fp16 div

Eight independent scalar f16 chains of x = a / x + b per thread, unrolled 4x — the fp16 scalar FMA test with divide in place of multiply, so the gap against f16-fma-scalar isolates the cost of division in half precision. The loop-carried value is the divisor, so the compiler cannot hoist a reciprocal and turn it back into an FMA.

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