Misc fixes and docs update

[rnett]

Fixes#168 (makes ndarray and framework respect -DskipTests), gitignores bazel config files, and adds a CONTRIBUTING.md file as discussed in #176 (comment). Also adds a link to the gitter channel.

If there's anything else you think should be added to CONTRIBUTING.md let me know, these are all of the tips that I'm aware of.

[rnett]

To add when I'm not building:

Currently, do to build time constraints, the GPU binaries only support compute capacities 3.5 and 7.0. To use with un-supported GPUs, change the value [here](tensorflow-core/tensorflow-core-api/build.sh#L27) build the binaries yourself.
To build for GPU, pass `-Djavacpp.platform.extension=-gpu` to maven. Tensorflow's configure script can be used to set options (like cuda locations) for the native build. Make sure the `TF_CUDA_COMPUTE_CAPABILITIES` value in your `.tf_configure.bazelrc` matches the value set in `build.sh`.

[karllessard]

Thank you @rnett , that's valuable documentation that was definitely missing

[rnett]

I changed build.sh to use a set TF_CUDA_COMPUTE_CAPABILITIES by default as per #200, @karllessard can you add the CI build tag just in case?

[karllessard]

@rnett , the GPU build failed with that change:

Cuda Configuration Error: Invalid compute capability: compute_'35
INFO: Elapsed time: 15.494s
INFO: 0 processes.
FAILED: Build did NOT complete successfully (0 packages loaded)

https://github.com/tensorflow/java/runs/1826080574?check_suite_focus=true

[rnett]

It's the single quotes, I think. Should be fixed.

[karllessard]

@rnett , failed again for a very awkward reason this time. I can simply retrigger a build, sometimes things get wrong on GitHub Action hosted servers but I'll you validate first:

Analyzing: 6 targets (180 packages loaded, 9563 targets configured)
Analyzing: 6 targets (199 packages loaded, 14836 targets configured)
WARNING: Download from https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/7e825abd5704ce28b166f9463d4bd304348fd2a9.tar.gz failed: class com.google.devtools.build.lib.bazel.repository.downloader.UnrecoverableHttpException GET returned 404 Not Found
INFO: Analyzed 6 targets (204 packages loaded, 18538 targets configured).
INFO: Found 6 targets...
[0 / 1,314] [Prepa] BazelWorkspaceStatusAction stable-status.txt
[2,727 / 6,848] Creating runfiles tree bazel-out/k8-opt-exec-50AE0418/bin/external/org_tensorflow/tensorflow/cc/ops/array_ops_gen_cc.runfiles; 0s local ... (2 actions, 1 running)
ERROR: /root/.cache/bazel/_bazel_root/47b4c4ffab8583ddc67d4f1b3159b1ed/external/nccl_archive/BUILD.bazel:53:1: C++ compilation of rule '@nccl_archive//:device_lib' failed (Exit 1)
/opt/rh/devtoolset-7/root/usr/lib/gcc/x86_64-redhat-linux/7/../../../../include/c++/7/limits: In static member function 'static constexpr long double std::numeric_limits<long double>::denorm_min()':
/opt/rh/devtoolset-7/root/usr/lib/gcc/x86_64-redhat-linux/7/../../../../include/c++/7/limits:1799:65: error: expected ')' before numeric constant
denorm_min() _GLIBCXX_USE_NOEXCEPT { return __LDBL_DENORM_MIN__; }
^~~
/opt/rh/devtoolset-7/root/usr/lib/gcc/x86_64-redhat-linux/7/../../../../include/c++/7/limits:1799:68: error: cannot convert 'double (*)(const char*) throw ()' to 'long double' in return
denorm_min() _GLIBCXX_USE_NOEXCEPT { return __LDBL_DENORM_MIN__; }

https://github.com/tensorflow/java/runs/1827956486?check_suite_focus=true

[rnett]

Yeah that doesn't sound like anything I changed.

[karllessard]

You are right, master is also failing now for the same reasons... (sigh)... @saudet , do you think it is a wrong GCC installation or something like that?

[saudet]

@karllessard Tried upgrading to GCC 8 and that didn't change anything, so probably not an issue with GCC:
https://github.com/saudet/tensorflow-java/actions/runs/541859636

[saudet]

@karllessard I'm not having any more luck with TensorFlow 2.4.1 and CUDA 11.0 either, so it's not something that used to be broken for sure. It actually seems like something recently broke. It's obviously downloading something somewhere for CUDA that isn't pinned to a specific version, and that got upgraded unintentionally somewhere upstream, which broke TF's build, at least on CentOS. I assume this is going to get fixed upstream, eventually...

[karllessard]

Thanks @saudet , maybe we can go back to this container with Cuda we used to have before we start installing it manually? It stopped working as well but might have been fixed upstream since? Also that shortened the build time for Linux, which seems less stable these days.

[saudet]

Using the container isn't going to change anything about CUDA. That's not the issue here.

[karllessard]

@rnett , were you able at least to build your version with GPU support on your machine? It might take awhile before this issue gets resolved by itself, I'll merge this if you confirm that it worked on your side.

[rnett]

Yeah, although I had TF_CUDA_COMPUTE_CAPABILITIES set in my bazelrc so I'm not 100% it was tested. I'll run one tonight or tomorrow without it and let you know.

[rnett]

Ok, it works locally.