pkg/tflite-micro: Update tflite-micro to latest version. requires update of pkg/flatbuffers as well.

[FlapKap]

Contribution description

This PR updates tflite-micro to the the latest version. Since the current version in Riot OS is quite old, this has mandated some other changes and updates:
The included flatbuffers pkg and dist/tools have been updated to v23.5.26 as thats the one tflite-micro uses.
The patch files to fix cast-align errors have been updated
The mnist example have been updated with new locations for headers and i've replaced the AllOpsResolver with mutable_ops_resolver, since the former has been deprecated. This requires the user to specify the exact operators needed, with an added bonus of a much smaller binary size.

Testing procedure

run the tests/pkg/tflite-micro test

Issues/PRs references

[aabadie]

I tried to build locally and got this error:

"make" -C /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro -f /work/riot/RIOT/pkg/tflite-micro/tflite-micro.mk
In file included from /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro/micro_interpreter.cc:33:
/work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/schema/schema_utils.h:29:38: error: 'OperatorCodeT' does not name a type; did you mean 'OperatorCode'?
   29 | BuiltinOperator GetBuiltinCode(const OperatorCodeT *op_code);
      |                                      ^~~~~~~~~~~~~
      |                                      OperatorCode

Any idea what it could be?

[aabadie]

Using make clean all fixes the problem.

[aabadie]

Tested on native and nrf52840dk and it works! Good job!

[aabadie]

Can you prepend your 4 last commit messages with the right context ("pkg/tflite-micro: ", "pkg/flatbuffers: ") ?

[FlapKap]

Can you prepend your 4 last commit messages with the right context ("pkg/tflite-micro: ", "pkg/flatbuffers: ") ?

done!

I tried to build locally and got this error:

"make" -C /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro -f /work/riot/RIOT/pkg/tflite-micro/tflite-micro.mk
In file included from /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro/micro_interpreter.cc:33:
/work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/schema/schema_utils.h:29:38: error: 'OperatorCodeT' does not name a type; did you mean 'OperatorCode'?
   29 | BuiltinOperator GetBuiltinCode(const OperatorCodeT *op_code);
      |                                      ^~~~~~~~~~~~~
      |                                      OperatorCode

Any idea what it could be?

I encountered the same issue when i had generated the schema using a different version of flatc than required. Just using the schema_generated already in the repo fixed it.

[riot-ci]

Murdock results

✔️ PASSED

6d51de0 pkg/tflite-micro: remove trailing spaces

Success	Failures	Total	Runtime
10104	0	10105	17m:35s

Artifacts

• Documentation preview

[aabadie]

Can you fix the check-commits check? https://github.com/RIOT-OS/RIOT/actions/runs/9207542826/job/25327831907?pr=20683#step:4:5.
A multi line commit message should make it:

pkg/flatbuffers: downgrade to v23.5.26

because that's what tflite-micro depends on

[aabadie]

Sorry to bother you again with 57abe1c
The commit message should be:

pkg/flatbuffers: downgrade to v23.5.26

because that's what tflite-micro depends on

format is: short description on the first line, blank line, extended description.

[FlapKap]

Sorry to bother you again with 57abe1c The commit message should be:

pkg/flatbuffers: downgrade to v23.5.26

because that's what tflite-micro depends on

format is: short description on the first line, blank line, extended description.

Done. No worries. Maybe I just shouldn't have done this past midnight :)

[aabadie]

This is removing too many boards to be right. I doubt all those removed boards have enough memory. nrf51dk has only 8kB of RAM.

[FlapKap]

I generated it through the RIOT/dist/tools/insufficient_memory/create_makefile.ci.sh script, but i must admit i don't have that board to test it on. Running BOARD=nrf51dk make -C tests/pkg/tflite-micro/ it builds fine, but I can't actually run it. The closest i have access to is a microbit, and this compiles and works on that

[FlapKap]

Murdock timed out on compiling the test for arduino-mkrwan1300 after 300 seconds. It does compile on my machine, but it takes 6,5 minutes

[maribu]

I tried to build locally and got this error:

"make" -C /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro -f /work/riot/RIOT/pkg/tflite-micro/tflite-micro.mk
In file included from /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro/micro_interpreter.cc:33:
/work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/schema/schema_utils.h:29:38: error: 'OperatorCodeT' does not name a type; did you mean 'OperatorCode'?
   29 | BuiltinOperator GetBuiltinCode(const OperatorCodeT *op_code);
      |                                      ^~~~~~~~~~~~~
      |                                      OperatorCode

Any idea what it could be?

This is now also popping up in the CI every now and then. There seems to be a race condition.

[FlapKap]

I tried to build locally and got this error:

"make" -C /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro -f /work/riot/RIOT/pkg/tflite-micro/tflite-micro.mk
In file included from /work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/micro/micro_interpreter.cc:33:
/work/riot/RIOT/build/pkg/tflite-micro/tensorflow/lite/schema/schema_utils.h:29:38: error: 'OperatorCodeT' does not name a type; did you mean 'OperatorCode'?
   29 | BuiltinOperator GetBuiltinCode(const OperatorCodeT *op_code);
      |                                      ^~~~~~~~~~~~~
      |                                      OperatorCode

Any idea what it could be?

This is now also popping up in the CI every now and then. There seems to be a race condition.

The cause is that schema_generated.h is not using the proper flatc flags to generate the appropriate types. However, the tflite-micro repo already has the correct schema_generated.h so there is probably some funny stuff going on when make decide to regenerate it. I'll create a new PR with the fix.

EDIT: PR #20703