PR openxla#21134: [XLA:GPU] Add profiler annotation for sequential th…

…unk. Imported from GitHub PR openxla#21134 This PR wraps sequential thunk with profiler annotations, which will make loop iterations, and conditional branch more easy to read in the profiler. The nsys profile looks like this: ![image](https://github.com/user-attachments/assets/8a3dd0be-4e1a-4516-ae64-b376336799bd) Copybara import of the project: -- eea74b8 by Shawn Wang <[email protected]>: add nvtx marker for sequential thunk Merging this change closes openxla#21134 COPYBARA_INTEGRATE_REVIEW=openxla#21134 from shawnwang18:shawnw/add_profiler_for_while_body_cond eea74b8 PiperOrigin-RevId: 714296359
Intel-tensorflow · Jan 11, 2025 · 45e6f1c · 45e6f1c
1 parent f78a406
commit 45e6f1c
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 5 deletions.
diff --git a/xla/service/gpu/ir_emitter_unnested.cc b/xla/service/gpu/ir_emitter_unnested.cc
@@ -232,7 +232,12 @@ absl::Status IrEmitterUnnested::EmitConditional(const HloInstruction* instr) {
   for (auto comp : instr->branch_computations()) {
     auto ir_emitter = IrEmitterUnnested::Create(ir_emitter_context_);
     TF_RETURN_IF_ERROR(ir_emitter->EmitHloComputation(comp));
-    branch_thunks.push_back(ir_emitter->ConsumeThunkSequence());
+    Thunk::ThunkInfo branch_thunk_info =
+        Thunk::ThunkInfo::WithProfileAnnotation(instr);
+    branch_thunk_info.profile_annotation +=
+        absl::StrCat("_branch_", comp->name());
+    branch_thunks.push_back(
+        ir_emitter->ConsumeThunkSequence(branch_thunk_info));
   }
 
   ConditionalThunkConfig config =
@@ -2263,9 +2268,17 @@ absl::StatusOr<std::unique_ptr<Thunk>> IrEmitterUnnested::BuildWhileThunk(
   TF_ASSIGN_OR_RETURN(
       auto pred, GetAllocationSliceForHlo(condition->root_instruction(), {}));
 
+  Thunk::ThunkInfo cond_thunk_info =
+      Thunk::ThunkInfo::WithProfileAnnotation(instr);
+  cond_thunk_info.profile_annotation += "_condition";
+  Thunk::ThunkInfo body_thunk_info =
+      Thunk::ThunkInfo::WithProfileAnnotation(instr);
+  body_thunk_info.profile_annotation += "_body";
+
   return std::unique_ptr<Thunk>(new WhileThunk(
-      thunk_info, pred, ir_emitter_condition->ConsumeThunkSequence(),
-      ir_emitter_body->ConsumeThunkSequence(), trip_count));
+      thunk_info, pred,
+      ir_emitter_condition->ConsumeThunkSequence(cond_thunk_info),
+      ir_emitter_body->ConsumeThunkSequence(body_thunk_info), trip_count));
 }
 
 absl::Status IrEmitterUnnested::EmitTargetElementLoop(

diff --git a/xla/service/gpu/ir_emitter_unnested.h b/xla/service/gpu/ir_emitter_unnested.h
@@ -89,8 +89,9 @@ class IrEmitterUnnested : public IrEmitter {
       IrEmitterContext* ir_emitter_context);
 
   // Transfers the ownship of thunk_sequence_ out.
-  std::unique_ptr<SequentialThunk> ConsumeThunkSequence() {
-    return std::make_unique<SequentialThunk>(Thunk::ThunkInfo{},
+  std::unique_ptr<SequentialThunk> ConsumeThunkSequence(
+      Thunk::ThunkInfo thunk_info = Thunk::ThunkInfo{}) {
+    return std::make_unique<SequentialThunk>(thunk_info,
                                              std::move(thunk_sequence_));
   }
 

diff --git a/xla/service/gpu/runtime/sequential_thunk.cc b/xla/service/gpu/runtime/sequential_thunk.cc
@@ -75,6 +75,8 @@ absl::Status SequentialThunk::Initialize(const InitializeParams& params) {
 }
 
 absl::Status SequentialThunk::ExecuteOnStream(const ExecuteParams& params) {
+  std::optional<tsl::profiler::ScopedAnnotation> seq_annotation =
+      GetKernelAnnotation(profile_annotation());
   for (const std::unique_ptr<Thunk>& thunk : thunks_) {
     std::optional<tsl::profiler::ScopedAnnotation> annotation =
         GetKernelAnnotation(thunk->profile_annotation());