Inverting ComplexF32 with Metal.jl throws an error

giordano · July 7, 2026, 10:10pm

The explanation of the error is right there in the stacktrace: inv calls widen:

base/complex.jl

6fe1d5110


      
          function inv(z::Complex{T}) where T<:Union{Float16,Float32}
              c, d = reim(widen(z))
              (isinf(c) | isinf(d)) && return complex(copysign(zero(T), c), flipsign(-zero(T), d))
              mag = inv(muladd(c, c, d^2))
              return oftype(z, Complex(c*mag, -d*mag))
          end

Metal.jl would need to overly that method with a different one which doesn’t involve going through double precision.

Topic		Replies	Views
Metal.jl does not speed up FFT GPU	8	2300	August 13, 2025
Factorize changes the float type from Float32 to Float64 Numerics	3	453	June 22, 2021
In-place multiplication is too much slower for ComplexF64 Performance complex-numbers	19	1462	February 10, 2022
Why `inv(Complex(0.0, 0.0))` returns `Complex(NaN, NaN)`? Internals & Design complex-numbers	12	979	January 10, 2023
InvalidIRError with CuArray broadcast when complex values are involved in Bessel functions GPU specialfunctions	2	374	December 8, 2022

Inverting ComplexF32 with Metal.jl throws an error

Related topics