diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2017-06-20 21:19:26 +0200 |
---|---|---|
committer | Cedric Nugteren <web@cedricnugteren.nl> | 2017-06-20 21:19:26 +0200 |
commit | e44feb85763c5dbae66beb51892d8dda2126e04c (patch) | |
tree | 965ed12eee66d34069a700c1310ddf8d298709f9 /src/database/kernels/xgemm_direct.hpp | |
parent | 48f2682eb7ee72b0f9e6f2922569fcf352f8ce5f (diff) |
Changed the structure of the database to reduce compilation time and save memory
Diffstat (limited to 'src/database/kernels/xgemm_direct.hpp')
-rw-r--r-- | src/database/kernels/xgemm_direct.hpp | 160 |
1 files changed, 80 insertions, 80 deletions
diff --git a/src/database/kernels/xgemm_direct.hpp b/src/database/kernels/xgemm_direct.hpp index f3e1bbe0..a2e45e07 100644 --- a/src/database/kernels/xgemm_direct.hpp +++ b/src/database/kernels/xgemm_direct.hpp @@ -16,22 +16,22 @@ namespace database { // ================================================================================================= const Database::DatabaseEntry XgemmDirectHalf = { - "XgemmDirect", Precision::kHalf, { + "XgemmDirect", Precision::kHalf, {"KWID" "MDIMAD" "MDIMCD" "NDIMBD" "NDIMCD" "PADA" "PADB" "VWMD" "VWND" "WGD"}, { { // AMD GPUs kDeviceTypeGPU, "AMD", { - { "Ellesmere", { {"KWID",8}, {"MDIMAD",32}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",32}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, - { "default", { {"KWID",8}, {"MDIMAD",32}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",32}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + { "Ellesmere", { 8, 32, 8, 8, 32, 1, 1, 1, 1, 32 } }, + { "default", { 8, 32, 8, 8, 32, 1, 1, 1, 1, 32 } }, } }, { // 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} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, + { "Intel(R) HD Graphics Skylake ULT GT2", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 8 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 8 } }, } }, { // 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", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, } }, } @@ -40,46 +40,46 @@ const Database::DatabaseEntry XgemmDirectHalf = { // ================================================================================================= const Database::DatabaseEntry XgemmDirectSingle = { - "XgemmDirect", Precision::kSingle, { + "XgemmDirect", Precision::kSingle, {"KWID" "MDIMAD" "MDIMCD" "NDIMBD" "NDIMCD" "PADA" "PADB" "VWMD" "VWND" "WGD"}, { { // AMD GPUs 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} } }, - { "ATI Radeon HD 6750M", { {"KWID",8}, {"MDIMAD",8}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",0}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, - { "Ellesmere", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",32}, {"NDIMCD",32}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",32} } }, - { "Fiji", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, - { "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",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "AMD Radeon R9 M370X Compute Engine", { 2, 8, 8, 8, 8, 1, 1, 2, 2, 32 } }, + { "ATI Radeon HD 6750M", { 8, 8, 16, 8, 8, 1, 0, 2, 2, 32 } }, + { "Ellesmere", { 2, 8, 8, 32, 32, 1, 1, 2, 1, 32 } }, + { "Fiji", { 2, 16, 16, 8, 8, 1, 1, 1, 1, 16 } }, + { "Tonga", { 16, 16, 16, 32, 8, 0, 1, 1, 1, 32 } }, + { "Turks", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 16 } }, + { "default", { 2, 16, 16, 8, 8, 1, 1, 1, 1, 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} } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { {"KWID",16}, {"MDIMAD",16}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",2}, {"VWND",4}, {"WGD",32} } }, - { "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",4}, {"VWND",2}, {"WGD",32} } }, + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 2, 8, 8, 8, 8, 0, 0, 1, 8, 64 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 16, 16, 8, 8, 8, 0, 0, 2, 4, 32 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 2, 8, 8, 8, 8, 0, 0, 2, 2, 64 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 4, 2, 32 } }, } }, { // 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} } }, - { "Iris Pro", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",4}, {"WGD",32} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, + { "Intel(R) HD Graphics Skylake ULT GT2", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 8 } }, + { "Iris Pro", { 2, 16, 16, 8, 8, 1, 1, 2, 4, 32 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 8 } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { - { "GeForce GT 650M", { {"KWID",16}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",16}, {"PADA",1}, {"PADB",0}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, - { "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} } }, - { "TITAN X (Pascal)", { {"KWID",8}, {"MDIMAD",32}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, + { "GeForce GT 650M", { 16, 16, 16, 8, 16, 1, 0, 2, 2, 32 } }, + { "GeForce GTX 1080", { 16, 16, 8, 16, 8, 1, 1, 1, 1, 32 } }, + { "GeForce GTX 750 Ti", { 2, 8, 8, 8, 8, 1, 1, 4, 2, 32 } }, + { "GeForce GTX TITAN Black", { 2, 8, 8, 16, 16, 1, 1, 4, 2, 32 } }, + { "TITAN X (Pascal)", { 8, 32, 8, 8, 16, 1, 1, 1, 1, 32 } }, + { "default", { 2, 8, 8, 16, 16, 1, 1, 4, 2, 32 } }, } }, { // 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", { 2, 8, 8, 8, 8, 1, 1, 4, 2, 32 } }, } }, } @@ -88,44 +88,44 @@ const Database::DatabaseEntry XgemmDirectSingle = { // ================================================================================================= const Database::DatabaseEntry XgemmDirectComplexSingle = { - "XgemmDirect", Precision::kComplexSingle, { + "XgemmDirect", Precision::kComplexSingle, {"KWID" "MDIMAD" "MDIMCD" "NDIMBD" "NDIMCD" "PADA" "PADB" "VWMD" "VWND" "WGD"}, { { // AMD GPUs 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} } }, - { "ATI Radeon HD 6750M", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, - { "Fiji", { {"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} } }, + { "AMD Radeon R9 M370X Compute Engine", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, + { "ATI Radeon HD 6750M", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 8 } }, + { "Fiji", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, + { "Tonga", { 2, 16, 16, 16, 16, 1, 1, 2, 2, 32 } }, + { "Turks", { 2, 8, 8, 8, 8, 1, 1, 2, 2, 16 } }, + { "default", { 2, 16, 16, 16, 16, 1, 1, 2, 2, 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} } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"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(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 2, 8, 8, 8, 8, 0, 0, 4, 4, 32 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 2, 16, 16, 8, 8, 1, 1, 1, 4, 32 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 2, 8, 8, 16, 8, 1, 1, 2, 1, 32 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 4, 4, 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} } }, - { "Iris Pro", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, - { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "Intel(R) HD Graphics Skylake ULT GT2", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, + { "Iris Pro", { 2, 16, 16, 8, 8, 1, 1, 2, 2, 32 } }, + { "default", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, } }, { // 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} } }, - { "TITAN X (Pascal)", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"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} } }, + { "GeForce GTX 1080", { 8, 8, 16, 16, 8, 1, 1, 2, 2, 32 } }, + { "GeForce GTX 750 Ti", { 16, 8, 8, 16, 8, 1, 1, 2, 1, 16 } }, + { "GeForce GTX TITAN Black", { 2, 8, 8, 16, 16, 1, 1, 1, 1, 16 } }, + { "TITAN X (Pascal)", { 2, 16, 16, 8, 8, 1, 1, 1, 1, 16 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 1, 2, 16 } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",32}, {"MDIMCD",32}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + { "default", { 2, 32, 32, 8, 8, 1, 1, 1, 1, 32 } }, } }, } @@ -134,36 +134,36 @@ const Database::DatabaseEntry XgemmDirectComplexSingle = { // ================================================================================================= const Database::DatabaseEntry XgemmDirectDouble = { - "XgemmDirect", Precision::kDouble, { + "XgemmDirect", Precision::kDouble, {"KWID" "MDIMAD" "MDIMCD" "NDIMBD" "NDIMCD" "PADA" "PADB" "VWMD" "VWND" "WGD"}, { { // AMD GPUs kDeviceTypeGPU, "AMD", { - { "AMD Radeon R9 M370X Compute Engine", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, - { "Ellesmere", { {"KWID",8}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",32} } }, - { "Fiji", { {"KWID",16}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"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",1}, {"VWND",1}, {"WGD",32} } }, - { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "AMD Radeon R9 M370X Compute Engine", { 2, 16, 16, 8, 8, 1, 1, 1, 1, 16 } }, + { "Ellesmere", { 8, 16, 16, 8, 16, 1, 1, 2, 1, 32 } }, + { "Fiji", { 16, 8, 8, 8, 16, 1, 1, 1, 1, 16 } }, + { "Tonga", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 32 } }, + { "default", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 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} } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { {"KWID",8}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"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",2}, {"WGD",32} } }, + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 2, 8, 8, 8, 8, 1, 1, 4, 4, 32 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 8, 8, 8, 8, 8, 0, 0, 1, 4, 32 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 2, 8, 8, 8, 8, 1, 1, 4, 4, 32 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 4, 2, 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} } }, - { "TITAN X (Pascal)", { {"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",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, + { "GeForce GTX 1080", { 2, 16, 16, 8, 8, 1, 1, 1, 2, 16 } }, + { "GeForce GTX 750 Ti", { 2, 8, 8, 8, 8, 1, 1, 2, 4, 32 } }, + { "GeForce GTX TITAN Black", { 8, 16, 16, 16, 8, 1, 0, 1, 1, 16 } }, + { "TITAN X (Pascal)", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 16 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 1, 2, 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", { 2, 8, 8, 8, 8, 1, 1, 2, 2, 16 } }, } }, } @@ -172,36 +172,36 @@ const Database::DatabaseEntry XgemmDirectDouble = { // ================================================================================================= const Database::DatabaseEntry XgemmDirectComplexDouble = { - "XgemmDirect", Precision::kComplexDouble, { + "XgemmDirect", Precision::kComplexDouble, {"KWID" "MDIMAD" "MDIMCD" "NDIMBD" "NDIMCD" "PADA" "PADB" "VWMD" "VWND" "WGD"}, { { // AMD GPUs 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} } }, - { "Ellesmere", { {"KWID",16}, {"MDIMAD",32}, {"MDIMCD",32}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, - { "Fiji", { {"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",1}, {"VWND",1}, {"WGD",16} } }, - { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "AMD Radeon R9 M370X Compute Engine", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, + { "Ellesmere", { 16, 32, 32, 16, 8, 0, 0, 1, 1, 32 } }, + { "Fiji", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, + { "Tonga", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 16 } }, + { "default", { 2, 16, 16, 16, 16, 1, 1, 1, 1, 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} } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { {"KWID",8}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",2}, {"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} } }, + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 2, 8, 8, 32, 8, 0, 0, 1, 1, 32 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 8, 16, 16, 8, 8, 0, 0, 2, 1, 32 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 8, 16, 8, 8, 8, 0, 0, 2, 2, 32 } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 2, 2, 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} } }, - { "TITAN X (Pascal)", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"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} } }, + { "GeForce GTX 1080", { 2, 16, 16, 8, 8, 1, 1, 1, 1, 16 } }, + { "GeForce GTX 750 Ti", { 2, 32, 32, 8, 8, 1, 1, 1, 1, 32 } }, + { "GeForce GTX TITAN Black", { 2, 8, 8, 8, 8, 1, 1, 1, 1, 8 } }, + { "TITAN X (Pascal)", { 2, 16, 16, 8, 8, 1, 1, 1, 2, 16 } }, + { "default", { 2, 16, 16, 8, 8, 1, 1, 1, 1, 16 } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, + { "default", { 2, 8, 8, 8, 8, 1, 1, 1, 2, 16 } }, } }, } |