Using fallback BLAS replacements for `cblas_zdotc_sub64_` and `dznrm2_64_`

[albertomercurio]

Hello,

the following code generates the BLAS warning

using Enzyme
using LinearAlgebra

function my_fidelity2(p)
  ψg = zeros(ComplexF64, 2)
  ψg[2] = 1.0
  ψe = zeros(ComplexF64, 2)
  ψe[1] = 1.0

  ψgg = kron(ψg, ψg)
  ψee = kron(ψe, ψe)
  ψ = normalize(p[1] * ψgg + p[2] * ψee)

  return abs(dot(ψ, ψgg))
end

##

p = [0.3, 0.4]
my_fidelity2(p) # 0.6

Enzyme.gradient(Reverse, my_fidelity2, p) # ([1.28, -0.96],)

And the warning

┌ Warning: Using fallback BLAS replacements for (["cblas_zdotc_sub64_", "dznrm2_64_"]), performance may be degraded
└ @ Enzyme.Compiler ~/.julia/packages/GPUCompiler/2MI6e/src/utils.jl:61

Also, why is it related to the GPUCompiler?

[danielwe]

Also, why is it related to the GPUCompiler?

I can answer this part. Enzyme.jl uses a lot of functionality from GPUCompiler.jl for interacting with the Julia compiler pipeline, regardless of what hardware you're running on. In this case, what's happening is simply that the warning you're seeing is emitted by the macro GPUCompiler.@safe_warn, which, unlike the regular @warn, is safe to use in a number of low-level contexts where standard task-switching IO would error.

[albertomercurio]

Thanks for the explanation. Does the mean that I can safely use this? Or a custom rule has to be set?

[danielwe]

Everything is safe and correct. The warning is just telling you that performance may not be optimal because Enzyme currently lacks custom rules to express the gradient of these particular BLAS functions in terms of other BLAS functions. Instead, the gradient will be computed by replacing them with straight, unoptimized implementations that Enzyme can differentiate directly. Basically, the gradient will be computed as if you had replaced dot(ψ, ψgg) with something like sum(conj(ψ[i]) * ψgg[i] for i in eachindex(ψ, ψgg)), and similar for the norm calculated inside normalize. See the FAQ here: https://enzyme.mit.edu/julia/stable/faq/#Linear-Algebra