diff --git a/src/impl/depthwiseconv_direct.jl b/src/impl/depthwiseconv_direct.jl
index b6822a488..b93aac6f4 100644
--- a/src/impl/depthwiseconv_direct.jl
+++ b/src/impl/depthwiseconv_direct.jl
@@ -81,12 +81,10 @@ function depthwiseconv_direct!(y::AbstractArray{yT,5}, x::AbstractArray{xT,5},
         h_idx in h_region,
         w_idx in w_region
 
-        # Probe for out-of-bounds accesses on `x` and `continue` if we hit one
         dotprod = yT(0)
         c_out = (c_in - 1)*channel_multiplier(cdims) + c_mult
-        for c_in in 1:channels_in(cdims),
-            kd in 1:kernel_d
-
+        for kd in 1:kernel_d
+            # Probe for out-of-bounds accesses on `x` and `continue` if we hit one
             x_d = project(d_idx, stride_d, pad_d_lo) + (kd - 1)*dil_d
             if x_d <= 0 || x_d > depth
                 continue
diff --git a/test/conv.jl b/test/conv.jl
index 6a84e5f5e..e0aed549f 100644
--- a/test/conv.jl
+++ b/test/conv.jl
@@ -242,6 +242,9 @@ conv_answer_dict = Dict(
     ),
 )
 
+# A "drop channels and batch dimension" helper
+ddims(x) = dropdims(x, dims=(ndims(x)-1, ndims(x)))
+
 @testset "Dense Convolution" begin
     # Start with some easy-to-debug cases that we have worked through and _know_ work
     for rank in (1,2,3)
@@ -272,9 +275,6 @@ conv_answer_dict = Dict(
             x = reshape(Float64[1:prod(size(dx));], size(dx)..., 1, 1)
             w = reshape(Float64[1:prod(size(dw));], size(dw)..., 1, 1)
 
-            # A "drop channels and batch dimension" helper
-            ddims(x) = dropdims(x, dims=(rank+1, rank+2))
-
             convs = [NNlib.conv, NNlib.conv_im2col, NNlib.conv_direct,]
             NNlib.is_nnpack_available() && push!(convs, NNlib.conv_nnpack)
             for conv in convs
@@ -450,8 +450,10 @@ else
 end
 
 @testset "Depthwise Convolution" begin
-    # Start with some easy-to-debug cases that we have worked through and _know_ work
-    for rank in (1,) #2,3)
+    # Start with some easy-to-debug cases that we have worked through and _know_ work.
+    # NOTE: these examples are all single-channel... which doesn't really stress test
+    # the important parts of depthwise convolution!
+    for rank in (1,2,3)
         @testset "depthwiseconv$(rank)d" begin
             # Pull out known-good answers for y = depthwiseconv(x, w)
             y_pad = conv_answer_dict[rank]["y_pad"]
@@ -479,9 +481,6 @@ end
             x = reshape(Float64[1:prod(size(dx));], size(dx)..., 1, 1)
             w = reshape(Float64[1:prod(size(dw));], size(dw)..., 1, 1)
 
-            # A "drop channels and batch dimension" helper
-            ddims(x) = dropdims(x, dims=(rank+1, rank+2))
-
             for conv in (NNlib.depthwiseconv, NNlib.depthwiseconv_im2col, NNlib.depthwiseconv_direct)
                 @testset "$(conv)" begin
                     # First, your basic convolution with no parameters
@@ -557,6 +556,14 @@ end
             end
         end
     end
+
+    # Do some real depthwise convolution tests
+    x = Float64.(reshape(1:2, (1,2,1)))
+    w = Float64.(reshape(1:6, (3,1,2)))
+    cdims = DepthwiseConvDims(x, w; padding=1)
+    for conv in (NNlib.depthwiseconv, NNlib.depthwiseconv_im2col, NNlib.depthwiseconv_direct)
+        @test conv(x, w, cdims)[:] ≈ [2, 10]  rtol=1e-7
+    end
 end