Bug 60396 – Add RAII-style DenormalDisabler class to disable denormals which hurt audio performance

• Component: New Bugs
• Status: RESOLVED FIXED

[Chris Rogers]

Add macros to disable denormals on x86 for audio-intensive algorithms

[Chris Rogers]

Created attachment 92627 [details]
Patch

[Alexey Proskuryakov]

Can this be a RIAA-style class instead of a pair of macros? This class could disable denormals in constructor, and enable them back in destructor, making sure that we don't accidentally leak the state by taking an unusual code path with an early return.

[Chris Rogers]

Sure, I can probably whip something up like that.  Just curious, what is RIAA?

[Eric Seidel (no email)]

http://en.wikipedia.org/wiki/Resource_Acquisition_Is_Initialization

[Chris Rogers]

Created attachment 92834 [details]
Patch

[Chris Rogers]

Latest patch turns this into an RIAA-style class.

[Eric Seidel (no email)]

Comment on attachment 92834 [details]
Patch

Looks nice.  Need someone who actually knows about this stuff to comment though.

[Eric Seidel (no email)]

CCing folks who might know something about denormal disabiling.

[Oliver Hunt]

In terms of actual data behaviour what does this do?  If I load a denormalised value into an xmm register what happens?

[Chris Rogers]

Oliver, I don't think a simple load will be any different.  Setting the denormal mask tells the processor to ignore any exceptions which occur due to any floating-point math which is dealing with denormals.  Not handling the exception causes a tiny amount of precision loss, which is acceptable for audio processing.  The performance benefit is very large because once these denormals get into a stateful audio processing algorithms they tend to stick around (in filter state, decaying delay lines, parameter slewing/de-glitching, etc.).  They can cause a very large number of exceptions to occur, causing enormous slowdown.

For all AudioUnits we used to disable them -- here's an old copy of an audio SDK file with this type of disabling (near the top of the file):
http://www.google.com/codesearch/p?hl=en#9WNLlqRP2nE/trunk/PublicUtility/AUPublic/AUBase/AUBase.cpp&q=AUPublic&sa=N&cd=2&ct=rc

[Kenneth Russell]

I think Alexey meant RAII, not RIAA. Very different concepts. :)

[Kenneth Russell]

Comment on attachment 92834 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=92834&action=review

Couple of minor comments; feel free to fix upon landing.

> Source/WebCore/ChangeLog:5
> +        Add RIAA-style DenormalDisabler class to disable denormals which hurt audio performance.

RIAA -> RAII

> Source/WebCore/platform/audio/DenormalDisabler.h:50
> +        asm volatile("stmxcsr %0" : "=m" (*&result));

It looks like the *& isn't needed here or on the ldmxcsr; see http://gcc.gnu.org/bugzilla/attachment.cgi?id=11062 .

[Chris Rogers]

Committed r86099: <http://trac.webkit.org/changeset/86099>

[Sam Weinig]

(In reply to comment #10)
> Oliver, I don't think a simple load will be any different.  Setting the denormal mask tells the processor to ignore any exceptions which occur due to any floating-point math which is dealing with denormals.  Not handling the exception causes a tiny amount of precision loss, which is acceptable for audio processing.  The performance benefit is very large because once these denormals get into a stateful audio processing algorithms they tend to stick around (in filter state, decaying delay lines, parameter slewing/de-glitching, etc.).  They can cause a very large number of exceptions to occur, causing enormous slowdown.
> 
> For all AudioUnits we used to disable them -- here's an old copy of an audio SDK file with this type of disabling (near the top of the file):
> http://www.google.com/codesearch/p?hl=en#9WNLlqRP2nE/trunk/PublicUtility/AUPublic/AUBase/AUBase.cpp&q=AUPublic&sa=N&cd=2&ct=rc

Can JavaScript run in the scope this is activated for?  If so, can the change in precision show up there?

[Chris Rogers]

Hi Sam, no this will only run in audio-specific threads.  It's per-thread state and will not affect the main JS thread or any worker threads.