Compile TensorFlow Lite for iOS Simulator on Apple Silicon

[bvirlet]

TensorFlowLite isn't available for iOS Simulator running on Apple Silicon.

It seems that this depends on Bazel supporting this, as well as on TensorFlowLite being built as a .xcframework (fat frameworks have an arm64 slice, but cannot have an arm64-mac slice).

[abattery]

@yyoon could you take a look?

[yyoon]

@bvirlet Yes, you are correct that we're blocked by the missing bazel support. Once bazel supports it, we will be migrating to .xcframework to incorporate the different target architectures, including the simulators on the new Apple silicon.

[bvirlet]

Thanks. For reference, I'm linking this issue here that's related/the same: #44609

[greenzeal]

+1

[bvirlet]

@yyoon do you think this PR is what we want from Bazel?

[yyoon]

@bvirlet I've looked at that PR before, but that's for building bazel itself to work natively on M1, which is different from building an ios_static_framework for arm64 simulator target. I think this would involve more work from apple build rules side.

Even better, if bazel + apple rules start to support building an XCFramework with multiple target architectures, we could switch to using that for releasing our iOS binaries.

[bvirlet]

Thanks for clarifying that!

[guillaume-tgl]

Looks like a new version of Bazel's rules_apple has been released with support for ios_sim_arm64 architecture: https://github.com/bazelbuild/rules_apple/releases/tag/0.32.0

Is it enough to be able to build TFLite for iOS Simulator on Apple Silicon?

[yyoon]

That's good to hear. Roughly speaking, this could be done in the following steps.

1. Update bazel to 5.0.0 or higher.
2. Update rules_apple version to 0.32.0 or higher. This version dependency is defined in:

   tensorflow/tensorflow/workspace2.bzl

   Lines 801 to 805 in 592f379

   	tf_http_archive(
   	name = "build_bazel_rules_apple",
   	sha256 = "0052d452af7742c8f3a4e0929763388a66403de363775db7e90adecb2ba4944b",
   	urls = tf_mirror_urls("https://github.com/bazelbuild/rules_apple/releases/download/0.31.3/rules_apple.0.31.3.tar.gz"),
   	)

3. Add a new .bazelrc config for ios_sim_arm64:

   tensorflow/.bazelrc

   Lines 193 to 206 in 592f379

   	# iOS configs for each architecture and the fat binary builds.
   	build:ios --apple_platform_type=ios
   	build:ios --apple_bitcode=embedded --copt=-fembed-bitcode
   	build:ios --copt=-Wno-c++11-narrowing
   	build:ios_armv7 --config=ios
   	build:ios_armv7 --cpu=ios_armv7
   	build:ios_arm64 --config=ios
   	build:ios_arm64 --cpu=ios_arm64
   	build:ios_i386 --config=ios
   	build:ios_i386 --cpu=ios_i386
   	build:ios_x86_64 --config=ios
   	build:ios_x86_64 --cpu=ios_x86_64
   	build:ios_fat --config=ios
   	build:ios_fat --ios_multi_cpus=armv7,arm64,i386,x86_64

4. Build the iOS static famework following this guide but with --config=ios_sim_arm64 option instead.

I didn't have time to try this out myself unfortunately and there might be other roadblocks ahead, but it'd be worth trying if you're in need.

[edritech93]

same issue for me