A Simple Tutorial

TaichiZoo supports a subset of the Taichi APIs. Here we list the supported Taichi features:

1. 32-bits scalar data types

   • ti.i32: A 32-bit signed integer
   • ti.f32: A 32-bit single precision floating point

2. Dense field:

   • ti.field(dtype, shape)
   • ti.Vector.field(n, dtype, shape)
   • ti.Matrix.field(n, m, dtype, shape)

3. Function decorators

   • @ti.kernel: Note that a TaichiZoo kernel only supports the aforementioned scalars as its paramter types. That is, a TaichiZoo kernel doesn't support types like ti.template() or ti.ext_arr().
   • @ti.func

4. Taichi internal functions

   • Basic mathematical functions: ti.sin(), ti.cos(), ti.exp(), etc. You can find these functions here.
   • ti.random()
   • Note that print is not supported yet

5. GUI draw functions:

   • ti.GUI(name, res, background_color)
   • gui.set_image(pixels)
   • gui.circles(array, radius, color)
   • gui.clear(color)
   • gui.show()

6. GUI events

   • gui.running
   • gui.get_event(event)
   • gui.is_pressed(*keys)
   • gui.get_cursor_pos()

7. GUI key alias

   • ti.GUI.LEFT, ti.GUI.RIGHT, ti.GUI.UP, ti.GUI.DOWN, ti.GUI.LMB, ti.GUI.RMB, ti.GUI.PRESS, ti.GUI.ESCAPE, ti.GUI.EXIT

More details can be found on and .

Here, we will walkthrough how code is organized in TaichiZoo using the Fractal example.

1. Initialize the Taichi runtime. Note that for TaichiZoo, the kernel will always run on the CPU backend (via WebAssembly), regardless of the specified arch.

2. Allocate a field, pixels, to record the results. dtype and shape need to be specified during this process. In taichi, field is a core notion representing a multi-dimensional array.

3. Define a Taichi function, complex_sqr(), to do the real computation. The types of the parameters are inferred from the calling kernel scope. complex_sqr() accepts a single parameter z, which represents a complex number: z[0] + z[1] * i. The function returns z * z.

4. Define a Taichi kernel, paint(). This kernel simplify fills in each entry in pixels.

5. Initialize the GUI system. The size of the canvas in the browser is determined by the res parameter.

6. Finally, we have some control flow code to invoke the Taichi kernels. TaichiZoo allows for control flows like if, for and while. From python scope, we can call into Taichi kernels and GUI-related methods. gui.set_image() directly set each pixel color. We only support grayscale for this. gui.show() will update the canvas in the browser.

Please make sure that all kernels are defined before calling any of them.