Jaroslav Sevcik

Replace top-k with first-k thresholding

1

Just for sharing a prototype, not for landing.

A prototype of capsule capsule collision detection primitive in Warp. This also contains a very rough plumbing to be able to call from Jax to Warp, but many things are...

Dynamic-slice-update is much slower than scatter on GPUs, e.g., on the 10 humanoid benchmark (270x270 matrix with 2430 sparse elements), factor_m is 3x slower with dynamic-slice-update (on H100).

NOT FOR COMMIT Depends on https://github.com/google/praxis/pull/51