00:05 esdrastarsis[d]: mhenning[d]: Maybe something like `NVK_PERFTEST=dlss`
00:31 mohamexiety[d]: wouldnt it be better for this to be handled at the proton level? :thonk:
00:32 mohamexiety[d]: but i guess it depends on how flaky/inconsistent it is currently
00:38 karolherbst[d]: phomes_[d]: are there CTS tests for `NVX_binary_import`? 😄
00:38 karolherbst[d]: (why am I even asking when I already know the answer)
00:38 zmike[d]: cts for vendor extensions haha
00:39 zmike[d]: karol you crazy
00:39 karolherbst[d]: look, sometimes the idealist and optimist shines through
00:40 karolherbst[d]: mhenning[d]: wondering if it would be helpful for shader constants
00:41 karolherbst[d]: or any ubo that's below 1k
00:46 esdrastarsis[d]: karolherbst[d]: misyl implemented a sample
00:47 mohamexiety[d]: yeah misyl wrote tests for it
00:47 esdrastarsis[d]: https://github.com/misyltoad/VK_NVX_binary_import-example
00:50 phomes_[d]: the sample is working fine. I also tested on bunch of games. Many work just fine but I found 3 games where specific presets are broken. Updating to newer DLSS versions fixed that. But then an even newer DLSS came out and broke a bunch of games again
00:52 phomes_[d]: so I don't think it is ready to release to all users yet, but behind an env var it should be easier to test and not break due to conflicts as much
00:55 mohamexiety[d]: https://gitlab.freedesktop.org/mesa/crucible/-/merge_requests/184 there's these too
00:55 mohamexiety[d]: but yeah i wonder why it keeps breaking
01:21 mhenning[d]: karolherbst[d]: you only get 256 bytes per bank x 8 banks, plus you can't load directly from memory, so it's a little niche for more general stuff
01:21 karolherbst[d]: ohh right..
01:21 karolherbst[d]: 256 ain't much indeed
01:21 karolherbst[d]: well 256 * 8
04:53 gfxstrand[d]: We probably only need 512 or so for root constants.
12:23 karolherbst[d]: phomes_[d]: I have a fun thing to performance test, but it's probably also entirely insignificant. Might checking the MR with and without that suggestion: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40293#note_3380547
12:30 phomes_[d]: I will give that a test in a short while
14:07 karolherbst[d]: should we remove the no_ugpr debug thing?
15:00 phomes_[d]: karolherbst[d]: I do see a small improvement. I want to repeat some of the tests to be sure, as there was a proton update in the middle
15:01 karolherbst[d]: mhhhh
15:02 karolherbst[d]: in theory accessing more memory _shouldn't_ be an issue because it's all cache lines anyway, but I'm also not entirely sure
15:03 karolherbst[d]: but also.. if it's small might just be random stuff or just within the normal error ranges
15:59 mhenning[d]: karolherbst[d]: why remove it?
16:00 karolherbst[d]: mhenning[d]: it's making the code I'm working on more annoying 😄 And I totally forgot it existence with the UGPR stuff
16:09 karolherbst[d]: okay.. I think `nak_nir_lower_non_uniform_ldcx` is too aggressive 🙃
16:10 karolherbst[d]: like we can pull from `cx[UGPR][GPR + immS16]`
16:10 mhenning[d]: I don't know what you mean
16:11 karolherbst[d]: like we can do `LDC cx[UGPR][GPR + immS16]`
16:11 karolherbst[d]: and instead of that it lowers to LDG
16:12 mhenning[d]: oh, right. yeah, that could make a difference
16:12 karolherbst[d]: like ULDC requires `cx[UGPR][UGPR + immS16]`, so for that it makes sense
16:12 karolherbst[d]: but then it's uniform anyway
16:12 karolherbst[d]: ehh
16:13 karolherbst[d]: `ULDC cx[UGPR][immS16]`
16:13 karolherbst[d]: U in that case
16:13 karolherbst[d]: but that goes into its own MR