Changed Makefile to support TF2.6.1.

[NobuoTsukamoto]

• Makefile compatible with TF2.6.1.
• Fixed undefined symbol issue.
  • Missing symbol after compilation #28
  • libedgetpu not working on raspberrypi4 NobuoTsukamoto/meta-tensorflow-lite#63

[stakach]

@NobuoTsukamoto could you update this for TFv2.16.1 ?

[NobuoTsukamoto]

@stakach
This PR is compatible with TensorFlow v2.16.1.
In my environment, I can build correctly with the combination of TensorFlow v2.16.1.

Are any build errors occurring?

[feranick]

This should be indeed pulled, as the Makefile in its current form is outdated. That said, when I try to build it, I have the following error, despite having flatbuffers installed in the version recommended (23.5.26):

In file included from ../tensorflow/tensorflow/lite/util.cc:30:
../tensorflow/tensorflow/lite/schema/schema_generated.h:25:41: error: static assertion failed: Non-compatible flatbuffers version included
   25 | static_assert(FLATBUFFERS_VERSION_MAJOR == 23 &&
        |                                                                                   ^

[NobuoTsukamoto]

Please attach all build logs.
What is reported in the line following the error in question?
Probably the wrong flatbuffers version is being reported.
Please install the correct flatbuffers version.

/tensorflow/tensorflow/lite/schema/schema_generated.h:25:41: note: the comparison reduces to ‘(x == 23)’

Also check the flatc version and installation path.

$ flatc --version
flatc version 23.5.26
$ which flatc
/usr/local/bin/flatc

[stakach]

@NobuoTsukamoto

In file included from /home/steve/projects/tensorflow/tensorflow/lite/util.cc:30:
/home/steve/projects/tensorflow/tensorflow/lite/schema/schema_generated.h:25:41: error: static assertion failed: Non-compatible flatbuffers version included
   25 | static_assert(FLATBUFFERS_VERSION_MAJOR == 23 &&
      |                                         ^
/home/steve/projects/tensorflow/tensorflow/lite/schema/schema_generated.h:25:41: note: the comparison reduces to ‘(2 == 23)’

flatc --version
=> flatc version 2.0.8

on the latest version of ubuntu

sudo apt install flatbuffers-compiler
Reading package lists... Done
Building dependency tree... Done
Reading state information... Done
flatbuffers-compiler is already the newest version (2.0.8+dfsg1-6build1).
flatbuffers-compiler set to manually installed.

[feranick]

Yes, flatc is in version 2.0.8 for me as well. Using pip to install the proper python package doesn't change the actual version of flatc.

[stakach]

This worked to install it

git clone https://github.com/google/flatbuffers.git
cd flatbuffers
git checkout tags/v23.5.26
cmake -G "Unix Makefiles" \
      -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_INSTALL_PREFIX=/usr/local \
      -DFLATBUFFERS_BUILD_SHAREDLIB:BOOL=ON
make
sudo make install

although the versioning is just based on release date and not semver so probably compatible with the 24 versions

[feranick]

TF requires a specific version of flatbuffers depending on the TF release as well. For TF 2.16.1 it's 23.5.26.

From: https://github.com/tensorflow/tensorflow/blob/v2.16.1/tensorflow/lite/schema/schema_generated.h

static_assert(FLATBUFFERS_VERSION_MAJOR == 23 &&
              FLATBUFFERS_VERSION_MINOR == 5 &&
              FLATBUFFERS_VERSION_REVISION == 26,
             "Non-compatible flatbuffers version included");

[feranick]

In any case, the README.md file should reflect that if someone wants to compile with Makefile, you need a specific version of flatbuffers as indicated (including compilation).

[stakach]

the other issue is that manually building flatbuffers only installs a static lib but the libedgetpu build expects a shared lib - is there any way to resolve that?

EDIT: updated my previous comment with an updated build - needed to toggle on the shared lib

[NobuoTsukamoto]

Please specify the FLATBUFFERS_BUILD_SHAREDLIB option.

https://github.com/google/flatbuffers/blob/v23.5.26/CMakeLists.txt#L28

[stakach]

yeah I did that and still seeing

Building libedgetpu.so
/usr/bin/ld.gold: error: cannot find -lflatbuffers
collect2: error: ld returned 1 exit status
make: *** [makefile_build/Makefile:202: libedgetpu] Error 1

even though the output of ldconfig -p | grep flatbuffers is

ldconfig -p | grep flatbuffers
	libflatbuffers.so.23.5.26 (libc6,x86-64) => /usr/local/lib/libflatbuffers.so.23.5.26
	libflatbuffers.so.2 (libc6,x86-64) => /lib/x86_64-linux-gnu/libflatbuffers.so.2
	libflatbuffers.so (libc6,x86-64) => /usr/local/lib/libflatbuffers.so

[NobuoTsukamoto]

Add the path to LDFLAGS,
e.g.
export LDFLAGS="-L/usr/local/lib"

[stakach]

nice, working for me
Are you able to update the README with these steps @NobuoTsukamoto ?

requires flatbuffers v23.5.26

git clone https://github.com/google/flatbuffers.git
cd flatbuffers
git checkout tags/v23.5.26
cmake -G "Unix Makefiles" \
      -DCMAKE_BUILD_TYPE=Release \
      -DFLATBUFFERS_BUILD_SHAREDLIB:BOOL=ON
make
sudo make install

# if libedgetpu build fails
export LDFLAGS="-L/usr/local/lib"

and hopefully we can get this merged

[NobuoTsukamoto]

Modified README with build instructions targeting TF v2.16.1 and FlatBuffers v23.5.26.

[mwprado]

new_makefile.patch.txt
this path will fix the makefile (without bazel) build (tested in fedora 40).