Skip to main content
Version: v1.4.0

Interacting with External Arrays

This document provides instructions on how to transfer data from external arrays to the Taichi scope and vice versa. For now, the external arrays supported by Taichi are NumPy arrays, PyTorch tensors, and Paddle tensors.

We use NumPy arrays as an example to illustrate the data transfer process because NumPy arrays are the most commonly used external arrays in Taichi. The same steps apply to PyTorch tensors and Paddle tensors.

There are two ways to import a NumPy array arr to the Taichi scope:

  • Create a Taichi field f, whose shape and dtype match the shape and dtype of arr, and call f.from_numpy(arr) to copy the data in arr into f. This approach is preferred when the original array is visited frequently from elsewhere in the Taichi scope (for example, in texture sampling).

  • Pass arr as an argument to a kernel or a Taichi function using ti.types.ndarray() as type hint. The argument is passed by reference without creating a copy of arr. Thus, any modification to this argument from inside a kernel or Taichi function also changes the original array arr. This approach is preferred when the kernel or Taichi function that takes in the argument needs to process the original array (for storage or filtering, for example).


from_numpy() / from_torch() can take in any numpy array or torch Tensor, no matter it's contiguous or not. Taichi will manage its own copy of data. However, when passing an argument to a Taichi kernel, only contiguous numpy arrays or torch Tensors are supported.

Data transfer between NumPy arrays and Taichi fields

To import data from a NumPy array to a Taichi field, first make sure that the field and the array have the same shape:

x = ti.field(float, shape=(3, 3))
a = np.arange(9).reshape(3, 3).astype(np.int32)
#[[0 1 2]
# [3 4 5]
# [6 7 8]]

In the example above, the scalar field x and the array a have the same shape (3, 3). This operation would fail if their shapes did not match. Shape matching of a vector or matrix field with a NumPy array is slightly different, which will be discussed in a later section.

The field should also have the same dtype as the array; otherwise, an implicit type casting would occur - see type system.

Conversely, to export the data in x to a NumPy array, call to_numpy():

arr = x.to_numpy()
#array([[0, 1, 2],
# [3, 4, 5],
# [6, 7, 8]], dtype=int32)

Data transfer between PyTorch/Paddle tensors and Taichi fields

Data transfer between a PyTorch tensor and a Taichi field is similar to the NumPy case above: Call from_torch() for data import and to_torch() for data export. But note that to_torch() requires one more argument device, which specifies the PyTorch device:

tensor = x.to_torch(device="cuda:0")
print(tensor.device) # device(type='cuda', index=0)

For Paddle, you need to specify the device by calling paddle.CPUPlace() or paddle.CUDAPlace(n), where n is an optional ID set to 0 by default.

device = paddle.CPUPlace()
tensor = x.to_paddle(device=device)

## External array shapes

When transferring data between a `ti.field/ti.Vector.field/ti.Matrix.field` and a NumPy array, you need to make sure that the shapes of both sides are aligned. The shape matching rules are summarized as below:

1. When importing data to or exporting data from a scalar field, ensure that **the shape of the corresponding NumPy array, PyTorch tensor, or Paddle tensor equals the shape of the scalar field**

field = ti.field(int, shape=(256, 512))
field.shape # (256, 512)

array = field.to_numpy()
array.shape # (256, 512)

field.from_numpy(array) # the input array must be of shape (256, 512)

An illustration is shown below:


┌ ┌───┬───┬───┬───┬───┬───┐ ┐
│ │ │ │ │ │ │ │ │
│ ├───┼───┼───┼───┼───┼───┤ │
field.shape[0]=array.shape[0] │ │ │ │ │ │ │ │ │
(=256) │ ├───┼───┼───┼───┼───┼───┤ │
│ │ │ │ │ │ │ │ │
└ └───┴───┴───┴───┴───┴───┘ ┘

2. When importing data to or exporting data from an `n`-dimensional vector field, ensure that **the shape of the corresponding NumPy array, PyTorch tensor, or Paddle tensor is set to** `(*field_shape, n)`:

field = ti.Vector.field(3, int, shape=(256, 512))
field.shape # (256, 512)
field.n # 3

array = field.to_numpy()
array.shape # (256, 512, 3)

field.from_numpy(array) # the input array must in the shape (256, 512, 3)

An illustration is shown below:


┌ ┌─────────┬─────────┬─────────┐ ┐
│ │[*, *, *][*, *, *][*, *, *]│ │
│ ├─────────┼─────────┼─────────┤ │
field.shape[0]=array.shape[0] │ │[*, *, *][*, *, *][*, *, *]│ │ [*, *, *]
(=256) │ ├─────────┼─────────┼─────────┤ │ └───────┘
│ │[*, *, *][*, *, *][*, *, *]│ │ n=array.shape[2]=3
└ └─────────┴─────────┴─────────┘ ┘

3. When importing data to or exporting data from an `n`-by-`m` (`n x m`) matrix field, ensure that **the shape of the corresponding NumPy array, PyTorch tensor, or Paddle tensor is set to** `(*field_shape, n, m)`:

field = ti.Matrix.field(3, 4, ti.i32, shape=(256, 512))
field.shape # (256, 512)
field.n # 3
field.m # 4

array = field.to_numpy()
array.shape # (256, 512, 3, 4)

field.from_numpy(array) # the input array must be of shape (256, 512, 3, 4)

4. When importing data to a struct field, export the data of the corresponding external array as **a dictionary of NumPy arrays, PyTorch tensors, or Paddle tensors** with keys being struct member names and values being struct member arrays. Nested structs are exported as nested dictionaries:

field = ti.Struct.field({'a': ti.i32, 'b': ti.types.vector(3, float)}, shape=(256, 512))
field.shape # (256, 512)

array_dict = field.to_numpy()
array_dict.keys() # dict_keys(['a', 'b'])
array_dict['a'].shape # (256, 512)
array_dict['b'].shape # (256, 512, 3)

field.from_numpy(array_dict) # the input array must have the same keys as the field

## Using external arrays as Taichi kernel arguments

Use type hint `ti.types.ndarray()` to pass external arrays as kernel arguments.

### An entry-level example

The following example shows the most basic way to call `ti.types.ndarray()`:

```python {10}
import taichi as ti
import numpy as np

a = np.zeros((5, 5))

def test(a: ti.types.ndarray()):
for i in range(a.shape[0]): # a parallel for loop
for j in range(a.shape[1]):
a[i, j] = i + j


Advanced usage

Assume that a and b are both 2D arrays of the same shape and dtype. For each cell (i, j) in a, we want to calculate the difference between its value and the average of its four neighboring cells while storing the result in the corresponding cell in b. In this case, cells on the boundary, which are cells with fewer than four neighbors, are ruled out for simplicity. This operation is usually denoted as the Discrete Laplace Operator:

b[i, j] = a[i, j] - (a[i-1, j] + a[i, j-1] + a[i+1, j] + a[i, j+1]) / 4

Such an operation is typically very slow, even with NumPy's vectorization as shown below:

b[1:-1, 1:-1] += (               a[ :-2, 1:-1] +
a[1:-1, :-2] + a[1:-1, 2:] +
a[2: , 1:-1])

But Taichi can meet the same purpose in one parallel for loop only:

def test(a: ti.types.ndarray(), b: ti.types.ndarray()): # assume a, b have the same shape
H, W = a.shape[0], a.shape[1]
for i, j in ti.ndrange(H, W): # one parallel for loop
if 0 < i < H - 1 and 0 < j < W - 1:
b[i, j] = a[i, j] - (a[i-1, j] + a[i, j-1] + a[i+1, j] + a[i, j+1]) / 4

Not only is this code snippet more readable than the NumPy version above, but it also runs way faster even on the CPU backend.


The elements in an external array must be indexed using a single square bracket. This contrasts with a Taichi vector field or matrix field where field members and elements are indexed separately:

x = ti.Vector.field(3, float, shape=(5, 5))
y = np.random.random((5, 5, 3))

def copy_vector(x: ti.template(), y: ti.types.ndarray()):
for i, j in ti.ndrange(5, 5):
for k in ti.static(range(3)):
y[i, j, k] = x[i, j][k] # correct
# y[i][j][k] = x[i, j][k] incorrect
# y[i, j][k] = x[i, j][k] incorrect

In addition, external arrays in a Taichi kernel are indexed using their physical memory layout. For PyTorch users, this means that a PyTorch tensor needs to be made contiguous before being passed into a Taichi kernel:

x = ti.field(dtype=int, shape=(3, 3))
y = torch.Tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
y = y.T # Transposing the tensor returns a view of the tensor which is not contiguous

def copy_scalar(x: ti.template(), y: ti.types.ndarray()):
for i, j in x:
y[i, j] = x[i, j]

copy(x, y) # error!
copy(x, y.clone()) # correct
copy(x, y.contiguous()) # correct


Can I use @ti.kernel to accelerate a NumPy function?

Unlike other Python acceleration frameworks, such as Numba, Taichi does not compile NumPy functions. Calling NumPy functions inside the Taichi scope is not supported, as the following example shows:

import numpy as np

def invalid_sum(arr: ti.types.ndarray()):
total = np.sum(arr) # Not supported!

If you want to use a NumPy function, which lacks a counterpart in Taichi, you can call the function in the Python scope as usual and pass the processed array to Taichi kernels via ti.types.ndarray(). For example:

indices = np.argsort(arr)  # arr is a Numpy.ndarray

def valid_example(arr: ti.types.ndarray(), indices: ti.types.ndarray()):
min_element = arr[indices[0]]