Re: [Bug-gnubg] Benchmarks, experiments, speedups

[Ingo Macherius]

Øystein Johansen <address@hidden> schrieb am 15.03.05 19:35:12:

> | one more on gcc. I've just compiled myself a gcc 4.0.0 from the
> | snapshot sources as of 20050220. There is a new "-ftree-vectorize"
> | method that mimics Intel's vectorizer very closely. Attached is code
> | produced by gcc 4.0.0 snapshot, you'll see the BIG difference to
> | 4.3.x code.

> I got a working snapshot of gcc-4.1 today. was tweaking the optimization
> parameters a bit and the ftree-vectorizer option gave me about 5%
> performance improvment. I still don't see the speed from your Inten build.

Hmmm. The ftree-vectorize is indeed the key flag and should boost performance.
See: http://gcc.gnu.org/projects/tree-ssa/vectorization.html
You should verify the workings of the vectorizer by running the verbose mode,
-ftree-vectorizer-verbose=5

NOTE: Both Intel and GCC require 16-Byte aligned data structures for maximum
speedup using the SIMD extensions. Moving the current parameters of NN methods
into a typedef and aligning this using a pragma AND using a malloc that 
allocates 
at 16-Byte aligned adresses is the single most performance booster out there.

It will even improve ICC times, as it will reduce the length of the pre-loop 
(which
is used to align arbitrary structures) to zero (as the data just IS aligned). 

> | So intrinsics would be a quick solution for speed improvement, but
> | maybe the GTK 2.6 update is more pressing?

> I don't feel anything is pressing.... ;-)

> "intrinsics" is my vote, but I don't feel I have the knowledge to
> implement that. Anyone?

I was traveling a lot the last weeks, but I'll take a look into it next week. I 
just can't believe 
gcc 4.1 does only a 5% speedup, given all the effort they spent into their new 
tree-SSA
optimization. But if this is so, intrinsics are really the way to go.

Ingo

______________________________________________________________
Verschicken Sie romantische, coole und witzige Bilder per SMS!
Jetzt bei WEB.DE FreeMail: http://f.web.de/?mc=021193