#
#asahi-gpu · 2022-10-26 02:20
alyssawhy would the same program use so many fewer registers on AGX?
#asahi-gpu · 2022-10-26 02:20
alyssawhereas AGX is using 50/2 = 25 dword registers
#asahi-gpu · 2022-10-26 02:20
alyssait's not explicitly printed for Mali-G57, but "1 thread" means that it's using at least 33 dword registers
#asahi-gpu · 2022-10-26 02:20
alyssaSimilarly, register pressure
#asahi-gpu · 2022-10-26 02:18
alyssaAGX uses variable length instructions, in this shader instructions are an average of 6.65 bytes
#asahi-gpu · 2022-10-26 02:18
alyssaWhy? mali-g57 uses fixed-length instructions, all are 8 bytes, and many are sparse
#asahi-gpu · 2022-10-26 02:18
alyssathere are fewer of them (1249 vs 1294), but they're bigger
#asahi-gpu · 2022-10-26 02:17
alyssaThis is interesting, a Mali-G57 instruction can actually do *more* than an AGX instruction
#asahi-gpu · 2022-10-26 02:17
alyssa(quadwords = 16 bytes)
#asahi-gpu · 2022-10-26 02:17
alyssaMali is ~10,000 bytes
#asahi-gpu · 2022-10-26 02:17
alyssaAGX is 8600 bytes
#asahi-gpu · 2022-10-26 02:16
alyssabut here's an obvious one:
#asahi-gpu · 2022-10-26 02:16
alyssa(AGX will get similar stuff in time~)
#asahi-gpu · 2022-10-26 02:16
alyssaoff the bat you'll notice that the statistics for Mali are a lot more advanced, because we actually understand the uarch there
#asahi-gpu · 2022-10-26 02:16
alyssatop is Apple M1, bottom is Mali-G57
#asahi-gpu · 2022-10-26 02:16
alyssashaders/glmark/1-22.shader_test - MESA_SHADER_FRAGMENT shader: 1249 inst, 16.015625 cycles, 16.015625 fma, 3.453125 cvt, 0.000000 sfu, 0.125000 v, 0.000000 t, 0.000000 ls, 632 quadwords, 1 threads, 0 loops, 0:0 spills:fills
#asahi-gpu · 2022-10-26 02:15
alyssashaders/glmark/1-22.shader_test - MESA_SHADER_FRAGMENT shader: 1294 inst, 8600 bytes, 50 halfregs, 1 threads, 0 loops, 0:0 spills:fills
#asahi-gpu · 2022-10-26 02:15
alyssahere's a chunky shader from glmark2 -bterrain
#asahi-gpu · 2022-10-26 02:15
alyssaIt's really interesting comparing statistics for a given compiler shader *across* instruction sets
#asahi-gpu · 2022-10-25 12:52
svensmells like this could be voltage swing and pre-emphasis
#asahi-gpu · 2022-10-25 12:49
_jannau_drivers//gpu/drm/nouveau/nvkm/engine/disp/* might be intersting as well
#asahi-gpu · 2022-10-25 12:49
svenoh.. interesting, nice find
#asahi-gpu · 2022-10-25 12:47
_jannau_have you looked at drivers/gpu/drm/amd/display/dc/core/dc_link_dp*
#asahi-gpu · 2022-10-25 12:37
svenshot in the dark, but does "drive settings" in the context of displayport (possibly link training) ring a bell for anyone here? maybe also 0x300000003 as "max drive settings"?
#asahi-gpu · 2022-10-25 12:31
ella-0:3
#asahi-gpu · 2022-10-25 00:54
alyssaella-0: ^^
#asahi-gpu · 2022-10-24 22:05
Venemohey guys. congrats on getting vkcube working :)
#asahi-gpu · 2022-10-24 14:12
alyssa(hardware reverse-engineering? probably easier on macOS. shader compiler optimizations? easier on linux.)
#asahi-gpu · 2022-10-24 14:12
alyssawhen I have more time on my hooves, I'll set up Lina's kernel on the Linux side and switch OS depending on what specific task I'm working on
#asahi-gpu · 2022-10-24 14:11
alyssafor other types, it's more convenient to use Linux
#asahi-gpu · 2022-10-24 14:11
alyssafor certain types of driver work, it's more convenient to use macOS right now
#asahi-gpu · 2022-10-24 14:09
Dementoralyssa, thanks for the answer !
#asahi-gpu · 2022-10-24 14:08
alyssamacOS, but linux results should be similar
#asahi-gpu · 2022-10-24 13:58
Dementoralyssa, are you running your test on MacOS or Linux ?
#asahi-gpu · 2022-10-24 13:51
alyssait's possible on linux too there's just massive amounts of overhead
#asahi-gpu · 2022-10-24 13:51
alyssaI haven't merged that commit
#asahi-gpu · 2022-10-24 13:51
alyssaon macOS at least wiring up the BO cache doubled fps or something crazy like that on terrain
#asahi-gpu · 2022-10-24 09:39
lina(Default is 0x10 so 2MiB)
#asahi-gpu · 2022-10-24 09:38
linaThe units are 128K buffer blocks
#asahi-gpu · 2022-10-24 09:37
linaI do get some improvement on terrain at small framebuffer sizes, but nothing to write home about, so I think most of the slowness there is just shader stuff, not TVB overflows
#asahi-gpu · 2022-10-24 09:37
linaJust added an initial_tvb_size module param (can be changed at runtime, only affects new GPU clients)
#asahi-gpu · 2022-10-24 09:26
linajannau: I think the crash should be fixed (turns out that tvb_something buffer does need to be dynamically sized based on resolution...)
#asahi-gpu · 2022-10-24 09:23
linaalyssa: I'm also pretty sure terrain is overflowing my fixed size TVB, possibly a lot, and I don't have any grow logic in yet...
#asahi-gpu · 2022-10-24 00:32
alyssaapparently even with what I've hardcoded that's still enough for a pile of tests :-p
#asahi-gpu · 2022-10-24 00:32
alyssa Passed: 378/420 (90.0%)
#asahi-gpu · 2022-10-24 00:22
alyssawhich is kinda what matters right now anyway
#asahi-gpu · 2022-10-24 00:22
alyssaso.. code isn't ready but I have a proof of understanding the hardware
#asahi-gpu · 2022-10-24 00:22
alyssalot of stuff hardcoded though
#asahi-gpu · 2022-10-24 00:22
alyssadEQP-GLES3.functional.fragment_out.array.fixed.rgba8_highp_vec4
#asahi-gpu · 2022-10-24 00:22
alyssajust passed her first MRT test

plus ›