codebook indices = 65536   k_inv = 0.006767272949218750   int8 step = 4*k_inv = 0.027069091796875000
[1] fp16 mul1 arm vs independent reference : bitwise mismatches = 0
[2] integer decode  sat((bytesum-510+A)>>2)  : mismatches = 0
[3] control 'no +2'     differs on 16374 indices  (must be > 0)
    control 'signed dp4a' differs on 62452 indices  (must be > 0)
    integer  rel-MSE=5.974e-05  max|e|=0.012695  bias=-0.001572 = -0.157% of sigma
    no +2    rel-MSE=1.267e-04  max|e|=0.019440  bias=-0.008335 = -0.833% of sigma
    signed   rel-MSE=1.118e+01  max|e|=6.812500  bias=-3.056919 = -305.597% of sigma
    vs round-half-even: 8188 disagreements of 16385 ties, max |delta| = 1; all inside the d=2 (mod 4) tie set: True
    sigma(codebook) = 1.0003   distinct fp16 values = 913   distinct int8 levels = 238   saturated = 0
GATE 1: PASS
