Complex{Float32} is not a subtype of Float32, breaking conv of complex valued input/weights

[dsweber2]

Just a simple convolution with two complex matrices breaks:
conv(randn(Complex{Float32}, 25,25,1,3), randn(Complex{Float32}, 5,5,1,3))

The stated reason is that this line

NNlib.jl/src/nnpack/interface.jl

Line 15 in 0d16973

T1.($(Symbol("$(front_name)$(backend)!"))(Float32.(out), Float32.(in1),

is trying to coerce a complex into a real by converting it to a Float32. Tweaking both

NNlib.jl/src/nnpack/impl.jl

Line 8 in 0d16973

function conv_nnpack!(y::A1, x::A1, w::A1, cdims::ConvDims;

and

NNlib.jl/src/nnpack/libnnpack.jl

Line 127 in 0d16973

function nnp_convolution_output(y::Array{Float32,4}, x::Array{Float32,4}, w::Array{Float32,4}, b::Array{Float32,1}; algo::nnp_convolution_algorithm = UInt32(0), workspace_buffer = nothing, workspace_size = 0, padding = 0, stride = 1, threadpool = C_NULL, profile = nothing)

to allow for Complex{Float32} as well allowed it to work as normal. I suspect a similar issue happens with any other functions that could work on complex inputs, but I haven't done extensive testing.

[CarloLucibello]

We are likely removing the forced conversion in #212. That said I don't think NNPACK has Complex support, right? We could support complex convolution in conv_im2col I guess

[aksuhton]

OP's code works now on a CPU but fails when I try to do this on the GPU. I'd like this functionality but as a relatively new programmer I'm not sure where to begin in order to implement this enhancement.

[ToucheSir]

Flux's GPU support for convs currently relies on cuDNN, which to my knowledge doesn't support complex numbers. If you can figure out a way to make it support them or to write GPU-friendly conv kernels we can use as a substitute, those would be the main paths for moving this forward.