PR (Peformance Regression) #2172

[kirbypuff]

The latest PR test build (based on 4.0-5944) is up to 60% slower:

4.0-5944 master : 63 / 72 / 59 / 43
4.0-5944 PR test: 45 / 54 / 45 / 27

[degasus]

After lots of testing by David3k, he noticed a much lower VPS at very GPU bottlenecked scene with this PR. But the FPS wasn't affected, so did you test FPS or VPS?

I've fixed a race condition in our syncing code. It is possible that the CPU continues too early which leads to fps/vps divergence. Do you notice some strange fps/vps rates in this test cases on master? Through this race condition speeds up dolphin, our CPU emulation will get angry with a broken syncing.

I'm still surprised as nobody was able to confirm this huge slowdown. Did you use some uncommon settings like syncGpu or deterministic dual core?

[kirbypuff]

(04-03-2015, 08:25 PM)degasus Wrote: Did you test FPS or VPS?

Both.

(04-03-2015, 08:25 PM)degasus Wrote: Do you notice some strange fps/vps rates in this test cases on master?

No. FPS=VPS in all cases.

(04-03-2015, 08:25 PM)degasus Wrote: Did you use some uncommon settings?

I always use this config for testing (see attachment).
Win7 with Basic theme (Aero/DWM disabled). D3D backend. GPU in high-performance mode. Latest WHQL GPU drivers. Simulated Dual-Core CPU (2 cores turned off).


Another nsmb benchmark. This time in wndowed mode (render to main):

4.0-5944 master: 63 / 72 / 64 / 74 / 59 / 43 / 57 / 51 / 36
4.0-5944 PR test: 45 / 55 / 49 / 55 / 46 / 27 / 45 / 38 / 28

(6xIR, efb2ram, stable minmum FPS)

[mimimi]

kirbypuff can you please test pr builds against the master version it is based on? It always says which version it is based on in the title of the window, like 4.0-4944 dirty.

Do you have Mario Galaxy? I noticed a slowdown there. At the start of a new game, after the storybook section, there's a viewing platform overlooking the town square, lake and castle. What's your performance there, min/max?

[kirbypuff]

(04-04-2015, 08:35 PM)mimimi Wrote: kirbypuff can you please test pr builds against the master version it is based on? It always says which version it is based on in the title of the window.

The results are the same with build 4.0-5944 (no change).

[David3k]

kirbypuff, I have personally been unable to find any case of a of a significant regression on PR2172 vs 4.0-5952 with all my tests.

I also screwed up my explanation to degasus of what I was seeing:

In Master, I experience a peculiar problem where the VPS is constantly much higher than FPS, not the other way around. This goes mostly away when using EFB2TEX.

In PR2172, The VPS is always in line with the FPS, emulation is generally more stable, and when using EFB2TEX, significantly faster in a lot of cases. The is no notable regression in performance compared to master.

EDIT: correction

[degasus]

@kirbypuff: Sorry for not listening on your tests, but I wasn't able to get the same results at all. I've also tried NSMBW with efb2ram on my dual core, but this PR was as fast (slow) as master.

I hope the performance is fine for you if you'll use all 4 cores?

[kirbypuff]

Since this PR was merged to master, the performance has regressed to the point where it's even worse than 4.0.2 stable
An .ini option to disable this power-saving "feature" would be the best solution.

[JMC47]

People with similar hardware to you were unable to find a slowdown, I was unable to find a serious slowdown on my computer. Some games are faster, some games are 1 - 2% slower, thermal impact is much lower.

[mimimi]

Or in other words, find more games with issues that are easily reproducible.

Well, i found one situation with a slowdown:
Mario Galaxy, at start of the game on the viewing platform overlooking the quare, lake and castle. It was something like master: 100-110, pr: 85-95 (forgot the actual numbers).

I did not find any situation where it is faster, which is strange considering that i have a dual core, and this frees up a core, so the 2 other main threads(cpu thread and gpu driver thread) get more cpu time.