Add Queue profling param and toggle for HIP

Use signal timestamps if NDRange command takes forceProfile flag.

Change-Id: Ib7f187d781fd78a7346818afb3344a9378f4c104

rocclr/device/rocm/rocvirtual.cpp

@@ -763,7 +763,7 @@ bool VirtualGPU::create() {
   gpu_queue_ = roc_device_.acquireQueue(queue_size, cooperative_, cuMask_, priority_);
   if (!gpu_queue_) return false;
 
-  if (!initPool(dev().settings().kernargPoolSize_, (profiling_) ? queue_size : 0)) {
+  if (!initPool(dev().settings().kernargPoolSize_, (amd::IS_HIP) ? queue_size : 0)) {
     LogError("Couldn't allocate arguments/signals for the queue");
     return false;
   }

rocclr/platform/command.cpp

@@ -76,7 +76,7 @@ uint64_t Event::recordProfilingInfo(int32_t status, uint64_t timeStamp) {
       break;
     default:
       profilingInfo_.end_ = timeStamp;
-      if (profilingInfo_.callback_ != NULL) {
+      if (profilingInfo_.callback_ != nullptr) {
         profilingInfo_.callback_->callback(timeStamp - profilingInfo_.start_,
             profilingInfo_.waves_);
       }
@@ -270,7 +270,8 @@ NDRangeKernelCommand::NDRangeKernelCommand(HostQueue& queue, const EventWaitList
                                            Kernel& kernel, const NDRangeContainer& sizes,
                                            uint32_t sharedMemBytes, uint32_t extraParam,
                                            uint32_t gridId, uint32_t numGrids,
-                                           uint64_t prevGridSum, uint64_t allGridSum, uint32_t firstDevice) :
+                                           uint64_t prevGridSum, uint64_t allGridSum,
+                                           uint32_t firstDevice, bool forceProfiling) :
     Command(queue, CL_COMMAND_NDRANGE_KERNEL, eventWaitList, AMD_SERIALIZE_KERNEL),
     kernel_(kernel),
     sizes_(sizes),
@@ -285,6 +286,11 @@ NDRangeKernelCommand::NDRangeKernelCommand(HostQueue& queue, const EventWaitList
   auto devKernel = const_cast<device::Kernel*>(kernel.getDeviceKernel(device));
   profilingInfo_.setCallback(devKernel->getProfilingCallback(
     queue.vdev()), devKernel->getWavesPerSH(queue.vdev()));
+  if (forceProfiling) {
+    profilingInfo_.enabled_ = true;
+    profilingInfo_.clear();
+    profilingInfo_.callback_ = nullptr;
+  }
   kernel_.retain();
 }
 

rocclr/platform/command.hpp

@@ -100,7 +100,7 @@ class Event : public RuntimeObject {
     ProfilingInfo(bool enabled = false) : enabled_(enabled), waves_(0) {
       if (enabled) {
         clear();
-        callback_ = NULL;
+        callback_ = nullptr;
       }
     }
 
@@ -801,7 +801,8 @@ class NDRangeKernelCommand : public Command {
   NDRangeKernelCommand(HostQueue& queue, const EventWaitList& eventWaitList, Kernel& kernel,
                        const NDRangeContainer& sizes, uint32_t sharedMemBytes = 0,
                        uint32_t extraParam = 0, uint32_t gridId = 0, uint32_t numGrids = 0,
-                       uint64_t prevGridSum = 0, uint64_t allGridSum = 0, uint32_t firstDevice = 0);
+                       uint64_t prevGridSum = 0, uint64_t allGridSum = 0,
+                       uint32_t firstDevice = 0, bool forceProfiling = false);
 
   virtual void submit(device::VirtualDevice& device) { device.submitKernel(*this); }
 

rocclr/utils/flags.hpp

@@ -237,7 +237,9 @@ release(uint, HIP_HIDDEN_FREE_MEM, 0,                                         \
 release(size_t, GPU_FORCE_BLIT_COPY_SIZE, 0,                                  \
         "Size in KB of the threshold below which to force blit instead for sdma") \
 release(bool, ROC_ENABLE_LARGE_BAR, true,                                     \
-        "Enable Large Bar if supported by the device")
+        "Enable Large Bar if supported by the device")                        \
+release(bool, HIP_FORCE_QUEUE_PROFILING, false,                               \
+        "Force command queue profiling by default")
 
 namespace amd {