TSC Meeting Notes 2024-10-17
Attendance:
Guests:
Li Ji, ILM
Michael Smith, HT-J2K compressor
Discussion:
Li Ji presented dry-run of “Use cases for benchmarking design” slides
What will be a good test, what are criteria? Needs to reflect real-world usage, be reproducible, scalable, understandable.
Joseph: looked at submission from Michael Smith for HT-J2K compressor, trying to determine whether he was doing relevant work. https://jpeg.org/jpeg2000/htj2k.html .
Questions about real-world production usage. What common patterns of calls are in real world? Approximate combinations / importance of patterns.
When someone tries to commit, benchmarks would be invoked?
Larry: might want to think of these items more on a task basis. Trying to get to one metric? Let people look at the cases and decide for themselves.
Larry: interesting cases for benchmarking
Hardware
Patching as add a patch, version to version comparisons are we getting better
Application that wraps EXR - how far off the best bare bones EXR case is it?
Aras P tests setup as an example: https://aras-p.info/blog/2021/08/04/EXR-Lossless-Compression/
Make tests useful for studios
Kimball - 4 major cases, covers a lot of ground:
Open in editing tool - reading whole window - test exists
Playback - OpenRV type access - stream in some portion of image
Compositing workflows - Nuke/Shake - very heavily scanlined, diff threads reading scanlines, concurrent scanline access
Rendering - test render from OSL could be used, Larry: access pattern is very particular
Larry: writing simpler to test, no access patterns to untangle.
Larry: content types
photographic - samples in DPEL now?
renderer output
texture input, a few diff patterns
HDR Photographic survey http://markfairchild.org/HDR.html
Joseph - no command tool to convert from cameras to OpenEXR
Michael Smith: standard photographic tests https://dpel.aswf.io/asc-stem2/
More reference content: https://opencontent.netflix.com/