Only use custom SSE FMUL and FMAC with non-clang compilers.

Only use custom SSE FMUL and FMAC with non-clang compilers.

clang's auto-vectorized C version performs better according to the
Chrome Performance Dashboard.  Searching back through the logs, this
occurred when we switched over to clang by default.

We could try to microoptimize further, but it's less of a maintenance
burden to just let the compiler do its thing!

The main reason the clang version is faster is it does 2x 128bit
operations per loop. Simply copying these optimization yields ~97%
similar performance, but  the SIMD code a bit gnarlier. As such I
choose to simply use the C variant when clang is present.

BUG=none
TEST=none

Committed: https://crrev.com/657073100664172d14d61e92e1e4c36fc7024e5d
Cr-Commit-Position: refs/heads/master@{#297268}

[DaleCurtis, 2014-09-23 22:37:24]

dalecurtis@chromium.org changed reviewers:
+ rileya@chromium.org

[DaleCurtis, 2014-09-23 22:37:25]

Found this while reviewing your perf packet :) 

https://chromeperf.appspot.com/report?masters=ChromiumPerf&bots=chromium-rel-...

[rileya (GONE FROM CHROMIUM), 2014-09-23 22:46:53]

On 2014/09/23 22:37:25, DaleCurtis wrote:
> Found this while reviewing your perf packet :) 
> 
>
https://chromeperf.appspot.com/report?masters=ChromiumPerf&bots=chromium-rel-...

LGTM! I'm always happy to see less asm use in media/ :)

[DaleCurtis, 2014-09-23 22:54:07]

Looks like this benefit came when we switched to clang by default! +thakis to
confirm we're using that for official builds too.

[Nico, 2014-09-23 22:57:56]

On 2014/09/23 22:54:07, DaleCurtis wrote:
> Looks like this benefit came when we switched to clang by default! +thakis to
> confirm we're using that for official builds too.

I'm not sure what to look for in the perf graph, but starting from m38 clang is
used by default (including official) for chrome/linux. It's not used for
chrome/chromeos or chrome/android yet.

[DaleCurtis, 2014-09-23 23:05:16]

Hmm, sadly we don't have media_perftests benchmarks for ChromeOS. I'll test
locally with a CrOS chroot and see what it looks like with gcc 4.8+

I didn't think about an X86 Android version, but I suppose that's a valid combo.
I'll take a look before submitting this.

thakis: The graph shows "unoptimized" outperforming the "optimized" test cases;
which is C versus intrinsics. The root cause is the autovectorized version does
2x MM128 FMUL/FMAC operations per loop cycle.

[DaleCurtis, 2014-09-29 20:11:50]

The SSE versions are faster when GCC is used, so I've switched the #if over to
check __clang__ rather than OS_WIN.

[DaleCurtis, 2014-09-29 20:12:59]

The CQ bit was checked by dalecurtis@chromium.org

[commit-bot: I haz the power, 2014-09-29 20:14:22]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/599693002/20001

[commit-bot: I haz the power, 2014-09-29 21:17:47]

Committed patchset #2 (id:20001) as 0dc9320185be31bceb0c48cda534c1c4cb12e2af

[commit-bot: I haz the power, 2014-09-29 21:19:32]

Patchset 2 (id:??) landed as
https://crrev.com/657073100664172d14d61e92e1e4c36fc7024e5d
Cr-Commit-Position: refs/heads/master@{#297268}