See neon on armv8 manual vaddq_f16, vfmaq_f16 such instructions, thinking and compare the efficiency of f32, read the same amount of data, read twice the number of times is the f16 f32, VLD final test - fma - VST time-consuming, found f16 slower than f32 3 times,,, the board is rk3399, compile the NDK, don't a72 support f1_ (: з) _6 <)