Incorrect kernel argument validation and error message with ti.field

[rafmudaf]

Describe the bug
The kernel type validation and error reporting system may have an error.

On the error message, the runtime error in kernel_impl.py's set_arg_ext_array function requires the type-hint to have the .to_string method:

raise TaichiRuntimeTypeError(
    f"Argument {needed.to_string()} cannot be converted into required type {v}"
)

However, it looks like Matrix and Vector are the only types that have it. Notably, it's missing in Field. This causes the error message to say this:

AttributeError: 'NdarrayType' object has no attribute 'to_string'

Instead of something like this:

Argument NdarrayType(dtype=None, ndim=None, layout=Layout.AOS, needs_grad=None) cannot be converted into required type <class 'taichi.lang.field.ScalarField'>

See the short code below to reproduce.

Additionally, I believe the checks for PyTorch and Paddle could result in an incorrect and difficult to track error. If the user has a Python environment activate with both Paddle and PyTorch installed and they are intending to use Paddle arrays, this flow control will error prior to executing the Paddle section:

        def set_arg_ext_array(indices, v, needed):
            ... redacted ...

            if isinstance(v, np.ndarray):
                if v.flags.c_contiguous:
                    ...
                elif v.flags.f_contiguous:
                    ...
                else:
                    raise ValueError(
                        "Non contiguous numpy arrays are not supported, please call np.ascontiguousarray(arr) "
                        "before passing it into taichi kernel."
                    )
            elif has_pytorch():
                if isinstance(v, torch.Tensor):
                    ...
-- bug? -->     else:
                    raise TaichiRuntimeTypeError(
                        f"Argument {needed.to_string()} cannot asdf be converted into required type {v}"
                    )
            elif has_paddle():
                if isinstance(v, paddle.Tensor):
                    ...
                else:
                    raise TaichiRuntimeTypeError(
                        f"Argument {needed.to_string()} cannot be converted into required type {v}"
                    )
            else:
                raise TaichiRuntimeTypeError(
                    # f"Argument {needed.to_string()} cannot be converted into required type {v}"
                    f"Argument {needed} cannot be converted into required type {type(v)}"
                )

To Reproduce

import taichi as ti

ti.init(arch=ti.cpu)

@ti.kernel
def kernel_example(arr: ti.types.ndarray()):
    pass

# arr = ti.ndarray(dtype=ti.f32, shape=(10,)) # Correct arg type to kernel
arr = ti.field(dtype=ti.f32, shape=(10,)) # Incorrect arg type to kernel
arr.fill(0.0)

kernel_example(arr)

Log/Screenshots

(raf) >> (gpu *$)$ python taichi_bug_demo.py 
[Taichi] version 1.7.3, llvm 15.0.7, commit 5ec301be, osx, python 3.12.1
[Taichi] Starting on arch=arm64
Traceback (most recent call last):
  File "taichi_bug_demo.py", line 14, in <module>
    kernel_example(arr)
  File "<removed>/miniforge3/envs/raf/lib/python3.12/site-packages/taichi/lang/kernel_impl.py", line 1114, in wrapped
    return primal(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^
  File "<removed>/miniforge3/envs/raf/lib/python3.12/site-packages/taichi/lang/kernel_impl.py", line 1046, in __call__
    return self.launch_kernel(kernel_cpp, *args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "<removed>/miniforge3/envs/raf/lib/python3.12/site-packages/taichi/lang/kernel_impl.py", line 957, in launch_kernel
    recursive_set_args(needed_, type(val), val, (i - template_num,))
  File "<removed>/miniforge3/envs/raf/lib/python3.12/site-packages/taichi/lang/kernel_impl.py", line 935, in recursive_set_args
    set_arg_ext_array(indices, v, needed)
  File "<removed>/miniforge3/envs/raf/lib/python3.12/site-packages/taichi/lang/kernel_impl.py", line 835, in set_arg_ext_array
    f"Argument {needed.to_string()} cannot be converted into required type {v}"
                ^^^^^^^^^^^^^^^^
AttributeError: 'NdarrayType' object has no attribute 'to_string'

Additional comments
I'm happy to propose a fix and submit a pull request. Prior to that, though, I'd like to confirm with the development team that this is indeed an issue and not user error. Thanks!

[rafmudaf]

One more point on the error message is that the language seems backwards to me. As I understand it, the needed variable in set_arg_ext_array is the type hint to the kernel and v is the value of the argument actually passed. The current error message is:

f"Argument {needed.to_string()} cannot be converted into required type {v}"

If I'm understanding the code correctly, I think it should be:

f"Argument {v} cannot be converted into required type {needed.to_string()}"

And if we want to have this actually print instead of error with Field then it might be:

f"Argument of type {type(v)} cannot be converted into required type {needed}"

Again, I'm happy to make this change if its valid.