Building TensorFlow Lite - macos_arm64 support using cmake

[simonmaurer]

with respect to this issue in the XNNPACK repository: bazel builds for macos_arm64 platform are not yet supported.
however building with cmake works successfully.

this change allows to build the TensorFlow Lite library and the benchmark_model on a macos_arm64 host platform when using -DTFLITE_ENABLE_XNNPACK=ON (default). see instructions.

[google-cla]

Thanks for your pull request. It looks like this may be your first contribution to a Google open source project (if not, look below for help). Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

📝 Please visit https://cla.developers.google.com/ to sign.

Once you've signed (or fixed any issues), please reply here with @googlebot I signed it! and we'll verify it.

---

What to do if you already signed the CLA

Individual signers

• It's possible we don't have your GitHub username or you're using a different email address on your commit. Check your existing CLA data and verify that your email is set on your git commits.

Corporate signers

• Your company has a Point of Contact who decides which employees are authorized to participate. Ask your POC to be added to the group of authorized contributors. If you don't know who your Point of Contact is, direct the Google project maintainer to go/cla#troubleshoot (Public version).
• The email used to register you as an authorized contributor must be the email used for the Git commit. Check your existing CLA data and verify that your email is set on your git commits.
• The email used to register you as an authorized contributor must also be attached to your GitHub account.

ℹ️ Googlers: Go here for more info.

[abattery]

@terryheo could you review this PR?

[simonmaurer]

@googlebot I signed it!

[simonmaurer]

oneDNN unit test failed?

[simonmaurer]

once this PR is accepted in XNNPACK repo, we can then also proceed with a new PR so TensorFlow Liteand benchmark_modelcan be built for macos_arm64using:

bazel build --config=macos_arm64 -c opt //tensorflow/lite:libtensorflowlite.so
bazel build --config=macos_arm64 -c opt //tensorflow/lite/tools/benchmark:benchmark_model

[terryheo]

Thanks for your contribution.

LGTM

[terryheo]

Oh, are you going to send another PR when google/XNNPACK#1327 is merged?
Otherwise, it might be better to wait?

[simonmaurer]

Oh, are you going to send another PR when google/XNNPACK#1327 is merged?
Otherwise, it might be better to wait?

thank you. mainly this PR was for updating the cmake build files for macos_arm64, but you're right.. it might indeed be better to wait for the XNNPACK PR to be accepted. so we can pull in all changes at once.

if you're ok with the PR naming I will wait and add the remaining commits of XNNPACK and pthreadpool too here. would be nice if you could review again to be safe.

there also needs to be a change for cpuinfo. did some testing and pulled the latest master branch in , but the cpuinfo.patch is not compatible so I deactivated applying this patch in WORKSPACE for verification. this went fine for macos_arm64 and host builds on macos_x86-64 but I'm not sure if this is the right way to go.

what do you think?

[terryheo]

I see. It might take time to wait for all the XNNPack related changes are merged.
I'm fine with having the second PR so we have can more time to verify all XNNPack related patches.

[terryheo]

This PR is merged and the XNNPack PR got LGTM.

@simonmaurer I have one question. Does the "--config=macos_arm64" option only works on M1 machine?
I can't use it on my Ubuntu desktop and x86 Mac machine.

[simonmaurer]

@terryheo thanks. if you're talking about the TensorFlow repo: this accepted PR here only allows the host build for macos_arm64 using cmake as described here.
I'll make the additional PR here to pull in the latest pthreadpool (PR accepted & merged), XNNPACK (PR accepted) and cpuinfo once they are both merged.

however if you're talking about the XNNPACK repo (the PR is indeed accepted, but not yet merged):

bazel build --config=macos_arm64 -c opt :end2end_bench

will work on a host M1 machine. on different platforms (like Linux or even macos x86-64) that does not work (yet) because of missing toolchains
at least I get the following errors when building on a macOS_x86-64 platform for macos_arm64:

ERROR: /private/var/tmp/**/**/external/local_config_cc/BUILD:48:19: in cc_toolchain_suite rule @local_config_cc//:toolchain: cc_toolchain_suite '@local_config_cc//:toolchain' does not contain a toolchain for cpu 'darwin_arm64' INFO: Repository cpuinfo instantiated at: /**/**/**/**/xnnpack_simonmaurer/WORKSPACE:66:13: in <toplevel> Repository rule http_archive defined at: /private/var/tmp/**/**/external/bazel_tools/tools/build_defs/repo/http.bzl:336:31: in <toplevel> ERROR: Analysis of target '//:end2end_bench' failed; build aborted: Analysis of target '@local_config_cc//:toolchain' failed

[simonmaurer]

@terryheo another note:
ideally we would want bazel to work like this:

M1 host machine compilation (without specifying the platform):

bazel build -c opt :end2end_bench (XNNPACK)
bazel build -c opt //tensorflow/lite/tools/benchmark:benchmark_model (TF)

Cross compilation on a different platform (e.g. Linux x86):

bazel build -c opt --config=macos_arm64 :end2end_bench (XNNPACK)
bazel build -c opt --config=macos_arm64 //tensorflow/lite/tools/benchmark:benchmark_model (TF)

the pip_package should work already according to this PR, which builds fine on a M1 host but I got errors about missing symbols when importing the tensorflow package in python3

[simonmaurer]

the pip_package should work already according to this PR, which builds fine on a M1 host but I got errors about missing symbols when importing the tensorflow package in python3

with this PR the pip package can be built using a native build on M1 host machine. the only dependencies that need manual installation are grpcio/h5py

[simonmaurer]

@terryheo / @gbaned also thanks for reviewing #47639. this includes the changes on the follow-up PR we discussed:

once this PR is accepted in XNNPACK repo, we can then also proceed with a new PR so TensorFlow Liteand benchmark_modelcan be built for macos_arm64using:

bazel build --config=macos_arm64 -c opt //tensorflow/lite:libtensorflowlite.so
bazel build --config=macos_arm64 -c opt //tensorflow/lite/tools/benchmark:benchmark_model

that being said, with #47639 we can now also build using bazel for macos_arm64 platform:

Cross compilation on a M1 host or a different platform (e.g. Linux x86):

bazel build -c opt --config=macos_arm64 //tensorflow/lite:libtensorflowlite.so
bazel build -c opt --config=macos_arm64 //tensorflow/lite/tools/benchmark:benchmark_model

and due to the additional changes #47594 by @freedomtan (thanks!) when using a more recent version of bazel (bazel #4928295):

M1 host machine compilation (without specifying the platform):

bazel build -c opt //tensorflow/lite:libtensorflowlite.so
bazel build -c opt //tensorflow/lite/tools/benchmark:benchmark_model

To summarize: TensorFlow Lite library and benchmark_model, but also the full TensorFlow pip package can be built using bazel