#
✕ nick:s0ullight[m]
#asahi-dev · 2026-11-16 15:12
s0ullight[m]Was m2 pmp support the only blocker for enabling pmp by default?
#asahi-dev · 2026-11-15 19:55
s0ullight[m]render updates last_compute and vice versa
#asahi-dev · 2026-11-15 19:54
s0ullight[m]That looks inverted to my untrained eyes
#asahi-dev · 2026-11-15 19:54
s0ullight[m]https://github.com/AsahiLinux/linux/blob/6ba0637d1ed9a2e3639dd3fc9eca6cd0483a6434/drivers/gpu/drm/asahi/queue/mod.rs#L720
#asahi-dev · 2026-11-15 01:57
s0ullight[m]uploaded an image: Took an ungodly amount of time to organize (107KiB) < https://matrix.org/oftc/media/v1/media/download/ActehkMNsnRNEt_0HXdK2RDdFBIteOFJdLkIBVV-DxtLFoT-bZwwlupifLJMYjpVVXFZgfUTf7xSEZvj7Bu3NqNCeiJQVnMgAG1hdHJpeC5vcmcvaGJmZWxZWWFqU0xGcFh3dGJlbVpJckhQ >
#asahi-dev · 2026-11-15 01:54
s0ullight[m]I should also figure out a more efficient way that is not "bunch of unrelated but potentially codependent changes. Now let's figure out how to group em and commit"
#asahi-dev · 2026-11-15 01:52
s0ullight[m]I'm surprised by how much time goes into writing comments, messages.
#asahi-dev · 2026-11-15 01:51
s0ullight[m]https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/46312
#asahi-dev · 2026-11-14 11:04
s0ullight[m]And yeah p3s might be interesting as well
#asahi-dev · 2026-11-14 11:03
s0ullight[m]<JamesCalligeros[m]> "Its a 22 year old d3d9 game :p" <- virtio implicit sync? Try with HK_PERFTEST=batch
#asahi-dev · 2026-11-14 10:58
s0ullight[m]My machine == mbp m1 pro
#asahi-dev · 2026-11-14 10:57
s0ullight[m]But gpu write bandwidth was affected by hooking pmp up to the gpu on my machine
#asahi-dev · 2026-11-14 10:56
s0ullight[m]chadmed: yeah it's gonna depend on where you're bottlenecked
#asahi-dev · 2026-11-10 12:00
s0ullight[m]Alrightie, I'll polish up my hk improvements and try to upstream. Also will try to get the conversation going regarding fence passing
#asahi-dev · 2026-11-10 11:51
s0ullight[m]chadmed: Is your work also covering this?
#asahi-dev · 2026-11-10 11:50
s0ullight[m]I think after virtio implicit sync that's the biggest cost for higher resolutions
#asahi-dev · 2026-11-10 11:48
s0ullight[m]I've been trying to figure out the fps tanking with higher resolutions and just realized it matches macos if the game runs windowed
#asahi-dev · 2026-11-10 11:46
s0ullight[m]Games yes
#asahi-dev · 2026-11-10 11:42
s0ullight[m]We still pay for the compositor in Fullscreen apps right?
#asahi-dev · 2026-11-08 19:50
s0ullight[m]Anyone who can take a gpu trace from macos for me?
#asahi-dev · 2026-11-08 15:54
s0ullight[m]I am hoping to do some draft pr/mr's at some point
#asahi-dev · 2026-11-08 15:53
s0ullight[m]For local testing?
#asahi-dev · 2026-11-08 15:51
s0ullight[m]Erm, no. I've mostly been looking at where we loose performance in certain games. Ended up being the virtio implicit sync stuff. Tried some things alyssa didn't like very much but did improve fps by a lot. Ended up stumbling upon lina's comment about fence passing. I got something working locally.
#asahi-dev · 2026-11-08 15:48
s0ullight[m]I am
#asahi-dev · 2026-11-08 15:40
s0ullight[m]I want to complete figuring out fence passing first. At some point should upstream the hk fixes I have as well
#asahi-dev · 2026-11-08 15:34
s0ullight[m]I'm glad the bandwidth benchmark results with and without pmp ended up relevant.
#asahi-dev · 2026-11-08 15:18
s0ullight[m][chaos_princess](https://matrix.to/#/@_oftc_chaos_princess:matrix.org) a bit of a hail mary but could there be a similar entry for the cpu side?
#asahi-dev · 2026-11-08 14:47
s0ullight[m]There's no crash log
#asahi-dev · 2026-11-08 14:39
s0ullight[m]Didn't help
#asahi-dev · 2026-11-08 14:28
s0ullight[m]Rebooted when I tried to bring back from sleep
#asahi-dev · 2026-11-08 14:24
s0ullight[m]93 was with pmp disabled. It then halved with pmp without entry
#asahi-dev · 2026-11-08 14:21
s0ullight[m]93
#asahi-dev · 2026-11-08 14:21
s0ullight[m]Metal does 212
#asahi-dev · 2026-11-08 14:16
s0ullight[m][chaos_princess](https://matrix.to/#/@_oftc_chaos_princess:matrix.org) it went up to 164 GB/s
#asahi-dev · 2026-11-08 13:05
s0ullight[m]Ok, I'll do that in a bit
#asahi-dev · 2026-11-08 12:58
s0ullight[m]Mbp m1 pro
#asahi-dev · 2026-11-08 10:55
s0ullight[m]Don't judge too hard please
#asahi-dev · 2026-11-08 10:55
s0ullight[m]https://paste.centos.org/view/42c8de59
#asahi-dev · 2026-11-08 10:49
s0ullight[m]Ye
#asahi-dev · 2026-11-08 10:48
s0ullight[m]As in metal code?
#asahi-dev · 2026-11-08 10:47
s0ullight[m][chaos_princess](https://matrix.to/#/@_oftc_chaos_princess:matrix.org) should I bother with the setup for tracing pmp traffic under gpu load or is that something you'll look at eventually?
#asahi-dev · 2026-11-06 10:13
s0ullight[m]OpenCL gives similar results fyi
#asahi-dev · 2026-11-06 10:11
s0ullight[m]I don't have the setup for that
#asahi-dev · 2026-11-06 10:09
s0ullight[m]Also 20 dispatches in one command buffer and submit
#asahi-dev · 2026-11-06 10:08
s0ullight[m]Ah, no a 256mb surface without overlap
#asahi-dev · 2026-11-06 10:04
s0ullight[m]v2, what do you mean by different one?
#asahi-dev · 2026-11-06 09:58
s0ullight[m]a 16 byte vec4 in 65536 workgroups and 256 threads each. gpu local memory, and 20 submits.
#asahi-dev · 2026-11-06 09:31
s0ullight[m]Came across "but it likely involves controlling Apple Fabric power and clocking, among other things" regarding pmp in the 7.0 progress report. Enabled pmp and it impacted gpu write throughput, as in it got reduced. I guess it's currently not fully understood and needs more work ¯_(ツ)_/¯
#asahi-dev · 2026-11-05 16:09
s0ullight[m]Any idea why the gpu write throughput is about %43 of macos? 93GB/s vs 212 on m1 pro
#asahi-dev · 2026-11-05 01:42
s0ullight[m]But looks like libkrunfw will remove the guest patch :(

plus ›