Add sk_float_rsqrt with SSE + NEON fast paths.

Add sk_float_rsqrt with SSE + NEON fast paths.

Current numbers:

N4:
running bench [640 480]      math_fastIsqrt    NONRENDERING: cmsecs =      3.12
running bench [640 480]      math_slowIsqrt    NONRENDERING: cmsecs =      4.82
running bench [640 480] math_sk_float_rsqrt    NONRENDERING: cmsecs =      1.99

Desktop:
running bench [640 480]      math_fastIsqrt    NONRENDERING: cmsecs =      0.89
running bench [640 480]      math_slowIsqrt    NONRENDERING: cmsecs =      0.94
running bench [640 480] math_sk_float_rsqrt    NONRENDERING: cmsecs =      0.09

Haven't found any other benches where this is a significant effect yet.

BUG=

Committed: http://code.google.com/p/skia/source/detail?r=12203

[mtklein, 2013-11-07 22:24:48]

for your early perusal.  NEON tomorrow.

[reed1, 2013-11-07 22:28:44]

https://codereview.chromium.org/60083014/diff/150001/include/core/SkFloatingP...
File include/core/SkFloatingPoint.h (right):

https://codereview.chromium.org/60083014/diff/150001/include/core/SkFloatingP...
include/core/SkFloatingPoint.h:106: inline float sk_float_rsqrt(float x) {
static ...

https://codereview.chromium.org/60083014/diff/150001/src/core/SkPoint.cpp
File src/core/SkPoint.cpp (right):

https://codereview.chromium.org/60083014/diff/150001/src/core/SkPoint.cpp#new...
src/core/SkPoint.cpp:166: scale = length * sk_float_rsqrt(mag2);
not prepared to land this, until we know more about the assumption of *all* of
our callers.

consider adding setFastLength() for now?

[mtklein, 2013-11-08 16:47:44]

On 2013/11/07 22:28:44, reed1 wrote:
>
https://codereview.chromium.org/60083014/diff/150001/include/core/SkFloatingP...
> File include/core/SkFloatingPoint.h (right):
> 
>
https://codereview.chromium.org/60083014/diff/150001/include/core/SkFloatingP...
> include/core/SkFloatingPoint.h:106: inline float sk_float_rsqrt(float x) {
> static ...
> 
> https://codereview.chromium.org/60083014/diff/150001/src/core/SkPoint.cpp
> File src/core/SkPoint.cpp (right):
> 
>
https://codereview.chromium.org/60083014/diff/150001/src/core/SkPoint.cpp#new...
> src/core/SkPoint.cpp:166: scale = length * sk_float_rsqrt(mag2);
> not prepared to land this, until we know more about the assumption of *all* of
> our callers.
> 
> consider adding setFastLength() for now?

PTAL?

[reed1, 2013-11-08 20:02:30]

lgtm

[commit-bot: I haz the power, 2013-11-08 20:05:44]

CQ is trying da patch. Follow status at
https://skia-tree-status.appspot.com/cq/mtklein@google.com/60083014/220001

[commit-bot: I haz the power, 2013-11-08 20:14:33]

Change committed as 12203