are at most 52 bits. func addMul19(v uint128, a, b uint64) uint128 { hi, lo := bits.Mul64(mul19(a), b) lo, c := bits.Add64(lo, v.lo, 0) hi, _ = bits.Add64(hi, v.hi, c) return uint128{lo, hi} } // addMul38 returns v + 38 * a * b, where a and b are at most 52 bits. func addMul38(v uint128, a, b uint64) uint128 { hi, lo := bits.Mul64(mul19(a), b*2) lo, c := bits.Add64(lo, v.lo, 0) hi, _ = bits.Add64(hi, v.hi, c) return uint128{lo, hi} } // shiftRightBy51 returns a >> 51. a is assumed to be at most 115...

H2L VSLDOI $8, H2, ZERO, H2H STXVD2X VH2L, (XIP+R8) // save H^2 MOVD $0x70, R8 STXVD2X VH2, (XIP+R9) MOVD $0x80, R9 STXVD2X VH2H, (XIP+R10) MOVD $0x90, R10 VPMSUMD IN, H2L, XL // H.lo路H^2.lo VPMSUMD IN1, H2L, XL1 // H^2.lo路H^2.lo VPMSUMD IN, H2, XM // H.hi路H^2.lo+H.lo路H^2.hi VPMSUMD IN1, H2, XM1 // H^2.hi路H^2.lo+H^2.lo路H^2.hi VPMSUMD IN, H2H, XH // H.hi路H^2.hi VPMSUMD IN1, H2H, XH1 // H^2.hi路H^2.hi VPMSUMD XL, XC2, T2 // 1st reduction phase VPMSUMD XL1, XC2, HH // 1st reduction phase VSLDOI $8, XM,...

    *   XLA:GPU reductions are deterministic by default (reductions within
        `jit_compile=True` are now deterministic).
    *   XLA:GPU works with Horovod (OSS contribution by Trent Lo from NVidia)
    *   XLA:CPU and XLA:GPU can compile tf.unique and tf.where when shapes are
        provably correct at compile time.
*   `tf.saved_model.save`:

0EA1..0EA3    ; valid                                  # 1.1  LAO LETTER MO..LAO LETTER LO LING
0EA4          ; disallowed                             # NA   <reserved-0EA4>
0EA5          ; valid                                  # 1.1  LAO LETTER LO LOOT
0EA6          ; disallowed                             # NA   <reserved-0EA6>