[Bug] Ban passing torch view tensors into taichi kernel

[ailzhang]

fixes #4208

Related issue = #

[netlify]

✔️ Deploy Preview for docsite-preview ready!

🔨 Explore the source changes: 4e07bd5

🔍 Inspect the deploy log: https://app.netlify.com/sites/docsite-preview/deploys/62012d0c3f51f600073e13f7

😎 Browse the preview: https://deploy-preview-4225--docsite-preview.netlify.app

[victoriacity]

LGTM Thanks!

[ifsheldon]

This is a breaking change, so I suggest at least bump the version by +0.1 according to semantic versioning

[ifsheldon]

Besides, this commit affects field.from_torch(), which impacts beyond the scope of ti.ndarray mentioned in #4225.

Moreover, the impact is weird. Can you test this case?

@ti.kernel
def ti_print(a: ti.template()):
    for i in ti.grouped(a):
        print(f"{i[0]} {i[1]}, {a[i]}")


def test_view():
    ti.init(ti.cpu)
    a = torch.tensor([[1, 2], [3, 4]], dtype=torch.float32)
    a_t = a.T
    assert a_t._is_view()
    a_field = ti.field(ti.f32, shape=a.shape)
    a_field.from_torch(a_t)  # ok, but why?
    ti_print(a_field)
    a_complex = torch.ones(2, 2, dtype=torch.cfloat)
    a_real_view = torch.view_as_real(a_complex)
    assert a_real_view._is_view()
    complex_field = ti.field(ti.f32, shape=a_real_view.shape)
    complex_field.from_torch(a_real_view)  # not ok
    ti_print(complex_field)

This impact on field.from_torch() breaks my library that depends on Taichi, but I don't know if it's expected.

Can you please check the commit? @ailzhang

My test log is

tests/test_ti.py:36 (test_view)
def test_view():
        ti.init(ti.cpu)
        a = torch.tensor([[1, 2], [3, 4]], dtype=torch.float32)
        a_t = a.T
        assert a_t._is_view()
        a_field = ti.field(ti.f32, shape=a.shape)
        a_field.from_torch(a_t)  # ok, but why?
        ti_print(a_field)
        a_complex = torch.ones(2, 2, dtype=torch.cfloat)
        a_real_view = torch.view_as_real(a_complex)
        assert a_real_view._is_view()
        complex_field = ti.field(ti.f32, shape=a_real_view.shape)
>       complex_field.from_torch(a_real_view)  # not ok

test_ti.py:49: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/util.py:226: in wrapped
    return func(*args, **kwargs)
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/field.py:146: in from_torch
    self.from_numpy(arr.contiguous())
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/util.py:226: in wrapped
    return func(*args, **kwargs)
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/field.py:264: in from_numpy
    ext_arr_to_tensor(arr, self)
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/kernel_impl.py:740: in wrapped
    return primal(*args, **kwargs)
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/shell.py:37: in new_call
    ret = old_call(*args, **kwargs)
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/kernel_impl.py:671: in __call__
    return self.compiled_functions[key](*args)
../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/kernel_impl.py:579: in func__
    tmp, torch_callbacks = self.get_torch_callbacks(
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <taichi.lang.kernel_impl.Kernel object at 0x7fbfcc516610>
v = tensor([[[1., 0.],
         [1., 0.]],

        [[1., 0.],
         [1., 0.]]])
has_torch = True, is_ndarray = False

    def get_torch_callbacks(self, v, has_torch, is_ndarray=True):
        callbacks = []
    
        def get_call_back(u, v):
            def call_back():
                u.copy_(v)
    
            return call_back
    
        assert has_torch
        assert isinstance(v, torch.Tensor)
        if v._is_view():
>           raise ValueError(
                "Torch view tensors are not supported, please call tensor.clone() before passing it into taichi kernel."
            )
E           ValueError: Torch view tensors are not supported, please call tensor.clone() before passing it into taichi kernel.

../../../miniconda3/envs/build_ti/lib/python3.9/site-packages/taichi-0.8.12-py3.9-linux-x86_64.egg/taichi/lang/kernel_impl.py:491: ValueError