[Webkit-unassigned] [Bug 77950] SSE optimization for vsvesq and vmaxmgv

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=77950





--- Comment #9 from Raymond Toy <rtoy at chromium.org>  2012-02-23 09:25:45 PST ---
(From update of attachment 128369)
View in context: https://bugs.webkit.org/attachment.cgi?id=128369&action=review

Looks good overall, except for the comment at line 508.

> Source/WebCore/ChangeLog:8
> +        Achieved the performance of 3.7x on vsvesq and 4.1x on vmaxmgv.

You mentioned that the udpated version of vmaxmgv is now slower(?) than before.  Can you update the numbers here?

> Source/WebCore/platform/audio/VectorMath.cpp:500
> +            source = _mm_and_ps(source, mMask);

You mentioned that the performance with and_ps is less than the previous version with min_ps.  Is this because and_ps is slower or maybe because of some pipelining isues?  I'm just curious as to why this is slower.

> Source/WebCore/platform/audio/VectorMath.cpp:508
> +        max = std::max(groupMaxP[2], groupMaxP[3]);

The assignment to max here at line 508 overwrites the max computed in line 485 when the source is not aligned.  This needs to be rearranged a little so as not to destroy that.

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.