remove NNlib

[CarloLucibello]

remove dependence on NNlib.jl as discussed in different places.
Most stuff removed will end up in FluxML/NNlib.jl#286

@denizyuret should the batchnorm function stay or go?

[denizyuret]

@denizyuret should the batchnorm function stay or go?

cuDNN introduced a new interface which I wrapped in CUDA/lib/cudnn/normalization.jl. They still support the old interface though, so feel free to move batchnorm.jl to whereever needed.

[CarloLucibello]

this should be ready. @denizyuret could you take a look and see I didn't remove too much stuff?

[denizyuret]

this should be ready. @denizyuret could you take a look and see I didn't remove too much stuff?

Looks good to me. I assume you are moving these to whereever they are needed? (NNlibCUDA, Flux etc)

[CarloLucibello]

I assume you are moving these to whereever they are needed? (NNlibCUDA, Flux etc)

yes.

@maleadt this should be ready (assuming the buildkite's julia 1.6 debug fail is unrelated)

[maleadt]

Thanks! That failure is unrelated, and should be addressed by #756. Let's merge this after that PR.

[codecov]

Codecov Report

Merging #753 (117dac0) into master (fe62319) will decrease coverage by 0.57%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #753      +/-   ##
==========================================
- Coverage   78.78%   78.20%   -0.58%     
==========================================
  Files         124      121       -3     
  Lines        7512     7245     -267     
==========================================
- Hits         5918     5666     -252     
+ Misses       1594     1579      -15     

Impacted Files	Coverage Δ
lib/cudnn/CUDNN.jl	67.27% <ø> (ø)
lib/cudnn/convolution.jl	0.00% <ø> (-61.18%)	⬇️
lib/cudnn/pooling.jl	73.07% <ø> (+7.55%)	⬆️
lib/cudnn/util.jl	38.09% <ø> (-8.06%)	⬇️
src/CUDA.jl	100.00% <ø> (ø)
src/pool.jl	83.89% <0.00%> (-1.13%)	⬇️
src/array.jl	89.49% <0.00%> (-0.46%)	⬇️
lib/cudnn/tensor.jl	18.96% <0.00%> (-0.34%)	⬇️
src/broadcast.jl	86.53% <0.00%> (+1.92%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update fe62319...117dac0. Read the comment docs.