P4 to Git Change 1792217 by asalmanp@asalmanp-ocl-stg on 2019/06/05 12:26:00

SWDEV-132899 - [OCL][GFX10] increase the numScratchWavesPerCu in Wave32 mode and use the actual num of CUs not the total num of WGPs when calculating the scratch buffer size

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/17474/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#139 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#81 edit


[ROCm/clr commit: de52451edd]

projects/clr/rocclr/runtime/device/pal/paldevice.cpp

@@ -2003,7 +2003,7 @@ bool Device::allocScratch(uint regNum, const VirtualGPU* vgpu) {
         if (scratchBuf->regNum_ > 0) {
           scratchBuf->destroyMemory();
           // Calculate the size of the scratch buffer for a queue
-          uint32_t numTotalCUs = info().maxComputeUnits_;
+          uint32_t numTotalCUs = properties().gfxipProperties.shaderCore.numAvailableCus;
           uint32_t numMaxWaves = settings().numScratchWavesPerCu_ * numTotalCUs;
           scratchBuf->size_ = static_cast<uint64_t>(info().wavefrontWidth_) * scratchBuf->regNum_ *
               numMaxWaves * sizeof(uint32_t);

projects/clr/rocclr/runtime/device/pal/palsettings.cpp

@@ -211,6 +211,9 @@ bool Settings::create(const Pal::DeviceProperties& palProp,
         enableWave32Mode_ = GPU_ENABLE_WAVE32_MODE;
       }
       lcWavefrontSize64_ = !enableWave32Mode_;
+      if (enableWave32Mode_) {
+        numScratchWavesPerCu_ *= 2;
+      }
       // Fall through to AI (gfx9) ...
     case Pal::AsicRevision::Vega20:
       // Enable HW P2P path for Vega20+. Runtime still relies on KMD/PAL for support