Photoprism compilation issues with tensorflow

[vbachvarov]

I have been trying to compile photoprism on FreeBSD using photoprism-freebsd-port for a couple of days without success.
The commands I used follow:

cd /tmp
git clone 'https://github.com/huo-ju/photoprism-freebsd-port.git'
cd photoprism-freebsd-port
LD_LIBRARY_PATH=".:/lib:/usr/libexec:/usr/local/lib"  make  BATCH=yes OPTIONS_SET=AVX

I identified two issues:

1. The port depends on science/libtensorflow1, however there is no such port in the ports tree. I assume it is meant to build against science/py-tensorflow, so I applied the following patch to the Makefile and the compilation started:
   sed -i '' 's#science/libtensorflow1#science/py-tensorflow#g' Makefile

2. The compilation of science/py-tensorflow fails with the following build error:

ERROR: /usr/ports/science/py-tensorflow/work-py38/tensorflow-1.15.5/tensorflow/python/BUILD:329:1: C++ compilation of rule '//tensorflow/python:bfloat16_lib' failed (Exit 1)
tensorflow/python/lib/core/bfloat16.cc:633:8: error: no matching function for call to object of type '(lambda at tensorflow/python/lib/core/bfloat16.cc:607:25)'
  if (!register_ufunc("equal", CompareUFunc<Bfloat16EqFunctor>,
       ^~~~~~~~~~~~~~
tensorflow/python/lib/core/bfloat16.cc:607:25: note: candidate function not viable: no overload of 'CompareUFunc' matching 'PyUFuncGenericFunction' (aka 'void (*)(char **, const long *, const long *, void *)') for 2nd argument
  auto register_ufunc = [&](const char* name, PyUFuncGenericFunction fn,
                        ^
tensorflow/python/lib/core/bfloat16.cc:637:8: error: no matching function for call to object of type '(lambda at tensorflow/python/lib/core/bfloat16.cc:607:25)'
  if (!register_ufunc("not_equal", CompareUFunc<Bfloat16NeFunctor>,
       ^~~~~~~~~~~~~~
tensorflow/python/lib/core/bfloat16.cc:607:25: note: candidate function not viable: no overload of 'CompareUFunc' matching 'PyUFuncGenericFunction' (aka 'void (*)(char **, const long *, const long *, void *)') for 2nd argument
  auto register_ufunc = [&](const char* name, PyUFuncGenericFunction fn,
                        ^
tensorflow/python/lib/core/bfloat16.cc:641:8: error: no matching function for call to object of type '(lambda at tensorflow/python/lib/core/bfloat16.cc:607:25)'
  if (!register_ufunc("less", CompareUFunc<Bfloat16LtFunctor>, compare_types)) {
       ^~~~~~~~~~~~~~
tensorflow/python/lib/core/bfloat16.cc:607:25: note: candidate function not viable: no overload of 'CompareUFunc' matching 'PyUFuncGenericFunction' (aka 'void (*)(char **, const long *, const long *, void *)') for 2nd argument
  auto register_ufunc = [&](const char* name, PyUFuncGenericFunction fn,
                        ^
tensorflow/python/lib/core/bfloat16.cc:644:8: error: no matching function for call to object of type '(lambda at tensorflow/python/lib/core/bfloat16.cc:607:25)'
  if (!register_ufunc("greater", CompareUFunc<Bfloat16GtFunctor>,
       ^~~~~~~~~~~~~~
tensorflow/python/lib/core/bfloat16.cc:607:25: note: candidate function not viable: no overload of 'CompareUFunc' matching 'PyUFuncGenericFunction' (aka 'void (*)(char **, const long *, const long *, void *)') for 2nd argument
  auto register_ufunc = [&](const char* name, PyUFuncGenericFunction fn,
                        ^
tensorflow/python/lib/core/bfloat16.cc:648:8: error: no matching function for call to object of type '(lambda at tensorflow/python/lib/core/bfloat16.cc:607:25)'
  if (!register_ufunc("less_equal", CompareUFunc<Bfloat16LeFunctor>,
       ^~~~~~~~~~~~~~
tensorflow/python/lib/core/bfloat16.cc:607:25: note: candidate function not viable: no overload of 'CompareUFunc' matching 'PyUFuncGenericFunction' (aka 'void (*)(char **, const long *, const long *, void *)') for 2nd argument
  auto register_ufunc = [&](const char* name, PyUFuncGenericFunction fn,
                        ^
tensorflow/python/lib/core/bfloat16.cc:652:8: error: no matching function for call to object of type '(lambda at tensorflow/python/lib/core/bfloat16.cc:607:25)'
  if (!register_ufunc("greater_equal", CompareUFunc<Bfloat16GeFunctor>,
       ^~~~~~~~~~~~~~
tensorflow/python/lib/core/bfloat16.cc:607:25: note: candidate function not viable: no overload of 'CompareUFunc' matching 'PyUFuncGenericFunction' (aka 'void (*)(char **, const long *, const long *, void *)') for 2nd argument
  auto register_ufunc = [&](const char* name, PyUFuncGenericFunction fn,
                        ^
6 errors generated.

I tried checking out the ports tree to a version from December 2021 and Photoprism to a commit before the switch to external tensorflow but I got other linker issues (see below). The commits I tried were:

• Ports: 56db10a78fb31ee3b0c4900991c862b9c1733bbf (Dec 20 2021), 809810e045293e9757b10f24a6b0f0078c985002 (Nov 1 2021), ad43a663a820dbabe60adedebef440c8752475fd (Jan 25 2022)
• Photoprism: 7895eaa (Dec 3 2021), 66cf274 (Dec 26 2021), 8368d02 (Jan 19 2022, current at the time of writing)

The linker errors with the code before splitting out tensorflow:

ERROR: /tmp/lansnap-build/photoprism-freebsd-port/work/photoprism-93b26f197908dd825dfa302d17793185cc3eab8c/docker/tensorflow/tensorflow-1.15.2/tensorflow/BUILD:563:1: Linking of rule '//tensorflow:libtensorflow_framework.so.1.15.2' failed (Exit 1)
ld: error: duplicate symbol: adler32
>>> defined at adler32.c
>>>            adler32.pic.o:(adler32) in archive bazel-out/host/bin/external/zlib/libzlib.pic.a
>>> defined at adler32.c
>>>            adler32.pic.o:(.text.adler32+0x0) in archive bazel-out/host/bin/external/zlib_archive/libzlib.pic.a

ld: error: duplicate symbol: adler32_combine
>>> defined at adler32.c
>>>            adler32.pic.o:(adler32_combine) in archive bazel-out/host/bin/external/zlib/libzlib.pic.a
>>> defined at adler32.c
>>>            adler32.pic.o:(.text.adler32_combine+0x0) in archive bazel-out/host/bin/external/zlib_archive/libzlib.pic.a

... (repeated for multiple other functions)

My question is: Am I compiling Photoprism correctly, and if yes - why is tensorflow failing to compile?

I am using FreeBSD 13.0-RELEASE.

[huo-ju]

Please install libtensorflow1 port from this repo: https://github.com/psa/libtensorflow1-freebsd-port

It has been submitted to the official port tree, but it's still not available now.

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=260694

[vbachvarov]

Thank you, I also saw libtensorflow1 has an open issue identical to this one.