Merge pull request #1734 from arcaneframework/dev/gg-add-color-to-profile-region

Add support to choose color when profiling region

Author: grospelliergilles

arcane/src/arcane/accelerator/core/ProfileRegion.cc

@@ -31,7 +31,19 @@ ProfileRegion(const RunQueue& queue, const String& name)
   if (queue.isNull())
     return;
   m_runtime = queue._internalRuntime();
-  m_runtime->pushProfilerRange(name);
+  m_runtime->pushProfilerRange(name, -1);
+}
+
+/*---------------------------------------------------------------------------*/
+/*---------------------------------------------------------------------------*/
+
+ProfileRegion::
+ProfileRegion(const RunQueue& queue, const String& name, Int32 color_rgb)
+{
+  if (queue.isNull())
+    return;
+  m_runtime = queue._internalRuntime();
+  m_runtime->pushProfilerRange(name, color_rgb);
 }
 
 /*---------------------------------------------------------------------------*/

arcane/src/arcane/accelerator/core/internal/IRunnerRuntime.h

@@ -52,7 +52,7 @@ class ARCANE_ACCELERATOR_CORE_EXPORT IRunnerRuntime
   virtual void startProfiling() {}
   virtual void stopProfiling() {}
   virtual bool isProfilingActive() { return false; }
-  virtual void pushProfilerRange(const String&) {}
+  virtual void pushProfilerRange([[maybe_unused]] const String& name, [[maybe_unused]] Int32 color_gdb) {}
   virtual void popProfilerRange() {}
   virtual void finalize(ITraceMng*) {}
 

arcane/src/arcane/accelerator/core/internal/ProfileRegion.h

@@ -37,6 +37,7 @@ class ARCANE_ACCELERATOR_CORE_EXPORT ProfileRegion
  public:
 
   ProfileRegion(const RunQueue& queue, const String& name);
+  ProfileRegion(const RunQueue& queue, const String& name, Int32 color_rgb);
   ~ProfileRegion();
 
  private:

arcane/src/arcane/accelerator/cuda/runtime/CudaAcceleratorRuntime.cc

@@ -382,10 +382,24 @@ class CudaRunnerRuntime
                           ptr, ca.devicePointer, ca.hostPointer);
   }
 
-  void pushProfilerRange(const String& name) override
+  void pushProfilerRange(const String& name, Int32 color_rgb) override
   {
 #ifdef ARCANE_HAS_CUDA_NVTOOLSEXT
-    nvtxRangePush(name.localstr());
+    if (color_rgb >= 0) {
+      // NOTE: Il faudrait faire: nvtxEventAttributes_t eventAttrib = { 0 };
+      // mais cela provoque pleins d'avertissement de type 'missing initializer for member'
+      nvtxEventAttributes_t eventAttrib;
+      std::memset(&eventAttrib, 0, sizeof(nvtxEventAttributes_t));
+      eventAttrib.version = NVTX_VERSION;
+      eventAttrib.size = NVTX_EVENT_ATTRIB_STRUCT_SIZE;
+      eventAttrib.colorType = NVTX_COLOR_ARGB;
+      eventAttrib.color = color_rgb;
+      eventAttrib.messageType = NVTX_MESSAGE_TYPE_ASCII;
+      eventAttrib.message.ascii = name.localstr();
+      nvtxRangePushEx(&eventAttrib);
+    }
+    else
+      nvtxRangePush(name.localstr());
 #endif
   }
   void popProfilerRange() override

arcane/src/arcane/materials/IncrementalComponentModifier.cc

@@ -328,7 +328,7 @@ _resizeVariablesIndexer(Int32 var_index)
 {
   RunQueue& queue = m_material_mng->_internalApi()->runQueue();
   RunQueue::ScopedAsync sc(&m_queue);
-  Accelerator::ProfileRegion ps(queue, "ResizeVariableIndexer");
+  Accelerator::ProfileRegion ps(queue, "ResizeVariableIndexer", 0xFF00FF);
   ResizeVariableIndexerArgs resize_args(var_index, queue);
   // Regarde si on n'utilise qu'une seule commande pour les copies des vues.
   // Pour l'instant (novembre 2024) on ne l'utilise par défaut que si

arcane/src/arcane/materials/IncrementalComponentModifier_Accelerator.cc

@@ -44,8 +44,8 @@ apply(MaterialModifierOperation* operation)
   const char* str_add = "ApplyConstituentOperationAdd";
   const char* str_remove = "ApplyConstituentOperationRemove";
   bool is_add = operation->isAdd();
-
-  Accelerator::ProfileRegion ps(m_queue, is_add ? str_add : str_remove);
+  Int32 color = (is_add) ? 0x00FFFF : 0x007FFF;
+  Accelerator::ProfileRegion ps(m_queue, is_add ? str_add : str_remove, color);
 
   IMeshMaterial* mat = operation->material();
   SmallSpan<const Int32> orig_ids = operation->ids();
@@ -326,7 +326,7 @@ _addItemsToIndexer(MeshMaterialVariableIndexer* var_indexer,
     // TODO: Comme tout est indépendant par variable, on pourrait
     // éventuellement utiliser plusieurs files.
     if (m_do_init_new_items) {
-      Accelerator::ProfileRegion ps(m_queue, "InitializeNewItems");
+      Accelerator::ProfileRegion ps(m_queue, "InitializeNewItems", 0xFFFF00);
       RunQueue::ScopedAsync sc(&m_queue);
       IMeshMaterialMng* mm = m_material_mng;
       auto func = [&](IMeshMaterialVariable* mv) {

arcane/src/arcane/materials/MeshMaterialModifierImpl.cc

@@ -223,7 +223,7 @@ endUpdate()
 void MeshMaterialModifierImpl::
 _endUpdate()
 {
-  Accelerator::ProfileRegion ps(m_queue,"ConstituentsEndUpdate");
+  Accelerator::ProfileRegion ps(m_queue,"ConstituentsEndUpdate", 0x97599A);
 
   bool is_keep_value = m_material_mng->isKeepValuesAfterChange();
   Integer nb_operation = m_operations.values().size();