From 0643a29af51f9eb13e2b276d0a0e74590c699d3b Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Sat, 18 Feb 2017 13:59:10 +0100
Subject: Added tuning parameters for the AMD RX480 GPU (Ellesmere)

---
 src/database/kernels/xgemm_direct.hpp | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

(limited to 'src/database/kernels/xgemm_direct.hpp')

diff --git a/src/database/kernels/xgemm_direct.hpp b/src/database/kernels/xgemm_direct.hpp
index bec0164f..4984475f 100644
--- a/src/database/kernels/xgemm_direct.hpp
+++ b/src/database/kernels/xgemm_direct.hpp
@@ -17,6 +17,12 @@ namespace database {
 
 const Database::DatabaseEntry XgemmDirectHalf = {
   "XgemmDirect", Precision::kHalf, {
+    { // 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} } },
+      }
+    },
     { // 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} } },
@@ -25,7 +31,7 @@ const Database::DatabaseEntry XgemmDirectHalf = {
     },
     { // Default
       kDeviceTypeAll, "default", {
-        { "default",                                         { {"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",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } },
       }
     },
   }
@@ -38,9 +44,10 @@ const Database::DatabaseEntry XgemmDirectSingle = {
     { // 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} } },
+        { "Ellesmere",                                       { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",32}, {"NDIMCD",32}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"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} } },
+        { "default",                                         { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } },
       }
     },
     { // Intel CPUs
@@ -68,7 +75,7 @@ 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",4}, {"VWND",4}, {"WGD",32} } },
       }
     },
   }
@@ -124,6 +131,7 @@ const Database::DatabaseEntry XgemmDirectDouble = {
     { // 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} } },
         { "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} } },
       }
@@ -159,6 +167,7 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = {
     { // 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} } },
         { "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} } },
       }
@@ -181,7 +190,7 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = {
     },
     { // Default
       kDeviceTypeAll, "default", {
-        { "default",                                         { {"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} } },
       }
     },
   }
-- 
cgit v1.2.3