Structural nodes (SNodes)

After writing the computation code, the user needs to specify the internal data structure hierarchy. Specifying a data structure includes choices at both the macro level, dictating how the data structure components nest with each other and the way they represent sparsity, and the micro level, dictating how data are grouped together (e.g. structure of arrays vs. array of structures). Taichi provides Structural Nodes (SNodes) to compose the hierarchy and particular properties. These constructs and their semantics are listed below:

  • dense: A fixed-length contiguous array.
  • bitmasked: This is similar to dense, but it also uses a mask to maintain sparsity information, one bit per child.
  • pointer: Store pointers instead of the whole structure to save memory and maintain sparsity.
  • dynamic: Variable-length array, with a predefined maximum length. It serves the role of std::vector in C++ or list in Python, and can be used to maintain objects (e.g. particles) contained in a block.

See Advanced dense layouts for more details. ti.root is the root node of the data structure., ...)
  • snode – (SNode) where to place
  • x – (tensor) tensor(s) to be placed

(SNode) the snode itself

The following code places two 0-D tensors named x and y:

x = ti.var(dt=ti.i32)
y = ti.var(dt=ti.f32), y)
Parameters:tensor – (Tensor)
Returns:(tuple of integers) the shape of tensor

For example,

ti.root.dense(ti.ijk, (3, 5, 4)).place(x)
x.shape() # returns (3, 5, 4)
  • snode – (SNode)
  • index – axis (0 for i and 1 for j)

(scalar) the size of tensor along that axis

Equivalent to tensor.shape()[i].

ti.root.dense(ti.ijk, (3, 5, 4)).place(x)
x.snode().get_shape(0)  # 3
x.snode().get_shape(1)  # 5
x.snode().get_shape(2)  # 4
Parameters:tensor – (Tensor)
Returns:(scalar) the dimensionality of the tensor

Equivalent to len(tensor.shape()).

ti.root.dense(ti.ijk, (8, 9, 10)).place(x)
x.dim()  # 3
Parameters:snode – (SNode)
Returns:(SNode) the parent node of snode
blk1 = ti.root.dense(ti.i, 8)
blk2 = blk1.dense(ti.j, 4)
blk3 = blk2.bitmasked(ti.k, 6)
blk1.parent()  # ti.root
blk2.parent()  # blk1
blk3.parent()  # blk2

Node types

snode.dense(indices, shape)
  • snode – (SNode) parent node where the child is derived from
  • indices – (Index or Indices) indices used for this node
  • shape – (scalar or tuple) shape the tensor of vectors

(SNode) the derived child node

The following code places a 1-D tensor of size 3:

x = ti.var(dt=ti.i32)
ti.root.dense(ti.i, 3).place(x)

The following code places a 2-D tensor of shape (3, 4):

x = ti.var(dt=ti.i32)
ti.root.dense(ti.ij, (3, 4)).place(x)


If shape is a scalar and there are multiple indices, then shape will be automatically expanded to fit the number of indices. For example,

snode.dense(ti.ijk, 3)

is equivalent to

snode.dense(ti.ijk, (3, 3, 3))
snode.dynamic(index, size, chunk_size = None)
  • snode – (SNode) parent node where the child is derived from
  • index – (Index) the dynamic node indices
  • size – (scalar) the maximum size of the dynamic node
  • chunk_size – (optional, scalar) the number of elements in each dynamic memory allocation chunk

(SNode) the derived child node

dynamic nodes acts like std::vector in C++ or list in Python. Taichi’s dynamic memory allocation system allocates its memory on the fly.

The following places a 1-D dynamic tensor of maximum size 16:

ti.root.dynamic(ti.i, 16).place(x)

TODO: add descriptions here

Working with dynamic SNodes

ti.length(snode, indices)
  • snode – (SNode, dynamic)
  • indices – (scalar or tuple of scalars) the dynamic node indices

(scalar) the current size of the dynamic node

ti.append(snode, indices, val)
  • snode – (SNode, dynamic)
  • indices – (scalar or tuple of scalars) the dynamic node indices
  • val – (depends on SNode data type) value to store

(int32) the size of the dynamic node, before appending

Inserts val into the dynamic node with indices indices.

Taichi tensors like powers of two

Non-power-of-two tensor dimensions are promoted into powers of two and thus these tensors will occupy more virtual address space. For example, a (dense) tensor of size (18, 65) will be materialized as (32, 128).


ti.indices(a, b, ...)