Thomas Müller

The V100 should be supported. Could you shed more light on the strange results you experience? A self-contained example for reproduction would be great.

Hi there, unfortunately __tiny-cuda-nn__ doesn't support generating optimally efficient code for multiple architectures simultaneously -- although it _can_ target the lowest common denominator with the caveat that it'll run suboptimally...

The answer is the same as in #89 . You're trying to use tiny-cuda-nn's fully fused MLP, which doesn't support second derivatives through double backward, hence the crash. If you...

Could you let us know which host compiler (+ version) you are using? That might shed light on why `nvcc` reports that it doesn't support C++14.

On my machine, PyTorch bindings compile and run on `v1.12.0+cu116` -- unfortunately the runtime error sounds like some PyTorch-internal thing going wrong.

I think you're right about `N_FEATURES_PER_ENTRY` being the better choice in any case. It's clearer in its meaning + we're already using this terminology in the instant-ngp paper. The idea...

`glScissor` is still used on the bounding box of the masking container for efficiency purposes. Scissored-away pixels usually don't consume expensive fill rate just for the shader to return transparent...

I see. That's actually a good thing, since it means we can solve the problem more easily this way.

If profiling shows promise I think the portable approach is to use a smaller viewport + rescaling to achieve scissoring.

While I agree that the length bonus itself does not _absolutely need_ to be monotonic when adding hit objects to the end or beginning of a map, being so would...