P4 to Git Change 1367969 by gandryey@gera-lnx-rcf-lc on 2017/02/01 11:53:10

SWDEV-112171 - [ROCm CQE][OCLonLC][QR][G] System hangs/Failures observed with few WF conf tests, due to CL#1364923
	- Disable ROCr copy buffer and use runtime kernel to avoid  HW hangs in integer_ops.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/blit.hpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.cpp#12 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#29 edit


[ROCm/clr commit: c0d4edda9d]

projects/clr/rocclr/runtime/device/blit.hpp

@@ -37,6 +37,7 @@ public:
             uint    disableFillBuffer_          : 1;
             uint    disableFillImage_           : 1;
             uint    disableCopyBufferToImageOpt_: 1;
+            uint    disableHwlCopyBuffer_       : 1;
         };
         uint32_t    value_;
         Setup() : value_(0) {}

projects/clr/rocclr/runtime/device/rocm/rocblit.cpp

@@ -2106,8 +2106,9 @@ KernelBlitManager::copyBuffer(
     amd::ScopedLock k(lockXferOps_);
     bool    result = false;
 
-    if (!gpuMem(srcMemory).isHostMemDirectAccess() &&
-        !gpuMem(dstMemory).isHostMemDirectAccess()) {
+    if (setup_.disableHwlCopyBuffer_ ||
+        (!gpuMem(srcMemory).isHostMemDirectAccess() &&
+         !gpuMem(dstMemory).isHostMemDirectAccess())) {
         uint    blitType = BlitCopyBuffer;
         size_t  dim = 1;
         size_t  globalWorkOffset[3] = { 0, 0, 0 };

projects/clr/rocclr/runtime/device/rocm/rocvirtual.cpp

@@ -539,6 +539,8 @@ VirtualGPU::create(bool profilingEna)
     }
 
     device::BlitManager::Setup  blitSetup;
+    // todo ROCr hangs in a heavy MT environment(integer_ops long_math)
+    blitSetup.disableHwlCopyBuffer_ = true; 
     blitMgr_ = new KernelBlitManager(*this, blitSetup);
     if ((NULL == blitMgr_) || !blitMgr_->create(roc_device_)) {
         LogError("Could not create BlitManager!");