PR (Peformance Regression) #2172

[kirbypuff]

PR #2172 causes a huge drop in performance (up to 60% slower compared to the latest dev/master build).

A quick bench:
nsmb @ 6xIR, EFB2RAM, D3D, BFS
---------------------------------------------
normal dev. build: 60 / 59 / 64 / 74 / 54 / 40 / 51 / 50 / 35
PR2172 test build: 43 / 46 / 49 / 54 / 42 / 25 / 40 / 37 / 28

nsmb @ 6xIR, EFB2RAM, OGL, BFS
---------------------------------------------
normal dev. build: 43 / 54 / 56 / 63 / 50 / 31 / 48 / 36 / 24
PR2172 test build: 37 / 51 / 52 / 59 / 47 / 29 / 45 / 31 / 22

Legend: title / cake / menu1 / menu2 / w2-overview / w-select / w1-overview / w1-1 / w2-4
(stable minimum framerates)

----------------------
TESTING NEEDED !
----------------------

Here's the test build:
https://dl.dolphin-emu.org/prs/pr-2172-dolphin-latest-x64.7z

Compare the performance of this test build with the normal build from here:
https://dl.dolphin-emu.org/builds/dolphin-master-4.0-5849-x64.7z

Please test with Intel IGPs and ATI/AMD GPUs (not just with NVIDIA)
Also with AMD CPUs and true Dual-Core CPUs (Pentium G, Core2 Duo, AMD X2, etc.)
More testing needed with different titles at native res (1xIR), FHD (3xIR) and high IR (6x [EFBScale = 9])

JMC47 Wrote:This build is actually faster for me than master on a quad-core CPU w/ NVIDIA GPU (less CPU usage).

???
Is that with the latest test build (Mar-13-2015, 11:46 PM, based on 4.0-5836)?

[JMC47]

Post on the pull request itself if you want degasus to see it. Seriously, if you do tons of testing to show something and then don't bother actually posting it where he'll see it, then this is pretty useless. Not all devs frequent the forums.

Anyway, I mentioned my computer BECAUSE I only have so many computers, and they're not very varied.

[degasus]

I've updated the PR again. Is this slowdown still there?

[dephined]

I tested the links above (post of 7 hours ago) with a Core2 e8400, Nvidia 650gtx and from what I can see the slowdown affects almost only D3D (10-15% slowdown for me), while OpenGL is almost not affected (only 1%)

[kirbypuff]

(03-14-2015, 07:42 PM)degasus Wrote: I've updated the PR again. Is this slowdown still there?

Sorry for the delayed reply (I was testing another PR for regressions).

There's no improvement wth the latest test build (Mar-14-2015, 6:09 PM).
Same issue even with the OpenGL backend (see first post for the updated nsmb bench results)

[kirbypuff]

The test was done in simulated dual-core mode (2 cores disabled).

Performance in normal (quad-core) mode is unpredictable / erratic: sometimes it's the same as master, sometimes it's slower, but most of the time it starts fast and then slows down after 10~30 seconds.

In dual-core mode it's stable (always slow).

[degasus]

To be honest, I have no clue how this should slow down that much any more. Wakeups are limited, flushes are rare, so what is left?
I begin to think this is maybe because of a syncing fix which is broken on master. So I've build a patch which also fixes this on master without everything else. Is this build also slow? http://dl.dolphin-emu.org/wips/degasus-dolphin-master-4.0-5843+-x64.7z

[kirbypuff]

(03-15-2015, 11:58 PM)degasus Wrote: Is this build also slow?

No slowdown issues with that build (almost the same performance as dev/master):

nsmb @ 6xIR, EFB2RAM, D3D, BFS
---------------------------------------------
4.0-5839-normal: 60 / 59 / 64 / 74 / 54 / 40 / 51 / 50 / 35
4.0-5843-degasus: 59 / 58 / 64 / 73 / 53 / 40 / 50 / 50 / 35

[degasus]

So, after rewriting the PR for the forth times, I've fixed a small performance issue in this PR. I don't think it did slow down that much, but it's likely worth to recheck.

Why is concurrency that hard?

[kirbypuff]

(03-16-2015, 03:09 AM)degasus Wrote: So, after rewriting the PR for the forth times, I've fixed a small performance issue in this PR. I don't think it did slow down that much, but it's likely worth to recheck.

4.0-5843-degasus was almost as fast as master, but the latest PR #2172 test build (Mar-15-2015, 5:28 PM) still has the performance regression.