diff options
Diffstat (limited to 'src/database/kernels/xgemm')
-rw-r--r-- | src/database/kernels/xgemm/xgemm_16.hpp | 19 | ||||
-rw-r--r-- | src/database/kernels/xgemm/xgemm_32.hpp | 161 | ||||
-rw-r--r-- | src/database/kernels/xgemm/xgemm_3232.hpp | 153 | ||||
-rw-r--r-- | src/database/kernels/xgemm/xgemm_64.hpp | 125 | ||||
-rw-r--r-- | src/database/kernels/xgemm/xgemm_6464.hpp | 123 |
5 files changed, 406 insertions, 175 deletions
diff --git a/src/database/kernels/xgemm/xgemm_16.hpp b/src/database/kernels/xgemm/xgemm_16.hpp index 53a8ccd9..c9954a9e 100644 --- a/src/database/kernels/xgemm/xgemm_16.hpp +++ b/src/database/kernels/xgemm/xgemm_16.hpp @@ -14,19 +14,28 @@ const DatabaseEntry XgemmHalf = { "Xgemm", Precision::kHalf, {"KWG", "KWI", "MDIMA", "MDIMC", "MWG", "NDIMB", "NDIMC", "NWG", "SA", "SB", "STRM", "STRN", "VWM", "VWN"}, { { // AMD GPUs kDeviceTypeGPU, "AMD", { - { "Ellesmere", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, - { "default", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "Ellesmere", { + { "AMD Radeon RX 480", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "default", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + } }, + { "default", { + { "default", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + } }, } }, { // Intel GPUs kDeviceTypeGPU, "Intel", { - { "Intel(R) HD Graphics Skylake ULT GT2", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, - { "default", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "default", { + { "Intel(R) HD Graphics Skylake ULT GT2", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "default", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "default", { + { "default", { 32, 2, 8, 8, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + } }, } }, } diff --git a/src/database/kernels/xgemm/xgemm_32.hpp b/src/database/kernels/xgemm/xgemm_32.hpp index 2c0a63f2..ea713339 100644 --- a/src/database/kernels/xgemm/xgemm_32.hpp +++ b/src/database/kernels/xgemm/xgemm_32.hpp @@ -14,84 +14,145 @@ const DatabaseEntry XgemmSingle = { "Xgemm", Precision::kSingle, {"KWG", "KWI", "MDIMA", "MDIMC", "MWG", "NDIMB", "NDIMC", "NWG", "SA", "SB", "STRM", "STRN", "VWM", "VWN"}, { { // AMD GPUs kDeviceTypeGPU, "AMD", { - { "AMD Radeon R9 M370X Compute Engine", { 32, 2, 16, 16, 64, 8, 16, 128, 0, 0, 0, 0, 2, 8 } }, - { "ATI Radeon HD 6750M", { 32, 2, 8, 16, 128, 8, 8, 128, 0, 0, 1, 1, 8, 8 } }, - { "Ellesmere", { 32, 2, 8, 8, 16, 16, 16, 64, 1, 1, 0, 0, 1, 2 } }, - { "Fiji", { 32, 2, 16, 16, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, - { "Hawaii", { 16, 2, 16, 32, 128, 32, 8, 64, 1, 1, 1, 1, 4, 2 } }, - { "Oland", { 16, 2, 32, 16, 64, 32, 16, 128, 1, 1, 1, 0, 2, 4 } }, - { "Pitcairn", { 16, 2, 16, 8, 32, 16, 16, 128, 0, 0, 1, 0, 1, 1 } }, - { "Tahiti", { 32, 2, 16, 32, 128, 16, 8, 64, 0, 0, 0, 0, 4, 1 } }, - { "Tonga", { 16, 2, 16, 32, 64, 16, 8, 128, 1, 1, 0, 0, 2, 8 } }, - { "Turks", { 32, 2, 8, 8, 64, 8, 8, 64, 0, 0, 0, 0, 4, 4 } }, - { "default", { 32, 2, 8, 8, 32, 8, 8, 64, 0, 0, 0, 0, 4, 4 } }, + { "Ellesmere", { + { "AMD Radeon RX 480", { 32, 2, 8, 8, 16, 16, 16, 64, 1, 1, 0, 0, 1, 2 } }, + { "default", { 32, 2, 8, 8, 16, 16, 16, 64, 1, 1, 0, 0, 1, 2 } }, + } }, + { "Fiji", { + { "AMD Radeon R9 Fury X", { 32, 2, 16, 16, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "AMD Radeon R9 M370X Compute Engine", { 32, 2, 16, 16, 64, 8, 16, 128, 0, 0, 0, 0, 2, 8 } }, + { "default", { 32, 2, 16, 16, 64, 16, 16, 64, 1, 1, 0, 0, 2, 2 } }, + } }, + { "Hawaii", { + { "AMD Radeon R9 290X", { 16, 2, 16, 32, 128, 32, 8, 64, 1, 1, 1, 1, 4, 2 } }, + { "default", { 16, 2, 16, 32, 128, 32, 8, 64, 1, 1, 1, 1, 4, 2 } }, + } }, + { "Oland", { + { "Oland", { 16, 2, 32, 16, 64, 32, 16, 128, 1, 1, 1, 0, 2, 4 } }, + { "default", { 16, 2, 32, 16, 64, 32, 16, 128, 1, 1, 1, 0, 2, 4 } }, + } }, + { "Pitcairn", { + { "AMD Radeon R9 270X", { 16, 2, 16, 8, 32, 16, 16, 128, 0, 0, 1, 0, 1, 1 } }, + { "default", { 16, 2, 16, 8, 32, 16, 16, 128, 0, 0, 1, 0, 1, 1 } }, + } }, + { "Tahiti", { + { "AMD Radeon HD 7970", { 32, 2, 16, 32, 128, 16, 8, 64, 0, 0, 0, 0, 4, 1 } }, + { "default", { 32, 2, 16, 32, 128, 16, 8, 64, 0, 0, 0, 0, 4, 1 } }, + } }, + { "Tonga", { + { "AMD Radeon R9 380", { 16, 2, 16, 32, 64, 16, 8, 128, 1, 1, 0, 0, 2, 8 } }, + { "default", { 16, 2, 16, 32, 64, 16, 8, 128, 1, 1, 0, 0, 2, 8 } }, + } }, + { "Turks", { + { "AMD Radeon HD 6770M", { 32, 2, 8, 8, 64, 8, 8, 64, 0, 0, 0, 0, 4, 4 } }, + { "default", { 32, 2, 8, 8, 64, 8, 8, 64, 0, 0, 0, 0, 4, 4 } }, + } }, + { "Vancouver", { + { "ATI Radeon HD 6750M", { 32, 2, 8, 16, 128, 8, 8, 128, 0, 0, 1, 1, 8, 8 } }, + { "default", { 32, 2, 8, 16, 128, 8, 8, 128, 0, 0, 1, 1, 8, 8 } }, + } }, + { "default", { + { "default", { 32, 2, 8, 8, 32, 8, 8, 64, 0, 0, 0, 0, 4, 4 } }, + } }, } }, { // ARM GPUs kDeviceTypeGPU, "ARM", { - { "Mali-T628", { 16, 2, 8, 8, 64, 8, 16, 16, 0, 0, 1, 1, 8, 1 } }, - { "default", { 16, 2, 8, 8, 64, 8, 16, 16, 0, 0, 1, 1, 8, 1 } }, + { "default", { + { "Mali-T628", { 16, 2, 8, 8, 64, 8, 16, 16, 0, 0, 1, 1, 8, 1 } }, + { "default", { 16, 2, 8, 8, 64, 8, 16, 16, 0, 0, 1, 1, 8, 1 } }, + } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { - { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 8, 8, 128, 16, 8, 128, 0, 1, 1, 1, 1, 8 } }, - { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 8, 32, 32, 64, 32, 16, 64, 1, 1, 1, 0, 2, 2 } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 16, 8, 128, 16, 8, 64, 0, 0, 1, 0, 1, 2 } }, - { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 32, 8, 128, 8, 8, 128, 1, 1, 1, 1, 2, 8 } }, - { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 16, 2, 8, 8, 128, 8, 8, 128, 1, 1, 1, 0, 1, 8 } }, - { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 8, 16, 16, 64, 32, 32, 64, 0, 1, 1, 0, 1, 2 } }, - { "default", { 32, 2, 8, 8, 32, 8, 8, 64, 1, 1, 0, 0, 4, 4 } }, + { "default", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 8, 8, 128, 16, 8, 128, 0, 1, 1, 1, 1, 8 } }, + { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 8, 32, 32, 64, 32, 16, 64, 1, 1, 1, 0, 2, 2 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 16, 8, 128, 16, 8, 64, 0, 0, 1, 0, 1, 2 } }, + { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 32, 8, 128, 8, 8, 128, 1, 1, 1, 1, 2, 8 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 16, 2, 8, 8, 128, 8, 8, 128, 1, 1, 1, 0, 1, 8 } }, + { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 8, 16, 16, 64, 32, 32, 64, 0, 1, 1, 0, 1, 2 } }, + { "default", { 32, 2, 8, 8, 32, 8, 8, 64, 1, 1, 0, 0, 4, 4 } }, + } }, } }, { // Intel GPUs kDeviceTypeGPU, "Intel", { - { "Intel(R) HD Graphics 530", { 32, 2, 8, 8, 128, 32, 16, 64, 0, 0, 1, 0, 4, 2 } }, - { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { 32, 8, 8, 8, 64, 32, 16, 64, 1, 1, 1, 1, 4, 2 } }, - { "Intel(R) HD Graphics Haswell Ultrabook GT2 Mobile", { 16, 2, 16, 8, 32, 8, 16, 128, 1, 1, 1, 1, 2, 4 } }, - { "Intel(R) HD Graphics IvyBridge M GT2", { 32, 2, 16, 16, 64, 16, 8, 64, 1, 1, 1, 0, 2, 4 } }, - { "Intel(R) HD Graphics Skylake ULT GT2", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 4 } }, - { "Iris", { 16, 8, 16, 8, 128, 32, 16, 64, 1, 1, 1, 1, 4, 1 } }, - { "Iris Pro", { 16, 2, 16, 8, 64, 32, 32, 128, 1, 1, 1, 0, 4, 4 } }, - { "default", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 4 } }, + { "default", { + { "Intel(R) HD Graphics 530", { 32, 2, 8, 8, 128, 32, 16, 64, 0, 0, 1, 0, 4, 2 } }, + { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { 32, 8, 8, 8, 64, 32, 16, 64, 1, 1, 1, 1, 4, 2 } }, + { "Intel(R) HD Graphics Haswell Ultrabook GT2 Mobile", { 16, 2, 16, 8, 32, 8, 16, 128, 1, 1, 1, 1, 2, 4 } }, + { "Intel(R) HD Graphics IvyBridge M GT2", { 32, 2, 16, 16, 64, 16, 8, 64, 1, 1, 1, 0, 2, 4 } }, + { "Intel(R) HD Graphics Skylake ULT GT2", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 4 } }, + { "Iris", { 16, 8, 16, 8, 128, 32, 16, 64, 1, 1, 1, 1, 4, 1 } }, + { "Iris Pro", { 16, 2, 16, 8, 64, 32, 32, 128, 1, 1, 1, 0, 4, 4 } }, + { "default", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 4 } }, + } }, } }, { // Intel accelerators kDeviceTypeAccelerator, "Intel", { - { "Intel(R) Many Integrated Core Acceleration Card", { 32, 2, 32, 32, 32, 32, 8, 128, 0, 0, 1, 0, 1, 4 } }, - { "default", { 32, 2, 32, 32, 32, 32, 8, 128, 0, 0, 1, 0, 1, 4 } }, + { "default", { + { "Intel(R) Many Integrated Core Acceleration Card", { 32, 2, 32, 32, 32, 32, 8, 128, 0, 0, 1, 0, 1, 4 } }, + { "default", { 32, 2, 32, 32, 32, 32, 8, 128, 0, 0, 1, 0, 1, 4 } }, + } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { - { "GRID K520", { 16, 2, 16, 8, 32, 8, 16, 64, 1, 1, 1, 1, 2, 4 } }, - { "GeForce GT 650M", { 32, 2, 8, 8, 32, 32, 32, 64, 1, 1, 0, 0, 4, 2 } }, - { "GeForce GTX 1070", { 16, 2, 32, 16, 128, 32, 8, 128, 1, 1, 1, 0, 4, 1 } }, - { "GeForce GTX 1080", { 32, 2, 16, 8, 64, 8, 8, 64, 1, 1, 1, 1, 4, 8 } }, - { "GeForce GTX 480", { 16, 2, 16, 8, 64, 32, 16, 64, 1, 1, 1, 1, 2, 2 } }, - { "GeForce GTX 670", { 16, 2, 8, 8, 64, 16, 16, 64, 1, 1, 1, 0, 2, 4 } }, - { "GeForce GTX 680", { 32, 8, 8, 16, 64, 32, 16, 128, 1, 1, 0, 0, 4, 2 } }, - { "GeForce GTX 750", { 16, 2, 16, 16, 64, 32, 8, 128, 1, 1, 1, 1, 1, 2 } }, - { "GeForce GTX 750 Ti", { 16, 2, 16, 16, 128, 32, 8, 64, 1, 1, 0, 1, 8, 2 } }, - { "GeForce GTX 980", { 16, 2, 16, 16, 64, 16, 8, 128, 1, 1, 1, 0, 4, 8 } }, - { "GeForce GTX TITAN", { 16, 8, 32, 16, 64, 8, 8, 64, 1, 1, 1, 0, 2, 2 } }, - { "GeForce GTX TITAN Black", { 16, 2, 16, 8, 64, 16, 16, 64, 1, 1, 1, 0, 4, 1 } }, - { "GeForce GTX TITAN X", { 16, 2, 8, 16, 128, 8, 8, 128, 1, 1, 1, 1, 4, 8 } }, - { "TITAN X (Pascal)", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 1 } }, - { "Tesla K20m", { 16, 2, 32, 16, 64, 16, 8, 64, 1, 1, 1, 0, 2, 4 } }, - { "Tesla K40m", { 16, 8, 16, 8, 64, 16, 16, 128, 1, 1, 1, 0, 2, 4 } }, - { "default", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 2 } }, + { "SM2.0", { + { "GeForce GTX 480", { 16, 2, 16, 8, 64, 32, 16, 64, 1, 1, 1, 1, 2, 2 } }, + { "default", { 16, 2, 16, 8, 64, 32, 16, 64, 1, 1, 1, 1, 2, 2 } }, + } }, + { "SM3.0", { + { "GRID K520", { 16, 2, 16, 8, 32, 8, 16, 64, 1, 1, 1, 1, 2, 4 } }, + { "GeForce GT 650M", { 32, 2, 8, 8, 32, 32, 32, 64, 1, 1, 0, 0, 4, 2 } }, + { "GeForce GTX 670", { 16, 2, 8, 8, 64, 16, 16, 64, 1, 1, 1, 0, 2, 4 } }, + { "GeForce GTX 680", { 32, 8, 8, 16, 64, 32, 16, 128, 1, 1, 0, 0, 4, 2 } }, + { "default", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 4 } }, + } }, + { "SM3.5", { + { "GeForce GTX TITAN", { 16, 8, 32, 16, 64, 8, 8, 64, 1, 1, 1, 0, 2, 2 } }, + { "GeForce GTX TITAN Black", { 16, 2, 16, 8, 64, 16, 16, 64, 1, 1, 1, 0, 4, 1 } }, + { "Tesla K20m", { 16, 2, 32, 16, 64, 16, 8, 64, 1, 1, 1, 0, 2, 4 } }, + { "Tesla K40m", { 16, 8, 16, 8, 64, 16, 16, 128, 1, 1, 1, 0, 2, 4 } }, + { "default", { 16, 8, 32, 16, 64, 32, 16, 64, 1, 0, 1, 0, 2, 2 } }, + } }, + { "SM5.0", { + { "GeForce GTX 750", { 16, 2, 16, 16, 64, 32, 8, 128, 1, 1, 1, 1, 1, 2 } }, + { "GeForce GTX 750 Ti", { 16, 2, 16, 16, 128, 32, 8, 64, 1, 1, 0, 1, 8, 2 } }, + { "default", { 32, 2, 8, 8, 64, 32, 32, 64, 0, 0, 0, 0, 2, 1 } }, + } }, + { "SM5.2", { + { "GeForce GTX 980", { 16, 2, 16, 16, 64, 16, 8, 128, 1, 1, 1, 0, 4, 8 } }, + { "GeForce GTX TITAN X", { 16, 2, 8, 16, 128, 8, 8, 128, 1, 1, 1, 1, 4, 8 } }, + { "default", { 16, 2, 16, 16, 128, 16, 8, 128, 1, 1, 1, 0, 4, 8 } }, + } }, + { "SM6.1", { + { "GeForce GTX 1070", { 16, 2, 32, 16, 128, 32, 8, 128, 1, 1, 1, 0, 4, 1 } }, + { "GeForce GTX 1080", { 32, 2, 16, 8, 64, 8, 8, 64, 1, 1, 1, 1, 4, 8 } }, + { "TITAN X (Pascal)", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 1 } }, + { "default", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 1 } }, + } }, + { "default", { + { "default", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 4, 2 } }, + } }, } }, { // QUALCOMM GPUs kDeviceTypeGPU, "QUALCOMM", { - { "QUALCOMM Adreno(TM)", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 1 } }, - { "default", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 1 } }, + { "default", { + { "QUALCOMM Adreno(TM)", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 1 } }, + { "default", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 1 } }, + } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 2 } }, + { "default", { + { "default", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 2 } }, + } }, } }, } diff --git a/src/database/kernels/xgemm/xgemm_3232.hpp b/src/database/kernels/xgemm/xgemm_3232.hpp index 3ba015e2..6f29dda1 100644 --- a/src/database/kernels/xgemm/xgemm_3232.hpp +++ b/src/database/kernels/xgemm/xgemm_3232.hpp @@ -14,77 +14,136 @@ const DatabaseEntry XgemmComplexSingle = { "Xgemm", Precision::kComplexSingle, {"KWG", "KWI", "MDIMA", "MDIMC", "MWG", "NDIMB", "NDIMC", "NWG", "SA", "SB", "STRM", "STRN", "VWM", "VWN"}, { { // AMD GPUs kDeviceTypeGPU, "AMD", { - { "AMD Radeon R9 M370X Compute Engine", { 32, 2, 32, 32, 64, 8, 8, 64, 0, 0, 1, 1, 2, 8 } }, - { "ATI Radeon HD 6750M", { 32, 2, 8, 8, 32, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, - { "Ellesmere", { 32, 2, 16, 16, 32, 8, 8, 32, 1, 1, 0, 0, 1, 4 } }, - { "Fiji", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 1, 2 } }, - { "Hawaii", { 32, 2, 32, 8, 32, 8, 16, 32, 1, 0, 1, 0, 1, 1 } }, - { "Oland", { 32, 2, 16, 8, 32, 32, 32, 128, 1, 0, 0, 1, 2, 4 } }, - { "Pitcairn", { 16, 2, 8, 8, 32, 8, 8, 32, 0, 1, 1, 1, 4, 2 } }, - { "Tahiti", { 16, 2, 8, 8, 32, 8, 16, 32, 1, 0, 0, 1, 2, 1 } }, - { "Tonga", { 16, 2, 32, 8, 64, 16, 32, 64, 1, 1, 1, 0, 2, 1 } }, - { "Turks", { 16, 2, 8, 8, 32, 32, 8, 32, 0, 1, 0, 0, 2, 1 } }, - { "default", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 1, 2 } }, + { "Ellesmere", { + { "AMD Radeon RX 480", { 32, 2, 16, 16, 32, 8, 8, 32, 1, 1, 0, 0, 1, 4 } }, + { "default", { 32, 2, 16, 16, 32, 8, 8, 32, 1, 1, 0, 0, 1, 4 } }, + } }, + { "Fiji", { + { "AMD Radeon R9 Fury X", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 1, 2 } }, + { "AMD Radeon R9 M370X Compute Engine", { 32, 2, 32, 32, 64, 8, 8, 64, 0, 0, 1, 1, 2, 8 } }, + { "default", { 32, 2, 8, 8, 16, 32, 32, 32, 1, 1, 0, 0, 1, 1 } }, + } }, + { "Hawaii", { + { "AMD Radeon R9 290X", { 32, 2, 32, 8, 32, 8, 16, 32, 1, 0, 1, 0, 1, 1 } }, + { "default", { 32, 2, 32, 8, 32, 8, 16, 32, 1, 0, 1, 0, 1, 1 } }, + } }, + { "Oland", { + { "Oland", { 32, 2, 16, 8, 32, 32, 32, 128, 1, 0, 0, 1, 2, 4 } }, + { "default", { 32, 2, 16, 8, 32, 32, 32, 128, 1, 0, 0, 1, 2, 4 } }, + } }, + { "Pitcairn", { + { "AMD Radeon R9 270X", { 16, 2, 8, 8, 32, 8, 8, 32, 0, 1, 1, 1, 4, 2 } }, + { "default", { 16, 2, 8, 8, 32, 8, 8, 32, 0, 1, 1, 1, 4, 2 } }, + } }, + { "Tahiti", { + { "AMD Radeon HD 7970", { 16, 2, 8, 8, 32, 8, 16, 32, 1, 0, 0, 1, 2, 1 } }, + { "default", { 16, 2, 8, 8, 32, 8, 16, 32, 1, 0, 0, 1, 2, 1 } }, + } }, + { "Tonga", { + { "AMD Radeon R9 380", { 16, 2, 32, 8, 64, 16, 32, 64, 1, 1, 1, 0, 2, 1 } }, + { "default", { 16, 2, 32, 8, 64, 16, 32, 64, 1, 1, 1, 0, 2, 1 } }, + } }, + { "Turks", { + { "AMD Radeon HD 6770M", { 16, 2, 8, 8, 32, 32, 8, 32, 0, 1, 0, 0, 2, 1 } }, + { "default", { 16, 2, 8, 8, 32, 32, 8, 32, 0, 1, 0, 0, 2, 1 } }, + } }, + { "Vancouver", { + { "ATI Radeon HD 6750M", { 32, 2, 8, 8, 32, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "default", { 32, 2, 8, 8, 32, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + } }, + { "default", { + { "default", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 1, 2 } }, + } }, } }, { // ARM GPUs kDeviceTypeGPU, "ARM", { - { "Mali-T628", { 16, 2, 8, 16, 128, 16, 8, 128, 0, 0, 0, 1, 8, 1 } }, - { "default", { 16, 2, 8, 16, 128, 16, 8, 128, 0, 0, 0, 1, 8, 1 } }, + { "default", { + { "Mali-T628", { 16, 2, 8, 16, 128, 16, 8, 128, 0, 0, 0, 1, 8, 1 } }, + { "default", { 16, 2, 8, 16, 128, 16, 8, 128, 0, 0, 0, 1, 8, 1 } }, + } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { - { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 32, 8, 128, 16, 16, 128, 1, 1, 0, 1, 1, 2 } }, - { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 2, 32, 16, 32, 16, 16, 64, 0, 1, 1, 0, 1, 2 } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 16, 16, 64, 8, 16, 64, 0, 1, 0, 0, 4, 4 } }, - { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 8, 8, 128, 16, 32, 128, 0, 0, 0, 0, 1, 4 } }, - { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 32, 2, 8, 8, 128, 32, 8, 128, 0, 0, 0, 0, 1, 4 } }, - { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 2, 8, 16, 16, 16, 16, 128, 0, 0, 1, 1, 1, 4 } }, - { "default", { 32, 2, 16, 16, 64, 8, 8, 32, 0, 0, 0, 0, 4, 2 } }, + { "default", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 32, 8, 128, 16, 16, 128, 1, 1, 0, 1, 1, 2 } }, + { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 2, 32, 16, 32, 16, 16, 64, 0, 1, 1, 0, 1, 2 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 16, 16, 64, 8, 16, 64, 0, 1, 0, 0, 4, 4 } }, + { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 8, 8, 128, 16, 32, 128, 0, 0, 0, 0, 1, 4 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 32, 2, 8, 8, 128, 32, 8, 128, 0, 0, 0, 0, 1, 4 } }, + { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 2, 8, 16, 16, 16, 16, 128, 0, 0, 1, 1, 1, 4 } }, + { "default", { 32, 2, 16, 16, 64, 8, 8, 32, 0, 0, 0, 0, 4, 2 } }, + } }, } }, { // Intel GPUs kDeviceTypeGPU, "Intel", { - { "Intel(R) HD Graphics 530", { 16, 8, 8, 16, 64, 32, 8, 32, 0, 0, 0, 0, 2, 1 } }, - { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { 16, 8, 8, 8, 32, 16, 16, 64, 1, 0, 0, 0, 4, 4 } }, - { "Intel(R) HD Graphics Haswell Ultrabook GT2 Mobile", { 32, 8, 16, 16, 64, 16, 16, 64, 1, 1, 1, 1, 2, 1 } }, - { "Intel(R) HD Graphics IvyBridge M GT2", { 32, 2, 16, 16, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, - { "Intel(R) HD Graphics Skylake ULT GT2", { 32, 2, 16, 16, 64, 16, 16, 64, 0, 0, 0, 0, 4, 2 } }, - { "Iris", { 32, 8, 32, 16, 64, 8, 16, 64, 1, 0, 1, 0, 1, 1 } }, - { "Iris Pro", { 16, 2, 8, 8, 32, 32, 8, 32, 1, 1, 1, 1, 1, 1 } }, - { "default", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 2 } }, + { "default", { + { "Intel(R) HD Graphics 530", { 16, 8, 8, 16, 64, 32, 8, 32, 0, 0, 0, 0, 2, 1 } }, + { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { 16, 8, 8, 8, 32, 16, 16, 64, 1, 0, 0, 0, 4, 4 } }, + { "Intel(R) HD Graphics Haswell Ultrabook GT2 Mobile", { 32, 8, 16, 16, 64, 16, 16, 64, 1, 1, 1, 1, 2, 1 } }, + { "Intel(R) HD Graphics IvyBridge M GT2", { 32, 2, 16, 16, 64, 16, 16, 64, 1, 1, 0, 0, 4, 4 } }, + { "Intel(R) HD Graphics Skylake ULT GT2", { 32, 2, 16, 16, 64, 16, 16, 64, 0, 0, 0, 0, 4, 2 } }, + { "Iris", { 32, 8, 32, 16, 64, 8, 16, 64, 1, 0, 1, 0, 1, 1 } }, + { "Iris Pro", { 16, 2, 8, 8, 32, 32, 8, 32, 1, 1, 1, 1, 1, 1 } }, + { "default", { 32, 2, 8, 8, 32, 8, 8, 32, 1, 1, 0, 0, 4, 2 } }, + } }, } }, { // Intel accelerators kDeviceTypeAccelerator, "Intel", { - { "Intel(R) Many Integrated Core Acceleration Card", { 32, 2, 32, 32, 32, 32, 16, 128, 1, 0, 0, 0, 1, 4 } }, - { "default", { 32, 2, 32, 32, 32, 32, 16, 128, 1, 0, 0, 0, 1, 4 } }, + { "default", { + { "Intel(R) Many Integrated Core Acceleration Card", { 32, 2, 32, 32, 32, 32, 16, 128, 1, 0, 0, 0, 1, 4 } }, + { "default", { 32, 2, 32, 32, 32, 32, 16, 128, 1, 0, 0, 0, 1, 4 } }, + } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { - { "GRID K520", { 16, 8, 32, 32, 64, 32, 16, 128, 1, 0, 1, 0, 1, 4 } }, - { "GeForce GTX 1070", { 16, 2, 16, 16, 128, 16, 16, 64, 1, 1, 1, 1, 2, 4 } }, - { "GeForce GTX 1080", { 16, 2, 32, 16, 64, 32, 8, 64, 1, 1, 0, 0, 1, 2 } }, - { "GeForce GTX 480", { 16, 2, 16, 16, 32, 32, 16, 128, 0, 1, 1, 1, 2, 2 } }, - { "GeForce GTX 670", { 16, 2, 32, 32, 64, 32, 8, 32, 1, 1, 1, 1, 1, 1 } }, - { "GeForce GTX 680", { 16, 2, 32, 16, 64, 32, 32, 128, 1, 0, 0, 0, 2, 2 } }, - { "GeForce GTX 750", { 16, 8, 16, 16, 64, 16, 16, 64, 1, 1, 1, 0, 2, 2 } }, - { "GeForce GTX 750 Ti", { 16, 2, 16, 8, 32, 32, 16, 64, 1, 1, 1, 0, 1, 2 } }, - { "GeForce GTX 980", { 32, 8, 32, 32, 64, 16, 16, 64, 1, 1, 1, 0, 2, 1 } }, - { "GeForce GTX TITAN", { 16, 8, 16, 16, 64, 32, 16, 64, 1, 1, 1, 0, 1, 1 } }, - { "GeForce GTX TITAN Black", { 16, 2, 8, 16, 64, 8, 8, 32, 0, 1, 1, 0, 1, 2 } }, - { "GeForce GTX TITAN X", { 16, 2, 8, 8, 64, 8, 8, 32, 1, 0, 1, 1, 1, 4 } }, - { "TITAN X (Pascal)", { 32, 2, 32, 32, 64, 8, 8, 32, 1, 1, 0, 0, 2, 4 } }, - { "Tesla K20m", { 32, 2, 8, 16, 64, 8, 16, 64, 1, 0, 0, 0, 1, 4 } }, - { "Tesla K40m", { 16, 2, 32, 32, 32, 32, 8, 64, 0, 1, 0, 0, 1, 1 } }, - { "default", { 32, 2, 8, 8, 16, 32, 32, 64, 1, 1, 0, 0, 1, 1 } }, + { "SM2.0", { + { "GeForce GTX 480", { 16, 2, 16, 16, 32, 32, 16, 128, 0, 1, 1, 1, 2, 2 } }, + { "default", { 16, 2, 16, 16, 32, 32, 16, 128, 0, 1, 1, 1, 2, 2 } }, + } }, + { "SM3.0", { + { "GRID K520", { 16, 8, 32, 32, 64, 32, 16, 128, 1, 0, 1, 0, 1, 4 } }, + { "GeForce GTX 670", { 16, 2, 32, 32, 64, 32, 8, 32, 1, 1, 1, 1, 1, 1 } }, + { "GeForce GTX 680", { 16, 2, 32, 16, 64, 32, 32, 128, 1, 0, 0, 0, 2, 2 } }, + { "default", { 16, 2, 32, 16, 64, 32, 16, 128, 1, 0, 0, 0, 1, 1 } }, + } }, + { "SM3.5", { + { "GeForce GTX TITAN", { 16, 8, 16, 16, 64, 32, 16, 64, 1, 1, 1, 0, 1, 1 } }, + { "GeForce GTX TITAN Black", { 16, 2, 8, 16, 64, 8, 8, 32, 0, 1, 1, 0, 1, 2 } }, + { "Tesla K20m", { 32, 2, 8, 16, 64, 8, 16, 64, 1, 0, 0, 0, 1, 4 } }, + { "Tesla K40m", { 16, 2, 32, 32, 32, 32, 8, 64, 0, 1, 0, 0, 1, 1 } }, + { "default", { 32, 2, 16, 16, 128, 8, 8, 64, 0, 1, 0, 1, 8, 2 } }, + } }, + { "SM5.0", { + { "GeForce GTX 750", { 16, 8, 16, 16, 64, 16, 16, 64, 1, 1, 1, 0, 2, 2 } }, + { "GeForce GTX 750 Ti", { 16, 2, 16, 8, 32, 32, 16, 64, 1, 1, 1, 0, 1, 2 } }, + { "default", { 16, 2, 16, 16, 32, 16, 16, 64, 1, 1, 1, 0, 1, 2 } }, + } }, + { "SM5.2", { + { "GeForce GTX 980", { 32, 8, 32, 32, 64, 16, 16, 64, 1, 1, 1, 0, 2, 1 } }, + { "GeForce GTX TITAN X", { 16, 2, 8, 8, 64, 8, 8, 32, 1, 0, 1, 1, 1, 4 } }, + { "default", { 16, 2, 32, 32, 64, 16, 16, 32, 1, 0, 1, 0, 1, 1 } }, + } }, + { "SM6.1", { + { "GeForce GTX 1070", { 16, 2, 16, 16, 128, 16, 16, 64, 1, 1, 1, 1, 2, 4 } }, + { "GeForce GTX 1080", { 16, 2, 32, 16, 64, 32, 8, 64, 1, 1, 0, 0, 1, 2 } }, + { "TITAN X (Pascal)", { 32, 2, 32, 32, 64, 8, 8, 32, 1, 1, 0, 0, 2, 4 } }, + { "default", { 32, 2, 16, 16, 32, 16, 16, 64, 1, 1, 0, 0, 2, 4 } }, + } }, + { "default", { + { "default", { 32, 2, 8, 8, 16, 32, 32, 64, 1, 1, 0, 0, 1, 1 } }, + } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { 32, 2, 16, 16, 32, 8, 8, 32, 1, 1, 0, 0, 2, 1 } }, + { "default", { + { "default", { 32, 2, 16, 16, 32, 8, 8, 32, 1, 1, 0, 0, 2, 1 } }, + } }, } }, } diff --git a/src/database/kernels/xgemm/xgemm_64.hpp b/src/database/kernels/xgemm/xgemm_64.hpp index 45d04e80..625920b7 100644 --- a/src/database/kernels/xgemm/xgemm_64.hpp +++ b/src/database/kernels/xgemm/xgemm_64.hpp @@ -14,63 +14,114 @@ const DatabaseEntry XgemmDouble = { "Xgemm", Precision::kDouble, {"KWG", "KWI", "MDIMA", "MDIMC", "MWG", "NDIMB", "NDIMC", "NWG", "SA", "SB", "STRM", "STRN", "VWM", "VWN"}, { { // AMD GPUs kDeviceTypeGPU, "AMD", { - { "AMD Radeon R9 M370X Compute Engine", { 32, 2, 16, 16, 64, 8, 8, 32, 0, 0, 0, 0, 4, 4 } }, - { "Ellesmere", { 32, 2, 16, 16, 32, 16, 16, 64, 1, 1, 0, 0, 2, 2 } }, - { "Fiji", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 2, 2 } }, - { "Hawaii", { 16, 8, 32, 8, 128, 8, 8, 32, 0, 1, 0, 0, 1, 4 } }, - { "Oland", { 16, 2, 8, 16, 64, 16, 8, 16, 0, 0, 1, 1, 1, 1 } }, - { "Pitcairn", { 32, 2, 32, 16, 64, 8, 16, 32, 0, 0, 0, 0, 1, 2 } }, - { "Tahiti", { 32, 2, 16, 8, 16, 8, 8, 32, 0, 0, 0, 1, 1, 4 } }, - { "Tonga", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 2, 2 } }, - { "default", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 2, 2 } }, + { "Ellesmere", { + { "AMD Radeon RX 480", { 32, 2, 16, 16, 32, 16, 16, 64, 1, 1, 0, 0, 2, 2 } }, + { "default", { 32, 2, 16, 16, 32, 16, 16, 64, 1, 1, 0, 0, 2, 2 } }, + } }, + { "Fiji", { + { "AMD Radeon R9 Fury X", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 2, 2 } }, + { "AMD Radeon R9 M370X Compute Engine", { 32, 2, 16, 16, 64, 8, 8, 32, 0, 0, 0, 0, 4, 4 } }, + { "default", { 32, 2, 8, 8, 16, 32, 32, 32, 1, 1, 0, 0, 1, 1 } }, + } }, + { "Hawaii", { + { "AMD Radeon R9 290X", { 16, 8, 32, 8, 128, 8, 8, 32, 0, 1, 0, 0, 1, 4 } }, + { "default", { 16, 8, 32, 8, 128, 8, 8, 32, 0, 1, 0, 0, 1, 4 } }, + } }, + { "Oland", { + { "Oland", { 16, 2, 8, 16, 64, 16, 8, 16, 0, 0, 1, 1, 1, 1 } }, + { "default", { 16, 2, 8, 16, 64, 16, 8, 16, 0, 0, 1, 1, 1, 1 } }, + } }, + { "Pitcairn", { + { "AMD Radeon R9 270X", { 32, 2, 32, 16, 64, 8, 16, 32, 0, 0, 0, 0, 1, 2 } }, + { "default", { 32, 2, 32, 16, 64, 8, 16, 32, 0, 0, 0, 0, 1, 2 } }, + } }, + { "Tahiti", { + { "AMD Radeon HD 7970", { 32, 2, 16, 8, 16, 8, 8, 32, 0, 0, 0, 1, 1, 4 } }, + { "default", { 32, 2, 16, 8, 16, 8, 8, 32, 0, 0, 0, 1, 1, 4 } }, + } }, + { "Tonga", { + { "AMD Radeon R9 380", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 2, 2 } }, + { "default", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 2, 2 } }, + } }, + { "default", { + { "default", { 32, 2, 16, 16, 32, 16, 16, 32, 1, 1, 0, 0, 2, 2 } }, + } }, } }, { // ARM GPUs kDeviceTypeGPU, "ARM", { - { "Mali-T628", { 32, 2, 8, 8, 64, 8, 8, 16, 0, 1, 1, 0, 8, 2 } }, - { "default", { 32, 2, 8, 8, 64, 8, 8, 16, 0, 1, 1, 0, 8, 2 } }, + { "default", { + { "Mali-T628", { 32, 2, 8, 8, 64, 8, 8, 16, 0, 1, 1, 0, 8, 2 } }, + { "default", { 32, 2, 8, 8, 64, 8, 8, 16, 0, 1, 1, 0, 8, 2 } }, + } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { - { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 32, 8, 128, 16, 16, 128, 1, 1, 1, 1, 2, 8 } }, - { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 2, 32, 16, 128, 16, 16, 64, 0, 1, 1, 0, 1, 2 } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 32, 16, 128, 16, 16, 128, 0, 0, 1, 0, 1, 2 } }, - { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 16, 8, 128, 8, 8, 64, 1, 0, 0, 1, 2, 8 } }, - { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 32, 2, 16, 8, 128, 8, 8, 128, 1, 0, 0, 0, 2, 8 } }, - { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 2, 8, 16, 128, 16, 8, 128, 0, 0, 1, 1, 1, 8 } }, - { "default", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 1, 4 } }, + { "default", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 32, 8, 128, 16, 16, 128, 1, 1, 1, 1, 2, 8 } }, + { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 2, 32, 16, 128, 16, 16, 64, 0, 1, 1, 0, 1, 2 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 32, 16, 128, 16, 16, 128, 0, 0, 1, 0, 1, 2 } }, + { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 16, 8, 128, 8, 8, 64, 1, 0, 0, 1, 2, 8 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 32, 2, 16, 8, 128, 8, 8, 128, 1, 0, 0, 0, 2, 8 } }, + { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 2, 8, 16, 128, 16, 8, 128, 0, 0, 1, 1, 1, 8 } }, + { "default", { 32, 2, 16, 16, 64, 8, 8, 64, 1, 1, 0, 0, 1, 4 } }, + } }, } }, { // Intel accelerators kDeviceTypeAccelerator, "Intel", { - { "Intel(R) Many Integrated Core Acceleration Card", { 32, 8, 8, 16, 16, 16, 16, 128, 0, 0, 1, 0, 1, 4 } }, - { "default", { 32, 8, 8, 16, 16, 16, 16, 128, 0, 0, 1, 0, 1, 4 } }, + { "default", { + { "Intel(R) Many Integrated Core Acceleration Card", { 32, 8, 8, 16, 16, 16, 16, 128, 0, 0, 1, 0, 1, 4 } }, + { "default", { 32, 8, 8, 16, 16, 16, 16, 128, 0, 0, 1, 0, 1, 4 } }, + } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { - { "GRID K520", { 16, 2, 8, 8, 16, 8, 8, 32, 1, 0, 0, 1, 2, 2 } }, - { "GeForce GTX 1070", { 16, 2, 8, 16, 32, 8, 8, 64, 0, 0, 1, 1, 2, 8 } }, - { "GeForce GTX 1080", { 32, 2, 16, 16, 32, 16, 16, 64, 0, 0, 0, 0, 2, 4 } }, - { "GeForce GTX 480", { 16, 2, 8, 16, 32, 32, 8, 64, 1, 1, 1, 0, 1, 2 } }, - { "GeForce GTX 670", { 32, 8, 16, 32, 128, 16, 8, 32, 0, 1, 1, 0, 1, 1 } }, - { "GeForce GTX 680", { 32, 8, 8, 8, 32, 16, 32, 128, 1, 0, 0, 1, 2, 4 } }, - { "GeForce GTX 750", { 32, 8, 16, 32, 64, 16, 8, 128, 0, 0, 0, 1, 2, 1 } }, - { "GeForce GTX 750 Ti", { 32, 2, 8, 8, 32, 16, 16, 32, 0, 0, 0, 0, 4, 2 } }, - { "GeForce GTX 980", { 32, 8, 16, 8, 64, 32, 32, 128, 0, 0, 1, 0, 2, 4 } }, - { "GeForce GTX TITAN", { 16, 8, 16, 8, 32, 16, 32, 128, 1, 1, 1, 1, 2, 2 } }, - { "GeForce GTX TITAN Black", { 16, 2, 16, 8, 16, 16, 8, 16, 1, 1, 1, 0, 1, 1 } }, - { "GeForce GTX TITAN X", { 16, 8, 16, 16, 16, 16, 16, 16, 0, 0, 0, 0, 1, 1 } }, - { "TITAN X (Pascal)", { 32, 2, 32, 32, 32, 16, 16, 32, 0, 0, 0, 0, 1, 2 } }, - { "Tesla K20m", { 16, 2, 32, 8, 32, 16, 16, 64, 1, 0, 0, 0, 1, 1 } }, - { "Tesla K40m", { 32, 2, 16, 8, 64, 16, 32, 128, 1, 0, 1, 1, 2, 4 } }, - { "default", { 32, 2, 16, 16, 32, 16, 16, 64, 0, 0, 0, 0, 2, 4 } }, + { "SM2.0", { + { "GeForce GTX 480", { 16, 2, 8, 16, 32, 32, 8, 64, 1, 1, 1, 0, 1, 2 } }, + { "default", { 16, 2, 8, 16, 32, 32, 8, 64, 1, 1, 1, 0, 1, 2 } }, + } }, + { "SM3.0", { + { "GRID K520", { 16, 2, 8, 8, 16, 8, 8, 32, 1, 0, 0, 1, 2, 2 } }, + { "GeForce GTX 670", { 32, 8, 16, 32, 128, 16, 8, 32, 0, 1, 1, 0, 1, 1 } }, + { "GeForce GTX 680", { 32, 8, 8, 8, 32, 16, 32, 128, 1, 0, 0, 1, 2, 4 } }, + { "default", { 16, 2, 16, 32, 128, 16, 32, 128, 0, 0, 0, 0, 1, 1 } }, + } }, + { "SM3.5", { + { "GeForce GTX TITAN", { 16, 8, 16, 8, 32, 16, 32, 128, 1, 1, 1, 1, 2, 2 } }, + { "GeForce GTX TITAN Black", { 16, 2, 16, 8, 16, 16, 8, 16, 1, 1, 1, 0, 1, 1 } }, + { "Tesla K20m", { 16, 2, 32, 8, 32, 16, 16, 64, 1, 0, 0, 0, 1, 1 } }, + { "Tesla K40m", { 32, 2, 16, 8, 64, 16, 32, 128, 1, 0, 1, 1, 2, 4 } }, + { "default", { 16, 2, 16, 8, 16, 16, 16, 128, 1, 0, 0, 0, 1, 1 } }, + } }, + { "SM5.0", { + { "GeForce GTX 750", { 32, 8, 16, 32, 64, 16, 8, 128, 0, 0, 0, 1, 2, 1 } }, + { "GeForce GTX 750 Ti", { 32, 2, 8, 8, 32, 16, 16, 32, 0, 0, 0, 0, 4, 2 } }, + { "default", { 32, 2, 16, 32, 32, 16, 16, 128, 0, 0, 0, 0, 2, 1 } }, + } }, + { "SM5.2", { + { "GeForce GTX 980", { 32, 8, 16, 8, 64, 32, 32, 128, 0, 0, 1, 0, 2, 4 } }, + { "GeForce GTX TITAN X", { 16, 8, 16, 16, 16, 16, 16, 16, 0, 0, 0, 0, 1, 1 } }, + { "default", { 16, 8, 16, 16, 16, 16, 16, 128, 0, 0, 0, 0, 1, 1 } }, + } }, + { "SM6.1", { + { "GeForce GTX 1070", { 16, 2, 8, 16, 32, 8, 8, 64, 0, 0, 1, 1, 2, 8 } }, + { "GeForce GTX 1080", { 32, 2, 16, 16, 32, 16, 16, 64, 0, 0, 0, 0, 2, 4 } }, + { "TITAN X (Pascal)", { 32, 2, 32, 32, 32, 16, 16, 32, 0, 0, 0, 0, 1, 2 } }, + { "default", { 32, 2, 16, 16, 32, 16, 16, 64, 0, 0, 0, 0, 2, 4 } }, + } }, + { "default", { + { "default", { 32, 2, 16, 16, 32, 16, 16, 64, 0, 0, 0, 0, 2, 4 } }, + } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { 32, 2, 8, 8, 32, 8, 8, 64, 0, 0, 0, 0, 4, 4 } }, + { "default", { + { "default", { 32, 2, 8, 8, 32, 8, 8, 64, 0, 0, 0, 0, 4, 4 } }, + } }, } }, } diff --git a/src/database/kernels/xgemm/xgemm_6464.hpp b/src/database/kernels/xgemm/xgemm_6464.hpp index 14f47eff..1cddbfae 100644 --- a/src/database/kernels/xgemm/xgemm_6464.hpp +++ b/src/database/kernels/xgemm/xgemm_6464.hpp @@ -14,62 +14,113 @@ const DatabaseEntry XgemmComplexDouble = { "Xgemm", Precision::kComplexDouble, {"KWG", "KWI", "MDIMA", "MDIMC", "MWG", "NDIMB", "NDIMC", "NWG", "SA", "SB", "STRM", "STRN", "VWM", "VWN"}, { { // AMD GPUs kDeviceTypeGPU, "AMD", { - { "AMD Radeon R9 M370X Compute Engine", { 32, 8, 8, 16, 32, 16, 16, 32, 0, 0, 1, 1, 2, 2 } }, - { "Ellesmere", { 32, 2, 16, 16, 16, 16, 16, 16, 1, 1, 0, 0, 1, 1 } }, - { "Fiji", { 32, 2, 16, 16, 16, 16, 16, 16, 1, 1, 0, 0, 1, 1 } }, - { "Hawaii", { 16, 2, 16, 16, 16, 16, 16, 32, 1, 0, 0, 0, 1, 2 } }, - { "Oland", { 16, 2, 16, 8, 16, 16, 32, 128, 0, 0, 0, 0, 1, 4 } }, - { "Pitcairn", { 32, 2, 16, 8, 32, 8, 32, 32, 0, 1, 1, 0, 1, 1 } }, - { "Tahiti", { 16, 2, 16, 8, 16, 8, 8, 16, 0, 0, 1, 0, 1, 1 } }, - { "Tonga", { 16, 2, 32, 16, 32, 16, 16, 16, 1, 1, 1, 1, 1, 1 } }, - { "default", { 32, 2, 16, 16, 16, 16, 16, 16, 1, 1, 0, 0, 1, 1 } }, + { "Ellesmere", { + { "AMD Radeon RX 480", { 32, 2, 16, 16, 16, 16, 16, 16, 1, 1, 0, 0, 1, 1 } }, + { "default", { 32, 2, 16, 16, 16, 16, 16, 16, 1, 1, 0, 0, 1, 1 } }, + } }, + { "Fiji", { + { "AMD Radeon R9 Fury X", { 32, 2, 16, 16, 16, 16, 16, 16, 1, 1, 0, 0, 1, 1 } }, + { "AMD Radeon R9 M370X Compute Engine", { 32, 8, 8, 16, 32, 16, 16, 32, 0, 0, 1, 1, 2, 2 } }, + { "default", { 32, 2, 16, 16, 32, 8, 8, 32, 0, 0, 0, 0, 2, 4 } }, + } }, + { "Hawaii", { + { "AMD Radeon R9 290X", { 16, 2, 16, 16, 16, 16, 16, 32, 1, 0, 0, 0, 1, 2 } }, + { "default", { 16, 2, 16, 16, 16, 16, 16, 32, 1, 0, 0, 0, 1, 2 } }, + } }, + { "Oland", { + { "Oland", { 16, 2, 16, 8, 16, 16, 32, 128, 0, 0, 0, 0, 1, 4 } }, + { "default", { 16, 2, 16, 8, 16, 16, 32, 128, 0, 0, 0, 0, 1, 4 } }, + } }, + { "Pitcairn", { + { "AMD Radeon R9 270X", { 32, 2, 16, 8, 32, 8, 32, 32, 0, 1, 1, 0, 1, 1 } }, + { "default", { 32, 2, 16, 8, 32, 8, 32, 32, 0, 1, 1, 0, 1, 1 } }, + } }, + { "Tahiti", { + { "AMD Radeon HD 7970", { 16, 2, 16, 8, 16, 8, 8, 16, 0, 0, 1, 0, 1, 1 } }, + { "default", { 16, 2, 16, 8, 16, 8, 8, 16, 0, 0, 1, 0, 1, 1 } }, + } }, + { "Tonga", { + { "AMD Radeon R9 380", { 16, 2, 32, 16, 32, 16, 16, 16, 1, 1, 1, 1, 1, 1 } }, + { "default", { 16, 2, 32, 16, 32, 16, 16, 16, 1, 1, 1, 1, 1, 1 } }, + } }, + { "default", { + { "default", { 32, 2, 16, 16, 16, 16, 16, 16, 1, 1, 0, 0, 1, 1 } }, + } }, } }, { // ARM GPUs kDeviceTypeGPU, "ARM", { - { "Mali-T628", { 16, 2, 8, 8, 64, 32, 8, 64, 0, 0, 1, 0, 8, 1 } }, - { "default", { 16, 2, 8, 8, 64, 32, 8, 64, 0, 0, 1, 0, 8, 1 } }, + { "default", { + { "Mali-T628", { 16, 2, 8, 8, 64, 32, 8, 64, 0, 0, 1, 0, 8, 1 } }, + { "default", { 16, 2, 8, 8, 64, 32, 8, 64, 0, 0, 1, 0, 8, 1 } }, + } }, } }, { // Intel CPUs kDeviceTypeCPU, "Intel", { - { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 32, 8, 64, 16, 8, 128, 0, 1, 0, 1, 2, 1 } }, - { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 2, 16, 32, 128, 16, 16, 64, 0, 1, 0, 0, 2, 4 } }, - { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 16, 32, 128, 16, 8, 32, 0, 1, 0, 0, 4, 1 } }, - { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 8, 8, 128, 8, 16, 128, 0, 0, 0, 1, 1, 8 } }, - { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 32, 2, 8, 8, 128, 32, 8, 128, 0, 0, 0, 0, 1, 4 } }, - { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 8, 8, 32, 32, 8, 8, 32, 0, 1, 0, 0, 1, 2 } }, - { "default", { 32, 2, 8, 8, 16, 8, 8, 32, 1, 1, 0, 0, 1, 2 } }, + { "default", { + { "Intel(R) Core(TM) i7-2670QM CPU @ 2.20GHz", { 16, 2, 32, 8, 64, 16, 8, 128, 0, 1, 0, 1, 2, 1 } }, + { "Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz", { 32, 2, 16, 32, 128, 16, 16, 64, 0, 1, 0, 0, 2, 4 } }, + { "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz", { 32, 2, 16, 32, 128, 16, 8, 32, 0, 1, 0, 0, 4, 1 } }, + { "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz", { 32, 2, 8, 8, 128, 8, 16, 128, 0, 0, 0, 1, 1, 8 } }, + { "Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz", { 32, 2, 8, 8, 128, 32, 8, 128, 0, 0, 0, 0, 1, 4 } }, + { "Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz", { 32, 8, 8, 32, 32, 8, 8, 32, 0, 1, 0, 0, 1, 2 } }, + { "default", { 32, 2, 8, 8, 16, 8, 8, 32, 1, 1, 0, 0, 1, 2 } }, + } }, } }, { // Intel accelerators kDeviceTypeAccelerator, "Intel", { - { "Intel(R) Many Integrated Core Acceleration Card", { 32, 2, 16, 16, 16, 16, 8, 32, 0, 0, 1, 0, 1, 1 } }, - { "default", { 32, 2, 16, 16, 16, 16, 8, 32, 0, 0, 1, 0, 1, 1 } }, + { "default", { + { "Intel(R) Many Integrated Core Acceleration Card", { 32, 2, 16, 16, 16, 16, 8, 32, 0, 0, 1, 0, 1, 1 } }, + { "default", { 32, 2, 16, 16, 16, 16, 8, 32, 0, 0, 1, 0, 1, 1 } }, + } }, } }, { // NVIDIA GPUs kDeviceTypeGPU, "NVIDIA", { - { "GRID K520", { 32, 8, 16, 16, 16, 8, 16, 64, 1, 0, 1, 1, 1, 1 } }, - { "GeForce GTX 1070", { 32, 8, 32, 16, 32, 8, 8, 32, 0, 0, 0, 1, 1, 4 } }, - { "GeForce GTX 1080", { 32, 2, 16, 16, 16, 8, 8, 16, 0, 0, 0, 0, 1, 2 } }, - { "GeForce GTX 480", { 16, 2, 32, 32, 32, 32, 8, 32, 0, 0, 1, 0, 1, 1 } }, - { "GeForce GTX 670", { 32, 8, 16, 8, 16, 16, 32, 64, 1, 0, 0, 1, 1, 2 } }, - { "GeForce GTX 680", { 16, 8, 16, 8, 64, 16, 32, 32, 0, 1, 1, 0, 1, 1 } }, - { "GeForce GTX 750", { 32, 2, 8, 32, 32, 8, 8, 64, 0, 0, 1, 0, 1, 4 } }, - { "GeForce GTX 750 Ti", { 32, 2, 8, 8, 16, 8, 8, 32, 0, 0, 0, 0, 1, 1 } }, - { "GeForce GTX 980", { 16, 2, 16, 8, 32, 8, 16, 128, 0, 0, 1, 1, 2, 2 } }, - { "GeForce GTX TITAN Black", { 16, 2, 16, 16, 32, 16, 8, 32, 0, 1, 1, 1, 1, 1 } }, - { "GeForce GTX TITAN X", { 32, 8, 16, 16, 128, 16, 16, 32, 0, 0, 1, 0, 1, 1 } }, - { "TITAN X (Pascal)", { 32, 2, 16, 16, 16, 16, 16, 16, 0, 0, 0, 0, 1, 1 } }, - { "Tesla K20m", { 32, 2, 32, 8, 32, 16, 16, 64, 0, 0, 1, 0, 1, 1 } }, - { "Tesla K40m", { 16, 8, 8, 8, 32, 32, 16, 32, 0, 0, 1, 0, 1, 1 } }, - { "default", { 32, 2, 16, 16, 32, 16, 16, 32, 0, 0, 0, 0, 1, 1 } }, + { "SM2.0", { + { "GeForce GTX 480", { 16, 2, 32, 32, 32, 32, 8, 32, 0, 0, 1, 0, 1, 1 } }, + { "default", { 16, 2, 32, 32, 32, 32, 8, 32, 0, 0, 1, 0, 1, 1 } }, + } }, + { "SM3.0", { + { "GRID K520", { 32, 8, 16, 16, 16, 8, 16, 64, 1, 0, 1, 1, 1, 1 } }, + { "GeForce GTX 670", { 32, 8, 16, 8, 16, 16, 32, 64, 1, 0, 0, 1, 1, 2 } }, + { "GeForce GTX 680", { 16, 8, 16, 8, 64, 16, 32, 32, 0, 1, 1, 0, 1, 1 } }, + { "default", { 16, 8, 16, 16, 16, 16, 16, 32, 0, 0, 0, 0, 1, 1 } }, + } }, + { "SM3.5", { + { "GeForce GTX TITAN Black", { 16, 2, 16, 16, 32, 16, 8, 32, 0, 1, 1, 1, 1, 1 } }, + { "Tesla K20m", { 32, 2, 32, 8, 32, 16, 16, 64, 0, 0, 1, 0, 1, 1 } }, + { "Tesla K40m", { 16, 8, 8, 8, 32, 32, 16, 32, 0, 0, 1, 0, 1, 1 } }, + { "default", { 32, 2, 8, 8, 32, 16, 16, 64, 0, 0, 0, 0, 4, 1 } }, + } }, + { "SM5.0", { + { "GeForce GTX 750", { 32, 2, 8, 32, 32, 8, 8, 64, 0, 0, 1, 0, 1, 4 } }, + { "GeForce GTX 750 Ti", { 32, 2, 8, 8, 16, 8, 8, 32, 0, 0, 0, 0, 1, 1 } }, + { "default", { 32, 2, 8, 32, 16, 8, 8, 32, 0, 0, 0, 0, 1, 1 } }, + } }, + { "SM5.2", { + { "GeForce GTX 980", { 16, 2, 16, 8, 32, 8, 16, 128, 0, 0, 1, 1, 2, 2 } }, + { "GeForce GTX TITAN X", { 32, 8, 16, 16, 128, 16, 16, 32, 0, 0, 1, 0, 1, 1 } }, + { "default", { 16, 2, 16, 16, 128, 16, 16, 128, 0, 0, 1, 0, 1, 1 } }, + } }, + { "SM6.1", { + { "GeForce GTX 1070", { 32, 8, 32, 16, 32, 8, 8, 32, 0, 0, 0, 1, 1, 4 } }, + { "GeForce GTX 1080", { 32, 2, 16, 16, 16, 8, 8, 16, 0, 0, 0, 0, 1, 2 } }, + { "TITAN X (Pascal)", { 32, 2, 16, 16, 16, 16, 16, 16, 0, 0, 0, 0, 1, 1 } }, + { "default", { 32, 2, 32, 32, 32, 32, 32, 64, 0, 0, 0, 0, 1, 2 } }, + } }, + { "default", { + { "default", { 32, 2, 16, 16, 32, 16, 16, 32, 0, 0, 0, 0, 1, 1 } }, + } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { 32, 2, 32, 32, 32, 8, 8, 32, 1, 1, 0, 0, 1, 1 } }, + { "default", { + { "default", { 32, 2, 32, 32, 32, 8, 8, 32, 1, 1, 0, 0, 1, 1 } }, + } }, } }, } |