optimizations for small drawRect calls

optimizations for small drawRect calls

[reed1, 2012-05-10 20:25:00]

Likely I will commit these piece-meal, but here is the sum of my experiments so
far to speed up small drawRects (simulating dashing).

Bench runs are very noisy :( but I'm seeing ~25% faster on dash_4_rect
benchmark.

Will run some timings on linux (64bit) before I commit anything
Want to get some aggregate feeling of perf change on all benches (but how?)

[reed1, 2012-05-10 20:50:31]

~ 20% faster on linux (when I ran with -03 for both runs)

[TomH, 2012-05-10 20:53:50]

The templates add a *lot* of code complexity to BlitRow_D32. Do they really
yield that much more performance than a naive unrolling? The only thing I see
you're winning is that the odd-man-out part of the loop (1..3) is unrolled.

For noisy timings, what -repeat count are you using?
My rule of thumb is -repeat 50 for 5% noise, -repeat 150 for 1%.

[reed1, 2012-05-10 20:57:12]

Good question about the value of the 1-3 tail. I will try expanding the test to
time w/ and w/o those being templated.

[reed1, 2012-05-11 12:33:50]

removed template for catching the tail n%4 counts

[reed1, 2012-05-11 14:40:28]

committing the simpler pieces separately, to isolate these if there are any
problems found.