P4 to Git Change 1582433 by gandryey@gera-w8 on 2018/07/18 17:23:28

SWDEV-155438 - Produce RGP Queue Timings chunk for OpenCL RGP files
	- Register SDMA queue in order to get SDMA timing. The RGP trace capture with SDMA may cause a HW hang occasionally

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palgpuopen.cpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palgpuopen.hpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#115 edit


[ROCm/clr commit: 8c44442f90]

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

@@ -149,7 +149,8 @@ bool RgpCaptureMgr::Init(Pal::IPlatform* platform)
 // ================================================================================================
 // This function finds out all the queues in the device that we have to synchronize for RGP-traced
 // frames and initializes resources for them.
-bool RgpCaptureMgr::RegisterTimedQueue(VirtualGPU* gpu, bool* debug_vmid) const
+bool RgpCaptureMgr::RegisterTimedQueue(
+  uint32_t queue_id, Pal::IQueue* iQueue, bool* debug_vmid) const
 {
   bool result = true;
 
@@ -157,14 +158,14 @@ bool RgpCaptureMgr::RegisterTimedQueue(VirtualGPU* gpu, bool* debug_vmid) const
   // it may be optional for Vulkan, but we provide it anyway if available).
   Pal::KernelContextInfo kernelContextInfo = {};
 
-  Pal::Result palResult = gpu->queue(MainEngine).iQueue_->QueryKernelContextInfo(&kernelContextInfo);
+  Pal::Result palResult = iQueue->QueryKernelContextInfo(&kernelContextInfo);
 
   // Ensure we've acquired the debug VMID (note that some platforms do not
   // implement this function, so don't fail the whole trace if so)
   *debug_vmid = kernelContextInfo.flags.hasDebugVmid;
 
   // Register the queue with the GPA session class for timed queue operation support.
-  if (trace_.gpa_session_->RegisterTimedQueue(gpu->queue(MainEngine).iQueue_, gpu->index(),
+  if (trace_.gpa_session_->RegisterTimedQueue(iQueue, queue_id,
       kernelContextInfo.contextIdentifier) != Pal::Result::Success) {
     result = false;
   }

projects/clr/rocclr/runtime/device/pal/palgpuopen.hpp

@@ -312,7 +312,7 @@ public:
 
   void WriteBarrierStartMarker(const Pal::Developer::BarrierData& data) const;
   void WriteBarrierEndMarker(const Pal::Developer::BarrierData& data) const;
-  bool RegisterTimedQueue(VirtualGPU* gpu, bool* debug_vmid) const;
+  bool RegisterTimedQueue(uint32_t queue_id, Pal::IQueue* iQueue, bool* debug_vmid) const;
   Pal::Result TimedQueueSubmit(
     Pal::IQueue* queue, uint64_t cmdId, const Pal::SubmitInfo& submitInfo) const;
 

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

@@ -914,7 +914,8 @@ bool VirtualGPU::create(bool profiling, uint deviceQueueSize, uint rtCUs,
   if ((index() != 0) && dev().rgpCaptureMgr() != nullptr) {
     bool dbg_vmid = false;
     state_.rgpCaptureEnabled_ = true;
-    dev().rgpCaptureMgr()->RegisterTimedQueue(this, &dbg_vmid);
+    dev().rgpCaptureMgr()->RegisterTimedQueue(2 * index(), queue(MainEngine).iQueue_, &dbg_vmid);
+    dev().rgpCaptureMgr()->RegisterTimedQueue(2 * index() + 1, queue(SdmaEngine).iQueue_, &dbg_vmid);
   }
 
   return true;