Attributing and assigning to a ndarray

[fccoelho]

Here is the problematic code:

import numpy as np
import taichi as ti
import copy

ti.init(arch=ti.cpu,default_ip=ti.i32, default_fp=ti.f32)

@ti.data_oriented
class Test:
    def __init__(self, ini):
        self.ini = ti.ndarray(shape=(10,), dtype=ti.i32)
        self.ini.from_numpy(ini)

    @ti.kernel
    def run(self):
        ini = self.ini
        print(ini)

if __name__ == '__main__':
    ini = np.arange(10)
    test = Test(ini)
    test.run()
    print(test.ini.to_numpy())

In it, I am simply trying to create a copy of self.ini in side a kernel function.

I get the following stack trace:

Invalid constant scalar data type: <class 'taichi.lang._ndarray.ScalarNdarray'>

If instead I try to attribute to the ndarray within the kernel function,

self.ini[5] = 100

I get the following error:

AssertionError: __getitem__ cannot be called in Taichi-scope

What am I doing wrong?

[bobcao3]

Ndarrays can not be captured through self / data oriented at this moment

[fccoelho]

Yes, I managed to work around this by using ti.field defining my "temp" variable in the constructor and filling it with the values of self.ini in a for loop in my kernel function (aargh).

[johnnybigert]

If you want to use ndarrays in a class, you can work around it like this:

import numpy as np
import taichi as ti

ti.init(arch=ti.cpu,default_ip=ti.i32, default_fp=ti.f32)

NdarrayType = ti.types.ndarray(dtype=ti.i32, ndim=1) 


@ti.data_oriented
class Test:
    ini: NdarrayType
    
    def __init__(self, ini):
        self.ini = ti.ndarray(shape=(10,), dtype=ti.i32)
        self.ini.from_numpy(ini)

    @ti.kernel
    def run_does_not_work(self):
        self.ini[5] = 100
        
    def run(self):
        self._run(self.ini)

    @staticmethod
    @ti.kernel
    def _run(ti_ini: NdarrayType):
        ti_ini[5] = 100

ini = np.arange(10, dtype=np.int32)
test = Test(ini)
test.run()
# test.run_does_not_work()  # crashes with `__getitem__ cannot be called in Taichi-scope`