Incorrect display of logs for ARM bulds on FVP

[west9906]

I'm not sure, but it seems that the ARM GNU compiler (arm-none-eabi-gcc) does not support the c99 input/output format by default. I encountered a problem when I got the following output in FVP after running examples/arm/run.sh --model_name=add2 --target=ethos-u85-2048 --delegate

I [executorch:arm_executor_runner.cpp:zu ()] main
I [executorch:arm_executor_runner.cpp:zu ()] runner_init
I [executorch:arm_executor_runner.cpp:zu ()] runner_init
I [executorch:arm_executor_runner.cpp:zu ()] runner_init
I [executorch:arm_executor_runner.cpp:zu ()] runner_init
I [executorch:arm_executor_runner.cpp:zu ()] runner_init
...

After simply replacing %zu -> %lu in all ET_LOG() in file executorch/examples/arm/executor_runner/arm_executor_runner.cpp, I managed to get the following logs:

I [executorch:arm_executor_runner.cpp:749 main()] PTE in 0x74000000 $ Size: 2208 bytes
I [executorch:arm_executor_runner.cpp:480 runner_init()] PTE Model data loaded. Size: 2208 bytes.
I [executorch:arm_executor_runner.cpp:493 runner_init()] Model buffer loaded, has 1 methods
I [executorch:arm_executor_runner.cpp:500 runner_init()] Running method forward
I [executorch:arm_executor_runner.cpp:511 runner_init()] Setup Method allocator pool. Size: 62914560 bytes.
I [executorch:arm_executor_runner.cpp:528 runner_init()] Setting up planned buffer 0, size 96.

...

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218

[west9906]

similar thing has to be done everywhere in project. Got same issue with %zd -> %ld for file:
executorch/runtime/executor/method.cpp

[Gasoonjia]

@digantdesai mind take a look?

[west9906]

I also have a problem with example_input. In the file examples/arm/aot_arm_compiler.py for AddModule2, I changed one of the example tensors from torch.ones() to torch.zeros() and compiled it using ./examples/arm/run.sh --model_name=add2 --target=ethos-u85-2048, but I still get this:
Output[0][0]: (int) 2
Output[0][1]: (int) 2
Output[0][2]: (int) 2
Output[0][3]: (int) 2
Output[0][4]: (int) 2

[digantdesai]

%zd -> %ld
%zu -> %lu

Can you try passing -std=c99 flag? What's your compiler version? And if that works do you mind submitting a PR?

[zingo]

Hi, nice to see you are playing with the Ethos-Ubackend and the FVP :)

I also have a problem with example_input. In the file examples/arm/aot_arm_compiler.py for AddModule2, I changed one of the example tensors from torch.ones() to torch.zeros() and compiled it using ./examples/arm/run.sh --model_name=add2 --target=ethos-u85-2048, but I still get this:
Output[0][0]: (int) 2
Output[0][1]: (int) 2
Output[0][2]: (int) 2
Output[0][3]: (int) 2
Output[0][4]: (int) 2

Regarding the inputs from the python side, they are not propagated to the embeded executable, the default it that it fills all input with "1" (so you probably always end up with 1+1->2) This is (or at least was) the case with the genric example runner also, so we keept it the same. The idea is that in the runner you will connect the input to your "proper input" like cam/sound or other sensor output.

But for testing there is a few ways around it :)

1. (Recomended) Easy ways as you are already using run.sh: If you use BundelIO the pte will "bundle" example input and "pytorch" generated outputput together with the model in the pte file. When running it will use this input and compare the output when done with the "proper" one. to use this just add --bundleio to the command line for run.sh
   if you use stuff by hand or other script you can check run.sh code to see how to pass it.

You can also generate/save inputs .pt files and reuse them, there is a flag to both run.sh (and aot_arm_compiler) so you do not need to use only the input from the modelfile used (you still need to use bundleIO)

2. Change the code filling the input with "1" to something else in the runner.

3. Run FVP in semihosting mode, this is a special build of the FVP simulator where it can access the filesystem on your dev computer and you can pass the model and input as files. We use this in the unit tests flow but it would need some digging and tinkering with to get working.

I hope this helps in your case.

[west9906]

%zd -> %ld
%zu -> %lu

Can you try passing -std=c99 flag? What's your compiler version? And if that works do you mind submitting a PR?

after I added -std=c99 to add_compile_options() in examples/arm/ethos-u-setup/arm-none-eabi-gcc.cmake I`ve got this error:

[ 9%] Building C object third-party/flatcc/src/runtime/CMakeFiles/flatccrt.dir/json_printer.c.obj
[ 9%] Built target program_schema
cc1plus: warning: command-line option '-std=c99' is valid for C/ObjC but not for C++
cc1plus: warning: command-line option '-std=c99' is valid for C/ObjC but not for C++
cc1plus: warning: command-line option '-std=c99' is valid for C/ObjC but not for C++
cc1plus: error: command-line option '-std=c99' is valid for C/ObjC but not for C++ [-Werror]
cc1plus: error: command-line option '-std=c99' is valid for C/ObjC but not for C++ [-Werror]

and if I switch change it to c++11 I get:
[ 9%] Building C object third-party/flatcc/src/runtime/CMakeFiles/flatccrt.dir/json_printer.c.obj
[ 9%] Built target program_schema
cc1: error: command-line option '-std=c++11' is valid for C++/ObjC++ but not for C [-Werror]
cc1: error: command-line option '-std=c++11' is valid for C++/ObjC++ but not for C [-Werror]

I use compiler version installed by examples/arm/setup.sh :
arm-gnu-toolchain-13.3.rel1-x86_64-arm-none-eabi

The compiler standards are also already set in the same .cmake file:

set(CMAKE_C_STANDARD 11)
set(CMAKE_CXX_STANDARD 17)

And c11 is compatible with c99, so there shouldn't be any problems. I guess the issue is with FVP itself. I remember having a similar problem with printf when using the ml-embedded-evaluation-kit from arm for FVP.