P4 to Git Change 1497050 by gandryey@gera-w8 on 2017/12/20 17:43:56

SWDEV-140410 - [CQE OCL][PAL][Vega10] Observed ~ 12 - 91 % performance drop while running OpenCL version SHOC - BusSpeedReadback - Switch FlushDMA() to a Barrier() call for SDMA synchronization. Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#42 edit [ROCm/clr commit: 4c27bfe641]
2017-12-20 17:55:46 -05:00
parent 21213abfda
commit 308b0efec3
1 changed files with 2 additions and 1 deletions
@@ -1264,7 +1264,8 @@ bool Resource::partialMemCopyTo(VirtualGPU& gpu, const amd::Coord3D& srcOrigin,
  dstResource.wait(gpu, waitOnBusyEngine);

  if (gpu.validateSdmaOverlap(*this, dstResource)) {
-    gpu.flushDMA(SdmaEngine);
+    // Note: PAL should insert a NOP into the command buffer for synchronization
+    gpu.flushCUCaches();
  }

  Pal::ImageLayout imgLayout = {};