#
#asahi-gpu · 2023-03-19 16:07
alyssa02345 is the only interesting thing here but, I'm printing it and it's not telling me what I need
#asahi-gpu · 2023-03-19 16:07
alyssabut that, doesn't tell me where the actual samplers live!
#asahi-gpu · 2023-03-19 16:06
alyssaand corresponding index in the 02345 buffer gets the second word replaced by 0x780 or 0x781
#asahi-gpu · 2023-03-19 16:06
alyssa(seeing as they each bump up when I create a bindless sampler even if it's not actually used)
#asahi-gpu · 2023-03-19 16:06
alyssa??
#asahi-gpu · 2023-03-19 16:06
alyssaIOGPU Compute word 39: # of bindless samplers
#asahi-gpu · 2023-03-19 16:06
alyssaIOGPU Compute word 38: # of bindless samplers + 1
#asahi-gpu · 2023-03-19 15:59
alyssabut, if changing the sampler descriptor is not changing memory as far as I can see, IDK what to make of that other than not mapping memory hard enough
#asahi-gpu · 2023-03-19 15:29
alyssaI'm clearly missing something dumb
#asahi-gpu · 2023-03-19 15:17
alyssaeither my Metal code is wrong or something is horribly broken in my decoder ..
#asahi-gpu · 2023-03-19 15:15
alyssai feel like i'm getting warmer but this isn't clicking yet
#asahi-gpu · 2023-03-19 14:47
alyssabut it isn't at all clear to me where the sampler descriptors themselves live. presumably this is indexing into a global heap but I can't figure out where.
#asahi-gpu · 2023-03-19 14:46
alyssathe sampler register is a 16-bit index, not a byte offset (!)
#asahi-gpu · 2023-03-19 14:46
alyssaas for the samplers though
#asahi-gpu · 2023-03-19 14:46
alyssamultiples of 24
#asahi-gpu · 2023-03-19 14:46
alyssabyte offsets
#asahi-gpu · 2023-03-19 14:46
alyssaMetal's texture handles are 32-bit offsets into the heap or something like that
#asahi-gpu · 2023-03-19 14:45
alyssaadd the register (if present) to the 64-bit uniform (must be present) and that's the address of the texture descriptor in memory, bam, done.
#asahi-gpu · 2023-03-19 14:45
alyssanamely
#asahi-gpu · 2023-03-19 14:44
alyssastill mystified by bindless samplers.
#asahi-gpu · 2023-03-19 14:44
alyssaOK. I think I have a more or less complete picture of bindless texture handles then
#asahi-gpu · 2023-03-19 14:31
alyssaI think I'm figuring out what the 02345 buffer is...
#asahi-gpu · 2023-03-19 14:31
alyssalina: uh oh
#asahi-gpu · 2023-03-18 22:45
alyssabecause my compressed 2d arrays aren't working properly, boo
#asahi-gpu · 2023-03-18 22:13
alyssanote to self: massive 35%+ perf boost in dolphin on this trace if I can get it to use compression more, it seems to work so next step is figuring out why that's not already happening
#asahi-gpu · 2023-03-18 21:45
alyssahttps://gitlab.freedesktop.org/mesa/mesa/-/issues/8649
#asahi-gpu · 2023-03-18 21:33
alyssaor AGX_MESA_DEBUG=nopreamble
#asahi-gpu · 2023-03-18 21:33
alyssaIf you suppress faults in the kernel it'll pass
#asahi-gpu · 2023-03-18 21:33
alyssalina: IDK where we're tracking known fails but "./arb_texture_buffer_object-max-size" is failing due to preamble brokenness
#asahi-gpu · 2023-03-18 21:02
alyssaunderallocating for block-compressed textures
#asahi-gpu · 2023-03-18 21:02
alyssaMESA: warning: Address 0x5fffae84c0 is 0x4c0 bytes beyond an object at 0x5fffac8000..0x5fffae7fff (Texture)
#asahi-gpu · 2023-03-18 21:02
alyssalooks like an ail bug
#asahi-gpu · 2023-03-18 21:01
alyssa./fbo-generatemipmap-formats -auto GL_EXT_texture_compression_s3tc
#asahi-gpu · 2023-03-18 19:21
alyssa1D, dimension=twiddled for some reason, width as expected, height=1, metal bans mipmapping, page-aligned layers used
#asahi-gpu · 2023-03-18 19:20
alyssa1d textures in metal, lets see
#asahi-gpu · 2023-03-17 01:41
alyssa(Stuff that touches other parts of the tree obviously should get MR'd upstream earlier and split out from the bunch)
#asahi-gpu · 2023-03-17 01:41
alyssaThis means for Asahi vendor stuff, we only interact with upstream CI once every 'some amount of time'
#asahi-gpu · 2023-03-17 01:40
alyssa* but immediately before that happens, asahi/main is rebased on upstream and force-pushed. Ideally this happens when there are minimal outstanding MRs because force pushing makes a mess of MRs, but it's not the end of the world. We'll just have to be fast about review! :~P
#asahi-gpu · 2023-03-17 01:39
alyssaOnce every 'some amount of time' (maybe weekly? maybe fortnightly?), asahi/main is MR'd against upstream mesa/mesa (with the UAPI commits dropped)
#asahi-gpu · 2023-03-17 01:38
alyssaNo special branch for that. In particular, asahi/main contains the new UAPI work.
#asahi-gpu · 2023-03-17 01:38
alyssamarcan: If we do CI properly, then asahi/main will always be healthy so you can cut a release "whenever" from asahi/main
#asahi-gpu · 2023-03-17 01:37
alyssaasahi/main is never* pushed to directly, any commits to it must go through the MR and review process undergoing a similar level of scrutiny to mesa/mesa (no HACKs, no WIPs, no FIXMEs, no NEEDSTESTs, no fixup!, you get the idea)
#asahi-gpu · 2023-03-17 01:36
alyssaThe way I envision this working is that any new Asahi work is MR'd against asahi/mesa
#asahi-gpu · 2023-03-17 01:35
alyssalina: I'm not thrilled that it's come to this, but... https://gitlab.freedesktop.org/asahi/mesa/-/merge_requests
#asahi-gpu · 2023-03-16 18:18
alyssaeven though there are no frag->vertex deps and I thought things were synced within the queues?
#asahi-gpu · 2023-03-16 18:17
alyssalina: also, removing the BARRIERs in terrain causes broken rendering
#asahi-gpu · 2023-03-15 21:44
alyssalina: as the other significant copyright holder, if you are in favour of alyssa/mesa:agx/spdx, I would like a joint sign-off on the commit (not a reviewed-by tag)
#asahi-gpu · 2023-03-15 14:58
alyssagrumble grumble
#asahi-gpu · 2023-03-15 14:58
alyssaso apparently 64-bit doesn't need full 4x alignment always? what?
#asahi-gpu · 2023-03-15 14:58
alyssa 22: 0e03c26b3c000000 isub r0_r1, r1_r2.discard, r3_r4.discard

plus ›