-funroll-loops resulting in 3-5% performance increase?

Discussion of development and patch submission.
Post Reply
UPerschke
Posts: 47
Joined: Wed 25 May, 2016 7:43 pm

-funroll-loops resulting in 3-5% performance increase?

Post by UPerschke » Sun 23 Oct, 2016 3:48 pm

Seems to be the case at least in built from recent commit, seen in various benchmarks and fraps.
Can anyone confirm? GCC is 4.7.2.

User avatar
leilei
Posts: 697
Joined: Fri 25 Apr, 2014 4:47 pm

Re: -funroll-loops resulting in 3-5% performance increase?

Post by leilei » Sat 29 Oct, 2016 12:29 am

For making a quick build to try I've found that -flto helps a lot (link-time optimization). Not as good as profile generated optimization but better than nothing.

From my experience, -funroll-loops brings a lot of cache hiccups as it causes some functions to really bloat up in size

startmenu
Posts: 104
Joined: Sat 29 Nov, 2014 7:39 am

Re: -funroll-loops resulting in 3-5% performance increase?

Post by startmenu » Sat 29 Oct, 2016 9:43 am

PGO cannot optimize everything. voodoo.o is broken when PGO used.

User avatar
leilei
Posts: 697
Joined: Fri 25 Apr, 2014 4:47 pm

Re: -funroll-loops resulting in 3-5% performance increase?

Post by leilei » Sat 29 Oct, 2016 9:44 am

Then you're doing it wrong. You should profile it with 1 render thread

User avatar
SarahWalker
Site Admin
Posts: 1730
Joined: Thu 24 Apr, 2014 4:18 pm

Re: -funroll-loops resulting in 3-5% performance increase?

Post by SarahWalker » Sat 29 Oct, 2016 5:17 pm

Or build with '-fprofile-use -fprofile-correction'

Post Reply