Force t1_ update when profiling is enabled

Fixes issue where t1_ counters may not be updated when doing dispatch
profiling, causing a divide by 0.

Change-Id: I91060ac3f9fd2183d277e6e7cd810398a453a87f


[ROCm/ROCR-Runtime commit: 3d1563ee68]

projects/rocr-runtime/runtime/hsa-runtime/core/inc/amd_aql_queue.h

@@ -196,9 +196,12 @@ class AqlQueue : public core::Queue, private core::LocalSignal, public core::Doo
   /// @return hsa_status_t
   hsa_status_t GetCUMasking(uint32_t num_cu_mask_count, uint32_t* cu_mask) override;
 
-  // @brief Submits a block of PM4 and waits until it has been executed.
+  /// @brief Submits a block of PM4 and waits until it has been executed.
   void ExecutePM4(uint32_t* cmd_data, size_t cmd_size_b) override;
 
+  /// @brief Enables/Disables profiling overrides SetProfiling from core::Queue
+  void SetProfiling(bool enabled) override;
+
   /// @brief Update signal value using Relaxed semantics
   void StoreRelaxed(hsa_signal_value_t value) override;
 

projects/rocr-runtime/runtime/hsa-runtime/core/inc/amd_gpu_agent.h

@@ -373,6 +373,13 @@ class GpuAgent : public GpuAgentInt {
   // @brief returns the libdrm device handle
   __forceinline amdgpu_device_handle libDrmDev() const { return ldrm_dev_; }
 
+  __forceinline void CheckClockTicks() {
+    // If we did not update t1 since agent initialization, force a SyncClock. Otherwise computing
+    // the SystemClockCounter to GPUClockCounter ratio in TranslateTime(tick) results to a division
+    // by 0.
+    if (t0_.GPUClockCounter == t1_.GPUClockCounter) SyncClocks();
+  }
+
   void ReserveScratch();
 
   // @brief If agent supports it, release scratch memory for all AQL queues on this agent.

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_aql_queue.cpp

@@ -1339,6 +1339,13 @@ hsa_status_t AqlQueue::GetCUMasking(uint32_t num_cu_mask_count, uint32_t* cu_mas
   return HSA_STATUS_SUCCESS;
 }
 
+void AqlQueue::SetProfiling(bool enabled) {
+  Queue::SetProfiling(enabled);
+
+  if (enabled) agent_->CheckClockTicks();
+  return;
+}
+
 void AqlQueue::ExecutePM4(uint32_t* cmd_data, size_t cmd_size_b) {
   // pm4_ib_buf_ is a shared resource, so mutually exclude here.
   ScopedAcquire<KernelMutex> lock(&pm4_ib_mutex_);

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -1191,11 +1191,7 @@ hsa_status_t GpuAgent::EnableDmaProfiling(bool enable) {
     }
   }
 
-  // If we did not update t1 since agent initialization, force a SyncClock. Otherwise computing
-  // the SystemClockCounter to GPUClockCounter ratio in TranslateTime(tick) results to a division
-  // by 0. We perform the check here because we do not want to check everytime there is a call to
-  // TranslateTime(tick)
-  if (enable && t0_.GPUClockCounter == t1_.GPUClockCounter) SyncClocks();
+  if (enable) CheckClockTicks();
 
   return HSA_STATUS_SUCCESS;
 }