#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 · 2026-05-07 12:21
alyssawe don't know how#asahi-gpu · 2026-05-06 23:06
alyssaoof#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 · 2026-05-05 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?#asahi-gpu · 2026-04-22 20:55
alyssawe don't know#asahi-gpu · 2026-03-19 23:35
alyssano#asahi-gpu · 2026-03-01 02:20
alyssathen probably you can merge#asahi-gpu · 2026-03-01 02:20
alyssaoh#asahi-gpu · 2026-03-01 02:12
alyssakarolherbst: probably should get cts'd..#asahi-gpu · 2026-02-28 14:04
alyssa& no I'm not getting sniped into writing it :p#asahi-gpu · 2026-02-28 14:04
alyssano uapi change needed, should be just a small amount of code#asahi-gpu · 2026-02-28 14:00
alyssashould be straightforward for asahi#asahi-gpu · 2026-02-28 14:00
alyssajannau: 06b41ca5899 ("iris: add support for fence signal capability") enables both exts#asahi-gpu · 2026-02-22 15:03
alyssathat's the classic way to force a particular register pressure of your choosing#asahi-gpu · 2026-02-22 15:02
alyssaint temp0, ...., tempN; for(..) { temp0 = temp1; temp1 = temp2; ... tempN-1 = tempN; tempN = temp0; } foo(tempN)#asahi-gpu · 2026-02-22 15:02
alyssaThat you can confirm by varying the number of registers used, which is easy with an unrollable loop#asahi-gpu · 2026-02-22 15:01
alyssaPotentially one of the instructions adapts "USC Registers" - in this case, dynamically allocating GPRs from the cache for the rest of the program to use#asahi-gpu · 2026-02-22 15:00
alyssaIf you wanted to confirm that hypothesis, compare against shaders that use threadgroup memory (and different amounts of it)#asahi-gpu · 2026-02-22 15:00
alyssaBut if I did have to make a guess, one of those instructions does the equivalent of the M1 "USC Shared" struct.#asahi-gpu · 2026-02-22 14:59
alyssaI wouldn't worry about it too much#asahi-gpu · 2026-02-22 14:57
alyssaLikely the unknown instructions relate to dynamic caching / setting up other bits of state that traditionally USC control words managed but is now shader managed#asahi-gpu · 2026-02-22 14:56
alyssaAlso this looks like a preamble shader, which on M1 are independent binaries from the main shader, not sure if they merged that in M3#asahi-gpu · 2026-02-22 14:55
alyssaThe first two moves are hardbaking the address of the data to plumb in, which those last 4 moves (from the loaded data into the uniform registers) are dealing with#asahi-gpu · 2026-02-22 14:55
alyssaPresumably that mov.todo is a tail call#asahi-gpu · 2026-02-22 14:54
alyssaBased on the snippet you linked, the solution is pretty simple: dynamically generate a prolog that uses shader instructions to setup the state that would previously be done with USC control words.#asahi-gpu · 2026-02-22 14:53
alyssaBut that presents a chicken & egg problem - how do you get kernel arguments and things like that into the shader if you don't have the sideband?#asahi-gpu · 2026-02-22 14:52
alyssauniform registers are the spiciest - Likely they still exist in the form of something like AMD's SGPRs but you can't use them the traditional m1 way because that part of the control stream is gone#asahi-gpu · 2026-02-22 14:52
alyssaNot sure what hapens to coeff registers#asahi-gpu · 2026-02-22 14:52
alyssaTexture/sampler state regs likely go away by improving hw bindless support#asahi-gpu · 2026-02-22 14:51
alyssaWith M3, all of that is gone due to dynamic caching.#asahi-gpu · 2026-02-22 14:51
alyssaThe driver puts the right control words and then the program has the environment accessible easily.#asahi-gpu · 2026-02-22 14:51
alyssauniform registers, texture state registers, sampler state registers, coefficient registers#asahi-gpu · 2026-02-22 14:50
alyssaOn M1, there are special control words to load special registers with data from memory#asahi-gpu · 2026-02-22 14:50
alyssaIf i had to take a guess --#asahi-gpu · 2026-02-22 14:50
alyssapac85: those shaders would be fine here I believe#asahi-gpu · 2026-02-22 14:47
alyssagood work#asahi-gpu · 2026-02-22 14:47
alyssaoliver: that sounds believable#asahi-gpu · 2026-02-21 17:21
alyssaexactly#asahi-gpu · 2026-02-21 17:04
alyssachaos_princess: pretending i've never written a display controller driver intensifies