P4 to Git Change 1281189 by jatang@jatang-opencl-hsa-stg2 on 2016/06/17 15:16:29

SWDEV-1 - OCL PerfCounter on PAL.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palbe/src/core/hw/gfxip/computeCmdBuffer.cpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palbe/src/core/hw/gfxip/computeCmdBuffer.h#2 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palbe/src/core/hw/gfxip/gfx6/gfx6ComputeCmdBuffer.cpp#4 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palbe/src/core/hw/gfxip/gfx6/gfx6ComputeCmdBuffer.h#4 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palbe/src/core/hw/gfxip/gfx6/gfx6PerfCounter.cpp#2 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palbe/src/core/hw/gfxip/gfx6/gfx6PerfTrace.cpp#2 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palcounters.cpp#2 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palcounters.hpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#12 edit


[ROCm/clr commit: 6e099aabfb]

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

@@ -33,7 +33,8 @@ PalCounterReference::Create(
     return memRef;
 }
 
-PalCounterReference::~PalCounterReference() {
+PalCounterReference::~PalCounterReference()
+{
     // The counter object is always associated with a particular queue,
     // so we have to lock just this queue
     amd::ScopedLock lock(gpu_.execution());
@@ -43,7 +44,8 @@ PalCounterReference::~PalCounterReference() {
 }
 
 bool
-PalCounterReference::growResultArray(uint index) {
+PalCounterReference::growResultArray(uint index)
+{
     if (results_ != nullptr) {
         delete [] results_;
     }
@@ -54,6 +56,68 @@ PalCounterReference::growResultArray(uint index) {
     return true;
 }
 
+bool PalCounterReference::finalize()
+{
+    iPerf()->Finalize();
+
+    // Acquire GPU memory for the query from the pool and bind it.
+    Pal::GpuMemoryRequirements gpuMemReqs = {};
+    iPerf()->GetGpuMemoryRequirements(&gpuMemReqs);
+
+    Pal::GpuMemoryCreateInfo createInfo = {};
+    createInfo.size      = amd::alignUp(gpuMemReqs.size, gpuMemReqs.alignment);
+    createInfo.alignment = gpuMemReqs.alignment;
+    createInfo.vaRange   = Pal::VaRange::Default;
+    createInfo.heapCount = 1;
+    //createInfo.heaps[0]  = Pal::GpuHeapGartCacheable;
+    createInfo.heaps[0]  = gpuMemReqs.heaps[0];
+    createInfo.priority  = Pal::GpuMemPriority::Normal;
+
+    Pal::Result result;
+    size_t gpuMemSize = gpu().dev().iDev()->GetGpuMemorySize(createInfo, &result);
+    if (result != Pal::Result::Success) {
+        return false;
+    }
+
+    char* pMemory = new char[gpuMemSize];
+
+    if (pMemory != nullptr) {
+        result = gpu().dev().iDev()->CreateGpuMemory(createInfo, pMemory, &pGpuMemory);
+
+        if (result == Pal::Result::Success) {
+            // GPU profiler memory is perma-resident.
+            result = gpu().dev().iDev()->AddGpuMemoryReferences(1, &pGpuMemory, nullptr, Pal::GpuMemoryRefCantTrim);
+
+            if (result == Pal::Result::Success) {
+                // GPU profiler memory is perma-mapped.
+                result = pGpuMemory->Map(&pCpuAddr);
+            }
+        }
+        
+        if (result != Pal::Result::Success) {
+            if (pGpuMemory != nullptr) {
+                pGpuMemory->Destroy();
+                pGpuMemory = nullptr;
+                pCpuAddr   = nullptr;
+            }
+
+            delete pMemory;
+            return false;
+        }
+    }
+    else {
+        return false;
+    }
+
+    result = iPerf()->BindGpuMemory(pGpuMemory, 0);
+    if (result == Pal::Result::Success) { 
+        return true;
+    }
+    else {
+        return false;
+    }
+}
+
 PerfCounter::~PerfCounter()
 {
     if (calRef_ == nullptr) {
@@ -106,8 +170,9 @@ PerfCounter::getInfo(uint64_t infoType) const
         return info()->eventIndex_;
     }
     case CL_PERFCOUNTER_DATA: {
-        Unimplemented();
-        //gslCounter()->GetResult(gpu().cs(), reinterpret_cast<uint64*>(calRef_->results()));
+        Pal::GlobalCounterLayout layout = {};
+        counter_->GetGlobalCounterLayout(&layout);
+
         return calRef_->results()[index_];
     }
     default:

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

@@ -38,11 +38,7 @@ public:
         uint maxIndex   //!< the maximum HW counter index in the PAL counter
         );
 
-    void finalize() {
-        iPerf()->Finalize();
-        Pal::GlobalCounterLayout layout = {};
-        layout.sampleCount = referenceCount() - 1;
-        iPerf()->GetGlobalCounterLayout(&layout); }
+    bool finalize();
 
     //! Returns the PAL counter results
     uint64_t*  results() const { return results_; }
@@ -62,6 +58,8 @@ private:
 
     VirtualGPU&     gpu_;           //!< The virtual GPU device object
     uint64_t*       results_;       //!< Counter results
+    Pal::IGpuMemory*    pGpuMemory;
+    void*               pCpuAddr;
 };
 
 //! Performance counter implementation on GPU

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

@@ -2303,6 +2303,12 @@ VirtualGPU::submitPerfCounter(amd::PerfCounterCommand& vcmd)
 
     // Create performance experiment
     Pal::PerfExperimentCreateInfo   createInfo = {};
+
+    createInfo.optionFlags.sampleInternalOperations       = 1;
+    createInfo.optionFlags.cacheFlushOnCounterCollection  = 1;
+    createInfo.optionFlags.sqShaderMask  = 1;
+    createInfo.optionValues.sampleInternalOperations      = true;
+    createInfo.optionValues.cacheFlushOnCounterCollection = true;
     createInfo.optionValues.sqShaderMask = Pal::PerfShaderMaskCs;
 
     PalCounterReference* palRef = PalCounterReference::Create(*this, createInfo);