Floating Point Hashing
summary
Rasterizing non-uniform point clouds to uniform voxels uses a BVH nearest-neighbor check per corner, which repeats work; a cache keyed by the point coordinates in a Rust HashMap avoids recomputation but F32/F64 lack Hash/Eq. The implementation uses to_bits plus an ulps comparison that collapses every four consecutive float bit patterns into one hash, accepting harmless false negatives and relying on bounded scan extents and recursion depth. It works in practice for the constrained dataset but misses subnormals and unbounded error accumulation in the original proof.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:54Problem Context
- 5:38Rust Example
- 6:55What is a Floating Point
- 9:20Inherent Precision Loss
- 10:38Drift and Error
- 13:04NAN
- 18:23NAN Problems
- 19:20Floating Point Hashing
- 23:53Conclusion
related talks
06mA tour of the compiler guided by bitvec
55mUPDATED: Scalable and Low Latency Lock-free Data Structures in C++
51mAllocating Less: Really Thin Rust Cloud Apps
58mOptimizing Binary Search
37mTeaching myself Rust through the medium of coordinate system transforms and FFI
39mFearless Concurrency on the GPU
40mRust’s approach of getting things right
25mVector Streaming: The Memory Efficient Indexing for Vector Databases