From 32b850b12ba99d468e5a7e378253f8eae5df0ed9 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Tue, 3 Jan 2017 20:30:56 +0100 Subject: Added tuning results for the AMD Turks GPU and the Intel Core i7-2670QM CPU --- src/database/kernels/copy.hpp | 18 +++++++++++------ src/database/kernels/pad.hpp | 12 +++++++++--- src/database/kernels/padtranspose.hpp | 8 +++++++- src/database/kernels/transpose.hpp | 12 +++++++++--- src/database/kernels/xaxpy.hpp | 10 ++++++++-- src/database/kernels/xdot.hpp | 22 +++++++++++++-------- src/database/kernels/xgemm.hpp | 24 ++++++++++++++--------- src/database/kernels/xgemm_direct.hpp | 34 +++++++++++++++++++++++++++++---- src/database/kernels/xgemv.hpp | 8 +++++++- src/database/kernels/xgemv_fast.hpp | 12 +++++++++--- src/database/kernels/xgemv_fast_rot.hpp | 18 +++++++++++------ src/database/kernels/xger.hpp | 14 ++++++++++---- 12 files changed, 142 insertions(+), 50 deletions(-) (limited to 'src/database/kernels') diff --git a/src/database/kernels/copy.hpp b/src/database/kernels/copy.hpp index 1bc63691..cf746901 100644 --- a/src/database/kernels/copy.hpp +++ b/src/database/kernels/copy.hpp @@ -44,6 +44,7 @@ const Database::DatabaseEntry CopySingle = { { "Pitcairn", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "Tahiti", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, { "Tonga", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, + { "Turks", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",2} } }, { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, } }, @@ -55,10 +56,11 @@ const Database::DatabaseEntry CopySingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, } }, { // Intel GPUs @@ -99,7 +101,7 @@ const Database::DatabaseEntry CopySingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, } }, } @@ -117,11 +119,13 @@ const Database::DatabaseEntry CopyComplexSingle = { { "Pitcairn", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "Tahiti", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, { "Tonga", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",1}, {"COPY_WPT",2} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "Turks", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",4} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",4}, {"COPY_WPT",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, @@ -164,7 +168,7 @@ const Database::DatabaseEntry CopyComplexSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, } }, } @@ -193,10 +197,11 @@ const Database::DatabaseEntry CopyDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",2}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, - { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, } }, { // Intel accelerators @@ -225,7 +230,7 @@ const Database::DatabaseEntry CopyDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, } }, } @@ -254,6 +259,7 @@ const Database::DatabaseEntry CopyComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",32}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, diff --git a/src/database/kernels/pad.hpp b/src/database/kernels/pad.hpp index 310d3a78..5d594046 100644 --- a/src/database/kernels/pad.hpp +++ b/src/database/kernels/pad.hpp @@ -44,7 +44,8 @@ const Database::DatabaseEntry PadSingle = { { "Pitcairn", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Tahiti", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Tonga", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, - { "default", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",4} } }, + { "Turks", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, + { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, } }, { // ARM GPUs @@ -55,6 +56,7 @@ const Database::DatabaseEntry PadSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",16}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, @@ -117,7 +119,8 @@ const Database::DatabaseEntry PadComplexSingle = { { "Pitcairn", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Tahiti", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Tonga", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, - { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "Turks", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",4} } }, + { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, } }, { // ARM GPUs @@ -128,10 +131,11 @@ const Database::DatabaseEntry PadComplexSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, - { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, + { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",2} } }, } }, { // Intel GPUs @@ -201,6 +205,7 @@ const Database::DatabaseEntry PadDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, @@ -262,6 +267,7 @@ const Database::DatabaseEntry PadComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",16}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, diff --git a/src/database/kernels/padtranspose.hpp b/src/database/kernels/padtranspose.hpp index 8ef09e85..c009425e 100644 --- a/src/database/kernels/padtranspose.hpp +++ b/src/database/kernels/padtranspose.hpp @@ -44,6 +44,7 @@ const Database::DatabaseEntry PadtransposeSingle = { { "Pitcairn", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, { "Tahiti", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, { "Tonga", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, + { "Turks", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, } }, @@ -55,6 +56,7 @@ const Database::DatabaseEntry PadtransposeSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, @@ -117,6 +119,7 @@ const Database::DatabaseEntry PadtransposeComplexSingle = { { "Pitcairn", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Tahiti", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "Tonga", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, + { "Turks", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, } }, @@ -128,6 +131,7 @@ const Database::DatabaseEntry PadtransposeComplexSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, @@ -201,6 +205,7 @@ const Database::DatabaseEntry PadtransposeDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, @@ -262,10 +267,11 @@ const Database::DatabaseEntry PadtransposeComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, - { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, + { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, } }, { // Intel accelerators diff --git a/src/database/kernels/transpose.hpp b/src/database/kernels/transpose.hpp index 23fecb49..56ce93a9 100644 --- a/src/database/kernels/transpose.hpp +++ b/src/database/kernels/transpose.hpp @@ -44,7 +44,8 @@ const Database::DatabaseEntry TransposeSingle = { { "Pitcairn", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Tahiti", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "Tonga", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, - { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, + { "Turks", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, + { "default", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, } }, { // ARM GPUs @@ -55,6 +56,7 @@ const Database::DatabaseEntry TransposeSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",16} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, @@ -117,6 +119,7 @@ const Database::DatabaseEntry TransposeComplexSingle = { { "Pitcairn", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Tahiti", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Tonga", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, + { "Turks", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "default", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, } }, @@ -128,6 +131,7 @@ const Database::DatabaseEntry TransposeComplexSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, @@ -166,7 +170,7 @@ const Database::DatabaseEntry TransposeComplexSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, + { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, } }, } @@ -195,10 +199,11 @@ const Database::DatabaseEntry TransposeDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, - { "default", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, + { "default", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, } }, { // Intel accelerators @@ -256,6 +261,7 @@ const Database::DatabaseEntry TransposeComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, diff --git a/src/database/kernels/xaxpy.hpp b/src/database/kernels/xaxpy.hpp index 52845e96..a852a895 100644 --- a/src/database/kernels/xaxpy.hpp +++ b/src/database/kernels/xaxpy.hpp @@ -44,6 +44,7 @@ const Database::DatabaseEntry XaxpySingle = { { "Pitcairn", { {"VW",2}, {"WGS",128}, {"WPT",1} } }, { "Tahiti", { {"VW",2}, {"WGS",64}, {"WPT",1} } }, { "Tonga", { {"VW",1}, {"WGS",256}, {"WPT",8} } }, + { "Turks", { {"VW",2}, {"WGS",256}, {"WPT",1} } }, { "default", { {"VW",2}, {"WGS",64}, {"WPT",2} } }, } }, @@ -55,10 +56,11 @@ const Database::DatabaseEntry XaxpySingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",8}, {"WGS",512}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",1}, {"WGS",512}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",4}, {"WGS",256}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, - { "default", { {"VW",2}, {"WGS",256}, {"WPT",1} } }, + { "default", { {"VW",8}, {"WGS",1024}, {"WPT",2} } }, } }, { // Intel GPUs @@ -117,6 +119,7 @@ const Database::DatabaseEntry XaxpyComplexSingle = { { "Pitcairn", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "Tahiti", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "Tonga", { {"VW",1}, {"WGS",256}, {"WPT",8} } }, + { "Turks", { {"VW",2}, {"WGS",256}, {"WPT",1} } }, { "default", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, } }, @@ -128,6 +131,7 @@ const Database::DatabaseEntry XaxpyComplexSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",4}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",4}, {"WGS",256}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",1}, {"WGS",1024}, {"WPT",2} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",2}, {"WGS",1024}, {"WPT",1} } }, @@ -201,6 +205,7 @@ const Database::DatabaseEntry XaxpyDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",4}, {"WGS",64}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",1}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",8}, {"WGS",64}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",8}, {"WGS",2048}, {"WPT",1} } }, @@ -262,6 +267,7 @@ const Database::DatabaseEntry XaxpyComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",4}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",8}, {"WGS",128}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",8}, {"WGS",512}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, @@ -294,7 +300,7 @@ const Database::DatabaseEntry XaxpyComplexDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, + { "default", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, } }, } diff --git a/src/database/kernels/xdot.hpp b/src/database/kernels/xdot.hpp index 8b07c539..a7a4e105 100644 --- a/src/database/kernels/xdot.hpp +++ b/src/database/kernels/xdot.hpp @@ -43,13 +43,15 @@ const Database::DatabaseEntry XdotSingle = { { "Pitcairn", { {"WGS1",128}, {"WGS2",32} } }, { "Tahiti", { {"WGS1",128}, {"WGS2",32} } }, { "Tonga", { {"WGS1",64}, {"WGS2",32} } }, - { "default", { {"WGS1",128}, {"WGS2",32} } }, + { "Turks", { {"WGS1",128}, {"WGS2",64} } }, + { "default", { {"WGS1",128}, {"WGS2",64} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WGS2",32} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",1024}, {"WGS2",32} } }, - { "default", { {"WGS1",1024}, {"WGS2",32} } }, + { "default", { {"WGS1",256}, {"WGS2",32} } }, } }, { // Intel GPUs @@ -81,7 +83,7 @@ const Database::DatabaseEntry XdotSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"WGS1",256}, {"WGS2",32} } }, + { "default", { {"WGS1",128}, {"WGS2",32} } }, } }, } @@ -98,13 +100,15 @@ const Database::DatabaseEntry XdotComplexSingle = { { "Pitcairn", { {"WGS1",256}, {"WGS2",32} } }, { "Tahiti", { {"WGS1",64}, {"WGS2",32} } }, { "Tonga", { {"WGS1",256}, {"WGS2",64} } }, - { "default", { {"WGS1",256}, {"WGS2",64} } }, + { "Turks", { {"WGS1",128}, {"WGS2",32} } }, + { "default", { {"WGS1",256}, {"WGS2",32} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WGS2",64} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",1024}, {"WGS2",32} } }, - { "default", { {"WGS1",1024}, {"WGS2",32} } }, + { "default", { {"WGS1",128}, {"WGS2",64} } }, } }, { // Intel GPUs @@ -136,7 +140,7 @@ const Database::DatabaseEntry XdotComplexSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"WGS1",256}, {"WGS2",64} } }, + { "default", { {"WGS1",256}, {"WGS2",32} } }, } }, } @@ -158,8 +162,9 @@ const Database::DatabaseEntry XdotDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",64}, {"WGS2",128} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",512}, {"WGS2",64} } }, - { "default", { {"WGS1",512}, {"WGS2",64} } }, + { "default", { {"WGS1",128}, {"WGS2",64} } }, } }, { // NVIDIA GPUs @@ -202,8 +207,9 @@ const Database::DatabaseEntry XdotComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WGS2",128} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",1024}, {"WGS2",32} } }, - { "default", { {"WGS1",1024}, {"WGS2",32} } }, + { "default", { {"WGS1",128}, {"WGS2",32} } }, } }, { // NVIDIA GPUs diff --git a/src/database/kernels/xgemm.hpp b/src/database/kernels/xgemm.hpp index 66ac8a9f..06f0f26a 100644 --- a/src/database/kernels/xgemm.hpp +++ b/src/database/kernels/xgemm.hpp @@ -43,7 +43,8 @@ const Database::DatabaseEntry XgemmSingle = { { "Pitcairn", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tahiti", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",32}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",1} } }, { "Tonga", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",8} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, + { "Turks", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, } }, { // ARM GPUs @@ -54,10 +55,11 @@ const Database::DatabaseEntry XgemmSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",32}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, + { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, } }, { // Intel GPUs @@ -116,7 +118,8 @@ const Database::DatabaseEntry XgemmComplexSingle = { { "Pitcairn", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",2} } }, { "Tahiti", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",32}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",1} } }, { "Tonga", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",1} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",32}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",2} } }, + { "Turks", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, } }, { // ARM GPUs @@ -127,10 +130,11 @@ const Database::DatabaseEntry XgemmComplexSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",1}, {"VWM",1}, {"VWN",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",4} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, + { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, } }, { // Intel GPUs @@ -171,7 +175,7 @@ const Database::DatabaseEntry XgemmComplexSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, } }, } @@ -200,10 +204,11 @@ const Database::DatabaseEntry XgemmDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",8} } }, - { "default", { {"KWG",32}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, } }, { // Intel accelerators @@ -232,7 +237,7 @@ const Database::DatabaseEntry XgemmDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, } }, } @@ -261,10 +266,11 @@ const Database::DatabaseEntry XgemmComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",32}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",1}, {"VWN",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, + { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, } }, { // Intel accelerators @@ -292,7 +298,7 @@ const Database::DatabaseEntry XgemmComplexDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, } }, } diff --git a/src/database/kernels/xgemm_direct.hpp b/src/database/kernels/xgemm_direct.hpp index 4413cf1b..d7aacfa3 100644 --- a/src/database/kernels/xgemm_direct.hpp +++ b/src/database/kernels/xgemm_direct.hpp @@ -39,9 +39,16 @@ const Database::DatabaseEntry XgemmDirectSingle = { kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, { "Tonga", { {"KWID",16}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",32}, {"NDIMCD",8}, {"PADA",0}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + { "Turks", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } }, + { // Intel CPUs + kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",8}, {"WGD",64} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",8}, {"WGD",64} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics Skylake ULT GT2", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, @@ -58,7 +65,7 @@ const Database::DatabaseEntry XgemmDirectSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, } }, } @@ -72,9 +79,16 @@ const Database::DatabaseEntry XgemmDirectComplexSingle = { kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "Tonga", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, + { "Turks", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, } }, + { // Intel CPUs + kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics Skylake ULT GT2", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, @@ -91,7 +105,7 @@ const Database::DatabaseEntry XgemmDirectComplexSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, } }, } @@ -108,6 +122,12 @@ const Database::DatabaseEntry XgemmDirectDouble = { { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } }, + { // Intel CPUs + kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, + } + }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",4}, {"WGD",32} } }, @@ -117,7 +137,7 @@ const Database::DatabaseEntry XgemmDirectDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, } }, } @@ -134,6 +154,12 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = { { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } }, + { // Intel CPUs + kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",32}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",32}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + } + }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",32}, {"MDIMCD",32}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, @@ -143,7 +169,7 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, } }, } diff --git a/src/database/kernels/xgemv.hpp b/src/database/kernels/xgemv.hpp index 5f25f210..7d1781f9 100644 --- a/src/database/kernels/xgemv.hpp +++ b/src/database/kernels/xgemv.hpp @@ -44,11 +44,13 @@ const Database::DatabaseEntry XgemvSingle = { { "Pitcairn", { {"WGS1",256}, {"WPT1",1} } }, { "Tahiti", { {"WGS1",256}, {"WPT1",1} } }, { "Tonga", { {"WGS1",128}, {"WPT1",2} } }, + { "Turks", { {"WGS1",32}, {"WPT1",1} } }, { "default", { {"WGS1",128}, {"WPT1",1} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",64}, {"WPT1",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, { "default", { {"WGS1",64}, {"WPT1",4} } }, @@ -110,14 +112,16 @@ const Database::DatabaseEntry XgemvComplexSingle = { { "Pitcairn", { {"WGS1",64}, {"WPT1",1} } }, { "Tahiti", { {"WGS1",64}, {"WPT1",1} } }, { "Tonga", { {"WGS1",32}, {"WPT1",1} } }, + { "Turks", { {"WGS1",64}, {"WPT1",1} } }, { "default", { {"WGS1",64}, {"WPT1",1} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",128}, {"WPT1",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, - { "default", { {"WGS1",64}, {"WPT1",1} } }, + { "default", { {"WGS1",64}, {"WPT1",2} } }, } }, { // Intel GPUs @@ -177,6 +181,7 @@ const Database::DatabaseEntry XgemvDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",64}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",64}, {"WPT1",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, { "default", { {"WGS1",64}, {"WPT1",4} } }, @@ -231,6 +236,7 @@ const Database::DatabaseEntry XgemvComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",64}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",64}, {"WPT1",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, { "default", { {"WGS1",64}, {"WPT1",4} } }, diff --git a/src/database/kernels/xgemv_fast.hpp b/src/database/kernels/xgemv_fast.hpp index 994a220c..988687d5 100644 --- a/src/database/kernels/xgemv_fast.hpp +++ b/src/database/kernels/xgemv_fast.hpp @@ -44,14 +44,16 @@ const Database::DatabaseEntry XgemvFastSingle = { { "Pitcairn", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Tahiti", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Tonga", { {"VW2",1}, {"WGS2",16}, {"WPT2",4} } }, + { "Turks", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",1}, {"WGS2",32}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",4}, {"WGS2",128}, {"WPT2",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, - { "default", { {"VW2",4}, {"WGS2",64}, {"WPT2",4} } }, + { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, } }, { // Intel GPUs @@ -110,14 +112,16 @@ const Database::DatabaseEntry XgemvFastComplexSingle = { { "Pitcairn", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Tahiti", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, { "Tonga", { {"VW2",2}, {"WGS2",32}, {"WPT2",2} } }, + { "Turks", { {"VW2",1}, {"WGS2",16}, {"WPT2",1} } }, { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",2}, {"WGS2",64}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",1}, {"WGS2",128}, {"WPT2",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",4}, {"WGS2",64}, {"WPT2",4} } }, - { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",2} } }, + { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, } }, { // Intel GPUs @@ -173,6 +177,7 @@ const Database::DatabaseEntry XgemvFastDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",4}, {"WGS2",128}, {"WPT2",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, @@ -227,9 +232,10 @@ const Database::DatabaseEntry XgemvFastComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",4}, {"WGS2",32}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",2}, {"WGS2",64}, {"WPT2",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",4}, {"WGS2",64}, {"WPT2",4} } }, - { "default", { {"VW2",2}, {"WGS2",64}, {"WPT2",4} } }, + { "default", { {"VW2",4}, {"WGS2",64}, {"WPT2",4} } }, } }, { // Intel accelerators diff --git a/src/database/kernels/xgemv_fast_rot.hpp b/src/database/kernels/xgemv_fast_rot.hpp index da8bcfeb..6fbb4619 100644 --- a/src/database/kernels/xgemv_fast_rot.hpp +++ b/src/database/kernels/xgemv_fast_rot.hpp @@ -39,11 +39,13 @@ const Database::DatabaseEntry XgemvFastRotSingle = { kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW3",8}, {"WGS3",64}, {"WPT3",32} } }, { "Tonga", { {"VW3",8}, {"WGS3",128}, {"WPT3",32} } }, + { "Turks", { {"VW3",8}, {"WGS3",128}, {"WPT3",16} } }, { "default", { {"VW3",8}, {"WGS3",128}, {"WPT3",32} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",8}, {"WGS3",16}, {"WPT3",8} } }, { "default", { {"VW3",8}, {"WGS3",16}, {"WPT3",8} } }, } @@ -68,7 +70,7 @@ const Database::DatabaseEntry XgemvFastRotSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"VW3",8}, {"WGS3",128}, {"WPT3",32} } }, + { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, } }, } @@ -82,13 +84,15 @@ const Database::DatabaseEntry XgemvFastRotComplexSingle = { kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW3",8}, {"WGS3",16}, {"WPT3",16} } }, { "Tonga", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, - { "default", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, + { "Turks", { {"VW3",4}, {"WGS3",32}, {"WPT3",8} } }, + { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, - { "default", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, + { "default", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, } }, { // Intel GPUs @@ -122,8 +126,9 @@ const Database::DatabaseEntry XgemvFastRotDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",8}, {"WGS3",16}, {"WPT3",8} } }, - { "default", { {"VW3",8}, {"WGS3",16}, {"WPT3",8} } }, + { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, } }, { // NVIDIA GPUs @@ -155,13 +160,14 @@ const Database::DatabaseEntry XgemvFastRotComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",2}, {"WGS3",16}, {"WPT3",16} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",8}, {"WGS3",16}, {"WPT3",16} } }, - { "default", { {"VW3",8}, {"WGS3",16}, {"WPT3",16} } }, + { "default", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",16} } }, + { "default", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, } }, } diff --git a/src/database/kernels/xger.hpp b/src/database/kernels/xger.hpp index 5e2be6a9..f7295ff4 100644 --- a/src/database/kernels/xger.hpp +++ b/src/database/kernels/xger.hpp @@ -44,6 +44,7 @@ const Database::DatabaseEntry XgerSingle = { { "Pitcairn", { {"WGS1",64}, {"WGS2",1}, {"WPT",1} } }, { "Tahiti", { {"WGS1",256}, {"WGS2",1}, {"WPT",1} } }, { "Tonga", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, + { "Turks", { {"WGS1",64}, {"WGS2",4}, {"WPT",2} } }, { "default", { {"WGS1",32}, {"WGS2",8}, {"WPT",1} } }, } }, @@ -55,9 +56,10 @@ const Database::DatabaseEntry XgerSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WGS2",4}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",128}, {"WGS2",2}, {"WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",128}, {"WGS2",1}, {"WPT",4} } }, - { "default", { {"WGS1",128}, {"WGS2",8}, {"WPT",4} } }, + { "default", { {"WGS1",64}, {"WGS2",1}, {"WPT",4} } }, } }, { // Intel GPUs @@ -105,7 +107,8 @@ const Database::DatabaseEntry XgerComplexSingle = { { "Pitcairn", { {"WGS1",128}, {"WGS2",2}, {"WPT",1} } }, { "Tahiti", { {"WGS1",64}, {"WGS2",2}, {"WPT",1} } }, { "Tonga", { {"WGS1",64}, {"WGS2",1}, {"WPT",1} } }, - { "default", { {"WGS1",64}, {"WGS2",1}, {"WPT",1} } }, + { "Turks", { {"WGS1",128}, {"WGS2",2}, {"WPT",1} } }, + { "default", { {"WGS1",128}, {"WGS2",2}, {"WPT",1} } }, } }, { // ARM GPUs @@ -116,9 +119,10 @@ const Database::DatabaseEntry XgerComplexSingle = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WGS2",2}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",512}, {"WGS2",4}, {"WPT",2} } }, - { "default", { {"WGS1",512}, {"WGS2",4}, {"WPT",2} } }, + { "default", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, } }, { // Intel GPUs @@ -177,9 +181,10 @@ const Database::DatabaseEntry XgerDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",512}, {"WGS2",16}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",512}, {"WGS2",8}, {"WPT",2} } }, - { "default", { {"WGS1",512}, {"WGS2",8}, {"WPT",2} } }, + { "default", { {"WGS1",256}, {"WGS2",4}, {"WPT",2} } }, } }, { // NVIDIA GPUs @@ -227,6 +232,7 @@ const Database::DatabaseEntry XgerComplexDouble = { }, { // Intel CPUs kDeviceTypeCPU, "Intel", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WGS2",4}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",512}, {"WGS2",4}, {"WPT",2} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, { "default", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, -- cgit v1.2.3 From 2e4f6e16098d36d0572769a1092d1e54cdb6d4ea Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Thu, 19 Jan 2017 19:42:31 +0100 Subject: Added tuning results for NVIDIA GTX 1080 and Intel Core i7-4790K --- README.md | 2 ++ src/database/kernels/copy.hpp | 14 +++++++++++--- src/database/kernels/pad.hpp | 14 +++++++++++--- src/database/kernels/padtranspose.hpp | 10 +++++++++- src/database/kernels/transpose.hpp | 10 +++++++++- src/database/kernels/xaxpy.hpp | 14 +++++++++++--- src/database/kernels/xdot.hpp | 20 ++++++++++++++------ src/database/kernels/xgemm.hpp | 26 +++++++++++++++++--------- src/database/kernels/xgemm_direct.hpp | 22 +++++++++++++++------- src/database/kernels/xgemv.hpp | 9 ++++++++- src/database/kernels/xgemv_fast.hpp | 8 +++++++- src/database/kernels/xgemv_fast_rot.hpp | 10 ++++++++-- src/database/kernels/xger.hpp | 16 ++++++++++++---- 13 files changed, 134 insertions(+), 41 deletions(-) (limited to 'src/database/kernels') diff --git a/README.md b/README.md index 45069b58..35e79db8 100644 --- a/README.md +++ b/README.md @@ -115,6 +115,7 @@ The CLBlast library will be tuned in the future for the most commonly used OpenC - GeForce GTX 750 Ti - GeForce GTX 980 - GeForce GTX 1070 + - GeForce GTX 1080 - GeForce GTX TITAN - GeForce GTX TITAN Black - GeForce GTX TITAN X @@ -140,6 +141,7 @@ The CLBlast library will be tuned in the future for the most commonly used OpenC - Core i5-6200U - Core i7-2670QM - Core i7-3770K + - Core i7-4790K - Core i7-5930K * Other devices: - ARM Mali-T628 GPU diff --git a/src/database/kernels/copy.hpp b/src/database/kernels/copy.hpp index cf746901..f0431933 100644 --- a/src/database/kernels/copy.hpp +++ b/src/database/kernels/copy.hpp @@ -59,8 +59,9 @@ const Database::DatabaseEntry CopySingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",2} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",2} } }, } }, { // Intel GPUs @@ -85,6 +86,7 @@ const Database::DatabaseEntry CopySingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "GeForce GTX 1070", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",4}, {"COPY_WPT",1} } }, + { "GeForce GTX 1080", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "GeForce GTX 480", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "GeForce GTX 670", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "GeForce GTX 680", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",4}, {"COPY_WPT",1} } }, @@ -101,7 +103,7 @@ const Database::DatabaseEntry CopySingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, } }, } @@ -128,6 +130,7 @@ const Database::DatabaseEntry CopyComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",4}, {"COPY_WPT",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",32}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, } @@ -154,6 +157,7 @@ const Database::DatabaseEntry CopyComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX 1070", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "GeForce GTX 1080", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "GeForce GTX 480", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX 670", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX 750", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, @@ -200,6 +204,7 @@ const Database::DatabaseEntry CopyDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",2}, {"COPY_WPT",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, } @@ -214,6 +219,7 @@ const Database::DatabaseEntry CopyDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",2}, {"COPY_WPT",1} } }, { "GeForce GTX 1070", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, + { "GeForce GTX 1080", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "GeForce GTX 480", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",1} } }, { "GeForce GTX 670", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",1} } }, { "GeForce GTX 680", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",2}, {"COPY_WPT",1} } }, @@ -262,8 +268,9 @@ const Database::DatabaseEntry CopyComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",32}, {"COPY_VW",8}, {"COPY_WPT",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",32}, {"COPY_VW",8}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",8}, {"COPY_WPT",1} } }, } }, { // Intel accelerators @@ -276,6 +283,7 @@ const Database::DatabaseEntry CopyComplexDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX 1070", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",1}, {"COPY_WPT",4} } }, + { "GeForce GTX 1080", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX 480", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX 670", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX 680", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, diff --git a/src/database/kernels/pad.hpp b/src/database/kernels/pad.hpp index 5d594046..3378709c 100644 --- a/src/database/kernels/pad.hpp +++ b/src/database/kernels/pad.hpp @@ -59,6 +59,7 @@ const Database::DatabaseEntry PadSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",16}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",2} } }, } @@ -85,6 +86,7 @@ const Database::DatabaseEntry PadSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "GeForce GTX 1070", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "GeForce GTX 1080", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 480", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",4} } }, { "GeForce GTX 670", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, { "GeForce GTX 680", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, @@ -96,7 +98,7 @@ const Database::DatabaseEntry PadSingle = { { "GeForce GTX TITAN X", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Tesla K20m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "Tesla K40m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, - { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, + { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, } }, { // Default @@ -134,6 +136,7 @@ const Database::DatabaseEntry PadComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",2} } }, } @@ -160,6 +163,7 @@ const Database::DatabaseEntry PadComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 1070", { {"PAD_DIMX",8}, {"PAD_DIMY",32}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "GeForce GTX 1080", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 480", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "GeForce GTX 670", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "GeForce GTX 680", { {"PAD_DIMX",16}, {"PAD_DIMY",32}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, @@ -176,7 +180,7 @@ const Database::DatabaseEntry PadComplexSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, } }, } @@ -208,6 +212,7 @@ const Database::DatabaseEntry PadDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, } @@ -222,6 +227,7 @@ const Database::DatabaseEntry PadDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 1070", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "GeForce GTX 1080", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "GeForce GTX 480", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 670", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "GeForce GTX 680", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, @@ -270,8 +276,9 @@ const Database::DatabaseEntry PadComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PAD_DIMX",16}, {"PAD_DIMY",32}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",32}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, - { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",16}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, + { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, } }, { // Intel accelerators @@ -284,6 +291,7 @@ const Database::DatabaseEntry PadComplexDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 1070", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, + { "GeForce GTX 1080", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 480", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 670", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX 680", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, diff --git a/src/database/kernels/padtranspose.hpp b/src/database/kernels/padtranspose.hpp index c009425e..212723c7 100644 --- a/src/database/kernels/padtranspose.hpp +++ b/src/database/kernels/padtranspose.hpp @@ -59,6 +59,7 @@ const Database::DatabaseEntry PadtransposeSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, } @@ -85,6 +86,7 @@ const Database::DatabaseEntry PadtransposeSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",2} } }, { "GeForce GTX 1070", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, + { "GeForce GTX 1080", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "GeForce GTX 480", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "GeForce GTX 670", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",2} } }, { "GeForce GTX 680", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, @@ -134,8 +136,9 @@ const Database::DatabaseEntry PadtransposeComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, - { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, + { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, } }, { // Intel GPUs @@ -160,6 +163,7 @@ const Database::DatabaseEntry PadtransposeComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 1070", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, + { "GeForce GTX 1080", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",1} } }, { "GeForce GTX 480", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 670", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 680", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, @@ -208,6 +212,7 @@ const Database::DatabaseEntry PadtransposeDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, } @@ -222,6 +227,7 @@ const Database::DatabaseEntry PadtransposeDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 1070", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, + { "GeForce GTX 1080", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",1} } }, { "GeForce GTX 480", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 670", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 680", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, @@ -270,6 +276,7 @@ const Database::DatabaseEntry PadtransposeComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, } @@ -284,6 +291,7 @@ const Database::DatabaseEntry PadtransposeComplexDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 1070", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, + { "GeForce GTX 1080", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",1} } }, { "GeForce GTX 480", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 670", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX 680", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, diff --git a/src/database/kernels/transpose.hpp b/src/database/kernels/transpose.hpp index 56ce93a9..f33f2a04 100644 --- a/src/database/kernels/transpose.hpp +++ b/src/database/kernels/transpose.hpp @@ -59,6 +59,7 @@ const Database::DatabaseEntry TransposeSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",16} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",16} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "default", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, } @@ -85,6 +86,7 @@ const Database::DatabaseEntry TransposeSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"TRA_DIM",32}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "GeForce GTX 1070", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, + { "GeForce GTX 1080", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "GeForce GTX 480", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "GeForce GTX 670", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "GeForce GTX 680", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, @@ -134,6 +136,7 @@ const Database::DatabaseEntry TransposeComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "default", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, } @@ -154,6 +157,7 @@ const Database::DatabaseEntry TransposeComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "GeForce GTX 1070", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, + { "GeForce GTX 1080", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX 480", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX 670", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "GeForce GTX 680", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, @@ -202,6 +206,7 @@ const Database::DatabaseEntry TransposeDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",16} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, { "default", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, } @@ -216,6 +221,7 @@ const Database::DatabaseEntry TransposeDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "GeForce GTX 1070", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, + { "GeForce GTX 1080", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "GeForce GTX 480", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "GeForce GTX 670", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "GeForce GTX 680", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, @@ -264,14 +270,16 @@ const Database::DatabaseEntry TransposeComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, - { "default", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, + { "default", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "GeForce GTX 1070", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, + { "GeForce GTX 1080", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX 480", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX 670", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "GeForce GTX 680", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, diff --git a/src/database/kernels/xaxpy.hpp b/src/database/kernels/xaxpy.hpp index a852a895..e4e3c621 100644 --- a/src/database/kernels/xaxpy.hpp +++ b/src/database/kernels/xaxpy.hpp @@ -59,8 +59,9 @@ const Database::DatabaseEntry XaxpySingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",8}, {"WGS",512}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",1}, {"WGS",512}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",4}, {"WGS",256}, {"WPT",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW",2}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, - { "default", { {"VW",8}, {"WGS",1024}, {"WPT",2} } }, + { "default", { {"VW",8}, {"WGS",512}, {"WPT",1} } }, } }, { // Intel GPUs @@ -85,6 +86,7 @@ const Database::DatabaseEntry XaxpySingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"VW",2}, {"WGS",64}, {"WPT",1} } }, { "GeForce GTX 1070", { {"VW",1}, {"WGS",64}, {"WPT",4} } }, + { "GeForce GTX 1080", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "GeForce GTX 480", { {"VW",2}, {"WGS",128}, {"WPT",1} } }, { "GeForce GTX 670", { {"VW",2}, {"WGS",64}, {"WPT",1} } }, { "GeForce GTX 680", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, @@ -96,7 +98,7 @@ const Database::DatabaseEntry XaxpySingle = { { "GeForce GTX TITAN X", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "Tesla K20m", { {"VW",4}, {"WGS",128}, {"WPT",1} } }, { "Tesla K40m", { {"VW",4}, {"WGS",128}, {"WPT",1} } }, - { "default", { {"VW",4}, {"WGS",64}, {"WPT",1} } }, + { "default", { {"VW",4}, {"WGS",256}, {"WPT",1} } }, } }, { // Default @@ -134,6 +136,7 @@ const Database::DatabaseEntry XaxpyComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",4}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",4}, {"WGS",256}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",1}, {"WGS",1024}, {"WPT",2} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW",4}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",2}, {"WGS",1024}, {"WPT",1} } }, { "default", { {"VW",8}, {"WGS",1024}, {"WPT",1} } }, } @@ -160,6 +163,7 @@ const Database::DatabaseEntry XaxpyComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"VW",1}, {"WGS",512}, {"WPT",1} } }, { "GeForce GTX 1070", { {"VW",1}, {"WGS",64}, {"WPT",2} } }, + { "GeForce GTX 1080", { {"VW",2}, {"WGS",64}, {"WPT",1} } }, { "GeForce GTX 480", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "GeForce GTX 670", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "GeForce GTX 680", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, @@ -208,8 +212,9 @@ const Database::DatabaseEntry XaxpyDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",4}, {"WGS",64}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",1}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",8}, {"WGS",64}, {"WPT",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW",8}, {"WGS",256}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",8}, {"WGS",2048}, {"WPT",1} } }, - { "default", { {"VW",8}, {"WGS",512}, {"WPT",1} } }, + { "default", { {"VW",8}, {"WGS",64}, {"WPT",1} } }, } }, { // Intel accelerators @@ -222,6 +227,7 @@ const Database::DatabaseEntry XaxpyDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "GeForce GTX 1070", { {"VW",1}, {"WGS",64}, {"WPT",8} } }, + { "GeForce GTX 1080", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "GeForce GTX 480", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "GeForce GTX 670", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "GeForce GTX 680", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, @@ -270,6 +276,7 @@ const Database::DatabaseEntry XaxpyComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW",4}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW",8}, {"WGS",128}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW",8}, {"WGS",512}, {"WPT",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW",8}, {"WGS",1024}, {"WPT",1} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "default", { {"VW",4}, {"WGS",1024}, {"WPT",1} } }, } @@ -284,6 +291,7 @@ const Database::DatabaseEntry XaxpyComplexDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "GeForce GTX 1070", { {"VW",1}, {"WGS",64}, {"WPT",2} } }, + { "GeForce GTX 1080", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "GeForce GTX 480", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "GeForce GTX 670", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "GeForce GTX 680", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, diff --git a/src/database/kernels/xdot.hpp b/src/database/kernels/xdot.hpp index a7a4e105..30d98e5d 100644 --- a/src/database/kernels/xdot.hpp +++ b/src/database/kernels/xdot.hpp @@ -51,7 +51,8 @@ const Database::DatabaseEntry XdotSingle = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WGS2",32} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",1024}, {"WGS2",32} } }, - { "default", { {"WGS1",256}, {"WGS2",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",64}, {"WGS2",32} } }, + { "default", { {"WGS1",64}, {"WGS2",32} } }, } }, { // Intel GPUs @@ -69,6 +70,7 @@ const Database::DatabaseEntry XdotSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",128}, {"WGS2",32} } }, { "GeForce GTX 1070", { {"WGS1",128}, {"WGS2",1024} } }, + { "GeForce GTX 1080", { {"WGS1",512}, {"WGS2",64} } }, { "GeForce GTX 480", { {"WGS1",512}, {"WGS2",32} } }, { "GeForce GTX 670", { {"WGS1",512}, {"WGS2",1024} } }, { "GeForce GTX 680", { {"WGS1",128}, {"WGS2",128} } }, @@ -78,7 +80,7 @@ const Database::DatabaseEntry XdotSingle = { { "GeForce GTX TITAN Black", { {"WGS1",512}, {"WGS2",64} } }, { "GeForce GTX TITAN X", { {"WGS1",256}, {"WGS2",32} } }, { "Tesla K20m", { {"WGS1",1024}, {"WGS2",32} } }, - { "default", { {"WGS1",256}, {"WGS2",256} } }, + { "default", { {"WGS1",256}, {"WGS2",64} } }, } }, { // Default @@ -108,7 +110,8 @@ const Database::DatabaseEntry XdotComplexSingle = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WGS2",64} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",1024}, {"WGS2",32} } }, - { "default", { {"WGS1",128}, {"WGS2",64} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",256}, {"WGS2",32} } }, + { "default", { {"WGS1",256}, {"WGS2",32} } }, } }, { // Intel GPUs @@ -126,6 +129,7 @@ const Database::DatabaseEntry XdotComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",64}, {"WGS2",32} } }, { "GeForce GTX 1070", { {"WGS1",128}, {"WGS2",32} } }, + { "GeForce GTX 1080", { {"WGS1",128}, {"WGS2",64} } }, { "GeForce GTX 480", { {"WGS1",512}, {"WGS2",32} } }, { "GeForce GTX 670", { {"WGS1",256}, {"WGS2",32} } }, { "GeForce GTX 680", { {"WGS1",128}, {"WGS2",64} } }, @@ -164,13 +168,15 @@ const Database::DatabaseEntry XdotDouble = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",64}, {"WGS2",128} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",512}, {"WGS2",64} } }, - { "default", { {"WGS1",128}, {"WGS2",64} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",256}, {"WGS2",32} } }, + { "default", { {"WGS1",256}, {"WGS2",64} } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",128}, {"WGS2",32} } }, { "GeForce GTX 1070", { {"WGS1",128}, {"WGS2",512} } }, + { "GeForce GTX 1080", { {"WGS1",128}, {"WGS2",128} } }, { "GeForce GTX 480", { {"WGS1",512}, {"WGS2",32} } }, { "GeForce GTX 670", { {"WGS1",256}, {"WGS2",32} } }, { "GeForce GTX 680", { {"WGS1",128}, {"WGS2",64} } }, @@ -180,7 +186,7 @@ const Database::DatabaseEntry XdotDouble = { { "GeForce GTX TITAN Black", { {"WGS1",128}, {"WGS2",64} } }, { "GeForce GTX TITAN X", { {"WGS1",256}, {"WGS2",32} } }, { "Tesla K20m", { {"WGS1",512}, {"WGS2",32} } }, - { "default", { {"WGS1",128}, {"WGS2",64} } }, + { "default", { {"WGS1",128}, {"WGS2",128} } }, } }, { // Default @@ -209,6 +215,7 @@ const Database::DatabaseEntry XdotComplexDouble = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WGS2",128} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",1024}, {"WGS2",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",32}, {"WGS2",32} } }, { "default", { {"WGS1",128}, {"WGS2",32} } }, } }, @@ -216,6 +223,7 @@ const Database::DatabaseEntry XdotComplexDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",64}, {"WGS2",32} } }, { "GeForce GTX 1070", { {"WGS1",128}, {"WGS2",64} } }, + { "GeForce GTX 1080", { {"WGS1",128}, {"WGS2",32} } }, { "GeForce GTX 480", { {"WGS1",512}, {"WGS2",32} } }, { "GeForce GTX 670", { {"WGS1",512}, {"WGS2",128} } }, { "GeForce GTX 680", { {"WGS1",256}, {"WGS2",64} } }, @@ -230,7 +238,7 @@ const Database::DatabaseEntry XdotComplexDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"WGS1",256}, {"WGS2",64} } }, + { "default", { {"WGS1",128}, {"WGS2",32} } }, } }, } diff --git a/src/database/kernels/xgemm.hpp b/src/database/kernels/xgemm.hpp index 06f0f26a..d9414f8b 100644 --- a/src/database/kernels/xgemm.hpp +++ b/src/database/kernels/xgemm.hpp @@ -58,8 +58,9 @@ const Database::DatabaseEntry XgemmSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",32}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, - { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, } }, { // Intel GPUs @@ -84,6 +85,7 @@ const Database::DatabaseEntry XgemmSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",4} } }, { "GeForce GTX 1070", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",4}, {"VWN",1} } }, + { "GeForce GTX 1080", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",8} } }, { "GeForce GTX 480", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",2} } }, { "GeForce GTX 670", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, { "GeForce GTX 680", { {"KWG",32}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",2} } }, @@ -95,7 +97,7 @@ const Database::DatabaseEntry XgemmSingle = { { "GeForce GTX TITAN X", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",8} } }, { "Tesla K20m", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, { "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",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",2} } }, } }, { // Default @@ -133,8 +135,9 @@ const Database::DatabaseEntry XgemmComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",1}, {"VWM",1}, {"VWN",2} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",4} } }, - { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, } }, { // Intel GPUs @@ -159,6 +162,7 @@ const Database::DatabaseEntry XgemmComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"KWG",16}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "GeForce GTX 1070", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",4} } }, + { "GeForce GTX 1080", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "GeForce GTX 480", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",2} } }, { "GeForce GTX 670", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",1} } }, { "GeForce GTX 680", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, @@ -207,8 +211,9 @@ const Database::DatabaseEntry XgemmDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",8} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",8} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, } }, { // Intel accelerators @@ -221,6 +226,7 @@ const Database::DatabaseEntry XgemmDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",2} } }, { "GeForce GTX 1070", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, + { "GeForce GTX 1080", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, { "GeForce GTX 480", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "GeForce GTX 670", { {"KWG",32}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",32}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "GeForce GTX 680", { {"KWG",32}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",4} } }, @@ -232,12 +238,12 @@ const Database::DatabaseEntry XgemmDouble = { { "GeForce GTX TITAN X", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tesla K20m", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "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",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, } }, } @@ -269,8 +275,9 @@ const Database::DatabaseEntry XgemmComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",1} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",32}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",1}, {"VWN",8} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"KWG",32}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, - { "default", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, } }, { // Intel accelerators @@ -283,6 +290,7 @@ const Database::DatabaseEntry XgemmComplexDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"KWG",32}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",1} } }, { "GeForce GTX 1070", { {"KWG",32}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",1}, {"VWN",4} } }, + { "GeForce GTX 1080", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "GeForce GTX 480", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "GeForce GTX 670", { {"KWG",32}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",1}, {"VWN",2} } }, { "GeForce GTX 680", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, @@ -293,12 +301,12 @@ const Database::DatabaseEntry XgemmComplexDouble = { { "GeForce GTX TITAN X", { {"KWG",32}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tesla K20m", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "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",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, } }, } diff --git a/src/database/kernels/xgemm_direct.hpp b/src/database/kernels/xgemm_direct.hpp index d7aacfa3..c0cd2c04 100644 --- a/src/database/kernels/xgemm_direct.hpp +++ b/src/database/kernels/xgemm_direct.hpp @@ -46,7 +46,8 @@ const Database::DatabaseEntry XgemmDirectSingle = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",8}, {"WGD",64} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",8}, {"WGD",64} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",2}, {"VWND",2}, {"WGD",64} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",4}, {"WGD",32} } }, } }, { // Intel GPUs @@ -58,6 +59,7 @@ const Database::DatabaseEntry XgemmDirectSingle = { }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { + { "GeForce GTX 1080", { {"KWID",16}, {"MDIMAD",16}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, { "GeForce GTX TITAN Black", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, @@ -65,7 +67,7 @@ const Database::DatabaseEntry XgemmDirectSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, } }, } @@ -86,7 +88,8 @@ const Database::DatabaseEntry XgemmDirectComplexSingle = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, } }, { // Intel GPUs @@ -98,6 +101,7 @@ const Database::DatabaseEntry XgemmDirectComplexSingle = { }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { + { "GeForce GTX 1080", { {"KWID",8}, {"MDIMAD",8}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, { "GeForce GTX 750 Ti", { {"KWID",16}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",16} } }, { "GeForce GTX TITAN Black", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, @@ -125,19 +129,21 @@ const Database::DatabaseEntry XgemmDirectDouble = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { + { "GeForce GTX 1080", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",4}, {"WGD",32} } }, { "GeForce GTX TITAN Black", { {"KWID",8}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, } }, } @@ -157,11 +163,13 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",32}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",32}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"KWID",8}, {"MDIMAD",16}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { + { "GeForce GTX 1080", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",32}, {"MDIMCD",32}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "GeForce GTX TITAN Black", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, @@ -169,7 +177,7 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } }, } diff --git a/src/database/kernels/xgemv.hpp b/src/database/kernels/xgemv.hpp index 7d1781f9..52b17d94 100644 --- a/src/database/kernels/xgemv.hpp +++ b/src/database/kernels/xgemv.hpp @@ -53,6 +53,7 @@ const Database::DatabaseEntry XgemvSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",64}, {"WPT1",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",64}, {"WPT1",4} } }, { "default", { {"WGS1",64}, {"WPT1",4} } }, } }, @@ -78,6 +79,7 @@ const Database::DatabaseEntry XgemvSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",256}, {"WPT1",1} } }, { "GeForce GTX 1070", { {"WGS1",128}, {"WPT1",1} } }, + { "GeForce GTX 1080", { {"WGS1",32}, {"WPT1",1} } }, { "GeForce GTX 480", { {"WGS1",64}, {"WPT1",1} } }, { "GeForce GTX 670", { {"WGS1",64}, {"WPT1",1} } }, { "GeForce GTX 680", { {"WGS1",256}, {"WPT1",1} } }, @@ -121,6 +123,7 @@ const Database::DatabaseEntry XgemvComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",128}, {"WPT1",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",64}, {"WPT1",4} } }, { "default", { {"WGS1",64}, {"WPT1",2} } }, } }, @@ -146,6 +149,7 @@ const Database::DatabaseEntry XgemvComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",256}, {"WPT1",1} } }, { "GeForce GTX 1070", { {"WGS1",64}, {"WPT1",1} } }, + { "GeForce GTX 1080", { {"WGS1",32}, {"WPT1",1} } }, { "GeForce GTX 480", { {"WGS1",64}, {"WPT1",1} } }, { "GeForce GTX 670", { {"WGS1",64}, {"WPT1",1} } }, { "GeForce GTX 680", { {"WGS1",64}, {"WPT1",1} } }, @@ -184,6 +188,7 @@ const Database::DatabaseEntry XgemvDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",64}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",64}, {"WPT1",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",64}, {"WPT1",4} } }, { "default", { {"WGS1",64}, {"WPT1",4} } }, } }, @@ -197,6 +202,7 @@ const Database::DatabaseEntry XgemvDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",128}, {"WPT1",1} } }, { "GeForce GTX 1070", { {"WGS1",64}, {"WPT1",1} } }, + { "GeForce GTX 1080", { {"WGS1",32}, {"WPT1",1} } }, { "GeForce GTX 480", { {"WGS1",256}, {"WPT1",1} } }, { "GeForce GTX 670", { {"WGS1",128}, {"WPT1",1} } }, { "GeForce GTX 680", { {"WGS1",128}, {"WPT1",1} } }, @@ -213,7 +219,7 @@ const Database::DatabaseEntry XgemvDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"WGS1",128}, {"WPT1",1} } }, + { "default", { {"WGS1",64}, {"WPT1",1} } }, } }, } @@ -239,6 +245,7 @@ const Database::DatabaseEntry XgemvComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",64}, {"WPT1",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",64}, {"WPT1",1} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"WGS1",64}, {"WPT1",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",32}, {"WPT1",4} } }, { "default", { {"WGS1",64}, {"WPT1",4} } }, } }, diff --git a/src/database/kernels/xgemv_fast.hpp b/src/database/kernels/xgemv_fast.hpp index 988687d5..2dd400bc 100644 --- a/src/database/kernels/xgemv_fast.hpp +++ b/src/database/kernels/xgemv_fast.hpp @@ -53,6 +53,7 @@ const Database::DatabaseEntry XgemvFastSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",1}, {"WGS2",32}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",4}, {"WGS2",128}, {"WPT2",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW2",2}, {"WGS2",16}, {"WPT2",4} } }, { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, } }, @@ -78,6 +79,7 @@ const Database::DatabaseEntry XgemvFastSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"VW2",2}, {"WGS2",256}, {"WPT2",2} } }, { "GeForce GTX 1070", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, + { "GeForce GTX 1080", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, { "GeForce GTX 480", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, { "GeForce GTX 670", { {"VW2",2}, {"WGS2",256}, {"WPT2",2} } }, { "GeForce GTX 680", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, @@ -121,7 +123,8 @@ const Database::DatabaseEntry XgemvFastComplexSingle = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",2}, {"WGS2",64}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",1}, {"WGS2",128}, {"WPT2",2} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",4}, {"WGS2",64}, {"WPT2",4} } }, - { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW2",4}, {"WGS2",16}, {"WPT2",4} } }, + { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",2} } }, } }, { // Intel GPUs @@ -180,6 +183,7 @@ const Database::DatabaseEntry XgemvFastDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",4}, {"WGS2",128}, {"WPT2",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW2",1}, {"WGS2",16}, {"WPT2",4} } }, { "default", { {"VW2",1}, {"WGS2",64}, {"WPT2",4} } }, } }, @@ -193,6 +197,7 @@ const Database::DatabaseEntry XgemvFastDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "GeForce GTX 1070", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, + { "GeForce GTX 1080", { {"VW2",1}, {"WGS2",32}, {"WPT2",2} } }, { "GeForce GTX 480", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "GeForce GTX 670", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, { "GeForce GTX 680", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, @@ -235,6 +240,7 @@ const Database::DatabaseEntry XgemvFastComplexDouble = { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW2",4}, {"WGS2",32}, {"WPT2",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"VW2",2}, {"WGS2",64}, {"WPT2",4} } }, { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { {"VW2",4}, {"WGS2",64}, {"WPT2",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW2",1}, {"WGS2",16}, {"WPT2",2} } }, { "default", { {"VW2",4}, {"WGS2",64}, {"WPT2",4} } }, } }, diff --git a/src/database/kernels/xgemv_fast_rot.hpp b/src/database/kernels/xgemv_fast_rot.hpp index 6fbb4619..36a435b5 100644 --- a/src/database/kernels/xgemv_fast_rot.hpp +++ b/src/database/kernels/xgemv_fast_rot.hpp @@ -46,8 +46,9 @@ const Database::DatabaseEntry XgemvFastRotSingle = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",8}, {"WGS3",16}, {"WPT3",8} } }, - { "default", { {"VW3",8}, {"WGS3",16}, {"WPT3",8} } }, + { "default", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, } }, { // Intel GPUs @@ -62,6 +63,7 @@ const Database::DatabaseEntry XgemvFastRotSingle = { }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { + { "GeForce GTX 1080", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, { "GeForce GTX 750 Ti", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, { "GeForce GTX TITAN", { {"VW3",1}, {"WGS3",16}, {"WPT3",16} } }, { "GeForce GTX TITAN Black", { {"VW3",4}, {"WGS3",128}, {"WPT3",16} } }, @@ -91,6 +93,7 @@ const Database::DatabaseEntry XgemvFastRotComplexSingle = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, { "default", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, } @@ -127,12 +130,14 @@ const Database::DatabaseEntry XgemvFastRotDouble = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",8}, {"WGS3",16}, {"WPT3",8} } }, { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { + { "GeForce GTX 1080", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, { "GeForce GTX 750 Ti", { {"VW3",4}, {"WGS3",32}, {"WPT3",16} } }, { "GeForce GTX TITAN", { {"VW3",1}, {"WGS3",16}, {"WPT3",16} } }, { "GeForce GTX TITAN Black", { {"VW3",1}, {"WGS3",16}, {"WPT3",16} } }, @@ -161,8 +166,9 @@ const Database::DatabaseEntry XgemvFastRotComplexDouble = { { // Intel CPUs kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"VW3",2}, {"WGS3",16}, {"WPT3",16} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"VW3",2}, {"WGS3",16}, {"WPT3",16} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"VW3",8}, {"WGS3",16}, {"WPT3",16} } }, - { "default", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, + { "default", { {"VW3",2}, {"WGS3",16}, {"WPT3",16} } }, } }, { // Default diff --git a/src/database/kernels/xger.hpp b/src/database/kernels/xger.hpp index f7295ff4..f99b7632 100644 --- a/src/database/kernels/xger.hpp +++ b/src/database/kernels/xger.hpp @@ -58,8 +58,9 @@ const Database::DatabaseEntry XgerSingle = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",32}, {"WGS2",4}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",128}, {"WGS2",2}, {"WPT",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",256}, {"WGS2",4}, {"WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",128}, {"WGS2",1}, {"WPT",4} } }, - { "default", { {"WGS1",64}, {"WGS2",1}, {"WPT",4} } }, + { "default", { {"WGS1",128}, {"WGS2",8}, {"WPT",4} } }, } }, { // Intel GPUs @@ -77,6 +78,7 @@ const Database::DatabaseEntry XgerSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",128}, {"WGS2",1}, {"WPT",2} } }, { "GeForce GTX 1070", { {"WGS1",512}, {"WGS2",1}, {"WPT",1} } }, + { "GeForce GTX 1080", { {"WGS1",16}, {"WGS2",4}, {"WPT",1} } }, { "GeForce GTX 480", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, { "GeForce GTX 670", { {"WGS1",32}, {"WGS2",8}, {"WPT",2} } }, { "GeForce GTX 680", { {"WGS1",128}, {"WGS2",1}, {"WPT",4} } }, @@ -121,8 +123,9 @@ const Database::DatabaseEntry XgerComplexSingle = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WGS2",2}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",256}, {"WGS2",2}, {"WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",512}, {"WGS2",4}, {"WPT",2} } }, - { "default", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, + { "default", { {"WGS1",256}, {"WGS2",2}, {"WPT",4} } }, } }, { // Intel GPUs @@ -140,6 +143,7 @@ const Database::DatabaseEntry XgerComplexSingle = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",64}, {"WGS2",4}, {"WPT",2} } }, { "GeForce GTX 1070", { {"WGS1",16}, {"WGS2",64}, {"WPT",2} } }, + { "GeForce GTX 1080", { {"WGS1",32}, {"WGS2",2}, {"WPT",1} } }, { "GeForce GTX 480", { {"WGS1",128}, {"WGS2",2}, {"WPT",2} } }, { "GeForce GTX 670", { {"WGS1",16}, {"WGS2",32}, {"WPT",2} } }, { "GeForce GTX 680", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, @@ -147,7 +151,7 @@ const Database::DatabaseEntry XgerComplexSingle = { { "GeForce GTX 750 Ti", { {"WGS1",32}, {"WGS2",8}, {"WPT",2} } }, { "GeForce GTX TITAN", { {"WGS1",16}, {"WGS2",16}, {"WPT",2} } }, { "GeForce GTX TITAN Black", { {"WGS1",16}, {"WGS2",16}, {"WPT",2} } }, - { "default", { {"WGS1",128}, {"WGS2",2}, {"WPT",2} } }, + { "default", { {"WGS1",128}, {"WGS2",1}, {"WPT",2} } }, } }, { // Default @@ -183,14 +187,16 @@ const Database::DatabaseEntry XgerDouble = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",512}, {"WGS2",16}, {"WPT",1} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",256}, {"WGS2",4}, {"WPT",4} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",512}, {"WGS2",8}, {"WPT",2} } }, - { "default", { {"WGS1",256}, {"WGS2",4}, {"WPT",2} } }, + { "default", { {"WGS1",256}, {"WGS2",1}, {"WPT",4} } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",128}, {"WGS2",8}, {"WPT",2} } }, { "GeForce GTX 1070", { {"WGS1",32}, {"WGS2",8}, {"WPT",1} } }, + { "GeForce GTX 1080", { {"WGS1",32}, {"WGS2",2}, {"WPT",1} } }, { "GeForce GTX 480", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, { "GeForce GTX 670", { {"WGS1",32}, {"WGS2",32}, {"WPT",2} } }, { "GeForce GTX 680", { {"WGS1",128}, {"WGS2",4}, {"WPT",2} } }, @@ -234,6 +240,7 @@ const Database::DatabaseEntry XgerComplexDouble = { kDeviceTypeCPU, "Intel", { { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { {"WGS1",128}, {"WGS2",4}, {"WPT",4} } }, { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { {"WGS1",512}, {"WGS2",4}, {"WPT",2} } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { {"WGS1",512}, {"WGS2",2}, {"WPT",2} } }, { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, { "default", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, } @@ -242,6 +249,7 @@ const Database::DatabaseEntry XgerComplexDouble = { kDeviceTypeGPU, "NVIDIA", { { "GRID K520", { {"WGS1",16}, {"WGS2",8}, {"WPT",2} } }, { "GeForce GTX 1070", { {"WGS1",8}, {"WGS2",128}, {"WPT",1} } }, + { "GeForce GTX 1080", { {"WGS1",8}, {"WGS2",4}, {"WPT",1} } }, { "GeForce GTX 480", { {"WGS1",64}, {"WGS2",2}, {"WPT",2} } }, { "GeForce GTX 670", { {"WGS1",8}, {"WGS2",16}, {"WPT",2} } }, { "GeForce GTX 680", { {"WGS1",8}, {"WGS2",16}, {"WPT",1} } }, -- cgit v1.2.3