COPY OF #22641 [SYCL][UR][OpenCL] Fix profiling tag ordering on OpenCL GPU

[crystarm]

Supersedes #22641.
Resolves #22229.

This PR reapplies the two previously reviewed OpenCL profiling tag changes on top of the current sycl branch and adds a separate fix for profiling tag fallback timestamps.

#22810 changed profiling tag timestamp handling so that MSubmitTime is no longer initialized for profiling tag events. The previous Native CPU fallback fix relied on that field, causing command_submit to remain zero when CI tested the change together with the current sycl branch.

The new fix follows the timestamp model introduced by #22810: profiling tag events remain marked as tags and use their device-recorded completion timestamp for command_submit, command_start, and command_end.

Validation:

• ExtensionsTests: 259/259 passed.
• Native CPU E2E: passed 100/100 consecutive local runs.

The Native CPU E2E was built locally with the existing diagnostic libspirv stub and -O0; CI remains the authoritative full-toolchain validation.

For the background on why this PR replaces #22641, see #22641 (comment).

[crystarm]

@KornevNikita
Could you please run the native CPU E2E workflow?

[Copilot]

🟡 Changes recommended

Unresolved event contract, argument-validation, and regression-test issues remain.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Fixes OpenCL GPU profiling-tag ordering and fallback timestamp handling.

Changes:

• Adds OpenCL timestamp recording via profiled buffer fills.
• Updates SYCL profiling and scheduler fallback behavior.
• Expands unit and end-to-end test coverage.

File summaries

File	Summary
unified-runtime/source/adapters/opencl/event.cpp	Implements timestamp recording; event-type and argument-validation issues remain.
unified-runtime/source/adapters/opencl/context.hpp	Manages the internal timestamp buffer.
sycl/unittests/Extensions/ProfilingTag.cpp	Adds profiling timestamp and fallback coverage.
sycl/test-e2e/ProfilingTag/profiling_queue.cpp	Exercises OpenCL profiling queues.
sycl/test-e2e/ProfilingTag/in_order_profiling_queue.cpp	Exercises OpenCL in-order profiling queues.
sycl/test-e2e/ProfilingTag/common.hpp	Validates timestamp ordering; the original submit-time assertion should be retained.
sycl/source/event.cpp	Waits before profiling-tag submit-time queries.
sycl/source/detail/scheduler/commands.cpp	Adds timestamp-recording fallback behavior.
sycl/source/detail/event_impl.cpp	Uses completion timestamps for profiling tags.
sycl/include/sycl/ext/oneapi/experimental/profiling_tag.hpp	Routes supported fallback queues through the scheduler.

Review details

Suppressed comments (2)

sycl/test-e2e/ProfilingTag/common.hpp:66

• The test no longer checks that StartTagSubmit is non-zero: this line replaced the original assertion instead of adding the new EndTagEnd check. A regression that returns zero for the first tag's submit timestamp would now pass the non-zero checks, so keep both assertions.

  CHECK(Failures, EndTagEnd != 0)

unified-runtime/source/adapters/opencl/event.cpp:314

• The UR contract for urEnqueueTimestampRecordingExp requires a non-null phEvent and specifies UR_RESULT_ERROR_INVALID_NULL_POINTER otherwise (unified-runtime/include/unified-runtime/ur_api.h:11205). Because ifUrEvent treats this pointer as optional, this implementation returns success without an event for that invalid call; validate phEvent before querying the queue or enqueueing the fill.

    const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent) {

• Files reviewed: 10/10 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[KornevNikita]

@crystarm looks like the pre-commit finally passes