Blender Git Commit Log

Cycles: workaround for noise performance regression in CUDA 5.5

Use manual ternary operation widening in grad(). Without it nvcc 5.5 produces multiple branch splits with very big branches (because of inlining). This solves 19% performance regression for BMW1M-MikePan.blend.

Also remove one redundant instruction in perlin SSE (when h == 12 or h == 14, then h is always >= 4).

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D190