branch uniform

The branch-none body with every unrolled step wrapped in an if/else whose two sides do equal work with different constants. The condition alternates each loop trip but is identical for every lane, so every invocation chooses the same source branch. The gap against branch-none includes condition handling and any compiler transformations; it does not isolate a native jump cost.

Measured in FLOP/s. Higher values mean better performance. Benchmark your device.