ChanServ changed the topic of #lima to: Development channel for open source lima driver for ARM Mali4** GPUs - Kernel has landed in mainline, userspace driver is part of mesa - Logs at and - Contact ARM for binary driver support!
mripard has joined #lima
kaspter has quit [Quit: kaspter]
kaspter has joined #lima
kaspter has quit [Ping timeout: 264 seconds]
kaspter has joined #lima
chewitt has quit [Quit: Adios!]
kaspter has quit [Ping timeout: 246 seconds]
<linkmauve> I live in CET btw, soon CEST. :)
<linkmauve> It works!
kaspter has joined #lima
kaspter has quit [Remote host closed the connection]
kaspter has joined #lima
kaspter has quit [Ping timeout: 260 seconds]
_whitelogger has joined #lima
zombah has quit [Ping timeout: 260 seconds]
zombah has joined #lima
egbert has quit [Ping timeout: 256 seconds]
egbert has joined #lima
<anarsoul|2> linkmauve: PDT here
enunes has joined #lima
<enunes> anarsoul|2: ping
<anarsoul|2> enunes: hey
<anarsoul|2> did you look into what needs to be done to enable shareable shaders on lima?
<enunes> hmm probably not, what do you mean by shareable shaders?
<enunes> I never looked into that
<anarsoul|2> OK, I'll look into it
<anarsoul|2> also I'll work on disk cache for shaders
<enunes> ok
<enunes> so some days ago I generated this:
<enunes> it was not easy, but the mali results are on the same platform, same kernel, same OS everything, only graphics stack changing
<enunes> 'effect2d' can be fixed relatibely easily
<enunes> 'loop3' seems to be because we dont take uniform loads out of loops, seems to be a nir thing thats why vc4 is also pretty bad there
<enunes> 'shadow' is what has been puzzling me
<anarsoul|2> enunes: IIRC blob uses Z16 in shadow while lima always uses Z24X8
<anarsoul|2> but we don't know how to reload Z16 so it's disabled
<enunes> yeah this is what I have been considering
<enunes> I also thought it could be the varying size thing that we never implemented
<anarsoul|2> yes, also this
<enunes> but overall better than what I expected
<anarsoul|2> that also affects refract (and anything with complex geometry)
<enunes> is it too hard to figure out Z16, or just needed someone to spend time in that?
<anarsoul|2> enunes: blob never attempts to reload Z16 and uses Z24X8 in any tests that requires depth reload
<anarsoul|2> and sampling it as Z16 in reload shader doesn't work
<anarsoul|2> so it's not a low-hanging fruit for sure
<enunes> hmm ok, I'll give it a try, if that doesnt work I'll try fp16 varying next to see if that helps
<anarsoul|2> I wonder what's wrong with loop-3
<enunes> loop-3 is loading uniform inside a loop
<anarsoul|2> oh
<enunes> blob compiler puts the uniform load outside of the loop, nir doesnt
<anarsoul|2> I guess we copy it into the loop ourselves :)
<enunes> I looked into that a bit but I think its not too much our fault
<anarsoul|2> but I wonder why it misses the cache
<enunes> I did a local hack to that shader to force it to be loaded outside, and it already improved the results
<anarsoul|2> what's wrong with effect2d?
<enunes> effect2d uses a RGB textures to which we apply a swizzle to make alpha '1.0' and that complicates the shader quite a bit
<enunes> this hack helps I was considering to submit it
<enunes> it comes from mesa which defines the swizzle to alpha 1.0
<anarsoul|2> wouldn't it be out of spec?
<anarsoul|2> some shaders may depend on alpha being 1.0
<enunes> I agree, which is why I didnt submit it so far
<enunes> I dont know if all mesa drivers do this swizzle though, neither does the blob, I think they just treat alpha as dontcare
<anarsoul|2> actually blob supports RGB format
<anarsoul|2> one that is 3 bytes per pixel
<anarsoul|2> we used to support it but since it's not renderable (and GL requires it to be renderable) we dropped it
<enunes> I didnt go as far as to check what format it uses, just checked the shader on that one, so its possible it was using that format too
<anarsoul|2> btw, are you using deqp-runner?
<anarsoul|2> it's very convenient, but looks like our skips and fails lists were outdated
<enunes> I have it set up but I also use the regular old deqp-gles2
<anarsoul|2> it's just a runner, it still requires deqp-gles2
<enunes> makes sense, I also have a similar script
<enunes> results were outdated since we are not in CI anymore, I want to enable it again with the stack of test boards I got but then I need to finish my secure setup to expose them and power control to the boards
anarsoul|c has joined #lima
<anarsoul|2> sounds good
<anarsoul|2> anyway, I sent an MR to update the lists since I use them with deqp-runner
egbert has quit [Ping timeout: 264 seconds]
egbert has joined #lima
megi has quit [Quit: WeeChat 3.0.1]
megi has joined #lima
mmind00 has joined #lima