Creating aliases for global variables and functions with cumbersome names can sometimes improve readability. In Taichi, this can be done by assigning kernel and function local variables with
ti.static(), which forces Taichi to use standard python pointer assignment.
For example, consider the simple kernel:
@ti.kernel def my_kernel(): for i, j in tensor_a: tensor_b[i, j] = some_function(tensor_a[i, j])
The tensors and function be aliased to new names with
@ti.kernel def my_kernel(): a, b, fun = ti.static(tensor_a, tensor_b, some_function) for i,j in a: b[i,j] = fun(a[i,j])
Aliases can also be created for class members and methods, which can help prevent cluttering objective data-oriented programming code with
For example, consider class kernel to compute the 2-D laplacian of some tensor:
@ti.kernel def compute_laplacian(self): for i, j in a: self.b[i, j] = (self.a[i + 1,j] - 2.0*self.a[i, j] + self.a[i-1, j])/(self.dx**2) \ + (self.a[i,j + 1] - 2.0*self.a[i, j] + self.a[i, j-1])/(self.dy**2)
ti.static(), it can be simplified to:
@ti.kernel def compute_laplacian(self): a,b,dx,dy = ti.static(self.a,self.b,self.dx,self.dy) for i,j in a: b[i,j] = (a[i+1, j] - 2.0*a[i, j] + a[i-1, j])/(dx**2) \ + (a[i, j+1] - 2.0*a[i, j] + a[i, j-1])/(dy**2)
ti.static can also be used in combination with
if (compile-time branching) and
for (compile-time unrolling). See Metaprogramming for more details.
Here, we are using it for compile-time const values, i.e. the tensor/function handles are constants at compile time.