subgroups-f16

Earlier experimental implementations exposed subgroups-f16 separately for subgroup operations on half-precision values. It was deprecated during the subgroup feature cleanup. This page retains the historical token so reports containing it still have an explanation and an availability history.

For current implementations, check and request both subgroups and shader-f16, then use enable subgroups; and enable f16; in WGSL. Together they allow applicable subgroup operations to operate on f16 values, useful for reductions and machine-learning kernels that can tolerate half precision.

A low or zero percentage here measures the disappearance of the legacy token, not the absence of half-precision subgroup functionality. Do not request subgroups-f16 in new code or use it as the support test. Fall back to f32 subgroup operations when half precision is unavailable, or to workgroup-memory algorithms when subgroups are unavailable.