Convolution should conjugate one of its parameters

Author: zsoerenm

src/conv.jl

@@ -100,7 +100,7 @@ for backend in (Symbol(), :_direct, :_im2col, :_nnpack)
                             dy::AbstractArray{yT,N}, w::AbstractArray{wT,N},
                             cdims::C; kwargs...) where {yT, wT, N, C <: ConvDims}
                 dx = similar(dy, input_size(cdims)..., channels_in(cdims), size(dy, N))
-                return $(Symbol("$(name)$(backend)!"))(dx, dy, w, cdims; kwargs...)
+                return conj($(Symbol("$(name)$(backend)!"))(dx, dy, w, cdims; kwargs...))
             end
         end
     end
@@ -113,7 +113,7 @@ for backend in (Symbol(), :_direct, :_im2col, :_nnpack)
                         cdims::ConvDims; kwargs...) where {xT, yT, N}
             dw = similar(dy, kernel_size(cdims)..., channels_in(cdims) ÷ groupcount(cdims),
                                                     channels_out(cdims))
-            return $(Symbol("∇conv_filter$(backend)!"))(dw, x, dy, cdims; kwargs...)
+            return conj($(Symbol("∇conv_filter$(backend)!"))(dw, x, dy, cdims; kwargs...))
         end
     end
 

src/gemm.jl

@@ -35,7 +35,7 @@ for (gemm, elt) in gemm_datatype_mappings
                                beta::$(elt), C::Ptr{$elt})
             # Convert our compile-time transpose marker to a char for BLAS
             convtrans(V::Val{false}) = 'N'
-            convtrans(V::Val{true})  = 'C'
+            convtrans(V::Val{true})  = $elt <: Complex ? 'C' : 'T'
 
             if transA == Val(false)
                 lda = M

src/impl/conv_direct.jl

@@ -107,7 +107,7 @@ function conv_direct!(
                     kproj(kh, kernel_h, fk),
                     kproj(kd, kernel_d, fk),
                     c_in, c_out]
-            dotprod = muladd(x_val, w_val, dotprod)
+            dotprod = muladd(x_val, conj(w_val), dotprod)
         end
         y[w_idx, h_idx, d_idx, c_out, batch] = alpha*dotprod + beta*y[w_idx, h_idx, d_idx, c_out, batch]
     end
@@ -147,7 +147,7 @@ function conv_direct!(
                             kproj(kh, kernel_h, fk),
                             kproj(kd, kernel_d, fk),
                             c_in, c_out]
-                    dotprod = muladd(x_val, w_val, dotprod)
+                    dotprod = muladd(x_val, conj(w_val), dotprod)
                 end
             end
         end

src/impl/conv_im2col.jl

@@ -51,7 +51,7 @@ function conv_im2col!(
         im2col!(col_slice, view(x, :, :, :, :, batch_idx), cdims)
         GC.@preserve col_slice w y begin
             col_ptr = pointer(col_slice)
-            w_ptr = pointer(w)
+            w_ptr = pointer(copy(conj(w)))
             y_ptr = pointer(y, (batch_idx - 1)*M*N + 1)
             gemm!(Val(false), Val(false), M, N, K, alpha, col_ptr, w_ptr, beta, y_ptr)
         end

test/conv.jl

@@ -359,6 +359,36 @@ ddims(x) = dropdims(x, dims=(ndims(x)-1, ndims(x)))
     end
 end
 
+@testset "Complex Dense Convolution" begin
+    # For now only 1 dimensional 1x1 convolution
+    x = reshape(complex.(Float64[1:4;], Float64[1:4;] .+ 1), 1, 4, 1)
+    w = reshape(complex.(Float64[1:4;] .+ 2, Float64[1:4;] .+ 3), 1, 4, 1)
+    cdims = DenseConvDims(x, w)
+    convs = [NNlib.conv, NNlib.conv_im2col, NNlib.conv_direct,]
+    NNlib.is_nnpack_available() && push!(convs, NNlib.conv_nnpack)
+    for conv in convs
+        if NNlib.is_nnpack_available()
+            if conv == NNlib.conv_nnpack && !NNlib.nnpack_supported_operation(cdims)
+                continue
+            end
+        end
+        @testset "$(conv)" begin
+            @test isapprox(ddims(conv(x, w, cdims)), [vec(w)' * vec(x)], rtol = 1.0e-7)
+        end
+    end
+    dy = NNlib.conv(x, w, cdims)
+    for (∇conv_filter, ∇conv_data) in (
+        (NNlib.∇conv_filter,        NNlib.∇conv_data),
+        (NNlib.∇conv_filter_im2col, NNlib.∇conv_data_im2col),
+        (NNlib.∇conv_filter_direct, NNlib.∇conv_data_direct),
+    )
+        @testset "$(∇conv_filter)/$(∇conv_data)" begin
+            @test isapprox(∇conv_filter(x, dy, cdims), x .* conj(dy), rtol = 1.0e-7)
+            @test isapprox(∇conv_data(dy, w, cdims), conj(dy) .* w, rtol = 1.0e-7)
+        end
+    end
+end
+
 if get(ENV, "NNLIB_TEST_FUZZING", "false") == "true"
     # @info("Skipping Convolutional fuzzing tests, set NNLIB_TEST_FUZZING=true to run them")
     @testset "fuzzing" begin