Fixed NaN error in dilation

[tejank10]

@staticfloat please review

[staticfloat]

Confirmed that this suppresses the symptoms I was seeing. Let's go ahead and add this workload to test suite (since the previous test didn't seem to catch this), remove the commented out dilation_dims() calls, and call it good.

As to why this was happening, I think this was due to us passing in the incorrect dimensions to the gemm() calls, which caused OpenBLAS to read from the wrong memory locations.

[staticfloat]

Uh oh, we need to fix gradients as well:

# 2-length convolution with 1 channel in and 1 channel out
w = reshape(Float64[1, 1], (2, 1, 1))

# 128-length input vector with 1 channel and 1 batch.
x = reshape(Float64[1:128...], (128, 1, 1))

# Do the convolution (this doesn't have any NaN's)
y = NNlib.conv(x, w; dilation = 2)
@test !any(isnan.(y))

# Generate random upstream derivative
dy = randn(size(y))

# Calculate backprop for w
dws = []
for idx in 1:1000
    dw = NNlib.∇conv_filter(dy, x, w; dilation=2)
    push!(dws, dw)
end

# Calculate backprop for x
dxs = []
for idx in 1:1000
    dx = NNlib.∇conv_data(dy, x, w; dilation=2)
    push!(dxs, dx)
end

@show any([any(isnan.(dw)) for dw in dws])
@show any([any(isnan.(dx)) for dx in dxs])

In my experiments, dx gets NaN's sometimes, dw has not, but that could just be because my w is very small, so whatever memory region the BLAS routines are reaching into to get uninitialized memory might not be getting read from if w and x are too small.

[tejank10]

im2col_dims was returning extra memory size. Also, I had goofed up the order of dimensions in dilation parameter of col2im2d!. Sorry about that.

[staticfloat]

Yep, that's great! Thanks Tejan! Can you add some of these workloads as test cases in the test suite, so that we can be sure to not run into these problems in the future? I can confirm that these changes fix my NaN issues.

[MikeInnes]

Thanks! I'm getting segfaults on 0.7, so hopefully this will address those.