#
#asahi-gpu · 2023-08-22 19:44
i509vcbI guess I wonder where do I throw the passes to replace the intrinsics for things like load_num_workgroups with loads from uniform registers
#asahi-gpu · 2023-08-22 19:44
alyssaonce you hook that up in lower_descriptor_sets, textureSize() should Just Work
#asahi-gpu · 2023-08-22 19:43
alyssain its place, AGXV is required to use bindless texture/image instructions (rather than non-bindless ones using the index to denote a texture state register) when the backend says so
#asahi-gpu · 2023-08-22 19:43
alyssabtw after the rebase, the load_texture_base_agx sysval goes away, because it's unimplementable on Vulkan
#asahi-gpu · 2023-08-22 19:42
alyssai509vcb: nice!
#asahi-gpu · 2023-08-22 19:41
i509vcbagxv seemed to be fine after the upgrade, although I am a bit deep into sysval stuff in agxv
#asahi-gpu · 2023-08-22 18:03
alyssa:)
#asahi-gpu · 2023-08-22 17:46
Marydeqp-vk seems to be running fine so I guess everything landed on fedora ^^
#asahi-gpu · 2023-08-22 17:21
MaryTime for some kernel upgrade~
#asahi-gpu · 2023-08-22 16:53
alyssano need to build your own kernel, I hope
#asahi-gpu · 2023-08-22 16:53
alyssaso bump to the latest agxv/main and upgrade your system and you should be synced
#asahi-gpu · 2023-08-22 16:53
alyssaella-0: i509vcb: Mary: agxv/main is now rebased on top of the latest UAPI, which is now shipping in both the fedora remix and the arch
#asahi-gpu · 2023-08-22 12:02
alyssai509vcb: nir_Shader_instructions_pass and just return fals
#asahi-gpu · 2023-08-22 03:56
i509vcbOver nir as in the entire shader
#asahi-gpu · 2023-08-22 03:56
i509vcbWell... I already reinvented the wheel, but it could be square
#asahi-gpu · 2023-08-22 03:56
i509vcbBefore I reinvent the wheel, is there some sort of accepted way to do a sort of read only pass over nir?
#asahi-gpu · 2023-08-22 03:21
alyssawhen the new kernel ships (very soon) I'll push that to agxv/main
#asahi-gpu · 2023-08-22 03:21
alyssai509vcb: ella-0: Mary: I've rebased AGXV with the new uapi, it's in alyssa/mesa:agxv/rebase
#asahi-gpu · 2023-08-22 02:26
alyssa~= if the pass made any change to the shader
#asahi-gpu · 2023-08-22 02:26
alyssai509vcb: Typically, whether the pass made "progress"
#asahi-gpu · 2023-08-22 02:16
i509vcbwhat does the bool return value mean for nir passes?
#asahi-gpu · 2023-08-21 22:55
alyssanod
#asahi-gpu · 2023-08-21 22:55
i509vcbElla and I did discuss this a bit and it seems we agreed that agxv would probably differ a bit there
#asahi-gpu · 2023-08-21 22:55
alyssae.g. you will never see a load_ssbo_address in Vulkan
#asahi-gpu · 2023-08-21 22:55
alyssaPart of it is that a big chunk of agx_nir_lower_sysvals is descriptor-related, and VK doesn't want any of that code because the equivalent lowering happens in the descriptor set code
#asahi-gpu · 2023-08-21 22:54
alyssaagx_nir_lower_sysvals.c was designed for vk's benefit :)
#asahi-gpu · 2023-08-21 22:54
alyssathe whole reason that sysvals get lowered in the gl driver is *because* i'm under the impression agxv will want to do something different than gl
#asahi-gpu · 2023-08-21 22:54
i509vcbok
#asahi-gpu · 2023-08-21 22:53
alyssai509vcb: Almost certainly should *not* be common
#asahi-gpu · 2023-08-21 21:54
i509vcbalyssa: I imagine some of the stuff in agx_nir_lower_sysvals.c will need to be replicated in agxv, is that something you'd require or really like to be common between vk and gallium or is agxv free to do what it wants?
#asahi-gpu · 2023-08-21 13:42
eric_engestromthe two options were 1) wait for the next CTS release that includes the fix (it will take a while) and have this expectation change drowned in all the other changes, or 2) backport the fix and update the expectations now, so that when the next version comes around hopefully not too many things change on Mesa side
#asahi-gpu · 2023-08-21 13:40
eric_engestromjannau: sure :)
#asahi-gpu · 2023-08-21 11:48
jannaueric_engestrom: thanks for https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24808 - I was going to look/ask how that should be handled
#asahi-gpu · 2023-08-20 15:35
alyssajannau: Alright. Worth looking into why mesa/st is falling down into software, I guess.
#asahi-gpu · 2023-08-20 15:32
jannaualyssa: still terrible with asahi/main, still _mesa_GetTexSubImage_sw/util_format_b10g10r10x2_unorm_unpack_rgba_8unorm
#asahi-gpu · 2023-08-20 14:16
alyssaoh no apple's compiler strongly disagrees with that
#asahi-gpu · 2023-08-20 14:08
alyssaooh, for the specific range of [0, 31) it might be doable in 1 IC + 1 SCIB with the `extr` instruction and some magic constants
#asahi-gpu · 2023-08-20 14:07
alyssathis is not relevant :~)
#asahi-gpu · 2023-08-20 14:04
alyssawhich is actually faster? who knows
#asahi-gpu · 2023-08-20 14:04
alyssathe two obvious approaches are multiply-and-shift (2 IC operations) or convert-multiply-convert (2 SCIB + 1 F16)
#asahi-gpu · 2023-08-20 14:04
alyssameanwhile i'm trying to figure out the fastest way for AGX assembly to divide/modulo a small integer by 3
#asahi-gpu · 2023-08-20 14:00
jannaunot a problem, kernel build is underway
#asahi-gpu · 2023-08-20 13:58
alyssaok, umm
#asahi-gpu · 2023-08-20 13:58
alyssajannau: ..Right, I also forgot the uapi bump. Uff.
#asahi-gpu · 2023-08-20 13:54
jannaualyssa: better but still bad with llvmpipe. I forgot/missed the uapi bump
#asahi-gpu · 2023-08-20 13:37
alyssa:+1:
#asahi-gpu · 2023-08-20 13:36
jannaualyssa: I'll try, building asahi/main right now
#asahi-gpu · 2023-08-20 13:15
alyssaotherwise, I guess try to figure out why st_GetTexSubImage is bailing to software
#asahi-gpu · 2023-08-20 13:15
alyssaI don't /think/ so, but..
#asahi-gpu · 2023-08-20 13:15
alyssaI'm wondering if the fast path needs GLES3.1

plus ›