CuFFT `Base.:(*)` overload types are too generic

[abussy]

In lib/cufft/fft.jl, the Base.:(*) operator is overloaded:

CUDA.jl/lib/cufft/fft.jl

Lines 359 to 377 in 9528a33

	function Base.:(*)(p::CuFFTPlan{T,S,K,false}, x::DenseCuArray{S1,M}) where {T,S,K,S1,M}
	if T<:Real
	# Out-of-place complex-to-real FFT will always overwrite input x.
	# We copy the input x in an auxiliary buffer.
	z = p.buffer
	copyto!(z, x)
	else
	if S1 != S
	# Convert to the expected input type.
	z = copy1(S, x)
	else
	z = x
	end
	end
	assert_applicable(p, z)
	y = CuArray{T,M}(undef, p.output_size)
	unsafe_execute_trailing!(p, z, y)
	y
	end

The type parametrization is probably too broad, as x::DenseCuArray{S1,M} where {S1, M} will cover any dense CuArray of any element type. In particular, S1 can stand for types not supported by CuFFT.

In DFTK, we overload the Base.:(*) operator for abstract FFT plans and abstract arrays of ForwardDiff.Dual (here). Unfortunately, that fails on NVIDIA GPUs because CuFFT's overload takes precedence, even though ForwardDiff.Dual is not a type that CuFFT should support.

@mfherbst for reference.

[maleadt]

The type parametrization is probably too broad, as x::DenseCuArray{S1,M} where {S1, M} will cover any dense CuArray of any element type. In particular, S1 can stand for types not supported by CuFFT.

There's a convertor in there, tough:

         if S1 != S 
             # Convert to the expected input type. 
             z = copy1(S, x) 

[abussy]

Indeed, but then it is assumed that the convertor and subsequent FFT applies to whatever type is captured by S1. This will not always be the case.

[maleadt]

What alternative do you suggest? Not supporting any conversion seems restrictive, as does limiting to what cuFFT supports.

Why can't you have your overrides be specific to ForwardDiff.Dual for S1?

[abussy]

Why can't you have your overrides be specific to ForwardDiff.Dual for S1?

We can, and that is what we will start doing as a workaround.

It seems to me that CuFFT is overreaching at the moment, but I have to admit I have no clear alternative in mind.