v156 · webgpu atomic-vec2u-min-max · point-cloud splatting & migration
Point-cloud splatting & CAS contention analyzer
Rendering millions of LiDAR or 3D Gaussian Splatting points in a compute shader requires updating both pixel depth and RGBA8 color atomically. Previously, developers had to choose between packing 16-bit depth + 16-bit RGB565 into atomic<u32> (causing severe Z-fighting) or running a two-buffer Compare-And-Swap (CAS) spin-loop that serializes SIMT threads under high overdraw. atomicMin on atomic<vec2u> executes a lock-free 64-bit update in 1 instruction per thread.
Overdraw contention & Z-precision simulator
Selected WGSL compute kernel
| metric | 32-bit packed u32 (16+16) | 32-bit CAS spin-loop | 64-bit atomic<vec2u> (Chrome 156) |
|---|