A Deep Dive Into The Performance-Focused AMDGPU "Bulk Moves" Functionality

[phoronix]

Phoronix: A Deep Dive Into The Performance-Focused AMDGPU "Bulk Moves" Functionality

Recently on Phoronix you have likely heard a lot about the LRU "bulk moves" functionality for the AMDGPU driver after it was talked up by a Valve Linux developer for the performance help to Linux games and then the change landing in Linux 5.4 as a "fix"...

http://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-Bulk-Moves-Dive

[Ardje]

What I miss in all this GPU talk and optimizations, is the PCIe bandwidth and . Especially since I have an eGPU now, and my PCIe bandwidth is 1/32 of the cards full potential. I want to see the bandwidth usage measured. The AMDGPU driver does have a dirty hack since 5.2 that shows the amount of PCIe messages, by resetting the message counters, waiting a full second in kernel and then reading out those counters. I actually want to either have a log of the last 100 measurements in 10 ms or 100ms intervals.
The fun thing is: sometimes I am amazed at how much bandwidht something uses, and sometimes I am looking at major stutter while PCIe is not active at all (maybe shader compilation?).
Michael : what kind statistics would you measure, how do you want it to see? Maybe I just try and take a stab at that bw_pcie code and give you something that you can use in your tests?
Hacking 100 readouts per 10 ms in a second would be an easy patch, because it doesn't change the current flow, it only changes the amount of data returned. It must fit one page though, so for 100 readouts it's about 39 bytes per measurement interval. There are 2 code paths that deliver bw_pcie: one for GCN1.0 and one for > GCN1.2 I think.

[Azpegath]

No videos from this year's conference?

[skeevy420]

I'm glad it backports easily enough to 5.2 & 5.3 because 5.4-rc2 hasn't had a successful compile on my PC.

[perpetually high]

I'm glad it backports easily enough to 5.2 & 5.3 because 5.4-rc2 hasn't had a successful compile on my PC.

Are you using those schedutil patches still? That was breaking the build with 5.4-rc1/rc2 for me as well. Without them, compiles just fine.

[tg--]

No videos from this year's conference?

Only compound recordings of the whole day.
Bulk moving: https://youtu.be/HYa4UvVtMOE?t=24222

[coder]

Okay, I flipped through the slide deck, and still I'm wondering why the patch seemingly caused a performance regression on Radeon VII.

https://www.phoronix.com/scan.php?pa...lk-moves&num=1