science/tensorflow-core: Work-in-progress port for TensorFlow core library.

[arrowd]

Summary

At this moment the port doesn't even build due to strange compiler/linker bug. The Diff is created to make it easier to reproduce that bug.
I'll proceed shaping the port after it is fixed.

[arrowd]

The blocking bug: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=229155

[tcberner]

science/tensorflow-core/Makefile

6

^ wrong order, ...PREFIX should come before ...VERSION

14

^ it seems this is already needed during fetch as it is now?

19

^ this is the default, you can drop it

53

^ this sounds like you want HAS_CONFIGURE

[arrowd]

• Update to 1.11.0 and address comments.

[arrowd]

Currently blocked by https://github.com/tensorflow/tensorflow/issues/22665

[0mp]

science/tensorflow-core/Makefile

12

Missing LICENSE_FILE=${WRKSRC}/LICENSE.

[mat]

science/tensorflow-core/Makefile

15

A port can never depend on lang/python. (or lang/python2 or lang/python3.)

44

Why the include?

58

This is empty, remove.

[arrowd]

science/tensorflow-core/Makefile

15

What should be done then? There is no USES=python:fetch.

[mat]

science/tensorflow-core/Makefile

15

Mmm, right, in any way, not lang/python, so python${PYTHON_DEFAULT}:lang/python${PYTHON_DEFAULT:S/.//} would probably be the right one.

49–52

You don't have access to WRKDIR during the fetch phase. (and you cannot access the network during the extract phase.)

[arrowd]

science/tensorflow-core/Makefile

49–52

You don't have access to WRKDIR during the fetch phase. (and you cannot access the network during the extract phase.)

I'm doomed, then? For build systems like Bazel or Haskell Stack it is extremely difficult to prepare all required distfiles, because they tend to download them themselves. I can try going hardcore way by examining what and where to gets downloaded and then make up distinfo from that, but maybe there is another way?

I was suspecting I can't access network after fetch, but I'm really surprised that I can't touch WRKDIR during fetch. Why is that?

[mat]

science/tensorflow-core/Makefile

49–52

Because it is created during the extract phase.

I think the correct way to go is to create a USES=bazel that works the same way as USES=cargo, in two steps, one for the maintainer to generate a list of stuff to download, and add to the Makefile, and then for the build system to use that to download stuff.

[arrowd]

science/tensorflow-core/Makefile

49–52

Unlike Cargo, Bazel fetches distfiles from various sources (like, https://github.com/grpc/grpc/archive/v1.13.0.tar.gz and https://curl.haxx.se/download/curl-7.60.0.tar.gz). We'll end up with huge MASTER_SITES list and huge distinfo. Should we really follow this path?

If we can't unpack to WRKDIR, maybe some other temporary dir will do? DISTDIR even?

[mat]

science/tensorflow-core/Makefile

49–52

I'm not sure having a huge MASTER_SITES is a problem, as it would be automatically generated.

[arrowd]

science/tensorflow-core/Makefile

49–52

As I said, automatic generation isn't easily implemetable for Bazel's BUILD files. For instance, TensorFlow has clauses like

tf_http_archive(
        name = "grpc",
        sha256 = "50db9cf2221354485eb7c3bd55a4c27190caef7048a2a1a15fbe60a498f98b44",
        strip_prefix = "grpc-1.13.0",
        system_build_file = clean_dep("//third_party/systemlibs:grpc.BUILD"),
        urls = [
            "https://mirror.bazel.build/github.com/grpc/grpc/archive/v1.13.0.tar.gz",
            "https://github.com/grpc/grpc/archive/v1.13.0.tar.gz",
        ],
)

scattered,possibly, over several .bzl files. We can parse it out, of course, but this would be TensorFlow-specific anyways. And since it is me who will be doing all the heavy-lifting, I'd like to hear if extracting into some temporary dir during fetch step is an absolute "no-no", before going this gross (in my opinion) way.

[arrowd]

Take repackaging route, like wine-i386.

I was able to produce libtensorflow.so and this port installs repackaged binary. This approach is also working for Python bindings, but this are currently failing due to external deps.

[arrowd]

Mark old comments as done.

[mat]

science/tensorflow-core/Makefile

30

The 3.6 here should probably not be hardcoded.

46–47

This can be done in one step:

git clone -b v${DISTVERSION} https://github.com/tensorflow/tensorflow tensorflow-${DISTVERSION}

Also, if you don't care about the history, you could add a --depth 1 to speed up the download (goes from 350MB down to 45MB).

[arrowd]

• Add ONLY_FOR_ARCHS, address comments.

[tobik]

Is this still relevant now that science/py-tensorflow was resurrected in rP508257?

[arrowd]

The current port doesn't provide libtensorflow.so, which is required for Haskell bindings, but it's better to build upon that, rather that this solution, which requires repackaging.