Repository navigation
spektrafilm: tile the Young-van Vliet row pass through local memory - #22825
Conversation
Sure about this? We only have the max group size defined by the standard. Couldn't find a ref. |
8e6954e to
035bc0d
Compare
The row kernels ran one work-item per row, each walking its own row, so neighbouring items accessed memory one row apart, and the row pass was far slower than the column pass over the same data. A work-group now covers several rows and walks them in chunks of 64 columns: each chunk is loaded and stored by the whole group in contiguous segments through local memory, and one item per row runs the unchanged recursion on the chunk. The work-group is 8 x 16 items, or smaller as dt_opencl_local_buffer_opt() finds the device's work-group and local memory limits require. Output is bit-identical for every tile size. Row pass on 6000x4000, median of 10, built with -cl-fast-relaxed-math, before -> after: device / driver float4 float Apple M1 Pro, rusticl (asahi) 100.3 -> 22.1 ms 13.8 -> 6.0 ms RX 6700 XT, rusticl (radeonsi) 8.8 -> 6.7 ms 5.4 -> 1.5 ms RX 6700 XT, ROCm 9.1 -> 6.3 ms 5.4 -> 1.7 ms
035bc0d to
00fc90b
Compare
|
You're right, the standard only guarantees 1. I dropped I ran a tile sweep on M1 (rusticl) and RX 6700 XT (rusticl and ROCm) that put 8x16 at or near the best on all three; updated numbers are in the commit message. So this is the starting point, shrinking to what the device and kernel allow, CPU fallback if nothing fits. Output is bit-identical for every tile size tried (16x16 down to 1x1, with and without |
|
Yes, that's the correct way. The possible workgroup size depends on what's inside the kernel, locals, ... It even makes a difference if you allow the compiler to re-use valiable like floats/ints ... even making them |
TurboGit
added
priority: medium
Summary
The YvV row kernels ran one work-item per row, each walking its own row, so neighboring work-items accessed memory a full row apart. The row pass was 1.8–10x slower than the column pass over the same data, and the largest kernel cost of spektrafilm on the GPU.
A work-group of 16x16 items now covers 16 rows and walks them in chunks of 64 columns. Each chunk is loaded and stored by the whole group in contiguous segments through local memory, and one item per row runs the unchanged recursion on the chunk. The arithmetic and its order are untouched, and the output is bit-identical to before. That was checked on all three drivers below, built with and without
-cl-fast-relaxed-math, at 6000x4000 and at 6001x4001 (sizes that aren't multiples of the tile size).Row pass, 6000x4000, median of 10–20 runs, standalone benchmark of both kernel versions:
The row kernels now require a work-group size of 256 (
reqd_work_group_size(16, 16, 1)), which every supported GPU provides.Checklist
git rebase --exec 'cmake --build buildmaster to check this.src/tests/integration/where the pixelpipe is touched, ordarktable-clias a headless smoke test._(), new preferences are registered indata/darktableconfig.xml.in.RELEASE_NOTES.mdentry was added (only needed if fixing an issue in a release). Do not reference GitHub issues.AI assistance
Co-created with Claude Opus 5.5