Send Scitech mailing list submissions to
email@hidden
To subscribe or unsubscribe via the World Wide Web, visit
http://lists.apple.com/mailman/listinfo/scitech
or, via email, send a message with subject or body 'help' to
email@hidden
You can reach the person managing the list at
email@hidden
When replying, please edit your Subject line so it is more specific
than "Re: Contents of Scitech digest..."
Today's Topics:
1. SGEMV slow in Leopard? (Tim Schmielau)
2. Re: SGEMV slow in Leopard? (Serge Cohen)
----------------------------------------------------------------------
Message: 1
Date: Wed, 28 Nov 2007 19:32:14 +0000
From: Tim Schmielau <email@hidden>
Subject: [apple scitech] SGEMV slow in Leopard?
To: <email@hidden> <email@hidden>
<email@hidden>
Message-ID: <email@hidden>
Content-Type: text/plain; charset=US-ASCII; format=flowed
After upgrading my iBook G4 from 10.3.9 to 10.5.1 (and XCode from 1.5
to 3.0) I noticed that my application's performance, which depends
almost exclusively on cblas_sgemv, has decreased by about a factor of
10.
This is about the same factor that the benefit of using VecLib's
AltiVec enabled implementation was in the first time. In some (rare)
cases I had been able to regain the performance by switching to a
different SDK, but usually after a few compiles I'm back at the slow
speed.
Once compiled the application always runs at the same (fast or slow)
speed. I've checked that all arrays are fully aligned even in the slow
version.
My guess would be that sometimes a non-AltiVec enables BLAS library
gets compiled in - but where does it come from?
Any advice would be appreciated.
Tim
------------------------------
Message: 2
Date: Thu, 29 Nov 2007 11:49:15 +0100
From: Serge Cohen <email@hidden>
Subject: Re: [apple scitech] SGEMV slow in Leopard?
To: Tim Schmielau <email@hidden>
Cc: Apple Scitech Mailing List <email@hidden>
Message-ID: <email@hidden>
Content-Type: text/plain; charset=ISO-8859-1; delsp=yes; format=flowed
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
I had the same problem going from 10.3 to 10.4 ...
In this case the problem was due to Accelerate not using "flush
denormals" anymore, but providing an IEEE compatible treatment of
denormals (which basically means that computation is made by the
scalar fpu). In other words, Accelerate would always turn the VE's
java Mode to ON, and it would NOT turn the mode back to the original
one.
In my case, I could get back to close the previous performance by
just explicitly turning java mode off after each call to accelerate
(as the rest of my code was hitting denormals more often than the
BLAS calls).
Hope this helps;
Serge.
PS : https://developer.apple.com/samplecode/vDSPExamples/
listing2.html explains how to turn java mode ON/OFF
Also one can look at :
http://developer.apple.com/documentation/Performance/Conceptual/
Accelerate_sse_migration/migration_sse_translation/
chapter_4_section_2.html
In this latter one, the text mentions that by default G5 (I presume
also G4 ?) have the default of handling denormals to OFF. Indeed this
is a bit misleading in the sense that this default will be changed as
soon as you call BLAS/LAPACK from Accelerate (as much as I
experienced).
Le 28 nov. 07 à 20:32, Tim Schmielau a écrit :
After upgrading my iBook G4 from 10.3.9 to 10.5.1 (and XCode from
1.5 to 3.0) I noticed that my application's performance, which
depends almost exclusively on cblas_sgemv, has decreased by about a
factor of 10.
This is about the same factor that the benefit of using VecLib's
AltiVec enabled implementation was in the first time. In some
(rare) cases I had been able to regain the performance by switching
to a different SDK, but usually after a few compiles I'm back at
the slow speed.
Once compiled the application always runs at the same (fast or
slow) speed. I've checked that all arrays are fully aligned even in
the slow version.
My guess would be that sometimes a non-AltiVec enables BLAS library
gets compiled in - but where does it come from?
Any advice would be appreciated.
Tim
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Scitech mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/scitech/email@hidden
This email sent to email@hidden
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.3 (Darwin)
iD8DBQFHTpkj5EPeG5y7WPsRAp24AKCfrCPiP0hHYTgDsf8V4YpGVxtXcwCeNrrx
WdKAFDtGEWzP5hR86k4Jzxc=
=Ujbi
-----END PGP SIGNATURE-----
------------------------------
_______________________________________________
Scitech mailing list
email@hidden
http://lists.apple.com/mailman/listinfo/scitech
End of Scitech Digest, Vol 4, Issue 180
***************************************
