05:34 <i509vcb> I 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)

05:35 <i509vcb> I think this might hint at some bug with the gpu recovery?

05:42 <i509vcb> Upon 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?

05:43 <i509vcb> Unless 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

05:45 <i509vcb> Although I assume there is a reason why vk-meta is currently using a fragment shader for that?

14:01 <alyssa> i509vcb: "sfloat" is indeed just float

14:02 <alyssa> possibly this is the issue of "copies are supposed to preserve nans/denorms but they get flushed when copying as floats"

14:03 <alyssa> casting to a size-equivalent safe format (e.g. r32_uint for r32_float) should deal with that

14:03 <alyssa> this means we can't compress rgb9e5 which is annoying but meh

14:04 <alyssa> i509vcb: 3d copies can be done in one draw call with a layered framebuffer, the vertex shader can write the layer on our hardware

14:04 <alyssa> (1 draw call, 8192 rectangles)

14:05 <alyssa> i509vcb: 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

14:05 <alyssa> this is the case as exposed in metal (imageStore disables compression)

14:05 <alyssa> and is how e.g. Mali works

14:06 <alyssa> I have since learned that imageStore works on compressed images in limited cases, but it's unclear if it's enough to build a blitter.

14:06 <alyssa> (without a 3d pipe fallback at least

14:06 <alyssa> agxv isn't doing compression yet but it will

14:07 <alyssa> at some point we need to see how metal does copies

14:07 <alyssa> i don't think there's a straight-up hw DMA engine but maybe there's something helpful

19:31 <i509vcb> hmm okay thanks for the info

19:31 <i509vcb> I guess I am forced to figure out layered rendering for 3d copies

19:32 <i509vcb> (I know there is stuff you pointed me to in the past, just need to try a few things)

19:48 <alyssa> or eat the draws for now, it's not the end of the world

19:48 <alyssa> i509vcb: meta->cmd_draw_volume

19:53 <alyssa> mm, I think that should be straightforward to wire up, you don't need GS or anything

19:56 <i509vcb> yeah I guess I'll try that first

19:56 <i509vcb> I did kind of vanish for a month lol

20:00 <alyssa> all good :-)

21:40 <i509vcb> I'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

21:40 <i509vcb> I'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

23:57 <ella-0> alyssa: as far as i could tell metal always just uses a compute pass for copies