#asahi-gpu on 2021-03-21 — irc logs at freenode.irclog.whitequark.org

01:18 <DarkShadow44> bloom: How exactly snooping GPU memory? Is that what comes out from wrap.c:dump_mappings?

01:19 <dougall> (to be clear you want g13 instructions, rather than llvm ir, right?)

01:19 <DarkShadow44> yes, for analysing the ISA

01:23 <dougall> i find this branch the easiest way to build and get G13 code from a metallib: https://github.com/Thog/gpu/tree/metallib (you can modify macho_disasm.c to write the code to a file instead of calling agx_disassemble... but metallib code does compile a bit differently to non-library code, so something like dump_mappings is preferable for that)

01:24 <dougall> (e.g. it'll never use the discard hint, as it can't assume all threads are active when the library function is called)

01:27 <DarkShadow44> ah thanks. so, to be clear, dump_mappings does indeed dump G13 code as well?

01:29 <dougall> i'm like 90% sure - yeah - the tricky bit is it'll dump code for about six other different shaders and you have to find the code you care aboout at some offset in one of the dumped mappings...

01:30 <DarkShadow44> oh I see...

01:30 <DarkShadow44> about the caching bit, I only knew about metallib shader files

01:31 <DarkShadow44> are shaders usually uploaded to the driver as source code? Or why is there a generation difference?

01:31 <DarkShadow44> not quite sure how the caching/threads work yet

01:31 <dougall> (i got it working with an old revision of asahilinux/gpu and never updated - iirc one of these hardcoded offsets worked for me https://github.com/AsahiLinux/gpu/blob/b9591908e19a737aaa60c28619fbd39a091ce0f1/wrap/wrap.c#L143 )

01:33 <dougall> and yeah, no idea - i don't really know how metal works, sorry

01:38 <dougall> (or, if I'm understanding correctly, I think they're usually given to the driver as LLVM IR? which it specialises because it knows "this is a top level compute or whatever function"? but that's mainly based on having used those two tools on the same functions and seen the code differences, so I'm not sure)

02:21 <DarkShadow44> hm, according to what I found, metallib seems to be LLVM IR code

02:21 <DarkShadow44> they are definitely not macho, the magic is different

02:22 <DarkShadow44> my I ask how you got your shader mdougall: acho?

02:22 <DarkShadow44> dougal: may I ask how you gor your shader-macho file?

02:26 <dougall> ah, yeah (heh, neither of those pings worked) - I use https://github.com/Thog/gpu/blob/metallib/metal_compiler.m which calls [MTLDynamicLibrary serializeToURL:...] to generate it

02:27 <dougall> (command to compile it is in a comment at the top of that file - sorry if that's pointing out the obvious, but i missed it myself)

02:31 <DarkShadow44> ah thanks, I kinda missed that

02:32 <DarkShadow44> probably won't need the compiler snipped, since I'm using XCode

04:16 <DarkShadow44> dougall: Is there a specific instruction where I could help fill in the gaps? As a beginner, and so.

06:01 <dougall> hmm, not that springs to mind... maybe the instructions that implement fast sin/cos, or simd-group ops like 'call/ret' and 'simd_shuffle_down'

11:24 <bloom> dougall: not that I'm discouraging more r/e, but fast sin/cos is surprisingly uninteresting on GPUs

11:24 <bloom> (nowadays)

11:24 <bloom> asin/acos/atan even more so

11:29 <bloom> Mesa doesn't even ave asin/acos/atan acceleration wired up, it's lowered to a (iirc polynomial..?) approximation unconditionally :-P

11:56 <dougall> heh - yeah, iirc they just use a couple of (hopefully) uninteresting instructions i haven't named yet, so it might be a good beginner instruction to re

17:31 <DarkShadow44> dougall: Do you already have documentation prepared to device_store/device_load?

17:31 <DarkShadow44> Because for simplicity reason, I need those first to start with proper testing

17:31 <DarkShadow44> if you don't have anything, maybe I could write something up?

17:32 <DarkShadow44> on your github it's still maked TODO(), that's why I ask