[QNN-EP] Enable Optrace from QNN into QNN EP #25987
Conversation
- Add optrace profiling level - Add profiling to compose graph - Add new qnn system profile serializer class - Add API versioning safeguards - Add backwards compatibility for QNN API < 2.28.1 - Use QNN System Profile API for QNN API >= 2.28.1 - Check for log file at end of profiling unit test - Ensure system libs are loaded when profiling is enabled
|
/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline |
|
Azure Pipelines successfully started running 5 pipeline(s). |
There was a problem hiding this comment.
Pull Request Overview
This PR adds optrace profiling level support to the QNN Execution Provider, enabling detailed operation tracing for debugging purposes. It introduces a new QNN System Profile API serializer that generates binary log files compatible with qnn-profile-viewer when using QNN API >= 2.28.1, falling back to CSV format for older versions.
- Adds optrace profiling level alongside existing basic and detailed levels
- Implements QNN System Profile API integration with backwards compatibility
- Adds profiling instrumentation to graph composition, finalization, and execution
Reviewed Changes
Copilot reviewed 10 out of 10 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
| qnn_basic_test.cc | Updates test to use .log extension and adds new optrace test |
| qnn_execution_provider.cc | Adds optrace parsing support to profiling level parser |
| qnn_utils.h/.cc | Adds timestamp utility function for profiling |
| qnn_profile_serializer.h/.cc | New serializer class for QNN System Profile API |
| qnn_model.cc | Adds profiling instrumentation to graph operations |
| qnn_def.h | Defines optrace level and profiling method types |
| qnn_backend_manager.h/.cc | Major refactoring of profiling infrastructure |
Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.
| #pragma once | ||
|
|
There was a problem hiding this comment.
This should be an include guard in a .cc file, not a .h file. This line should be removed from the implementation file.
| #pragma once |
Copilot uses AI. Check for mistakes.
| QnnProfile_EventData_t event_data) { | ||
| auto system_event = &(event_list.emplace_back()); | ||
|
|
||
| if (system_event != nullptr) { |
There was a problem hiding this comment.
This null check is unnecessary as emplace_back() returns a reference to the newly created object, which cannot be null. The check should be removed to improve code clarity.
Copilot uses AI. Check for mistakes.
| QnnProfile_ExtendedEventData_t event_data) { | ||
| auto system_event = &(event_list.emplace_back()); | ||
|
|
||
| if (system_event != nullptr) { |
There was a problem hiding this comment.
This null check is unnecessary as emplace_back() returns a reference to the newly created object, which cannot be null. The check should be removed to improve code clarity.
Copilot uses AI. Check for mistakes.
|
|
||
| // This name must be same with the EPContext node name | ||
| const auto& graph_name = fused_node.Name(); | ||
| ORT_RETURN_IF_ERROR(SetGraphInputOutputInfo(graph_viewer, fused_node, logger)); |
There was a problem hiding this comment.
[nitpick] Empty line was removed between logically separate operations. Consider adding the blank line back to improve code readability and maintain separation between setup and construction phases.
| ORT_RETURN_IF_ERROR(SetGraphInputOutputInfo(graph_viewer, fused_node, logger)); | |
| ORT_RETURN_IF_ERROR(SetGraphInputOutputInfo(graph_viewer, fused_node, logger)); |
Copilot uses AI. Check for mistakes.
Description
Motivation and Context
Adds optrace level profiling for debugging purposes. Utilizes new QNN System Profile API to generate a binary log file compatible with qnn-profile-viewer executable (or creates a .csv file as before if QNN API version is < 2.28.1).