P4 to Git Change 1233743 by bcrane@MSDNBCRANE99-OCL on 2016/02/02 20:45:13

SWDEV-86253 - [QR] 6 to 7% performance drop is observed in BasemarkCL test

	CL1226184 adds serialization to LHIO, because pxproxy accesses global state, without locks, leading to crash, etc.

	To fix OpenCL perf regressions, allow unserialized access to pxproxy.  It may be that OpenCL design leads away from racy behavior, so it may be safe.  This was checked in to gather information.  What software gets perf drops?  Does any software get fixed?

	The dangerous functions are CreateDevice, CreateContext, because the handles from the OS are saved in global caches, and if those global caches are modified in one thread, while other threads are looking-up from the caches - that is a problem.

	TC: http://ocltc:8111/viewModification.html?modId=66278&personal=true

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/rt/GSLDevice.cpp#158 edit

rocclr/runtime/device/gpu/gslbe/src/rt/GSLDevice.cpp

@@ -200,6 +200,9 @@ CALGSLDevice::open(uint32 gpuIndex, bool enableHighPerformanceState, bool report
         m_dcfg.nPatchDumpLevel.value |= NPATCHDUMPLEVEL_BITS_HW_DEBUG;
     }
 
+    // disable serialization in LHIO to get perf gains
+    m_dcfg.nPatchDumpLevel.value |= NPATCHDUMPLEVEL_FEATURE_ALLOWUNSERIALIZEDPXPROXY;
+
     //we can use environment variable CAL_ENABLE_ASYNC_DMA to force dma on or off when we need it
     char *s = NULL;
     if((s = getenv("CAL_ENABLE_ASYNC_DMA")))