P4 to Git Change 1973670 by gandryey@gera-win10 on 2019/07/25 11:27:11

SWDEV-79445 - OCL generic changes and code clean-up
	- Don't create an extra queue for DMA transfers when SDMA is disabled. That should allow to avoid useless sync operations.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.cpp#13 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palgpuopen.cpp#15 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#75 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paltimestamp.cpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paltimestamp.hpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#143 edit


[ROCm/clr commit: b9b798616f]

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

@@ -69,8 +69,10 @@ address ManagedBuffer::reserve(uint32_t size, uint64_t* gpu_address) {
     // Get the next buffer in the list
     ++activeBuffer_;
     activeBuffer_ %= MaxNumberOfBuffers;
-    // Make sure the buffer isn't busy
-    gpu().waitForEvent(&pool_[activeBuffer_].events[SdmaEngine]);
+    if (!gpu().dev().settings().disableSdma_) {
+      // Make sure the buffer isn't busy
+      gpu().waitForEvent(&pool_[activeBuffer_].events[SdmaEngine]);
+    }
     gpu().waitForEvent(&pool_[activeBuffer_].events[MainEngine]);
     wrtAddress_ = pool_[activeBuffer_].buf->data();
     wrtOffset_ = 0;

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

@@ -582,10 +582,11 @@ Pal::Result RgpCaptureMgr::EndRGPTrace(VirtualGPU* gpu) {
   // end the session itself
   if (result == Pal::Result::Success) {
     assert(trace_.gpa_session_ != nullptr);
+    EngineType engine = (gpu->dev().settings().disableSdma_) ? MainEngine : SdmaEngine;
     // Initiate SDMA copy
-    gpu->eventBegin(SdmaEngine);
-    result = trace_.gpa_session_->End(gpu->queue(SdmaEngine).iCmd());
-    gpu->eventEnd(SdmaEngine, trace_.end_event_);
+    gpu->eventBegin(engine);
+    result = trace_.gpa_session_->End(gpu->queue(engine).iCmd());
+    gpu->eventEnd(engine, trace_.end_event_);
   }
 
   // Submit the trace-end command buffer

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

@@ -1332,7 +1332,12 @@ bool Resource::partialMemCopyTo(VirtualGPU& gpu, const amd::Coord3D& srcOrigin,
     }
   }
 
-  gpu.engineID_ = SdmaEngine;
+  if (dev().settings().disableSdma_) {
+    // Make sure compute is done before CP DMA start
+    gpu.addBarrier(RgpSqqtBarrierReason::MemDependency);
+  } else {
+    gpu.engineID_ = SdmaEngine;
+  }
 
   // Wait for the resources, since runtime may use async transfers
   wait(gpu, waitOnBusyEngine);
@@ -1423,6 +1428,11 @@ bool Resource::partialMemCopyTo(VirtualGPU& gpu, const amd::Coord3D& srcOrigin,
     }
   }
 
+  if (dev().settings().disableSdma_) {
+    // Make sure CP dma is done
+    gpu.addBarrier(RgpSqqtBarrierReason::MemDependency);
+  }
+
   gpu.eventEnd(gpu.engineID_, event);
 
   // Mark source and destination as busy

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

@@ -17,7 +17,7 @@ TimeStamp::TimeStamp(const VirtualGPU& gpu, Pal::IGpuMemory* iMem, uint memOffse
 
 TimeStamp::~TimeStamp() {}
 
-void TimeStamp::begin(bool sdma) {
+void TimeStamp::begin() {
   if (!flags_.beginIssued_) {
     gpu().iCmd()->CmdWriteTimestamp(Pal::HwPipePoint::HwPipeTop, *iMem_,
                                     memOffset_ + CommandStartTime * sizeof(uint64_t));
@@ -25,12 +25,11 @@ void TimeStamp::begin(bool sdma) {
   }
 }
 
-void TimeStamp::end(bool sdma) {
+void TimeStamp::end() {
   CondLog(!flags_.beginIssued_, "We didn't issue a begin operation!");
   gpu().iCmd()->CmdWriteTimestamp(Pal::HwPipePoint::HwPipeTop, *iMem_,
                                   memOffset_ + CommandEndTime * sizeof(uint64_t));
   flags_.endIssued_ = true;
-  flags_.sdma_ = sdma;
 }
 
 inline void SetValue(uint64_t* time, uint64_t val, double nanos) {
@@ -57,7 +56,9 @@ TimeStampCache::~TimeStampCache() {
   for (uint i = 0; i < tsBuf_.size(); ++i) {
     tsBuf_[i]->unmap(&gpu_);
     gpu_.queue(MainEngine).removeMemRef(tsBuf_[i]->iMem());
-    gpu_.queue(SdmaEngine).removeMemRef(tsBuf_[i]->iMem());
+    if (!gpu_.dev().settings().disableSdma_) {
+      gpu_.queue(SdmaEngine).removeMemRef(tsBuf_[i]->iMem());
+    }
     delete tsBuf_[i];
   }
   tsBuf_.clear();
@@ -77,7 +78,9 @@ TimeStamp* TimeStampCache::allocTimeStamp() {
         return nullptr;
       }
       gpu_.queue(MainEngine).addMemRef(buf->iMem());
-      gpu_.queue(SdmaEngine).addMemRef(buf->iMem());
+      if (!gpu_.dev().settings().disableSdma_) {
+        gpu_.queue(SdmaEngine).addMemRef(buf->iMem());
+      }
       tsBufCpu_ = reinterpret_cast<address>(buf->map(&gpu_));
       memset(tsBufCpu_, 0, TimerBufSize);
       tsOffset_ = 0;

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

@@ -29,7 +29,6 @@ class TimeStamp : public amd::HeapObject {
     struct {
       uint32_t beginIssued_ : 1;
       uint32_t endIssued_ : 1;
-      uint32_t sdma_ : 1;
     };
     uint32_t value_;
     Flags() : value_(0) {}
@@ -46,10 +45,10 @@ class TimeStamp : public amd::HeapObject {
   ~TimeStamp();
 
   //! Starts the timestamp
-  void begin(bool sdma = false);
+  void begin();
 
   //! Ends the timestamp
-  void end(bool sdma = false);
+  void end();
 
   //! Returns the timestamp result in nano seconds
   void value(uint64_t* startTime, uint64_t* endTime);

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

@@ -838,13 +838,6 @@ bool VirtualGPU::create(bool profiling, uint deviceQueueSize, uint rtCUs,
       if (nullptr == queues_[SdmaEngine]) {
         return false;
       }
-    } else {
-      queues_[SdmaEngine] =
-          Queue::Create(*this, Pal::QueueTypeCompute, idx, cmdAllocator_, rtCUs,
-                        amd::CommandQueue::Priority::Normal, residency_limit, max_cmd_buffers);
-      if (nullptr == queues_[SdmaEngine]) {
-        return false;
-      }
     }
   } else {
     LogError("Runtme couldn't find compute queues!");
@@ -2484,7 +2477,9 @@ void VirtualGPU::submitMarker(amd::Marker& vcmd) {
 
 void VirtualGPU::releaseMemory(GpuMemoryReference* mem) {
   queues_[MainEngine]->removeCmdMemRef(mem);
-  queues_[SdmaEngine]->removeCmdMemRef(mem);
+  if (!dev().settings().disableSdma_) {
+    queues_[SdmaEngine]->removeCmdMemRef(mem);
+  }
 }
 
 void VirtualGPU::submitPerfCounter(amd::PerfCounterCommand& vcmd) {
@@ -3128,9 +3123,9 @@ void VirtualGPU::profileEvent(EngineType engine, bool type) const {
     return;
   }
   if (type) {
-    profileTs_->begin((engine == SdmaEngine) ? true : false);
+    profileTs_->begin();
   } else {
-    profileTs_->end((engine == SdmaEngine) ? true : false);
+    profileTs_->end();
   }
 }