fmsub Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h lines 663–677

template <>
Vectorized<c10::BFloat16> inline fmsub(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
#if BF16_ARITHMETIC_SUPPORTED()
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  bfloat16x8_t z = c;
  return x * y - z;
#else
  // See NOTE [BF16 FMA] above.
  return a * b - c;
#endif
}