mbabinski-at-google

Results 17 comments of mbabinski-at-google

rcRasterizeTriangles runs much faster if faces are pre-sorted.

I get about 16s.

rcRasterizeTriangles runs much faster if faces are pre-sorted.

Sorting moves the time

rcRasterizeTriangles runs much faster if faces are pre-sorted.

Looking at the new implementation, it seems like you could get rid of pools entirely, and directly allocate arrays of spans when needed, and if you're doing that, you might...

rcRasterizeTriangles runs much faster if faces are pre-sorted.

I think the reason sorting doesn't help with the new approach is because you traverse the whole list regardless of where the item goes -- if the correct insertion point...

rcRasterizeTriangles runs much faster if faces are pre-sorted.

Also, I don't see why using rcPermVector would be much slower than what you are doing? You still allocate in batches, double size when you run out, etc.

rcRasterizeTriangles runs much faster if faces are pre-sorted.

I wrote it up, and it is ~5% slower than using the pools. I wonder if it would make sense to add support for arena allocation to rcVector. https://github.com/mbabinski-at-google/recastnavigation/tree/heightfield-layout

rcRasterizeTriangles runs much faster if faces are pre-sorted.

That's true -- I haven't checked iterating backwards with pooling. The change to reservation size makes a pretty small impact. In my data set, the average number of spans in...