Add fast path for stripping rectangles #900
Conversation
Co-authored-by: Tom Churchman <[email protected]>
There was a problem hiding this comment.
Most tests with rectangles call RenderContext::fill_rect, which now uses the specialized path for rectangles. It would be good to continue exercising the general case as well.
I don't know what the best solution looks like. Perhaps performing multiple tests against the same snapshot could work, though we'd have to be robust against small rounding errors.
# Conflicts: # sparse_strips/vello_cpu/src/render.rs
|
I've added a few more test cases which check against the same snapshot as the rectangle case, I feel like this should already be an improvement? As you mentioned, there are few other test cases with rectangles, but there's also only so many ways to draw a rectangle. :P So I think this should already be a good coverage. |
| // and pattern fills. | ||
| let (x0, x1, y0, y1) = ( | ||
| rect.min_x().max(0.0) as f32, | ||
| rect.max_x().min(width as f64) as f32, |
There was a problem hiding this comment.
for min_x() == 30, max_x() == 40, width == 10, this sets x0 = 30 and x1 = 10. This propagates through to x_start and x_end. Later we check that x_end - x_start >= 1, but due to wraparound, that'll be true. The end result is that we write a lot of unnecessary alpha values.
|
Maybe we can revisit in the future, but I'm pretty sure that for performant fine rasterization with SIMD, we will want to uphold the assumption that strips are always a multiple of 4 in width, which isn't the case with this fast path, and it might be tricky to change the PR to uphold this. So I'll close this for now. |
This re-adds my previously implemented fast path for rectangles. In my test scene with 20000 rectangles, this reduces the rendering time from around 37ms to 19ms!