triton.language.fdiv

triton.language.fdiv(x, y, ieee_rounding=False, approx=False, _semantic=None)

Computes the element-wise division of x by y.

Parameters:

  • x – the numerator

  • y – the denominator

  • ieee_rounding – if True, use IEEE round-to-nearest division.

  • approx – if True, allow approximate division for float32 operands. On NVIDIA, this uses div.approx.f32, with a maximum error of 2 ULP for denominator magnitudes in [2**-126, 2**126]. Results outside this range may differ from full-range division. On AMD, this uses llvm.amdgcn.fdiv.fast, with a maximum error of 2.5 ULP and no support for denormal inputs or results. Cannot be combined with ieee_rounding=True.