Posts: 3,228
Threads: 25
Joined: Nov 2012
Reputation:
98
I thought the same thing, but then Henrik pointed out that x64 has more registers. This means the jit performs better, although not significantly. From my tests, maybe ~2% iirc.
That said, I think the x64 version compiles faster (at least with MSVC) which is a good enough reason to use it for development.
-[Unknown]
Posts: 953
Threads: 16
Joined: Feb 2013
Reputation:
15
This means that in some cases it's ~3,6 fps increase over the 32bit builds.
From now on I'll build my own builds I guess!
I'd provide 64bit autobuilds if I were you! :p
Those fps are enough to make a nealy perfect speed game go perfect with zero effort! :p
Posts: 3,228
Threads: 25
Joined: Nov 2012
Reputation:
98
Most games on Windows already run great. And if they don't, it's probably because your older processor isn't 64-bit.
I'm sure it varies, that's just from memory on I think it was Hexyz Force. I didn't check many games. It was something like 782 vps -> 800 vps with 64-bit when fast forwarding. No change, of course, without fast forward.
-[Unknown]
Posts: 953
Threads: 16
Joined: Feb 2013
Reputation:
15
03-29-2013, 01:35 AM
(This post was last modified: 03-29-2013, 01:36 AM by VIRGIN KLM.)
There are still games that struggle though to get 60vps even on a 64bit, I'm talking about them. It would make the issue a bit more easily fixable.
Posts: 3,228
Threads: 25
Joined: Nov 2012
Reputation:
98
Any example games?
Note that for example God of War is rendering 8x times, so it's spending (afaiu) almost all its time in the GPU. Jit being faster won't even affect that, so it'll still be slow.
-[Unknown]
Posts: 18
Threads: 3
Joined: Apr 2013
Reputation:
0
any news on (maybe even automated) x64 builds?
Posts: 42
Threads: 1
Joined: Nov 2012
Reputation:
5
I could easily build 64 bits build, but as Unknown said, the gain is very minimal. It's very unlikely that a game is limited by the CPU, this isn't a very powerful CPU we are emulating after all.
If using a faster CPU helped a game, that's more likely there's a problem elsewhere and there are better ways to fix it. Since some devs are focusing on mobiles, these fix will probably be made for these platforms and then help the desktop version to be even faster.
Posts: 953
Threads: 16
Joined: Feb 2013
Reputation:
15
The point now with x64 builds is not speed anymore but apparently x64 and x86 builds are not transparent and don't have the same compatibility and won't return the same results while debugging despite what was known/thought until today and the only way to correct that is testing both x86 and x64 builds incase you have a feeling something is worng and affects only one build.
Posts: 3,228
Threads: 25
Joined: Nov 2012
Reputation:
98
Well, there's some MSVC compiler bug that only happens when building 32-bit. It's kinda annoying, and seems to affect the VFPU. The solution I have so far is converting that to use jit, which we're gonna do anyway... although that means the interpreter will stay broken.
-[Unknown]
Posts: 42
Threads: 1
Joined: Nov 2012
Reputation:
5
If it's a bad optimization, we can always disable it for this function and compiler using some #pragma
Posts: 3,228
Threads: 25
Joined: Nov 2012
Reputation:
98
It happens even in debug mode on win32, and even with optimizations totally off everywhere in release or debug. Already tried that...
I even narrowed down one of them (the LittleBigPlanet video one) to Int_Vmmov iirc, but couldn't find a way of rewriting it or what it called to make the problem go away (but it did go away with my jit version of that instruction.)
-[Unknown]
