16x16 Tile Rasterizer

[rezajebeli97]

This is more of a question than an issue.
How did you end up choosing 16x16 tiles? Why not 32x32 or 8x8?
And why didn’t you collect features into shared memory?

I made some calculations, and it seems you are using 16×16×28B = 7.2KB of shared memory:

collected_Id = 4B
collected_xy = 8B
collected_conic_opacity = 16B

The GeForce RTX 3090 has 128KB of shared memory, so I’m wondering why you decided not to collect features into shared memory.
In that case, you would need 16×16×40B = 10.2KB, which is still far below the shared memory limit.

Also, how did you end up choosing 16x16 tiles? Why not 32x32 or 8x8?
Shared memory doesn’t seem to be an issue for either option, as 32×32×28B = 28.6KB is still far below the cap.

Am I missing something here?

[AsherJingkongChen]

16x16 is generally faster.