trace_processor: Parse adreno_cmdbatch_retired/sync events

[rossning92]

Add custom parsing for kgsl adreno_cmdbatch_sync and adreno_cmdbatch_retired ftrace events to produce GPU timeline slices in the trace processor.

This is a rework of #5519 (reverted in #5793) which fixes the GPU event drift in the timeline by using per-cmdbatch sync points instead of a single global clock snapshot.
TODO: Pending validation if this PR fixes all drift issues that @batesj has observed on his end.

Test run result:

$ tools/ninja -C out/host
[16/16] stamp obj/default.stamp

$ .venv/bin/python3 tools/diff_test_trace_processor.py out/host/trace_processor_shell --name-filter '.*adreno.*'
[==========] Running 1 tests.
[ RUN      ] AdrenoCmdbatch:adreno_cmdbatch_retired_slice
[       OK ] AdrenoCmdbatch:adreno_cmdbatch_retired_slice (ingest: 738.62 ms query: 0.44 ms)
[==========] Name filter selected 1 tests out of 1357.
[==========] 1 tests ran out of 1 total. (1115 ms total)
[  PASSED  ] 1 tests.

[batesj]

I'll try this out shortly. Give me a few days though - hopefully this is not in a rush.

[rossning92]

@batesj no rush many thanks for the help!

[batesj]

@rossning92 I'm getting a build error:

../../src/trace_processor/importers/ftrace/ftrace_tokenizer.cc:551:24: error: no viable conversion from 'TracePacketData' to 'FtraceData'

[rossning92]

@batesj My bad. I did a rebase on to main branch, but apparently some structure name changed.
I just fixed it and retested. All GPU trace event is still fully aligned with the sql query you have shared earlier.

[batesj]

@rossning92 This is working much better! I think there's only one remaining issue where initial clipped events are interpreted wrong. Looks like if an event was clipped at the start of ftrace events (like there's only an end event with no begin event), this CL will turn it into an even that starts at the end time instead. You can see in this screenshot the start of ftrace events there's an offset GPU event in Ctx=25. In my kgsl query I get no event for that:

Is there a way to drop that instead when there's no begin event?

[batesj]

Other examples of additional events showing at the start of the trace that don't all appear to be related to clipped events. Is it possible that this CL is being a bit too flexible in interpreting events with partial data or something like that?

The red square above is zoomed in below:

[rossning92]

@batesj thanks for trying it! Indeed, the initial clipped/drifted events are because they are not paired with adreno_cmdbatch_sync (the trace misses a few initial sync event). For such cases, we should simply skip generating GPU slices to avoid confusion.

I updated the PR and I'm pretty confident that it should now match the parity of your SQL query. On top of that, I also attached the process name to the track name (whenever adreno_cmdbatch_queue events are available)

Thanks so much again for such helpful feedback!