Revert to multiplication without conjugate

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

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, conj(w_val), dotprod)
+            dotprod = muladd(x_val, 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, conj(w_val), dotprod)
+                    dotprod = muladd(x_val, w_val, dotprod)
                 end
             end
         end
@@ -169,7 +169,7 @@ Calculate the gradient imposed upon `x` in the convolution `y = x * w`.
 function ∇conv_data_direct!(dx::AbstractArray{xT,5}, dy::AbstractArray{yT,5},
                             w::AbstractArray{wT,5}, cdims::DenseConvDims;
                             alpha::xT=xT(1), beta=false) where {xT, yT, wT}
-    w = transpose_swapbatch(w[end:-1:1, end:-1:1, end:-1:1, :, :])
+    w = conj(transpose_swapbatch(w[end:-1:1, end:-1:1, end:-1:1, :, :]))
     dy = predilate(dy, stride(cdims))
     ctdims = DenseConvDims(dy, w; padding=transpose_pad(cdims),
                                   dilation=dilation(cdims),
@@ -188,7 +188,7 @@ Calculate the gradient imposed upon `w` in the convolution `y = x * w`.
 function ∇conv_filter_direct!(dw::AbstractArray{wT,5}, x::AbstractArray{xT,5},
                               dy::AbstractArray{yT,5}, cdims::DenseConvDims;
                               alpha::wT=wT(1), beta=false) where {xT, yT, wT}
-    x = transpose_swapbatch(x[end:-1:1, end:-1:1, end:-1:1, :, :])
+    x = conj(transpose_swapbatch(x[end:-1:1, end:-1:1, end:-1:1, :, :]))
     dy = transpose_swapbatch(predilate(dy, stride(cdims)))
     ctdims = DenseConvDims(dy, x; padding=transpose_pad(cdims),
                                     stride=dilation(cdims))

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(copy(conj(w)))
+            w_ptr = pointer(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

@@ -373,7 +373,7 @@ end
             end
         end
         @testset "$(conv)" begin
-            @test isapprox(ddims(conv(x, w, cdims)), [vec(w)' * vec(x)], rtol = 1.0e-7)
+            @test isapprox(ddims(conv(x, w, cdims)), [transpose(vec(w)) * vec(x)], rtol = 1.0e-7)
         end
     end
     dy = NNlib.conv(x, w, cdims)
@@ -383,8 +383,8 @@ end
         (NNlib.∇conv_filter_direct, NNlib.∇conv_data_direct),
     )
         @testset "$(∇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)
+            @test isapprox(∇conv_filter(x, dy, cdims), conj(x) .* dy, rtol = 1.0e-7)
+            @test isapprox(∇conv_data(dy, w, cdims), dy .* conj(w), rtol = 1.0e-7)
         end
     end
 end