Remove float/double conversions from SkMatrix44's map2_pf

Remove float/double conversions from SkMatrix44's map2_pf

BUG=chromium:303811
Committed: http://code.google.com/p/skia/source/detail?r=11986

[enne (OOO), 2013-10-23 22:17:27]

What do you think about removing the doubles here? I am wondering if this makes
the float version of map2_pf even slower than the double version.

[tomhudson, 2013-10-24 07:20:33]

Test like crazy?

It feels a bit odd, since I think all the consumers of SkMatrix44 are in
Chromium instead of in Skia itself.

[reed1, 2013-10-24 16:06:49]

I think the fastest would be to avoid as many conversions as possible.

If the matrix is floats (i.e. SkMScalar == float) then this CL will be fastest.

If the matrix is doubles, then both the old and new versions are probably
equivalent, and probably slower than if the matrix was floats.

lgtm -- tho I agree that running our existing benchmarks make sense.

[enne (OOO), 2013-10-24 18:44:25]

The bench target appears to not call map2_pf.  Are there other benchmarks I
should be running?

It also appears that other map2_Xf functions (for various X) have double<->float
conversions too.  I want to land this first to see  if this is actually the
performance problem that's causing the bug.  If that makes performance better, I
can follow up and convert the others.

[enne (OOO), 2013-10-28 17:12:35]

Are there other Skia tests I should run other than bench?

[enne (OOO), 2013-10-28 20:01:19]

Running locally, this looks like a clear 25% win for the regressed test:
https://docs.google.com/a/chromium.org/spreadsheet/ccc?key=0AklwB21ZzpmDdEQ2a...

Given the lack of response, I will just go ahead and land this.

[commit-bot: I haz the power, 2013-10-28 20:03:23]

CQ is trying da patch. Follow status at
https://skia-tree-status.appspot.com/cq/enne@chromium.org/37693004/1

[commit-bot: I haz the power, 2013-10-28 20:20:57]

Change committed as 11986