RE: [Bug-gnubg] Vectorizing 3rd step

[macherius]

Oystein,

i just wanna give you a "ping", I was busy and had not much time looking
into it. I'll do so now, probably adding code for intel/msvc. Thanks!

Ingo


> -----Original Message-----
> From: address@hidden 
> [mailto:address@hidden On Behalf 
> Of Øystein Johansen
> Sent: Tuesday, April 19, 2005 12:01 AM
> To: address@hidden
> Subject: Re: [Bug-gnubg] Vectorizing 3rd step
> 
> 
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Øystein Johansen wrote:
> | Here's a patch for i386 / GCC vectorizing of the inner loops of 
> | Evaluate(). I see some improvement, but I believe this can 
> be improved 
> | even further.
> |
> | Some comments: I believe having a integer counter in the 
> loop slows it 
> | down. Can I exit the loop in an other way?
> |
> | I initialize a vector for scaling in the second loop. I 
> believe this 
> | can be made simpler. Any suggestions?
> |
> | Please comment on this two issues.
> 
> Today I've aligned the arrays in memory. That gained some 
> speed as well. I still believe it can be improved a bit 
> further. My GCC 3.4.2 builds gives me now 25200 eval/sec on 
> this computer. I used to have about 20500 eval/sec. I'm still 
> missing a bit 30000 I get with Ingos ICC build, but I believe 
> the way to go is with intrinsics.
> 
> Aligned patch attached. There is also a mm_malloc.h file you 
> need if you're not using a gcc snapshot.
> 
> - -Øystein
> 
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.2.4 (MingW32)
> Comment: Using GnuPG with Thunderbird - http://enigmail.mozdev.org
> 
> iD8DBQFCZC4G6kDTFPhwyqYRAhI8AKCLmdmXX3gJivdbutrmcJejaGaVZwCeKEif
> rfRCRCxPuE0EAyuuZLQ1NhU=
> =fNcM
> -----END PGP SIGNATURE-----
> 
>