GitLab

Add SSE2 optimization for the main processing loop for the a-rate
Oscillator and for the computation in WaveDataForFundamentalFrequency.
Profiling showed WaveDataForFundamentalFrequency was taking 27% of the
total time.  The optimized version takes about half as long by computing
4 results at a time.

The failed test is caused by changing the linear interpolation formula
from (1-f)*x0 + f*x1 to the mathematically equivalent x0+f*(x1-x0).
However, this isn't exactly the same in floating-point.  The updated
formula has one less operation so it's a useful speed up.

Using a float virtual_read_index in the SIMD loop causes significant
loss of precision (50 dB or less in the osc sweep tests instead of
100+ dB), so we need to do a slightly more complex SIMD version with
double virtual_read_index.

WebAudio bench results without this CL:

TEST	μs	MIN	Q1	MEDIAN	Q3	MAX	MEAN	STDDEV
Oscillator.frequency-linear-a-rate	1884	1884	1948	1969	1997	2598	1987.72	94.13

With this CL:
TEST	μs	MIN	Q1	MEDIAN	Q3	MAX	MEAN	STDDEV
Oscillator.frequency-linear-a-rate	1348	1348	1397	1418	1453	2028	1474.74	151.59

We see about a 25% improvement in speed (based on the mean).

Bug: 1013118
Change-Id: I818834c64c3529b26467dbaa7623354f92b47b2a
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2264204Reviewed-by: Hongchan Choi <hongchan@chromium.org>
Reviewed-by: Dale Curtis <dalecurtis@chromium.org>
Commit-Queue: Raymond Toy <rtoy@chromium.org>
Cr-Commit-Position: refs/heads/master@{#789097}