Merged in latest changes from 0.7.1 release

2 files changed, 17 insertions, 6 deletions

diff --git a/include/internal/database/xgemm.h b/include/internal/database/xgemm.h
index e24adb19..9ca2bff5 100644
--- a/include/internal/database/xgemm.h
+++ b/include/internal/database/xgemm.h
@@ -18,11 +18,11 @@ const Database::DatabaseEntry Database::XgemmSingle = {
   "Xgemm", Precision::kSingle, {
     { // AMD GPUs
       kDeviceTypeGPU, "AMD", {
-        { "AMD Radeon R9 M370X Compute Engine",              { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",32}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",1} } },
+        { "AMD Radeon R9 M370X Compute Engine",              { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",8} } },
         { "Hawaii",                                          { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",32}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",2} } },
         { "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} } },
-        { "default",                                         { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
+        { "default",                                         { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } },
       }
     },
     { // ARM GPUs
diff --git a/include/internal/tuning.h b/include/internal/tuning.h
index 6ba1db61..3eba6fdb 100644
--- a/include/internal/tuning.h
+++ b/include/internal/tuning.h
@@ -50,14 +50,18 @@ void Tuner(int argc, char* argv[]) {
   // Tests validity of the given arguments
   C::TestValidArguments(args);
 
-  // Tests for validity of the precision
+  // Tests for validity of the precision and retrieves properties
+  auto isAMD = false;
+  auto isGPU = false;
   {
-    auto platform = Platform(args.platform_id);
-    auto device = Device(platform, args.device_id);
+    const auto platform = Platform(args.platform_id);
+    const auto device = Device(platform, args.device_id);
     if (!PrecisionSupported<T>(device)) {
       printf("* Unsupported precision, skipping this tuning run\n\n");
       return;
     }
+    isAMD = device.Vendor() == "AMD" || device.Vendor() == "Advanced Micro Devices, Inc.";
+    isGPU = device.Type() == "GPU";
   }
 
   // Creates input buffers with random data
@@ -86,8 +90,15 @@ void Tuner(int argc, char* argv[]) {
     tuner.UseRandomSearch(1.0/args.fraction);
   }
 
+  // Set extra settings for specific defines. This mimics src/routine.cc.
+  auto defines = std::string{""};
+  if (isAMD && isGPU) {
+    defines += "#define USE_CL_MAD 1\n";
+    defines += "#define USE_STAGGERED_INDICES 1\n";
+  }
+
   // Loads the kernel sources and defines the kernel to tune
-  auto sources = C::GetSources();
+  auto sources = defines + C::GetSources();
   auto id = tuner.AddKernelFromString(sources, C::KernelName(), C::GlobalSize(args), C::LocalSize());
   tuner.SetReferenceFromString(sources, C::KernelName(), C::GlobalSizeRef(args), C::LocalSizeRef());