00:29 <steev> Seems like there’s a Lenovo system hardware update - 9/15/2023 in windows. Not sure what firmware changes there might be

01:46 <HdkR> https://store.avantek.co.uk/nv-grace-hopper-2u2n.html Just a casual $120k for Neoverse-V2

01:55 <clover[m]> Costs an ARM and a leg!

01:56 <HdkR> Guess I'm going to be waiting for AmpereOne eMag replacements

02:05 <robclark> HdkR: hmm, perf-record of the apitrace replay points at _mesa_LinkProgram() ... but not sure yet whether that just means the apitrace didn't cover the right sequence of gl calls or if something is different (less efficient?) about x86 emu vs native replay

02:05 <HdkR> Well this was aarch64 native apitrace capture

02:06 <HdkR> Since I had thunking enabled

02:06 <robclark> (basically I'm trying to figure out the callstack that leads to those hotspots so I can figure out why you are hitting that path)

02:06 <HdkR> Did it need to sit in the menu a bit longer?

02:07 <robclark> hmm, or just get me a `perf record -g` call stack

02:07 <robclark> I guess if you are hitting that path w/ thunking it isn't an emu issue

02:07 <HdkR> yea, that's all CPU time outside of the emulator

02:08 <robclark> there are several paths that lead to _mesa_apply_rgba_transfer_ops.. having a call stack would help track down why you are hitting that

02:08 <HdkR> Let's see if I can get one

02:08 <robclark> thx

02:17 <HdkR> robclark: https://drive.google.com/file/d/1Cnppq1_otSIRW9NrcyttLTm8QTtXWI4h/view?usp=sharing

02:18 <robclark> hmm, I probably need a copy/paste of the perf-report, since I don't think I have the same debug syms as you

02:21 <HdkR> https://drive.google.com/file/d/10CDr22szCio301LsvqxwYN-FAcVO7ddU/view?usp=sharing Should help

02:24 <robclark> ok.. so some sort of readpix going on.. I guess to get gpu rendered content back into browser(ish) thing? I'll see if I can figure out if there is some slow path we are hitting there although it seems like it is largely fault handling

02:25 <HdkR> I can only guess. CEF does as it do

02:27 <robclark> fault handling is going to suck no matter how you slice it, ie. the issue isn't what is accessing memory from cpu but that something is.. but that shouldn't be anything that is x86 vs arm related, I wouldn't expect

02:27 <robclark> what is CEF?

02:27 <HdkR> Chrome Embedded Framework

02:28 <robclark> hmm, ok

02:28 <robclark> looks like _mesa_GetTexSubImage_sw is a thing to look at closer

02:30 <robclark> I guess a dumb-thing-that-apps-can-do-that-we-haven't-optimized-yet

02:30 <HdkR> It's always spooky when the CPU time isn't dominated by the emulation :)

14:31 <robclark> HdkR: btw, did you look at `apitrace dump csgo.trace`? Not the most modern of gl usage..

14:33 <robclark> looks like it is just doing a boatload of texture upload

15:00 <robclark> HdkR: try this.. not sure if there is any more to it, but seems to help:

15:00 <robclark> https://www.irccloud.com/pastebin/OtERLfAx/

15:03 <Jasper[m]> <robclark> "HdkR: btw, did you look at `..." <- Source engine is from 2004, the game runs in dx9 mode on Windows :^)

15:04 <Jasper[m]> Not that it'll do for long though, csgo is going to be deprecated this year

15:04 <steev> oh no :(

15:08 <Jasper[m]> steev: eh, CS2 will do dx11 or Vulkan afaik

15:10 <steev> oh

17:10 <HdkR> robclark: Nice, I'll give it a try today

17:16 <robclark> perf says now it is bottlenecked on apitrace replay (decompression mostly).. but not sure how much faster that will make it in practice

17:18 <Jasper[m]> <Jasper[m]> "eh, CS2 will do dx11 or Vulkan..." <- (If I get early access I'll try it out, should have CEF aswell since that UI got backported)

17:47 <robclark> HdkR: hmm, seems like I should be somehow able to use FEXBash + chroot to get an env where I can build x86 mesa.. but they don't seem to play nicely together

17:48 <HdkR> The `unbreak_chroot.sh` script in the rootfs has some onf the finer details for how chrooting works

17:49 <HdkR> robclark: Patch takes CS:GO menu from sub 1FPS to 112FPS

17:49 <robclark> \o/

17:50 <HdkR> highest bits are now some atomic fetch_add and fd6_sampler_view_invalidate. atomic at 1.39%, sampler_view at 0.73%

17:51 <HdkR> Smattering split between JIT and msm_dri

17:51 <robclark> that isn't necessarily completely unreasonable.. what is gpu vs cpu load?

17:52 <HdkR> Primary thread at 65% usage, GPu at 60% usage

17:52 <HdkR> I guess a ping of ping-ponging CPU waiting on GPU

17:53 <robclark> yeah

17:53 <HdkR> Would need gpuvis to see where the stalls are

17:54 <HdkR> Not every day you can get single line changes that >100x the perf :)

17:55 <robclark> heheh

18:00 <HdkR> Looks like the hottest block is a < 32-byte memset. Rude

18:01 <robclark> `perf record -g` and `perf report -G` is useful to find where that is coming from

18:02 <HdkR> Well that one is in JIT code. Bit harder to trace back

18:02 <robclark> ahh

18:04 <HdkR> Sadly we don't have instructions yet that have ARM's new memset/memcpy instructions

18:04 <HdkR> Would have changed this three instruction loop in to a three instruction not-loop

18:05 <robclark> does a78c/x1c support this?

18:06 <HdkR> Nope. Nothing supports it yet, it's super fresh

18:07 <HdkR> Cortex-X1C/A78C memory subsystem is actually slightly worse off since it's an in-between core that glued on ARMv8.4 atomicity requirements

18:07 <robclark> ahh, if hw I have doesn't support it, it doesn't exist :-P

18:07 <HdkR> :)

18:07 <HdkR> Maybe with whatever comes after Cortex-A720 we'll have it

18:17 <HdkR> Oh interesting, the sampler view invalidate increases periodically

18:20 <robclark> sampler view invalidates might just be coming from the constant texture uploads?

18:21 <HdkR> Could be the case

19:02 <HdkR> Every time I see 4% CPU time in memcpy I think there is something wrong with memory clocking. But it could just be that set_constant_buffer is getting hammered by csgo

19:13 <robclark> pastebin decoded `perf report -G`?

19:18 <HdkR> robclark: https://drive.google.com/file/d/11yjSKky8ECZ8DCu5A5ODB8HEZxAYSjZU/view?usp=sharing Not sure how much it'll show

19:29 <HdkR> The good thing is that 75% of the CPU time is in the emulation now

20:27 <robclark> yeah, that doesn't look too bad.. I'd have to look at the man page, but you should be able to sort by dso to see the total amount of time spent in msm_dri.so

20:28 <robclark> but looks like most of it is in csgo_linux64

20:31 <robclark> sampler_view_invalidate looks like all sampler_view_destroy path so not hitting a lot of demote paths (ie. with some sampler view formats if the UBWC and/or tiled layout is different from the underlying texture format we need to demote to tiled-non-UBWC or linear, but doesn't look like you are hitting that sampler_view_invalidate() path)

21:21 <steev> clover[m]: have you seen [ 101.273817] MultiMedia1 Playback: ASoC: no backend DAIs enabled for MultiMedia1 Playback, possibly missing ALSA mixer-based routing or UCM profile ? someone asked me about it and i've never seen it