#
#asahi-gpu · 2023-01-15 17:10
alyssaI don't know if any real apps would even hit that. I'm doubful.
#asahi-gpu · 2023-01-15 17:10
alyssawhile that's almost certainly faster, it isn't worth the complexity ... vkCmdCopyImageToImage with compressed<--->compressed and incompatible formats is an abomination that should never have been added to the spec and bad perf is to be expected
#asahi-gpu · 2023-01-15 17:09
alyssapanvk 1.0 has some extremely delicate logic to try to do compressed->compressed image copies with format conversion in one shot with format packing/unpacking in the copy shaders
#asahi-gpu · 2023-01-15 17:08
alyssaMali is #1 on Valhall, but #2 on anything older
#asahi-gpu · 2023-01-15 17:08
alyssaI am unsure which class of hardware AGX is.
#asahi-gpu · 2023-01-15 17:08
alyssageneric on the fragment pipe.
#asahi-gpu · 2023-01-15 17:08
alyssa2. hw that can't. obviously this class of hw can still do uncompressed -- at worst, linear -- texture views. so the sane option here is to eat the bandwidth hit -- allocate an uncompressed staging resource and lower compressed->compressed copies to a pair of copies compressed->uncompressed + uncompressed->compressed because you can do views for each of those. and those simpler copies are easy and
#asahi-gpu · 2023-01-15 17:06
alyssa1. hw that can do compressed texture views. in this case, img2img copies are just blits! you might need to munge the dest format into something renderable with the same size, maybe munge the dimensions in some evil cases, but by and large this is generic blitting on the fragment pipe, where a texture view (with view format != original format) on the sampling image for casting
#asahi-gpu · 2023-01-15 17:04
alyssaright, so, there are basically 2 classes of hardware
#asahi-gpu · 2023-01-15 17:03
ella-0Yup it's pain :<
#asahi-gpu · 2023-01-15 17:02
alyssaimplementing that efficiently requires being able to "cast" framebuffer compressed images to different formats
#asahi-gpu · 2023-01-15 17:02
alyssathe only spicy thing is vkCmdCopyImageToImage
#asahi-gpu · 2023-01-15 17:02
alyssavkCmdCopyBufferToBuffer is also a totally generic compute kernel
#asahi-gpu · 2023-01-15 17:01
alyssaalso I would expect vk_meta can do those because they're pretty general
#asahi-gpu · 2023-01-15 17:01
alyssaexcept maybe some little details about ASTC/BCn formats? but I think you can just munge the dimensions?
#asahi-gpu · 2023-01-15 17:01
alyssaprobably
#asahi-gpu · 2023-01-15 17:01
alyssavkCmdCopyBufferToImage should be a fragment shader
#asahi-gpu · 2023-01-15 17:01
alyssavkCmdCopyImageToBuffer should probably be a compute kernel
#asahi-gpu · 2023-01-15 17:00
alyssavkCmdCopy* sucks.
#asahi-gpu · 2023-01-15 17:00
alyssaRight..
#asahi-gpu · 2023-01-15 16:58
ella-0That makes sense. I'm happy to work on compute stuff :3 currently the main thing holding agxv back is not having vkCmdCopy* and other transfer operations implemented. I tried and failed to implement them using the tilebuffer load/store shaders
#asahi-gpu · 2023-01-15 16:50
alyssao
#asahi-gpu · 2023-01-15 16:43
alyssa
#asahi-gpu · 2023-01-15 16:33
alyssamake sure my baby can grow without me, yknow
#asahi-gpu · 2023-01-15 16:30
alyssaI was hoping the combination of you, Lina, Karol, and Dougall would be able to finish off compute support without me ^.^
#asahi-gpu · 2023-01-15 16:29
alyssamakes sense
#asahi-gpu · 2023-01-15 16:29
ella-0I was hoping to get some basic compute running on agxv before that I think
#asahi-gpu · 2023-01-15 16:28
alyssakie
#asahi-gpu · 2023-01-15 16:28
ella-0uhh not sure
#asahi-gpu · 2023-01-15 16:22
alyssai.e. when should I pull your tree and start hacking on deqp-vk, versus continuing to support more and better gl in the interest of agxv
#asahi-gpu · 2023-01-15 16:22
alyssaby the way.. at what point will we be bottlenecked on agxv specific stuff
#asahi-gpu · 2023-01-15 16:15
alyssaand obviously it solves a problem I had in GL lol
#asahi-gpu · 2023-01-15 16:15
alyssabut if you say it's right and Jason says it's right that is good for me
#asahi-gpu · 2023-01-15 16:15
alyssaI still don't really get how
#asahi-gpu · 2023-01-15 16:15
alyssaHmmm ok
#asahi-gpu · 2023-01-15 16:14
ella-0It simplifies the command buffer code somewhat and makes it possible to allocate a chunk of the register file for push constants.
#asahi-gpu · 2023-01-15 16:12
ella-0yes
#asahi-gpu · 2023-01-15 16:07
alyssaI've never written a VK driver lol
#asahi-gpu · 2023-01-15 16:07
alyssaIs it, like... better than what we had before?
#asahi-gpu · 2023-01-15 16:07
alyssaella-0: awesome!
#asahi-gpu · 2023-01-15 15:56
ella-0alyssa: With respect to the system values MR I am currently porting AGXV to it. Will let you know if anything doesn't work well with vulkan but so far so good :3
#asahi-gpu · 2023-01-15 15:54
alyssathis should be predicated on stage == COMPUTE, you're writing into a union
#asahi-gpu · 2023-01-15 15:54
alyssahttps://gitlab.freedesktop.org/asahi/mesa/-/commit/e4a856e52ac8d88b3af6b5c4cad0534d01edac96#89b6e919adec3916dacf34377fe9e0036bd40140_89_89
#asahi-gpu · 2023-01-15 15:52
alyssathis should be _writes, not _reads
#asahi-gpu · 2023-01-15 15:52
alyssalina: https://gitlab.freedesktop.org/asahi/mesa/-/commit/fe3eed5f527cf51f8a45b2ad24ba5a79eb7e71c3#89b6e919adec3916dacf34377fe9e0036bd40140_42_47
#asahi-gpu · 2023-01-15 06:02
linaI think what happens is that since we're not flushing caches, the fault happens after the compute job is complete, in some state where the firmware doesn't implement timeouts properly...
#asahi-gpu · 2023-01-15 06:02
linaAaaa, I just tried doing a faulting write in that shader and it locks up the whole GPU... the compute part runs fine, but a subsequent 3D job hangs. I can see the fault status but the firmware is just sitting there...
#asahi-gpu · 2023-01-15 04:21
linaAnd then there's a magic store to address 0x2d822acc in here... is this some kind of MMIO or doorbell?
#asahi-gpu · 2023-01-15 04:12
linaBut at least one op, 0xf, just runs 5 unknown instructions. I wonder if that is cache maintenance...
#asahi-gpu · 2023-01-15 04:12
linaSome codepaths do save/restore stuff (and I think we might be missing some looping instruction decoding? I see unknown instructions right after the load/store groups...)

plus ›