00:00 <HdkR> agl7_: It's because it would be the same file. a690 is effectively just a bigger a660 so the same file works for both.

00:01 <mayco> hey, on the x13s, I'm having an audio issue with wired headphones where it seems like gain is being throttled if audio is too loud for more than a fraction of a second. it's a little nauseating to listen to. anyone know about this?

00:06 <agl7_> HdkR: Which is a nice thing: Cinnamon now doesn't say "Software rendering" with the a690_gmu.bin, so now it's the graphics hardware's turn.

00:14 <agl7_> The load is much lower!

02:03 <robclark> yeah, the sqe fw is generally the same within the "sub-generation" of a6xx.. but named after the first device of that sub-generation.. a660 and a690 (and others) are all subgen 4 so they all use a660_sqe.fw

06:11 <steev> quick hide everyone, johan is back

06:16 <HdkR> \o

06:48 <jhovold> :)

08:54 <HdkR> `Configured Memory Speed: 1555 MT/s` Memory reclocking on the Lenovo X13s not working?

08:55 <HdkR> `Speed: 2092 MT/s` Roughly correct speed

08:55 <HdkR> But configured incorrectly even under load?

08:56 <HdkR> Might explain my memory benchmark not getting expected numbers

08:59 <HdkR> Any way to force it to maximum clocks?

09:40 <HdkR> That looks to be around 50GB/s of its theoretical peak of 68GB/s. Which my memset benchmark is able to pull ~44GB/s out of

09:45 <Kelsar> After reading here for a while, I decided to stay on x86 for atleast one more generation ;)

15:16 <clover[m]> Blasphemy

17:11 <HdkR> bamse: Anything about the memory reclocking not working on X13s right now?

21:13 <bamse> HdkR: how much not working?

21:14 <HdkR> bamse: 1555MT/s instead of the expected maximum of 2092 MT/s, which is about 75%

21:16 <HdkR> According to dmidecode anyway

21:25 <bamse> HdkR: i've seen that number as well...

21:26 <bamse> HdkR: i apparently had a relevant pull request in my inbox...

21:27 <HdkR> :O

21:27 <bamse> HdkR: https://github.com/andersson/debugcc/commits/master (check the top one)

21:28 <HdkR> Let me give it a try

21:29 <HdkR> wait, wtf is debugcc

21:36 <HdkR> Oh, it's a query tools

21:38 <bamse> HdkR: sudo ./debugcc -p sc8280xp measure_only_mccc_clk

21:39 <HdkR> Neat, 2096Mhz plus change

21:39 <bamse> under load?

21:39 <HdkR> Under even not load

21:40 <bamse> but when you let it idle it goes back to 1018MHz, right?

21:41 <HdkR> nope

21:41 <bamse> hmm, it should

21:42 <HdkR> Well that's probably a different issue I guess :)

21:43 <bamse> HdkR: do ou have two devices in /sys/bus/platform/drivers/qcom-bwmon ?

21:43 <HdkR> I don't even have a qcom-bwmon

21:44 <bamse> then that's the reason why you don't see that number dance around

21:44 <bamse> the MT

21:44 <bamse> /s

21:44 <bamse> but that said, when i'm running my memory benchmark i get 2096 as well...

21:45 <HdkR> That's good, so just dmidecode not getting new information

21:45 <bamse> checking...

21:46 <bamse> no, it doesn't move as the ddr frequency changs

21:46 <bamse> changes*

21:47 <HdkR> Quirky but okay

21:48 <bamse> those layers are not involved in the frequncy change, so they would need to go the extra mile to update the data i think

21:48 <bamse> more than the extra mile...

22:10 <HdkR> bamse: Do you know if there is a way to get BW usage by client on the bus?

22:10 <HdkR> Curious about querying BW usage of the CPU clusters and GPU

22:19 <clover[m]> what do you guys do for chatgpt prompts? right now im using Bing Chat in the chromium browser

22:19 <HdkR> You say that like I would waste my time with an AI network

22:21 <steev> i don't even know what the question means

22:24 <clover[m]> are there any cool open linux software you use or know of for AI? preferably something not connected to the big tech companies or censored or anything

22:24 <HdkR> AMD's FSR I guess. It's better than most TAA implementations :P

22:25 <HdkR> Wwait, is that one AI yet or just a modified Lancos filter with temporal information? Maybe the 3.0 release was going to be "AI" based.

22:25 <HdkR> Lanczos even

22:33 <bamse> HdkR: we have two bwmon instances in the upstream dts, one measures the traffic in/out of the cpu subsystem, the other measures the traffic between llcc and ddr...

22:35 <HdkR> So just missing the GPU cluster which would be the most interesting bit

22:44 <robclark> currently we just tie the gpu's icc vote to opp, so gpu running at higher freq, we ask for more bw

22:48 <HdkR> Very reasonable for just ensuring the GPU has all the bandwidth available to it

22:49 <HdkR> I used to just keep tegrastats open when porting games to the SHIELD TV which had super nice GPU BW stats. If a scene starting crossing 50% memory BW of the device, it meant that I need to start looking at optimizing that

22:50 <robclark> we at least _used_ to do that for CPU as well, pre bwmon

23:00 <HdkR> Sounds like it's potentially viable. Just needs to be wired up somewhere

23:03 <steev> do it

23:03 <HdkR> As a testament to this statistic being useful. In one game we found a scene that wasn't 60fps. GPU load by itself was something like 90FPS, CPU Load by itself was like 120 "FPS", Together? like 50FPS. Tegrastats with its total memory BW utilized statistic pointed immediately towards a bottleneck since it was at like 80% utilization

23:04 <HdkR> Optimized the GPU to stop overdrawing so hard and it was fixed :D

23:04 <HdkR> Also optimized the crap out of a vertex wave generator, which didn't show any uplift in perf because memory BW bounded

23:06 <HdkR> steev: But that would require me to do kernel development when I'm already swamped with stuff that I know how to do. Rather than jumping in to the empty pool of kernel development

23:07 <steev> didn't stop me from shoving out the bluetooth driver!

23:08 <HdkR> steev: Sounds like you're willing to do kernel work :)

23:08 <steev> i don't even have a clue what you're talking about

23:09 <steev> maybe you could convince abelvesa to do it

23:09 <clover[m]> just ask chatgpt to do it

23:11 <HdkR> abelvesa: Can I get a way to get live BW consumption numbers pumped to userspace somehow? Total bus usage numbers, CPU, and GPU clients broken out. GB/s sort of thing :)

23:11 <HdkR> AMBA/AXI obviously has this information somewhere since Tegra X1 had the stats

23:12 <steev> note: i only picked on him because he did some bwmon stuff before

23:12 <HdkR> I'll poke at every thread possible