Merge pull request #101 from FluxML/sf/loop_ordering

Fix loop ordering problem in reference conv implementation

Author: MikeInnes

src/impl/conv_direct.jl

@@ -68,9 +68,9 @@ conv_direct!
     # explicit formulation of convolution.  Oh hoisting gods, hear my plea.
     @inbounds for batch in 1:size(x)[end],
         c_out in 1:out_c,
+        d_idx in 1:out_depth,
         h_idx in 1:out_height,
-        w_idx in 1:out_width,
-        d_idx in 1:out_depth
+        w_idx in 1:out_width
 
         # Starting points of the window of x we're going to grab
         x_w = project(w_idx, stride_w, pad_w_lo)
@@ -145,4 +145,4 @@ Calculate the gradient imposed upon `w` in the convolution `y = x * w`.
         dw .= dw[end:-1:1, end:-1:1, end:-1:1, :, :]
     end
     return dw
-end
+end