatabaseEntry Database::PadTraDouble = {
// =================================================================================================
-const Database::DatabaseEntry Database::PadTraComplexSingle = {
- "PadTranspose", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",1} } },
- { "Tesla K20m", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",1} } },
- { "Tesla K40m", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",0} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PADTRA_TILE",16}, {"PADTRA_WPT",2}, {"PADTRA_PAD",0} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"PADTRA_TILE",16}, {"PADTRA_WPT",2}, {"PADTRA_PAD",0} } },
+const Database::DatabaseEntry Database::PadtransposeDouble = {
+ "Padtranspose", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
}
},
}
@@ -100,27 +80,17 @@ const Database::DatabaseEntry Database::PadTraComplexSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::PadTraComplexDouble = {
- "PadTranspose", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",1} } },
- { "Tesla K20m", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",1} } },
- { "Tesla K40m", { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"PADTRA_TILE",8}, {"PADTRA_WPT",2}, {"PADTRA_PAD",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+const Database::DatabaseEntry Database::PadtransposeComplexDouble = {
+ "Padtranspose", Precision::kComplexDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"PADTRA_TILE",16}, {"PADTRA_WPT",1}, {"PADTRA_PAD",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } },
}
},
}
diff --git a/include/internal/database/transpose.h b/include/internal/database/transpose.h
index 1d12a13e..46a38bc2 100644
--- a/include/internal/database/transpose.h
+++ b/include/internal/database/transpose.h
@@ -5,37 +5,32 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the Transpose kernels.
+// This file populates the database with best-found tuning parameters for the 'Transpose' kernels.
//
// =================================================================================================
namespace clblast {
// =================================================================================================
-const Database::DatabaseEntry Database::TraSingle = {
+const Database::DatabaseEntry Database::TransposeSingle = {
"Transpose", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"TRA_DIM",16}, {"TRA_WPT",2}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- { "Tesla K20m", { {"TRA_DIM",16}, {"TRA_WPT",2}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_WPT",2}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"TRA_DIM",16}, {"TRA_WPT",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1} } },
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } },
+ { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"TRA_DIM",8}, {"TRA_WPT",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } },
+ { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } },
}
},
}
@@ -43,27 +38,23 @@ const Database::DatabaseEntry Database::TraSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::TraDouble = {
- "Transpose", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"TRA_DIM",8}, {"TRA_WPT",2}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- { "Tesla K20m", { {"TRA_DIM",16}, {"TRA_WPT",2}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_WPT",2}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"TRA_DIM",16}, {"TRA_WPT",2}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1} } },
+const Database::DatabaseEntry Database::TransposeComplexSingle = {
+ "Transpose", Precision::kComplexSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } },
+ { "default", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } },
+ { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } },
}
},
}
@@ -71,28 +62,17 @@ const Database::DatabaseEntry Database::TraDouble = {
// =================================================================================================
-const Database::DatabaseEntry Database::TraComplexSingle = {
- "Transpose", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- { "Tesla K20m", { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0} } },
- { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"TRA_DIM",16}, {"TRA_WPT",2}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
+const Database::DatabaseEntry Database::TransposeDouble = {
+ "Transpose", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } },
+ { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } },
}
},
}
@@ -100,27 +80,17 @@ const Database::DatabaseEntry Database::TraComplexSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::TraComplexDouble = {
+const Database::DatabaseEntry Database::TransposeComplexDouble = {
"Transpose", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"TRA_DIM",8}, {"TRA_WPT",1}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- { "Tesla K20m", { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } },
+ { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"TRA_DIM",16}, {"TRA_WPT",1}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } },
}
},
}
diff --git a/include/internal/database/xaxpy.h b/include/internal/database/xaxpy.h
index 058e3c0a..783e142d 100644
--- a/include/internal/database/xaxpy.h
+++ b/include/internal/database/xaxpy.h
@@ -5,9 +5,9 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the Xaxpy kernels.
+// This file populates the database with best-found tuning parameters for the 'Xaxpy' kernels.
//
// =================================================================================================
@@ -16,26 +16,21 @@ namespace clblast {
const Database::DatabaseEntry Database::XaxpySingle = {
"Xaxpy", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS",128}, {"WPT",1}, {"VW",2} } },
- { "Tesla K20m", { {"WGS",128}, {"WPT",2}, {"VW",2} } },
- { "Tesla K40m", { {"WGS",128}, {"WPT",1}, {"VW",4} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS",64}, {"WPT",1}, {"VW",2} } },
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"VW",1}, {"WGS",64}, {"WPT",1} } },
+ { "default", { {"VW",1}, {"WGS",64}, {"WPT",1} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"WGS",512}, {"WPT",1}, {"VW",1} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",4}, {"WGS",128}, {"WPT",1} } },
+ { "default", { {"VW",4}, {"WGS",128}, {"WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS",128}, {"WPT",1}, {"VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",1}, {"WGS",64}, {"WPT",1} } },
}
},
}
@@ -43,55 +38,41 @@ const Database::DatabaseEntry Database::XaxpySingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::XaxpyDouble = {
- "Xaxpy", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS",128}, {"WPT",1}, {"VW",1} } },
- { "Tesla K20m", { {"WGS",512}, {"WPT",1}, {"VW",2} } },
- { "Tesla K40m", { {"WGS",64}, {"WPT",1}, {"VW",2} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS",256}, {"WPT",1}, {"VW",1} } },
+const Database::DatabaseEntry Database::XaxpyComplexSingle = {
+ "Xaxpy", Precision::kComplexSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"VW",2}, {"WGS",128}, {"WPT",1} } },
+ { "default", { {"VW",2}, {"WGS",128}, {"WPT",1} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",1}, {"WGS",128}, {"WPT",1} } },
+ { "default", { {"VW",1}, {"WGS",128}, {"WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS",128}, {"WPT",1}, {"VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",1}, {"WGS",128}, {"WPT",1} } },
}
},
}
};
+
// =================================================================================================
-const Database::DatabaseEntry Database::XaxpyComplexSingle = {
- "Xaxpy", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS",256}, {"WPT",1}, {"VW",1} } },
- { "Tesla K20m", { {"WGS",128}, {"WPT",1}, {"VW",1} } },
- { "Tesla K40m", { {"WGS",128}, {"WPT",2}, {"VW",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS",64}, {"WPT",1}, {"VW",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"WGS",256}, {"WPT",1}, {"VW",1} } },
+const Database::DatabaseEntry Database::XaxpyDouble = {
+ "Xaxpy", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",2}, {"WGS",128}, {"WPT",1} } },
+ { "default", { {"VW",2}, {"WGS",128}, {"WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS",128}, {"WPT",1}, {"VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",2}, {"WGS",128}, {"WPT",1} } },
}
},
}
@@ -101,25 +82,15 @@ const Database::DatabaseEntry Database::XaxpyComplexSingle = {
const Database::DatabaseEntry Database::XaxpyComplexDouble = {
"Xaxpy", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS",128}, {"WPT",2}, {"VW",1} } },
- { "Tesla K20m", { {"WGS",256}, {"WPT",1}, {"VW",1} } },
- { "Tesla K40m", { {"WGS",64}, {"WPT",2}, {"VW",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS",64}, {"WPT",1}, {"VW",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",1}, {"WGS",64}, {"WPT",1} } },
+ { "default", { {"VW",1}, {"WGS",64}, {"WPT",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS",128}, {"WPT",1}, {"VW",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",1}, {"WGS",64}, {"WPT",1} } },
}
},
}
diff --git a/include/internal/database/xdot.h b/include/internal/database/xdot.h
index 05841eb7..66a5231e 100644
--- a/include/internal/database/xdot.h
+++ b/include/internal/database/xdot.h
@@ -5,9 +5,9 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the Xdot kernels.
+// This file populates the database with best-found tuning parameters for the 'Xdot' kernels.
//
// =================================================================================================
@@ -16,22 +16,21 @@ namespace clblast {
const Database::DatabaseEntry Database::XdotSingle = {
"Xdot", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"VW",1}, {"WGS1",512}, {"WGS2",32} } },
+ { "default", { {"VW",1}, {"WGS1",512}, {"WGS2",32} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"WGS1",512}, {"WGS2",512} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
+ { "default", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WGS2",64} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",1}, {"WGS1",128}, {"WGS2",32} } },
}
},
}
@@ -39,47 +38,41 @@ const Database::DatabaseEntry Database::XdotSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::XdotDouble = {
- "Xdot", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
+const Database::DatabaseEntry Database::XdotComplexSingle = {
+ "Xdot", Precision::kComplexSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"VW",1}, {"WGS1",512}, {"WGS2",1024} } },
+ { "default", { {"VW",1}, {"WGS1",512}, {"WGS2",1024} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
+ { "default", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WGS2",64} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
}
},
}
};
+
// =================================================================================================
-const Database::DatabaseEntry Database::XdotComplexSingle = {
- "Xdot", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"WGS1",512}, {"WGS2",512} } },
+const Database::DatabaseEntry Database::XdotDouble = {
+ "Xdot", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",1}, {"WGS1",256}, {"WGS2",1024} } },
+ { "default", { {"VW",1}, {"WGS1",256}, {"WGS2",1024} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WGS2",64} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",1}, {"WGS1",256}, {"WGS2",1024} } },
}
},
}
@@ -89,21 +82,15 @@ const Database::DatabaseEntry Database::XdotComplexSingle = {
const Database::DatabaseEntry Database::XdotComplexDouble = {
"Xdot", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
+ { "default", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WGS2",64} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"VW",1}, {"WGS1",128}, {"WGS2",1024} } },
}
},
}
diff --git a/include/internal/database/xgemm.h b/include/internal/database/xgemm.h
index 49598c8c..9fbd8fbb 100644
--- a/include/internal/database/xgemm.h
+++ b/include/internal/database/xgemm.h
@@ -5,9 +5,9 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the Xgemm kernels.
+// This file populates the database with best-found tuning parameters for the 'Xgemm' kernels.
//
// =================================================================================================
@@ -16,27 +16,21 @@ namespace clblast {
const Database::DatabaseEntry Database::XgemmSingle = {
"Xgemm", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"MWG",128}, {"NWG",64}, {"KWG",32}, {"MDIMC",16}, {"NDIMC",16}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",2}, {"VWM",2}, {"VWN",2}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",1} } },
- { "Tesla K20m", { {"MWG",128}, {"NWG",64}, {"KWG",16}, {"MDIMC",16}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",32}, {"KWI",2}, {"VWM",4}, {"VWN",1}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",1} } },
- { "Tesla K40m", { {"MWG",128}, {"NWG",128}, {"KWG",16}, {"MDIMC",16}, {"NDIMC",16}, {"MDIMA",32}, {"NDIMB",16}, {"KWI",8}, {"VWM",2}, {"VWN",1}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",1} } },
- { kDefaultDevice, { {"MWG",128}, {"NWG",64}, {"KWG",16}, {"MDIMC",16}, {"NDIMC",16}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",2}, {"VWM",2}, {"VWN",1}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"MWG",128}, {"NWG",128}, {"KWG",32}, {"MDIMC",16}, {"NDIMC",16}, {"MDIMA",32}, {"NDIMB",8}, {"KWI",2}, {"VWM",4}, {"VWN",4}, {"STRM",1}, {"STRN",1}, {"SA",1}, {"SB",1} } },
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",1} } },
+ { "default", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",1} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"MWG",64}, {"NWG",64}, {"KWG",32}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",8}, {"KWI",8}, {"VWM",4}, {"VWN",4}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",0} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",4} } },
+ { "default", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",4} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"MWG",32}, {"NWG",32}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",1}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",0}, {"SB",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",1} } },
}
},
}
@@ -44,28 +38,23 @@ const Database::DatabaseEntry Database::XgemmSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::XgemmDouble = {
- "Xgemm", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"MWG",32}, {"NWG",64}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",16}, {"MDIMA",16}, {"NDIMB",32}, {"KWI",2}, {"VWM",1}, {"VWN",2}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",1} } },
- { "Tesla K20m", { {"MWG",64}, {"NWG",128}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",32}, {"MDIMA",32}, {"NDIMB",32}, {"KWI",8}, {"VWM",2}, {"VWN",4}, {"STRM",1}, {"STRN",1}, {"SA",1}, {"SB",1} } },
- { "Tesla K40m", { {"MWG",64}, {"NWG",64}, {"KWG",16}, {"MDIMC",16}, {"NDIMC",16}, {"MDIMA",16}, {"NDIMB",32}, {"KWI",2}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",1} } },
- { kDefaultDevice, { {"MWG",32}, {"NWG",64}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",16}, {"MDIMA",16}, {"NDIMB",32}, {"KWI",2}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",1}, {"SB",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"MWG",128}, {"NWG",64}, {"KWG",16}, {"MDIMC",32}, {"NDIMC",8}, {"MDIMA",32}, {"NDIMB",16}, {"KWI",8}, {"VWM",1}, {"VWN",2}, {"STRM",1}, {"STRN",0}, {"SA",0}, {"SB",0} } },
+const Database::DatabaseEntry Database::XgemmComplexSingle = {
+ "Xgemm", Precision::kComplexSingle, {
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"KWG",32}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
+ { "default", { {"KWG",32}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
+ { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"MWG",32}, {"NWG",32}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",1}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",0}, {"SB",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
}
},
}
@@ -73,29 +62,17 @@ const Database::DatabaseEntry Database::XgemmDouble = {
// =================================================================================================
-const Database::DatabaseEntry Database::XgemmComplexSingle = {
- "Xgemm", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"MWG",32}, {"NWG",64}, {"KWG",16}, {"MDIMC",16}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",2}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",1}, {"SB",1} } },
- { "Tesla K20m", { {"MWG",32}, {"NWG",64}, {"KWG",16}, {"MDIMC",16}, {"NDIMC",8}, {"MDIMA",8}, {"NDIMB",8}, {"KWI",8}, {"VWM",2}, {"VWN",2}, {"STRM",1}, {"STRN",0}, {"SA",1}, {"SB",0} } },
- { "Tesla K40m", { {"MWG",32}, {"NWG",64}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",32}, {"MDIMA",32}, {"NDIMB",16}, {"KWI",8}, {"VWM",1}, {"VWN",1}, {"STRM",0}, {"STRN",1}, {"SA",1}, {"SB",1} } },
- { kDefaultDevice, { {"MWG",32}, {"NWG",64}, {"KWG",16}, {"MDIMC",16}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",2}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",1}, {"SB",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"MWG",16}, {"NWG",64}, {"KWG",32}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",8}, {"NDIMB",16}, {"KWI",2}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",1}, {"SB",0} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"MWG",32}, {"NWG",32}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",1}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",0}, {"SB",0} } },
+const Database::DatabaseEntry Database::XgemmDouble = {
+ "Xgemm", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",4} } },
+ { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",4} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"MWG",32}, {"NWG",32}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",1}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",0}, {"SB",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",4} } },
}
},
}
@@ -105,29 +82,19 @@ const Database::DatabaseEntry Database::XgemmComplexSingle = {
const Database::DatabaseEntry Database::XgemmComplexDouble = {
"Xgemm", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"MWG",16}, {"NWG",32}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",8}, {"KWI",2}, {"VWM",1}, {"VWN",4}, {"STRM",1}, {"STRN",0}, {"SA",0}, {"SB",0} } },
- { "Tesla K20m", { {"MWG",16}, {"NWG",128}, {"KWG",32}, {"MDIMC",8}, {"NDIMC",32}, {"MDIMA",8}, {"NDIMB",32}, {"KWI",2}, {"VWM",1}, {"VWN",4}, {"STRM",1}, {"STRN",1}, {"SA",1}, {"SB",0} } },
- { "Tesla K40m", { {"MWG",32}, {"NWG",32}, {"KWG",16}, {"MDIMC",32}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",32}, {"KWI",8}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",0}, {"SB",1} } },
- { kDefaultDevice, { {"MWG",16}, {"NWG",32}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",8}, {"KWI",2}, {"VWM",1}, {"VWN",4}, {"STRM",1}, {"STRN",0}, {"SA",0}, {"SB",0} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"MWG",128}, {"NWG",32}, {"KWG",16}, {"MDIMC",32}, {"NDIMC",8}, {"MDIMA",32}, {"NDIMB",16}, {"KWI",8}, {"VWM",2}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",0}, {"SB",0} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"KWG",16}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
+ { "default", { {"KWG",16}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"MWG",32}, {"NWG",32}, {"KWG",16}, {"MDIMC",8}, {"NDIMC",8}, {"MDIMA",16}, {"NDIMB",16}, {"KWI",1}, {"VWM",1}, {"VWN",1}, {"STRM",1}, {"STRN",1}, {"SA",0}, {"SB",0} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"KWG",16}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
}
},
}
};
+
// =================================================================================================
} // namespace clblast
diff --git a/include/internal/database/xgemv.h b/include/internal/database/xgemv.h
index c315500f..ae9fbf30 100644
--- a/include/internal/database/xgemv.h
+++ b/include/internal/database/xgemv.h
@@ -5,9 +5,9 @@
// width of 100 characters per line.
//
// Author(s):
-// Cedric Nugteren <www.cedricnugteren.nl>
+// Database generator <database.py>
//
-// This file populates the database with best-found tuning parameters for the Xgemv kernels.
+// This file populates the database with best-found tuning parameters for the 'Xgemv' kernels.
//
// =================================================================================================
@@ -16,26 +16,21 @@ namespace clblast {
const Database::DatabaseEntry Database::XgemvSingle = {
"Xgemv", Precision::kSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K20m", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K40m", { {"WGS1",256}, {"WPT1",1}, {"WGS2",256}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",4} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
+ { // Intel GPUs
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"WGS1",64}, {"WPT1",2}, {"VW2",1}, {"WGS2",128}, {"WPT2",2}, {"VW3",4}, {"WGS3",64}, {"WPT3",8} } },
+ { "default", { {"WGS1",64}, {"WPT1",2}, {"VW2",1}, {"WGS2",128}, {"WPT2",2}, {"VW3",4}, {"WGS3",64}, {"WPT3",8} } },
}
},
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"WGS1",256}, {"WPT1",2}, {"WGS2",64}, {"WPT2",4}, {"VW2",4}, {"WGS3",256}, {"WPT3",2}, {"VW3",8} } },
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",256}, {"WPT2",1}, {"VW3",1}, {"WGS3",128}, {"WPT3",1} } },
+ { "default", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",256}, {"WPT2",1}, {"VW3",1}, {"WGS3",128}, {"WPT3",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"WGS1",64}, {"WPT1",1}, {"VW2",1}, {"WGS2",128}, {"WPT2",1}, {"VW3",1}, {"WGS3",64}, {"WPT3",1} } },
}
},
}
@@ -43,55 +38,35 @@ const Database::DatabaseEntry Database::XgemvSingle = {
// =================================================================================================
-const Database::DatabaseEntry Database::XgemvDouble = {
- "Xgemv", Precision::kDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K20m", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K40m", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- }
- },
+const Database::DatabaseEntry Database::XgemvComplexSingle = {
+ "Xgemv", Precision::kComplexSingle, {
{ // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
+ kDeviceTypeGPU, "Intel", {
+ { "Iris", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",64}, {"WPT2",1}, {"VW3",1}, {"WGS3",64}, {"WPT3",1} } },
+ { "default", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",64}, {"WPT2",1}, {"VW3",1}, {"WGS3",64}, {"WPT3",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",64}, {"WPT2",1}, {"VW3",1}, {"WGS3",64}, {"WPT3",1} } },
}
},
}
};
+
// =================================================================================================
-const Database::DatabaseEntry Database::XgemvComplexSingle = {
- "Xgemv", Precision::kComplexSingle, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K20m", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K40m", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- { "Iris", { {"WGS1",256}, {"WPT1",1}, {"WGS2",64}, {"WPT2",4}, {"VW2",2}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
+const Database::DatabaseEntry Database::XgemvDouble = {
+ "Xgemv", Precision::kDouble, {
+ { // NVIDIA Corporation GPUs
+ kDeviceTypeGPU, "NVIDIA Corporation", {
+ { "Tesla K40m", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",256}, {"WPT2",1}, {"VW3",1}, {"WGS3",256}, {"WPT3",1} } },
+ { "default", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",256}, {"WPT2",1}, {"VW3",1}, {"WGS3",256}, {"WPT3",1} } },
}
},
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"WGS1",256}, {"WPT1",1}, {"VW2",1}, {"WGS2",256}, {"WPT2",1}, {"VW3",1}, {"WGS3",256}, {"WPT3",1} } },
}
},
}
@@ -101,25 +76,9 @@ const Database::DatabaseEntry Database::XgemvComplexSingle = {
const Database::DatabaseEntry Database::XgemvComplexDouble = {
"Xgemv", Precision::kComplexDouble, {
- { // NVIDIA GPUs
- kDeviceTypeGPU, kDeviceVendorNVIDIA, {
- { "GeForce GTX 480", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K20m", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- { "Tesla K40m", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- }
- },
- { // AMD GPUs
- kDeviceTypeGPU, kDeviceVendorAMD, {
- { "Tahiti", { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
- }
- },
- { // Intel GPUs
- kDeviceTypeGPU, kDeviceVendorIntel, {
- }
- },
{ // Default
- kDeviceTypeAll, kDeviceVendorAll, {
- { kDefaultDevice, { {"WGS1",64}, {"WPT1",1}, {"WGS2",64}, {"WPT2",1}, {"VW2",1}, {"WGS3",64}, {"WPT3",1}, {"VW3",1} } },
+ kDeviceTypeAll, "default", {
+ { "default", { {"WGS1",128}, {"WPT1",1}, {"VW2",1}, {"WGS2",128}, {"WPT2",1}, {"VW3",1}, {"WGS3",128}, {"WPT3",1} } },
}
},
}
diff --git a/include/internal/tuning.h b/include/internal/tuning.h
index 6ea530ba..5645a5e5 100644
--- a/include/internal/tuning.h
+++ b/include/internal/tuning.h
@@ -127,9 +127,11 @@ void Tuner(int argc, char* argv[]) {
{"precision", precision_string}
};
for (auto &o: C::GetOptions()) {
- if (o == kArgM) { metadata.push_back({"arg_m", std::to_string(args.m)}); }
- if (o == kArgN) { metadata.push_back({"arg_n", std::to_string(args.n)}); }
- if (o == kArgK) { metadata.push_back({"arg_k", std::to_string(args.k)}); }
+ if (o == kArgM) { metadata.push_back({"arg_m", std::to_string(args.m)}); }
+ if (o == kArgN) { metadata.push_back({"arg_n", std::to_string(args.n)}); }
+ if (o == kArgK) { metadata.push_back({"arg_k", std::to_string(args.k)}); }
+ if (o == kArgAlpha) { metadata.push_back({"arg_alpha", ToString(args.alpha)}); }
+ if (o == kArgBeta) { metadata.push_back({"arg_beta", ToString(args.beta)}); }
}
tuner.PrintJSON("clblast_"+C::KernelFamily()+"_"+precision_string+".json", metadata);
}
diff --git a/scripts/database/database.db b/scripts/database/database.db
new file mode 100644
index 00000000..bf793177
--- /dev/null
+++ b/scripts/database/database.db
Binary files differ
diff --git a/src/database.py b/scripts/database/database.py
index 2852b54c..01662a4b 100644
--- a/src/database.py
+++ b/scripts/database/database.py
@@ -20,7 +20,16 @@ import json
import pandas as pd
# Constants
-ATTRIBUTES = ["device", "type", "vendor", "precision", "kernel_family", "arg_m", "arg_n", "arg_k"]
+VENDOR_DEFAULT = "default"
+DEVICETYPE_DEFAULT = "All"
+DEVICENAME_DEFAULT = "default"
+
+# Attributes
+DEVICETYPE_ATTRIBUTES = ["device_vendor", "device_type"]
+DEVICE_ATTRIBUTES = ["device", "device_core_clock", "device_compute_units"]
+KERNEL_ATTRIBUTES = ["precision", "kernel_family",
+ "arg_m", "arg_n", "arg_k", "arg_alpha", "arg_beta"]
+ATTRIBUTES = DEVICE_ATTRIBUTES + DEVICETYPE_ATTRIBUTES + KERNEL_ATTRIBUTES
# Pandas options
pd.set_option('display.width', 1000)
@@ -60,15 +69,53 @@ def ConcatenateData(df1, df2):
def RemoveDuplicates(df):
return df.drop_duplicates()
-# Bests
+def RemoveEntriesByDevice(df, devicename):
+ return df[df["device"] != devicename]
+
+def GetEntriesByField(df, field, value):
+ return df[df[field] == value]
+
+# Retrieves the results with the lowest execution times
def GetBestResults(df):
dfbest = pd.DataFrame()
grouped = df.groupby(ATTRIBUTES+["kernel"])
for name, dfgroup in grouped:
- bestcase = dfgroup.loc[[dfgroup["time"].idxmin()]]
- dfbest = ConcatenateData(dfbest, bestcase)
+ besttime = dfgroup["time"].min()
+ bestcase = dfgroup[dfgroup["time"] == besttime].iloc[0]
+ dfbest = dfbest.append(bestcase, ignore_index=True)
return dfbest
+# Sets defaults for devices of the same type/vendor based on the smallest values of all know
+# entries. The average might be better for performance but some parameters might not be supported
+# on other devices.
+def CalculateDefaults(df):
+ dfdefault = pd.DataFrame()
+
+ # Defaults per type/vendor
+ groups = df.groupby(DEVICETYPE_ATTRIBUTES+KERNEL_ATTRIBUTES+["kernel"])
+ for name, dfgroup in groups:
+ default_values = dfgroup.min(axis=0)
+ default_values["device"] = DEVICENAME_DEFAULT
+ default_values["device_compute_units"] = 0
+ default_values["device_core_clock"] = 0
+ default_values["time"] = 0.0
+ dfdefault = dfdefault.append(default_values, ignore_index=True)
+
+ # Defaults in general
+ groups = df.groupby(KERNEL_ATTRIBUTES+["kernel"])
+ for name, dfgroup in groups:
+ default_values = dfgroup.min(axis=0)
+ default_values["device_vendor"] = VENDOR_DEFAULT
+ default_values["device_type"] = DEVICETYPE_DEFAULT
+ default_values["device"] = DEVICENAME_DEFAULT
+ default_values["device_compute_units"] = 0
+ default_values["device_core_clock"] = 0
+ default_values["time"] = 0.0
+ dfdefault = dfdefault.append(default_values, ignore_index=True)
+
+ # Database with both types of defaults only
+ return dfdefault
+
# ==================================================================================================
# C++ header generation
# ==================================================================================================
@@ -110,27 +157,28 @@ def GetPrecision(family, precision):
# The C++ device type and vendor
def GetDeviceVendor(vendor, devtype):
- return(" { // %s %ss\n kDeviceType%s, kDeviceVendor%s, {\n"
- % (vendor, devtype, devtype, vendor))
+ if vendor == VENDOR_DEFAULT and devtype == DEVICETYPE_DEFAULT:
+ return(" { // Default\n kDeviceType%s, \"%s\", {\n" % (devtype, vendor))
+ return(" { // %s %ss\n kDeviceType%s, \"%s\", {\n" % (vendor, devtype, devtype, vendor))
# Prints the data to a C++ database
-def PrintData(df):
+def PrintData(df, outputdir):
# Iterates over the kernel families: creates a new file per family
for family, dffamily in df.groupby(["kernel_family"]):
dffamily = dffamily.dropna(axis=1, how='all')
- f = open(family+'.h', 'w+')
+ f = open(os.path.join(outputdir, family+'.h'), 'w+')
f.write(GetHeader(family))
# Loops over the different entries for this family and prints their headers
for precision, dfprecision in dffamily.groupby(["precision"]):
f.write(GetPrecision(family, precision))
- for vendor, dfvendor in dfprecision.groupby(["vendor"]):
- for devtype, dfdevtype in dfvendor.groupby(["type"]):
+ for vendor, dfvendor in dfprecision.groupby(["device_vendor"]):
+ for devtype, dfdevtype in dfvendor.groupby(["device_type"]):
f.write(GetDeviceVendor(vendor, devtype))
for device, dfdevice in dfdevtype.groupby(["device"]):
devicename = "\"%s\"," % device
- f.write(" { %-20s { " % devicename)
+ f.write(" { %-48s { " % devicename)
# Collects the paramaters for this case and prints them
parameters = []
@@ -158,17 +206,16 @@ if len(sys.argv) != 3:
# Parses the command-line arguments
path_json = sys.argv[1]
path_clblast = sys.argv[2]
-file_db = path_clblast+"/src/database.db"
-glob_json = path_json+"/*.json"
+file_db = os.path.join(path_clblast, "scripts", "database", "database.db")
+glob_json = os.path.join(path_json, "*.json")
# Checks whether the command-line arguments are valid; exists otherwise
-clblast_h = path_clblast+"/include/clblast.h" # Not used but just for validation
+clblast_h = os.path.join(path_clblast, "include", "clblast.h") # Not used but just for validation
if not os.path.isfile(clblast_h):
print "[ERROR] The path '"+path_clblast+"' does not point to the root of the CLBlast library"
sys.exit()
if len(glob.glob(glob_json)) < 1:
- print "[ERROR] The path '"+path_json+"' does not contain any JSON files"
- sys.exit()
+ print "## The path '"+path_json+"' does not contain any JSON files"
# ==================================================================================================
# The main body of the script
@@ -198,11 +245,13 @@ SaveDatabase(database, file_db)
# Retrieves the best performing results
bests = GetBestResults(database)
-# TODO: Determines the defaults for other vendors and per vendor
-#defaults = CalculateDefaults(bests)
-#bests = ConcatenateData(bests, defaults)
+# Determines the defaults for other vendors and per vendor
+defaults = CalculateDefaults(bests)
+bests = ConcatenateData(bests, defaults)
# Outputs the data as a C++ database
-PrintData(bests)
+path_cpp_database = os.path.join(path_clblast, "include", "internal", "database")
+print "## Producing a C++ database in '"+path_cpp_database+"'"
+PrintData(bests, path_cpp_database)
# ==================================================================================================
diff --git a/src/database.cc b/src/database.cc
index b7275dad..7f5ac6eb 100644
--- a/src/database.cc
+++ b/src/database.cc
@@ -34,8 +34,8 @@ const std::vector<Database::DatabaseEntry> Database::database = {
XgemmSingle, XgemmDouble, XgemmComplexSingle, XgemmComplexDouble,
CopySingle, CopyDouble, CopyComplexSingle, CopyComplexDouble,
PadSingle, PadDouble, PadComplexSingle, PadComplexDouble,
- TraSingle, TraDouble, TraComplexSingle, TraComplexDouble,
- PadTraSingle, PadTraDouble, PadTraComplexSingle, PadTraComplexDouble
+ TransposeSingle, TransposeDouble, TransposeComplexSingle, TransposeComplexDouble,
+ PadtransposeSingle, PadtransposeDouble, PadtransposeComplexSingle, PadtransposeComplexDouble
};
// =================================================================================================
@@ -78,6 +78,7 @@ Database::Parameters Database::Search(const std::string &this_kernel,
const std::string &this_device,
const Precision this_precision) const {
for (auto &db: database) {
+
if (db.kernel == this_kernel && db.precision == this_precision) {
// Searches for the right vendor and device type, or selects the default if unavailable. This
@@ -89,7 +90,8 @@ Database::Parameters Database::Search(const std::string &this_kernel,
// Searches for the right device. If the current device is unavailable, selects the vendor
// default parameters. This assumes the default is last in the database.
for (auto &device: vendor.devices) {
- if (device.name == this_device || device.name == kDefaultDevice) {
+
+ if (device.name == this_device || device.name == "default") {
// Sets the parameters accordingly
return device.parameters;
diff --git a/src/routines/level3/xgemm.cc b/src/routines/level3/xgemm.cc
index 94aadcad..3961a3fd 100644
--- a/src/routines/level3/xgemm.cc
+++ b/src/routines/level3/xgemm.cc
@@ -30,7 +30,7 @@ template <> const Precision Xgemm<double2>::precision_ = Precision::kComplexDoub
// Constructor: forwards to base class constructor
template <typename T>
Xgemm<T>::Xgemm(Queue &queue, Event &event, const std::string &name):
- Routine<T>(queue, event, name, {"Copy","Pad","Transpose","PadTranspose","Xgemm"}, precision_) {
+ Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) {
source_string_ =
#include "../../kernels/level3/copy.opencl"
#include "../../kernels/level3/pad.opencl"
diff --git a/src/routines/level3/xher2k.cc b/src/routines/level3/xher2k.cc
index 5b84decb..e9970fd1 100644
--- a/src/routines/level3/xher2k.cc
+++ b/src/routines/level3/xher2k.cc
@@ -28,7 +28,7 @@ template <> const Precision Xher2k<double2,double>::precision_ = Precision::kCom
// Constructor: forwards to base class constructor
template <typename T, typename U>
Xher2k<T,U>::Xher2k(Queue &queue, Event &event, const std::string &name):
- Routine<T>(queue, event, name, {"Copy","Pad","Transpose","PadTranspose","Xgemm"}, precision_) {
+ Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) {
source_string_ =
#include "../../kernels/level3/copy.opencl"
#include "../../kernels/level3/pad.opencl"
diff --git a/src/routines/level3/xherk.cc b/src/routines/level3/xherk.cc
index 6a915c0b..49fd12af 100644
--- a/src/routines/level3/xherk.cc
+++ b/src/routines/level3/xherk.cc
@@ -28,7 +28,7 @@ template <> const Precision Xherk<double2,double>::precision_ = Precision::kComp
// Constructor: forwards to base class constructor
template <typename T, typename U>
Xherk<T,U>::Xherk(Queue &queue, Event &event, const std::string &name):
- Routine<T>(queue, event, name, {"Copy","Pad","Transpose","PadTranspose","Xgemm"}, precision_) {
+ Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) {
source_string_ =
#include "../../kernels/level3/copy.opencl"
#include "../../kernels/level3/pad.opencl"
diff --git a/src/routines/level3/xsyr2k.cc b/src/routines/level3/xsyr2k.cc
index de5f1afc..966a000f 100644
--- a/src/routines/level3/xsyr2k.cc
+++ b/src/routines/level3/xsyr2k.cc
@@ -30,7 +30,7 @@ template <> const Precision Xsyr2k<double2>::precision_ = Precision::kComplexDou
// Constructor: forwards to base class constructor
template <typename T>
Xsyr2k<T>::Xsyr2k(Queue &queue, Event &event, const std::string &name):
- Routine<T>(queue, event, name, {"Copy","Pad","Transpose","PadTranspose","Xgemm"}, precision_) {
+ Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) {
source_string_ =
#include "../../kernels/level3/copy.opencl"
#include "../../kernels/level3/pad.opencl"
diff --git a/src/routines/level3/xsyrk.cc b/src/routines/level3/xsyrk.cc
index d8fc6357..630cb731 100644
--- a/src/routines/level3/xsyrk.cc
+++ b/src/routines/level3/xsyrk.cc
@@ -30,7 +30,7 @@ template <> const Precision Xsyrk<double2>::precision_ = Precision::kComplexDoub
// Constructor: forwards to base class constructor
template <typename T>
Xsyrk<T>::Xsyrk(Queue &queue, Event &event, const std::string &name):
- Routine<T>(queue, event, name, {"Copy","Pad","Transpose","PadTranspose","Xgemm"}, precision_) {
+ Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) {
source_string_ =
#include "../../kernels/level3/copy.opencl"
#include "../../kernels/level3/pad.opencl"
diff --git a/src/tuning/xgemv.cc b/src/tuning/xgemv.cc
index 6a066518..60c73d76 100644
--- a/src/tuning/xgemv.cc
+++ b/src/tuning/xgemv.cc
@@ -72,7 +72,10 @@ class TuneXgemv {
tuner.AddConstraint(id, MultipleOfX, {"WPT"+std::to_string(V), "VW"+std::to_string(V)});
}
}
- static void SetLocalMemorySize(cltune::Tuner &, const size_t, const Arguments<T> &) { }
+ static void SetLocalMemorySize(cltune::Tuner &tuner, const size_t id, const Arguments<T> &args) {
+ auto LocalMemorySize = [args] (std::vector<size_t> v) { return v[0]*GetBytes(args.precision); };
+ tuner.SetLocalMemoryUsage(id, LocalMemorySize, {"WGS"+std::to_string(V)});
+ }
// Sets the base thread configuration
static std::vector<size_t> GlobalSize(const Arguments<T> &args) { return {args.m}; }
@@ -108,6 +111,9 @@ class TuneXgemv {
tuner.AddArgumentScalar(0);
tuner.AddArgumentScalar(1);
tuner.AddArgumentScalar(0); // Conjugate transpose
+ tuner.AddArgumentScalar(0); // Additional parameter
+ tuner.AddArgumentScalar(0); // Banded 'kl'
+ tuner.AddArgumentScalar(0); // Banded 'ku'
}
// Describes how to compute the performance metrics
generated by cgit v1.2.3 (git 2.39.1) at 2024-08-01 02:41:21 +0000