#asahi-gpu · 2026-05-31 15:06
karolherbstanybody is free to pick this commit and move it into a MR and test it's all working as expected: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36803/diffs?commit_id=01d5ebd0067b7e87dbb91602e21696d2fe86e971#asahi-gpu · 2026-05-31 14:57
chnoblouch[m]someone implements support for this in the GPU driver.#asahi-gpu · 2026-05-31 14:57
chnoblouch[m]ffx: The problem seems to be that Asahi's GPU driver doesn't implement EXT_semaphore_fd, which Firefox probably uses for sharing GPU images between processes, so it has to fall back to reading the images to RAM and sharing the bytes like that, which is of course much slower. I wouldn't call this a problem in the GPU driver, it's just that it doesn't implement an *optional* OpenGL extension. The only thing you can do here is wait until#asahi-gpu · 2026-05-31 08:55
ffx seem to support EXT_semaphore_fd, which is causing the issue. The linked webpage, which renders multiple canvases layered should be easily drawable even for a plain m1, even at retina at 144hz. For example my firefox on my pixel7a can render all layers at 90fps very smoothly. Again here is the bugreport: https://bugzilla.mozilla.org/show_bug.cgi?id=2019868#asahi-gpu · 2026-05-31 08:55
ffxHi again, my point was that something still seems to be not ok in the asahi 44 graphics driver using an m1pro. you can see that effect also in browser scrolling, which is already not smooth at power management setting medium. i would be happy to help here, but i still have no idea what is the problem, and also i cant code drivers or similar, i am a webdev. In my firefox bug report there are telling me that the graphics driver does not#asahi-gpu · 2026-05-30 16:43
ffxok, c or rust then#asahi-gpu · 2026-05-30 16:35
chaos_princesslinux using c++ is news to me. and well, haven't the previous conversations told you about the support status#asahi-gpu · 2026-05-30 16:27
ffxso is EXT_semaphore_fd supported?#asahi-gpu · 2026-05-30 16:27
ffxhm, sadly i am not a kernel c++ coder like you :)#asahi-gpu · 2026-05-27 22:21
StanfordAppliedCyber[m]i just don't want to duplicate work other people are doing#asahi-gpu · 2026-05-27 22:19
StanfordAppliedCyber[m]can anyone point me towards the most up to date resources on the m3 (or later) gpu? i looked through the project documentation and everything agx related in m1n1 and most of it appears to be quite old#asahi-gpu · 2026-05-26 19:08
chaos_princesspower management is unrelated, your observed effects are due to giving it more power to brute-force the results#asahi-gpu · 2026-05-26 19:05
ffxalyssa not sure if I can follow... From my understanding, this is a powermangement + gpu driver issue?#asahi-gpu · 2026-05-26 18:53
karolherbsthttps://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36803/diffs?commit_id=01d5ebd0067b7e87dbb91602e21696d2fe86e971#asahi-gpu · 2026-05-26 18:52
karolherbstohh I think I had patches for it...#asahi-gpu · 2026-05-26 17:45
alyssaand I believe I explained at the time what you needed to do to fix your issue#asahi-gpu · 2026-05-26 17:44
alyssaffx: we talked about this 3 months ago https://oftc.catirclogs.org/asahi-gpu/2026-02-28#asahi-gpu · 2026-05-26 16:35
ffxhttps://bugzilla.mozilla.org/show_bug.cgi?id=2019868#asahi-gpu · 2026-05-26 16:34
ffxEXT_semaphore_fd is required for presentation without reading back to the CPU. Until the driver implements that, we're always going to be pretty slow.#asahi-gpu · 2026-05-26 16:34
ffxfirefox devs wrote: P5: Performance.#asahi-gpu · 2026-05-26 16:33
ffxIn asahi linux 44 firefox, it is like such a fullscreen video was opened all the time.#asahi-gpu · 2026-05-26 16:32
ffxhi, some observation under macOS M1pro Firefox: If you have a video at fullscreen on the internal monitor and replaying the above chiptune player, the player will not replay at smooth fps until you quit fullscreen video of the other screen.#asahi-gpu · 2026-05-25 19:15
jannau@s u]#asahi-gpu · 2026-05-25 19:15
jannaubf./p;['}|#asahi-gpu · 2026-05-07 12:29
karolherbstyeah.. guess I'll just upstream the thing I also need for radeonsi, have the strided copy working now, so whatever#asahi-gpu · 2026-05-07 12:21
alyssawe don't know how#asahi-gpu · 2026-05-07 08:40
karolherbstbut also.. should probably have a way to increase the timeput duration for compute workloads..#asahi-gpu · 2026-05-07 07:11
karolherbstokay.. seems like I'm simply running into a timeout, but also kinda my fault for splitting a CopyBufferRect into 150 copies of 16 bytes 🙃 Yes, I should have a strided copy helper kernel for this...#asahi-gpu · 2026-05-06 23:06
alyssaoof#asahi-gpu · 2026-05-06 22:34
karolherbstokay.. "fun", I have a weird synchronization? or something bug, that is fixed by adding 0x70 to dispatch_upper_bound... and now I'm wondering if somewhere an overrun of the batch isn't caught.. but that's for tomorrow to look into#asahi-gpu · 2026-05-06 15:12
karolherbstcool thanks, that has helped#asahi-gpu · 2026-05-06 13:24
alyssastuff is supposed to get flushed#asahi-gpu · 2026-05-06 13:24
alyssasee the very end of agx_launch_grid#asahi-gpu · 2026-05-06 13:24
alyssathat opt never got merged#asahi-gpu · 2026-05-06 13:24
alyssaoh wait#asahi-gpu · 2026-05-06 13:23
alyssanot sure how that's getting hit though#asahi-gpu · 2026-05-06 13:23
alyssakarolherbst: agx_ensure_cmdbuf_has_space(...cdm, [size])#asahi-gpu · 2026-05-06 13:23
alyssakarolherbst: umm#asahi-gpu · 2026-05-06 13:00
karolherbsthow would I fix this error? ../src/gallium/drivers/asahi/agx_state.c:3013: agx_launch_internal: Assertion `batch->cdm.current <= batch->cdm.end && "Failed to reserve sufficient space in encoder"' failed.#asahi-gpu · 2026-05-05 22:02
karolherbstI only brought my macbook to iwocl, so I guess I'm gonna fix all CTS regressions for CL 3.1 support then 🙃#asahi-gpu · 2026-05-05 22:01
karolherbstwell something for tomorrow to fix#asahi-gpu · 2026-05-05 22:01
karolherbstokay.. the lower_fminmax_signed_zero lowering breaks it...#asahi-gpu · 2026-05-05 21:55
karolherbstI think something optimizes the flushing away..#asahi-gpu · 2026-05-05 21:54
karolherbstmhh something is funky here: ERROR: fmax: -8386560.000000 ulp error at {-0x1.ffcp-139 (0x800007ff), -0x1.fffffcp-127 (0x807fffff)}: *-0x1.ffcp-139 vs. -0x1.fffffcp-127 (0x807fffff) at index: 3763#asahi-gpu · 2026-05-05 21:43
karolherbstMhhhh... yeah I just remember that we were talking about it.. will debug more tomorrow and figure out if this is a new regression or something#asahi-gpu · 2026-05-05 19:44
alyssathe outstanding problem is signed zero stuff which you should talk to glehmann about and fix, thank you#asahi-gpu · 2026-05-05 19:43
alyssabut also I do not work on this driver anymore lol#asahi-gpu · 2026-05-05 19:43
alyssaagx_nir_lower_fminmax.c#asahi-gpu · 2026-05-05 19:42
alyssaMesa handles this correctly#asahi-gpu · 2026-05-05 19:42
alyssakarolherbst: what's the problem?