Optimize pixel overwrite

[RazrFalcon]

Skip blending when color is opaque or blending mode is Source/Clear. This is what skia and cairo do. And the results are pretty radical:

test opaque_fill_skia      ... bench:      47,662 ns/iter (+/- 1,119)
test opaque_fill_cairo     ... bench:      45,804 ns/iter (+/- 263)
test opaque_fill_raqote    ... bench:   2,432,875 ns/iter (+/- 51,948)

test source_fill_skia      ... bench:      50,663 ns/iter (+/- 207)
test source_fill_cairo     ... bench:      48,839 ns/iter (+/- 354)
test source_fill_raqote    ... bench:     678,795 ns/iter (+/- 10,330)

[jrmuizel]

Can you change the benchmark harness to use fill_rect() instead of fill(). Raqote intentionally doesn't try to detect rectangular paths like Cairo and Skia do.

[RazrFalcon]

Current benchmarks are unaffected by this optimization.

[jrmuizel]

Hmmm. That's a bit surprising. Can you commit the change and I'll take a look.

[RazrFalcon]

What change?

[jrmuizel]

Using fill_rect() instead of fill()

[RazrFalcon]

Yes, fill_rect will be faster, but cairo and skia results are independent from the shape in this case.

Here are results of filling a star:

test source_fill_skia      ... bench:      32,676 ns/iter (+/- 440)
test source_fill_cairo     ... bench:      43,970 ns/iter (+/- 858)
test source_fill_raqote    ... bench:     465,450 ns/iter (+/- 6,622)

They are even faster here, because there are less scanlines.

[jrmuizel]

Ah, ok this will probably be best to look into after #158 is resolved.