Import cpu-sparse prototype #818
Conversation
This brings in the cpu-sparse prototype from the piet-next branch of the piet repo. No substantive changes, but cpu-sparse is renamed vello_hybrid and piet-next is renamed vello_api. Quite a bit of editing to satisfy the lint monster. There was a half-written SIMD implementation of flattening, that's removed. It should be finished and re-added, as it's a good speedup.
| // Note: getting rid of this predicate might help with | ||
| // auto-vectorization. That said, just getting rid of | ||
| // it causes artifacts (which may be divide by zero). | ||
| if dy != 0.0 { | ||
| let xx0 = startx + (y0 - starty) * slope; | ||
| let xx1 = startx + (y1 - starty) * slope; | ||
| let xmin0 = xx0.min(xx1); | ||
| let xmax = xx0.max(xx1); | ||
| let xmin = xmin0.min(1.0) - 1e-6; | ||
| let b = xmax.min(1.0); | ||
| let c = b.max(0.0); | ||
| let d = xmin.max(0.0); | ||
| let a = (b + 0.5 * (d * d - c * c) - xmin) / (xmax - xmin); | ||
| areas[x as usize][y] += a * dy; | ||
| } |
There was a problem hiding this comment.
Removing this indeed appears to improve things, cutting the time needed to generate strips by ~50% on my platform. See: https://xi.zulipchat.com/#narrow/channel/197075-gpu/topic/CPU.20sparse.20strip.20rendering.20to.20pixels/near/500401632.
| // Note: getting rid of this predicate might help with | |
| // auto-vectorization. That said, just getting rid of | |
| // it causes artifacts (which may be divide by zero). | |
| if dy != 0.0 { | |
| let xx0 = startx + (y0 - starty) * slope; | |
| let xx1 = startx + (y1 - starty) * slope; | |
| let xmin0 = xx0.min(xx1); | |
| let xmax = xx0.max(xx1); | |
| let xmin = xmin0.min(1.0) - 1e-6; | |
| let b = xmax.min(1.0); | |
| let c = b.max(0.0); | |
| let d = xmin.max(0.0); | |
| let a = (b + 0.5 * (d * d - c * c) - xmin) / (xmax - xmin); | |
| areas[x as usize][y] += a * dy; | |
| } | |
| let xx0 = startx + (y0 - starty) * slope; | |
| let xx1 = startx + (y1 - starty) * slope; | |
| let xmin0 = xx0.min(xx1); | |
| let xmax = xx0.max(xx1); | |
| let xmin = xmin0.min(1.0) - 1e-6; | |
| let b = xmax.min(1.0); | |
| let c = b.max(0.0); | |
| let d = xmin.max(0.0); | |
| let a = (b + 0.5 * (d * d - c * c) - xmin) / (xmax - xmin); | |
| // This is (on x86 and ARM) a branchless way to set `a` to 0 if it is NaN. | |
| areas[x as usize][y] += a.abs().max(0.).copysign(a) * dy; |
There was a problem hiding this comment.
There was a problem hiding this comment.
Unfortunately just removing it causes artifacts, as some divide by zeros get through. If you look at the explicit SIMD version, there's an is_finite check. Very likely that could be adapted, but I haven't spent a lot of time on it.
There was a problem hiding this comment.
The change to areas[x as usize][y] += a.abs().max(0.).copysign(a) * dy; should prevent NaNs branchlessly (but not infinities). I'd have to try it with this branch explicitly, but running Ghostcript_Tiger.svg through an older revision of cpu-sparse-experiments with this patch results in an identical image.
An alternative is areas[x as usize][y] += (a * dy).abs().max(0.).copysign(a * dy), which would also handle setting infinities to 0.0.
Renders a simple scene to the GPU, first by doing coarse rasterization the same as cpu-sparse, then doing a single draw call.
This brings in the cpu-sparse prototype from the piet-next branch of the piet repo. No substantive changes, but cpu-sparse is renamed vello_hybrid and piet-next is renamed vello_api.
Quite a bit of editing to satisfy the lint monster.
There was a half-written SIMD implementation of flattening, that's removed. It should be finished and re-added, as it's a good speedup.