Taichi Programming Language

1.7.4 · active · verified Thu Apr 16

Taichi is an open-source parallel programming language designed for high-performance visual computing. It allows users to write highly parallelized computational kernels using Python syntax, which are then compiled to run efficiently on various backends like GPUs (CUDA, Metal, Vulkan, OpenGL) or CPUs. It's currently at version 1.7.4 and maintains an active release schedule with frequent minor updates.

Common errors

```
AttributeError: module 'taichi' has no attribute 'experimental'
```
cause Attempting to use `@ti.experimental.real_func` after Taichi v1.7.0.

fix Replace `@ti.experimental.real_func` with `@ti.real_func`.
```
TypeError: ndarray() got an unexpected keyword argument 'field_dim'
```
cause Using the `field_dim` argument with `ti.ndarray` after Taichi v1.5.0.

fix Change `field_dim` to `ndim` when defining your `ti.ndarray`, e.g., `ti.ndarray(ti.f32, ndim=2)`.
```
AttributeError: 'Matrix' object has no attribute 'rotation2d'
```
cause Calling `rotation2d()` directly on a `ti.Matrix` object after Taichi v1.4.0.

fix Use `ti.math.rotation2d()` for 2D rotation matrix generation.
```
TypeError: ti.init() got an unexpected keyword argument 'packed'
```
cause Using the `packed` argument in `ti.init()` after Taichi v1.4.0.

fix Remove `packed=True/False` from your `ti.init()` call. The SNode system manages memory packing automatically.

Warnings

breaking The `@ti.experimental.real_func` decorator was deprecated in Taichi v1.7.0 and replaced by `@ti.real_func`.
Fix: Update your code to use `@ti.real_func` instead of `@ti.experimental.real_func`.
breaking The `field_dim` argument for `ti.ndarray` has been removed. Use `ndim` instead.
Fix: Replace `field_dim` with `ndim` when creating `ti.ndarray` instances. For example, `ti.ndarray(ti.f32, ndim=2)`.
breaking The `ti.Matrix.rotation2d()` method was removed. Its functionality is now provided by `ti.math.rotation2d()`.
Fix: Migrate calls from `ti.Matrix.rotation2d()` to `ti.math.rotation2d()`.
breaking The `packed` and `dynamic_index` arguments in `ti.init()` have been removed. `packed` was removed in v1.4.0, `dynamic_index` in v1.5.0.
Fix: Remove these arguments from your `ti.init()` calls. Taichi's SNode system has evolved to handle these aspects more automatically or through different configuration mechanisms.
gotcha Slicing a single row or column of a `ti.Matrix` now returns a vector, not a 1xN or Nx1 matrix, potentially leading to dimension mismatches in subsequent matrix operations.
Fix: If a matrix type is explicitly required, you may need to explicitly construct a 1xN or Nx1 matrix from the resulting vector, e.g., `ti.Matrix.rows([my_vector])` or `ti.Matrix.cols([my_vector])`.
deprecated The `ti.cc` backend is deprecated in favor of the Taichi Runtime (TiRT) and its C API, especially for Ahead-of-Time (AOT) deployment.
Fix: Consider migrating AOT workflows to use TiRT. Refer to Taichi's AOT documentation for the latest deployment strategies.

Install

pip install taichi Install latest version

Imports

taichi
```
import taichi as ti
```

ti.math

import taichi as ti
# ... later in code ...
vec = ti.math.vec3(1, 2, 3)

ti.real_func
wrong:
```
@ti.experimental.real_func
```
correct:
```
@ti.real_func
```
The experimental decorator was deprecated in v1.7.0.

Quickstart

This example initializes Taichi, declares two 2D fields, and then defines a Taichi kernel to perform a simple computation over all elements. The `@ti.kernel` decorator compiles the Python function into a high-performance kernel that runs on the chosen backend (CPU or GPU).

import taichi as ti

ti.init(arch=ti.cpu) # Try ti.gpu for GPU acceleration if available

# Declare a 2D field (similar to a NumPy array)
N = 32
x = ti.field(ti.f32, shape=(N, N))
y = ti.field(ti.f32, shape=(N, N))

@ti.kernel
def compute():
    for i, j in x: # Iterate over all elements in the field x
        x[i, j] = float(i + j) / N
        y[i, j] = x[i, j] * 2.0

if __name__ == '__main__':
    compute()
    print(f"x[0,0]: {x[0,0]}")
    print(f"y[0,0]: {y[0,0]}")
    # Fields can be converted to NumPy arrays for further processing:
    # x_np = x.to_numpy()
    # print(x_np[0,0])

view raw JSON →