WebAudio: Add SSE2 optimization for DirectConvolver

WebAudio: Add SSE2 optimization for DirectConvolver

Benchmark result from
http://rtoy.github.io/webaudio-hacks/tests/time-convolver-short.html,
running Linux on a Z620:

M36: 1680 ms
Opt: 1060 ms

Time is now reduced by about 33%.

For Windows running on a Z600:
M38: 3005 ms
Opt: 2280 ms

A savings of about 25%.

BUG=395276
TESTS=convolution-mono-mono still passes.

Committed: https://src.chromium.org/viewvc/blink?view=rev&revision=178576

[Raymond Toy, 2014-07-18 23:11:38]

PTAL

[Ken Russell (switch to Gerrit), 2014-07-19 01:02:25]

LGTM

https://codereview.chromium.org/408563003/diff/20001/Source/platform/audio/Di...
File Source/platform/audio/DirectConvolver.cpp (right):

https://codereview.chromium.org/408563003/diff/20001/Source/platform/audio/Di...
Source/platform/audio/DirectConvolver.cpp:43: #include <emmintrin.h>
This header's available on all non-Mac OSs?

https://codereview.chromium.org/408563003/diff/20001/Source/platform/audio/Di...
Source/platform/audio/DirectConvolver.cpp:107: // are multiples of 4. If not,
use the straighforward loop below.
typo: straightforward

https://codereview.chromium.org/408563003/diff/20001/Source/platform/audio/Di...
Source/platform/audio/DirectConvolver.cpp:110: // AudioFloatArray's are always
aligned on at least a 16-byte boundary.
Sure hope that's true. I'm not verifying that assertion in this review.

[Raymond Toy, 2014-07-21 17:09:57]

https://codereview.chromium.org/408563003/diff/20001/Source/platform/audio/Di...
File Source/platform/audio/DirectConvolver.cpp (right):

https://codereview.chromium.org/408563003/diff/20001/Source/platform/audio/Di...
Source/platform/audio/DirectConvolver.cpp:43: #include <emmintrin.h>
On 2014/07/19 01:02:24, Ken Russell wrote:
> This header's available on all non-Mac OSs?

Yes, I believe so. But we don't need to include this on OSX since we use OSX's
direct convolver instead.

https://codereview.chromium.org/408563003/diff/20001/Source/platform/audio/Di...
Source/platform/audio/DirectConvolver.cpp:110: // AudioFloatArray's are always
aligned on at least a 16-byte boundary.
On 2014/07/19 01:02:24, Ken Russell wrote:
> Sure hope that's true. I'm not verifying that assertion in this review.

For the record, it's in Source/platform/audio/AudioArray.h. We get 32-byte
alignment using FFMPEG or OpenMAX DL.  Otherwise it's 16 bytes.

[Raymond Toy, 2014-07-21 18:18:13]

The CQ bit was checked by rtoy@chromium.org

[commit-bot: I haz the power, 2014-07-21 18:19:19]

CQ is trying da patch. Follow status at
 https://chromium-status.appspot.com/cq/rtoy@chromium.org/408563003/40001

[commit-bot: I haz the power, 2014-07-21 18:41:11]

Change committed as 178576