#
#asahi-gpu · 2024-01-02 04:45
alyssai509vcb: casting at the level of formats input/output to the copy
#asahi-gpu · 2024-01-02 04:45
alyssaella-0: fascinating
#asahi-gpu · 2024-01-01 23:57
ella-0alyssa: as far as i could tell metal always just uses a compute pass for copies
#asahi-gpu · 2024-01-01 21:40
i509vcbI've tried not to make a mess with what I was given: https://gitlab.freedesktop.org/i509VCB/mesa/-/blob/i5-vkCmdCopy/src/vulkan/runtime/vk_meta_copy.c?ref_type=heads#L104
#asahi-gpu · 2024-01-01 21:40
i509vcbI'm not 100% on how I'd do that casting. nir_bitcast_vector seems wrong since it wants the bitsize (probably for f32 to f64 casts or something like that) rather than a transmute of sorts
#asahi-gpu · 2024-01-01 20:00
alyssaall good :-)
#asahi-gpu · 2024-01-01 19:56
i509vcbI did kind of vanish for a month lol
#asahi-gpu · 2024-01-01 19:56
i509vcbyeah I guess I'll try that first
#asahi-gpu · 2024-01-01 19:53
alyssamm, I think that should be straightforward to wire up, you don't need GS or anything
#asahi-gpu · 2024-01-01 19:48
alyssai509vcb: meta->cmd_draw_volume
#asahi-gpu · 2024-01-01 19:48
alyssaor eat the draws for now, it's not the end of the world
#asahi-gpu · 2024-01-01 19:32
i509vcb(I know there is stuff you pointed me to in the past, just need to try a few things)
#asahi-gpu · 2024-01-01 19:31
i509vcbI guess I am forced to figure out layered rendering for 3d copies
#asahi-gpu · 2024-01-01 19:31
i509vcbhmm okay thanks for the info
#asahi-gpu · 2024-01-01 14:07
alyssai don't think there's a straight-up hw DMA engine but maybe there's something helpful
#asahi-gpu · 2024-01-01 14:07
alyssaat some point we need to see how metal does copies
#asahi-gpu · 2024-01-01 14:06
alyssaagxv isn't doing compression yet but it will
#asahi-gpu · 2024-01-01 14:06
alyssa(without a 3d pipe fallback at least
#asahi-gpu · 2024-01-01 14:06
alyssaI have since learned that imageStore works on compressed images in limited cases, but it's unclear if it's enough to build a blitter.
#asahi-gpu · 2024-01-01 14:05
alyssaand is how e.g. Mali works
#asahi-gpu · 2024-01-01 14:05
alyssathis is the case as exposed in metal (imageStore disables compression)
#asahi-gpu · 2024-01-01 14:05
alyssai509vcb: as for why fragment, it was originally thought that we needed to use the 3d pipe (fragment) in order to get apple's lossless compression
#asahi-gpu · 2024-01-01 14:04
alyssa(1 draw call, 8192 rectangles)
#asahi-gpu · 2024-01-01 14:04
alyssai509vcb: 3d copies can be done in one draw call with a layered framebuffer, the vertex shader can write the layer on our hardware
#asahi-gpu · 2024-01-01 14:03
alyssathis means we can't compress rgb9e5 which is annoying but meh
#asahi-gpu · 2024-01-01 14:03
alyssacasting to a size-equivalent safe format (e.g. r32_uint for r32_float) should deal with that
#asahi-gpu · 2024-01-01 14:02
alyssapossibly this is the issue of "copies are supposed to preserve nans/denorms but they get flushed when copying as floats"
#asahi-gpu · 2024-01-01 14:01
alyssai509vcb: "sfloat" is indeed just float
#asahi-gpu · 2024-01-01 05:45
i509vcbAlthough I assume there is a reason why vk-meta is currently using a fragment shader for that?
#asahi-gpu · 2024-01-01 05:43
i509vcbUnless I've missed something, we'd need to write in the worst case 8192 layers of a 2d copy for each layer of the 3d image
#asahi-gpu · 2024-01-01 05:42
i509vcbUpon further thought for vk-meta-copy to properly do 3d copies without doing something horrible for performance like submitting 8192 draw calls for a full 8192x8192x8192 copy for an image, compute shaders with imageLoad and imageStore might be best?
#asahi-gpu · 2024-01-01 05:35
i509vcbI think this might hint at some bug with the gpu recovery?
#asahi-gpu · 2024-01-01 05:34
i509vcbI wonder if this info would be helpful for figuring out how to prevent the unrecoverable timeouts I get. I noticed with deqp-runner if I set the maximum number of jobs to 1 or 2 (although I think the unrecoverable timeouts will occur with less frequency) I don't get unrecoverable timeouts. By default deqp-runner will use all my cores (8 on the M2 air)
#asahi-gpu · 2023-12-31 19:27
i509vcbUNORM and SNORM are pretty much SFLOAT but with a fixed range
#asahi-gpu · 2023-12-31 19:25
i509vcbI don't get any 2d_to_2d failures if I disable those formats, but I get a feeling there is a more precise issue at hand here
#asahi-gpu · 2023-12-31 19:24
i509vcbIs AGX_TEXTURE_TYPE_FLOAT actually UNORM/SNORM?
#asahi-gpu · 2023-12-31 06:35
i509vcbdEQP-VK.api.copy_and_blit.core.image_to_image.all_formats.color.2d_to_2d.r16_sint.r16_sfloat.* seems especially suspicious
#asahi-gpu · 2023-12-31 06:35
i509vcbhttps://gist.github.com/i509VCB/2dba6f60e8e73ccde116cb965de3739d
#asahi-gpu · 2023-12-31 06:35
i509vcbI noticed a pattern that no formats with a source sfloat numeric type (except for the exception of packed formats being the destination and depth/stencil) fail
#asahi-gpu · 2023-12-31 06:18
i509vcbI guess it seems valid to be Float then?
#asahi-gpu · 2023-12-31 06:18
i509vcbHmm pipe_format says SFLOAT is FLOAT except for stencil formats it seems
#asahi-gpu · 2023-12-31 06:16
i509vcbI grepped out 2d to 2d tests and noticed sfloat is the vast majority of the failures
#asahi-gpu · 2023-12-31 06:15
i509vcbI noticed that Texture Type in cmdbuf.xml does not list SFLOAT, I assume this means the hardware can't do anything with an SFLOAT numerical type?
#asahi-gpu · 2023-12-30 19:45
incogninto1-1Hello
#asahi-gpu · 2023-12-25 14:19
chaos_princesstesselation control shader
#asahi-gpu · 2023-12-25 14:19
alyssasure
#asahi-gpu · 2023-12-25 14:18
chaos_princessalyssa: can push the terrible wip branch
#asahi-gpu · 2023-12-25 14:18
alyssaalso, what you trying to do? :D
#asahi-gpu · 2023-12-25 14:18
alyssawould need to see the diff
#asahi-gpu · 2023-12-25 14:18
alyssachaos_princess: hmm

plus ›