04:19 airlied: glehmann: okay posted it, seems cleaner all right, so good idea, sorry it took me a while to get it :-)
07:34 MrCooper: hwentlan__ lileo_: when trying to enable an overlay plane fails with '[drm:amdgpu_dm_atomic_check [amdgpu]] DC global validation failure: Bandwidth validation failure (BW and Watermark) (13)', does the required bandwidth also depend on the size of the overlay planes, or only on their number?
11:23 karolherbst: Anyway.. I think I've decided that I don't want to use upstream libclc anymore. And I want to pull in the entire thing. Do we just want to have a mesa/libclc directory and compile that stuff with clang? Or do we want this to be a new git repository where we track upstream libclc + patches?
11:27 karolherbst: I kinda mildly prefer just cloning the llvm repo and add some instructions how to build it + change mesa to include the libclc blobs at build time, so we don't have any runtime selection of the binary anymore
11:44 karolherbst: Also.. anybody want to review the nir changes in https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/42646/diffs?commit_id=74c7a2a48b1f4725a71834ac767ba1853f8eebca or should I just merge it?
11:55 glehmann: rb
11:56 karolherbst: thanks
13:31 robclark: daniels, MrCooper: idk if you want to have a look at dri/gbm parts of https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/42171 ? (gbm allocation of yuv.. prep work for EXT_YUV_target)
13:32 MrCooper:adds to list
13:51 daniels: robclark: yes! thanks
15:37 alyssa: do we have a helper for "does this phi reconcile divergent control paths?"
15:37 alyssa: this is not the same as phi->def.divergent: a phi of divergent nir_defs after a uniform IF
15:38 alyssa: this is also not the same as predecessor->divergent: a uniform IF nested inside a divergent IF
15:39 alyssa: "if this phi is immediately after a nir_if, return condition->divergent. else, it's after a loop, so return loop->divergent_break" i think
15:41 alyssa: nir_divergence_analysis talks about gamma/mu/eta phis which corresponds to the issue here. but not seeing any nice helpers here
15:44 alyssa: I guess nir_cf_node_prev(phi->instr->block) would probably get what I need
15:44 glehmann: yes
15:44 glehmann: except for loop header phis before continue blocks are lowered
15:44 alyssa: (..or maybe I need something to traverse the tree version..?)
15:45 glehmann: or nvm, with continue blocks it's still always uniform in loop headers
15:46 glehmann: but not in the continue block potentially
16:05 alyssa: glehmann: https://alyssarosenzweig.ca/helper.txt is what I came up with
16:05 alyssa: not sure whether that deserves to live outside of src/intel
16:07 alyssa: handling this correctly (not checking predecessor->divergent) is worth -4.24% instructions in RDR2 so. yknow.
16:07 glehmann: this is too conservative for loop header phis. nir_loop_is_divergent can be true but the phi will never merge divergent control flow
16:08 alyssa: umm.. right
16:08 alyssa: if it breaks unconditionally I guess?
16:11 alyssa: or actually I guess loop header phis are never merging divergent if continues are gone
16:11 alyssa: which they are for jay
16:11 alyssa: thanks
16:11 glehmann: yes
16:11 glehmann: and even with continues, they all go to the continue block
16:12 glehmann: so the control flow at the loop header should still be uniform
16:14 alyssa: cool ok
16:17 glehmann: dschuermann: btw does this mean the divergent continue handling in nir_divergence's visit_loop_header_phi is dead code?
16:17 glehmann: and I don't think the analysis supports continue blocks at all, so I guess all of the continue handling is dead or should be fixed
16:18 alyssa: should be
20:39 anonymix007[m]: I think I found one definite issue with my win32 ANV build: it seems that vkCmdCopyBuffer for compute queue (family 1) is broken. Some assembly instructions from INTEL_DEBUG=bat also differ between a working Linux host run and this broken Windows one. Here's the diff: paste.debian.net/hidden/2c8f78fb
20:40 anonymix007[m]: What part of ANV should I look into for this?
22:31 anonymix007[m]: Here's the full compute shader dump for both broken and working cased: paste.debian.net/hidden/393c6429
22:31 anonymix007[m]: Any ideas what could be causing such a difference?