3 ms·
> Or you can do three operations using a single 16 x 16 x 16 x 16 lookup table, and introduce fewer data dependencies into your CPU pipeline. Gotcha. So to su
by dragontamer 6y ago
> Or you can do three operations using a single 16 x 16 x 16 x 16 lookup table, and introduce fewer data dependencies into your CPU pipeline.
Gotcha.
So to summarize: you've seen GF(2^4) calculations performed SIMD style. A singular 16-bit lookup can be 4x GF(2^4) lookups in parallel.
Seems legit to me. A non-obvious implementation, but clearly possible and faster.