Closed bitxsw93 closed 9 months ago
The reason for having a maximum resolution is that the rasterizer uses 32-bit fixed point math internally. Fixed-point representation is necessary to ensure exact coverage in rasterization, but it also forces a tradeoff between maximum viewport resolution and subpixel accuracy.
To answer the second question I would first need to know why the tile-based solution is not sufficient for your purposes. To the best of my knowledge, it produces the exact same image as a higher-resolution rasterization call would (up to floating-point rounding differences on the input side).
OK, Thanks. We try your suggestion from calling method first. If have any question, we give issues afterwards
Hello,I have the same trouble with issue 128(https://github.com/NVlabs/nvdiffrast/issues/128).
I have a model which includes some ops. One of ops adopts nvdiffrast CUDA rasterizer.
And, resolution is 4096. So, For me, I want to support CUDA rasterizer operation up to 4096, instead of tile the model.
My Question: 1、What is the reason for 2048 limitation, cuda shared memory limit or something else? 2、Is there some way to support rasterizer operation with higher resolution?
Besh wishes.