Fix Windows build

[SS-JIA]

Summary:

Context

Fix third party CMakeLists.txt to allow flatcc to build for Windows. Some CMake configuration settings need to be adjusted for windows platforms.

Test Plan:

Test Plan

python install_executorch.py

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/10946

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 New Failures

As of commit 25a7a8b with merge base 3032398 ():

NEW FAILURES - The following jobs have failed:

• trunk / test-arm-backend (test_models_ethos-u85) / linux-job (gh)
  RuntimeError: Command docker exec -t a156b3cf92ed9dd9d2d712bf153bde17ecb0531323af3cd96e2ca386b6b15d2a /exec failed with exit code 1
• trunk / test-llama-runner-mac (fp32, coreml) / macos-job (gh)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[zingo]

sorry about the fail in
pull / unittest-arm-backend-with-no-fvp (test_pytest_models) / linux-job
Possible fix is ongoing here
#10950
Please ignore this error it is unrelated to your PR

[SS-JIA]

Needed to prevent linker errors when linking the aot lib.

[jathu]

LGTM but lets make sure CI is happy

[jathu]

This doesn't look right. Why is the dependency different for Windows?

[SS-JIA]

I don't actually know... if I don't do this, then on Windows there are linker errors where the AOT lib cannot find the symbols for the actual operator implementations in quantized_ops_lib.

However, if I do this across the board, then I run into duplicate registration issues on non-Windows platforms. Could be platform dependent linker behaviour, I suppose?

[SS-JIA]

I also tested running a model on windows to make sure that the duplicate registration did not occur on Windows.

[SS-JIA]

Failures don't seem related. In both cases the build succeeds (which this PR would impact) but failures occur during specific tests.

Build Linux Wheels got cancelled because

"Canceling since a higher priority waiting request for Build Linux Wheels-10946-wheel-linux-pytorch/test-infra-main exists"

test-arm-backend failure is because CPU time was exceeded.

--------------------------------------------------------------------------------
Running /pytorch/executorch/arm_test/test_model/mv2_arm_delegate_ethos-u85-256/mv2_arm_delegate_ethos-u85-256/cmake-out/arm_executor_runner for ethos-u85-256 run with FVP:FVP_Corstone_SSE-320 num_macs:256 timeout:600
WARNING: Corstone FVP is not cycle accurate and should NOT be used to determine valid runtime
--------------------------------------------------------------------------------
telnetterminal0: Listening for serial connection on port 5000
telnetterminal1: Listening for serial connection on port 5001
telnetterminal2: Listening for serial connection on port 5002
telnetterminal5: Listening for serial connection on port 5003
I [executorch:arm_executor_runner.cpp:491] PTE in 0x78000000 , Size: 3541184 bytes

I [executorch:arm_executor_runner.cpp:516] PTE Model data loaded. Size: 2934768 bytes.

I [executorch:arm_executor_runner.cpp:529] Model buffer loaded, has 1 methods

I [executorch:arm_executor_runner.cpp:537] Running method forward

I [executorch:arm_executor_runner.cpp:548] Setup Method allocator pool. Size: 62914560 bytes.

Traceback (most recent call last):
  File "/home/ec2-user/actions-runner/_work/executorch/executorch/test-infra/.github/scripts/run_with_env_secrets.py", line 102, in <module>
I [executorch:arm_executor_runner.cpp:565] Setting up planned buffer 0, size 752640.

I [executorch:arm_executor_runner.cpp:595] Setting up ETDump

I [executorch:EthosUBackend.cpp:114] EthosUBackend::init 0x780000b0

I [executorch:arm_executor_runner.cpp:612] Method 'forward' loaded.

I [executorch:arm_executor_runner.cpp:614] Preparing inputs...

I [executorch:arm_executor_runner.cpp:621] Input testset[0] from bundled bpte

I [executorch:arm_executor_runner.cpp:682] Input prepared.

I [executorch:arm_executor_runner.cpp:684] Starting the model execution...


Info: Simulation is stopping. Reason: CPU time has been exceeded.

test-llama-runner-mac failure is because the LLM output doesn't look right.

PyTorchObserver {"prompt_tokens":1,"generated_tokens":8,"model_load_start_ms":1747689319901,"model_load_end_ms":1747689324733,"inference_start_ms":1747689324843,"inference_end_ms":1747689324936,"prompt_eval_end_ms":1747689324854,"first_token_ms":1747689324854,"aggregate_sampling_time_ms":0,"SCALING_FACTOR_UNITS_PER_SECOND":1000}'
+ EXPECTED_PREFIX='Once upon a time,'
+ [[ Once upon the end, she felt really happy and
PyTorchObserver {"prompt_tokens":1,"generated_tokens":8,"model_load_start_ms":1747689319901,"model_load_end_ms":1747689324733,"inference_start_ms":1747689324843,"inference_end_ms":1747689324936,"prompt_eval_end_ms":1747689324854,"first_token_ms":1747689324854,"aggregate_sampling_time_ms":0,"SCALING_FACTOR_UNITS_PER_SECOND":1000} == \O\n\c\e\ \u\p\o\n\ \a\ \t\i\m\e\,* ]]
+ echo 'Expected result prefix: Once upon a time,'
Expected result prefix: Once upon a time,
+ echo 'Actual result: Once upon the end, she felt really happy and
PyTorchObserver {"prompt_tokens":1,"generated_tokens":8,"model_load_start_ms":1747689319901,"model_load_end_ms":1747689324733,"inference_start_ms":1747689324843,"inference_end_ms":1747689324936,"prompt_eval_end_ms":1747689324854,"first_token_ms":1747689324854,"aggregate_sampling_time_ms":0,"SCALING_FACTOR_UNITS_PER_SECOND":1000}'
Actual result: Once upon the end, she felt really happy and
PyTorchObserver {"prompt_tokens":1,"generated_tokens":8,"model_load_start_ms":1747689319901,"model_load_end_ms":1747689324733,"inference_start_ms":1747689324843,"inference_end_ms":1747689324936,"prompt_eval_end_ms":1747689324854,"first_token_ms":1747689324854,"aggregate_sampling_time_ms":0,"SCALING_FACTOR_UNITS_PER_SECOND":1000}
+ echo 'Failure; results not the same'
Failure; results not the same

[zingo]

I just merged a PR that I expect/hope fixes. Feel free to ignore the Arm backend fails. If you rebase I expect it to go away.

This:
#10973