core: Optimize BMD.fillRect()

[adrian17]

This removed repeated range checks on each pixel, replacing them by possibly-memcpy.
I originally only added set_pixel32_row_raw, which improved 640x480 fills by 3x (FF) - 5x (desktop).
Then I tried adding a full-fill fast path, which (as expected) didn't do a noticeable difference on desktop, but somehow improved FF times by another 2x.

EDIT: I just remembered that I tested browser perf without wasm extensions, but bulk-memory might also do a difference here. Will check it soon.
With wasm-opt and bulk-memory enabled, the line-fill path and full-fill path are equally fast, and another 2x faster than the best result without extensions. Just slightly slower than desktop now (though weirdly, the FF profiler doesn't show memory.fill like it used to).

This only helps the CPU side, it won't help is most time is spent waiting for GPU.

[torokati44]

Ooh, really nice!

I think, instead of detecting a "full fill", you could, with a little bit more code, make it ever so slightly more generic, by detecting "full width fill of rows from i through j" - then the full fill would be just an accidental edge case of this, while still performing the same, I reckon.

nit: I think this is A-core, not A-rendering... :P Maybe both...? 👀 - See the discussion on Discord about this...

[adrian17]

I think, instead of detecting a "full fill", you could, with a little bit more code, make it ever so slightly more generic, by detecting "full width fill of rows from i through j" - then the full fill would be just an accidental edge case of this, while still performing the same, I reckon.

Yeah, but I'd say at this point the difference between 1 big copy and H width-sized copies is already quite small, so I'm not sure if it's worth it. I was already on a fence about adding the fill() path.

[torokati44]

Yeah, but I'd say at this point the difference between 1 big copy and H width-sized copies is already quite small, so I'm not sure if it's worth it. I was already on a fence about adding the fill() path.

That's fair - it's not like I can find any real world content right off the top of my head that does "fill all but the last 3 lines of a 5x1700 BitmapData repeatedly". And even if such content does exist ... hopefully we'll notice...?

[torokati44]

Thanks!