asc2.global_tensor

asc2.global_tensor(base: GlobalAddress, shape: Iterable[PlainValue | int]) → GlobalTensor

Define a new tensor descriptor for accessing data in global memory.

Tensors represent contiguous ND-arrays in global memory and are used to transfer data between global and local memory via copy_in() and copy_out() operations.

Parameters:

  • base – The base address of an array in global memory representing the tensor

  • shape – An iterable of integer-like values representing the number of elements for each dimension

Returns:

A new tensor descriptor

Return type:

GlobalTensor

Raises:

  • TypeError – If base is not a GlobalAddress or shape contains non-integer values

  • RuntimeError – If shape is empty

Examples

Create a 1D tensor with static shape:

x_gm = asc2.global_tensor(x_ptr, [1024])

Create a 2D tensor with static shape:

x_gm = asc2.global_tensor(x_ptr, [64, 128])

Create a tensor with dynamic shape (using runtime values):

x_gm = asc2.global_tensor(x_ptr, [num_rows, num_cols])