Reorder the separable blur passes into XXX/YYY, with an image transpose on the

Reorder the separable blur passes into XXX/YYY, with an image transpose on the last pass of each group.  This results in continuguous memory reads in all passes, giving a 22% speedup on theverge.skp over the previous separable implementation, and a 30%-50% improvement over the existing implementation (depending on platform).

Committed: https://code.google.com/p/skia/source/detail?r=6445

[Stephen White, 2012-11-15 19:43:14]

PTAL.  Thanks!

[reed1, 2012-11-15 20:14:17]

The speedup sounds great. I suggest adding some block-comments, perhaps on these
helper functions, would make it a lot easier for the next coder to understand
(a) what's going on, and (b) why.

[Stephen White, 2012-11-15 20:29:30]

On 2012/11/15 20:14:17, reed1 wrote:
> The speedup sounds great. I suggest adding some block-comments, perhaps on
these
> helper functions, would make it a lot easier for the next coder to understand
> (a) what's going on, and (b) why.

Fixed.  Also landed (Sorry!  Realized too late I didn't have official LG).

[TomH, 2012-11-15 22:50:42]

Sweet speedup!