perfetto: Add InstrumentedSamplingConfig to GpuCounterConfig#5358
perfetto: Add InstrumentedSamplingConfig to GpuCounterConfig#5358
Conversation
🎨 Perfetto UI Builds
|
|
@olehkuznetsov - this might need more discussion or at least comments for Android indicating what is supported etc. I think in general, this CL could also do with more description of which path you would want to choose, what options will be choose if you use the boolean option etc. Also I think we need to start thinking about adding documentation to docs.perfetto.dev about all the things you are adding. We didn't add it before but you're significantly expanding the surface, I think it's worth us properly looking into this now. |
Yes, I've been thinking about how to best do this. Just updating the docs/data_sources/gpu.md is a good start. I was thinking we have 3 sections in there:
I'll start this as a separate change if that sounds good? I was also thinking that the last bullet point could be added to the built-in examples so it's easy to see how perfetto UI can visualize these workloads. |
8fcc0a9 to
8faf550
Compare
8faf550 to
43aafa2
Compare
43aafa2 to
65988b5
Compare
Add counter_names field to GpuCounterConfig as an alternative to counter_ids for selecting counters by name. Add InstrumentedSamplingConfig with activity filtering support: - ActivityNameFilter message with per-filter NameBase enum and name_glob field for matching mangled or demangled kernel names. - TX range include/exclude globs for filtering by in-process annotations (e.g. NVTX ranges for CUDA). - ActivityRange for skip/count based sampling of matching activities. - Three-step filtering pipeline: activity name filters, TX range include/exclude, then range-based sampling. Add instrumented sampling documentation to docs/data-sources/gpu.md.
65988b5 to
0d9ada5
Compare
|
@LalitMaganti @olehkuznetsov Does the latest version of this look good to you? |
Add counter_names field to GpuCounterConfig as an alternative to
counter_ids for selecting counters by name.
Add InstrumentedSamplingConfig with activity filtering support:
name_glob field for matching mangled or demangled kernel names.
annotations (e.g. NVTX ranges for CUDA).
include/exclude, then range-based sampling.
Add instrumented sampling documentation to docs/data-sources/gpu.md.