[Webkit-unassigned] [Bug 98131] Add ARM-NEON support to VectorMath in WebAudio

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=98131





--- Comment #8 from Raymond Toy <rtoy at chromium.org>  2012-10-03 13:18:38 PST ---
(In reply to comment #6)
> (In reply to comment #4)
> > (From update of attachment 166684 [details] [details])
> > View in context: https://bugs.webkit.org/attachment.cgi?id=166684&action=review
> > 
> > Looks good.  Do you have any measurements to show what kind of improvements these instrinsics make?  I'm just curious.
> > 
> 
> Yes I have made independent tests on the functions.
> The vector size was 1000 with float numbers and I ran it 100000 times.
> The average results:
> 
> vsma:
> REF  elasped time is 3747 ms
> NEON elasped time is 3642 ms

Thank you very much for the measurements.

There wasn't much difference here.  Do you have any ideas on why there was very little improvement?

> 
> vsmul:
> REF  elasped time is 2654 ms
> NEON elasped time is 2375 ms
> 
> vadd:
> REF  elasped time is 3457 ms
> NEON elasped time is 2631 ms
> 
> vmul:
> REF  elasped time is 3528 ms
> NEON elasped time is 2635 ms
> 
> vmaxmgv:
> REF  elasped time is 5203 ms
> NEON elasped time is 2201 ms
> 
> vsvesq:
> REF  elasped time is 2234 ms
> NEON elasped time is 1840 ms
> 
> zvmul: (with vmla and vmls)
> REF  elasped time is 8778 ms
> NEON elasped time is 7963 ms
> 
> > > Source/WebCore/platform/audio/VectorMath.cpp:493
> > > +            float32x4_t imagResult = vaddq_f32(vmulq_f32(real1, imag2), vmulq_f32(imag1, real2));
> > 
> > Can the vmla and vmls instrinsics be used here to speed things up slightly?
> 
> Yes you're right. We can use it here I will change this.

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