NNPACK convolution issue

[Alexander-Barth]

The convolution with NNPACK do not seem to work reliably. The output of all these convolutions should be zero. Sometimes they are, but sometimes they have a strange value (maybe uninitialized memory)? Is this a known issue?

Thanks

julia> @time extrema(NNlib.conv(zeros(Float32,100,100,10,32),ones(Float32,4,4,10,10)))
  0.044346 seconds (44 allocations: 23.707 MiB)
(0.0f0, 0.0f0)

julia> @time extrema(NNlib.conv(zeros(Float32,100,100,10,32),ones(Float32,4,4,10,10)))
  0.043753 seconds (44 allocations: 23.707 MiB)
(0.0f0, 0.0f0)

julia> @time extrema(NNlib.conv(zeros(Float32,100,100,10,32),ones(Float32,4,4,10,20)))
  0.055470 seconds (44 allocations: 35.205 MiB)
(NaN32, NaN32)

julia> @time extrema(NNlib.conv(zeros(Float32,100,100,10,32),ones(Float32,4,4,10,30)))
  0.197568 seconds (144.16 k allocations: 54.368 MiB)
(-0.00033701234f0, 2.953125f0)

julia> NNlib.is_nnpack_available()
true

julia> NNlib.NNPACK_CPU_THREADS
0x0000000000000008

julia> NNlib.libnnpack
"/home/abarth/.julia/dev/NNlib/deps/usr/lib/libnnpack.so"

[ljluestc]

Ensure that NNPACK is installed correctly by checking that the library path is correct (/home/abarth/.julia/dev/NNlib/deps/usr/lib/libnnpack.so). Make sure you have the latest version of NNPACK installed, as newer versions may contain bug fixes and improvements.

NNlib.NNPACK_CPU_THREADS appears to be 0x0000000000000008. This value corresponds to 8 CPU threads. You might want to verify that it matches your system configuration.

[Alexander-Barth]

Thank you for your input!
For some reasons the loading now fails on my system:

julia> using NNPACK_jll

julia> using NNlib
┌ Warning: Error requiring `NNPACK_jll` from `NNlib`
│   exception =
│    LoadError: ArgumentError: Package NNlib does not have NNPACK_jll in its dependencies:
│    - You may have a partially installed environment. Try `Pkg.instantiate()`
│      to ensure all packages in the environment are installed.
│    - Or, if you have NNlib checked out for development and have
│      added NNPACK_jll as a dependency but haven't updated your primary
│      environment's manifest file, try `Pkg.resolve()`.
│    - Otherwise you may need to report an issue with NNlib
│    Stacktrace:
│      [1] macro expansion
│        @ ./loading.jl:1599 [inlined]
│      [2] macro expansion
│        @ ./lock.jl:267 [inlined]
│      [3] require(into::Module, mod::Symbol)
│        @ Base ./loading.jl:1576
│      [4] include(mod::Module, _path::String)
│        @ Base ./Base.jl:457
│      [5] include(x::String)
│        @ NNlib ~/.julia/packages/NNlib/uuZgJ/src/NNlib.jl:1

If I am not mistaken, I use the latested version (in an empty envorinement).

(foo2) pkg> st
Status `/tmp/foo2/Project.toml`
  [872c559c] NNlib v0.9.1
  [a6bfbf70] NNPACK_jll v2018.6.22+0

Yes, I used 8 CPU cores on the system where I submitted the bug report. I included this info, as it could be potentially a multi-threading issue.

I am wondering of NNLib still uses NNPACK as it is not mentioned here:

https://github.com/FluxML/NNlib.jl/blob/master/Project.toml

[ToucheSir]

Technically we still have some NNPACK routines hidden behind Requires, but I don't know of a single person who uses them and they probably should've been removed in v0.9.

[Alexander-Barth]

OK, so as NNlib.conv gives the correct results, I think we can close the issue.

julia> extrema(NNlib.conv(zeros(Float32,100,100,10,32),ones(Float32,4,4,10,10)))
(0.0f0, 0.0f0)