In a Triton kernel, a pointer identifies where tensor data begins in GPU memory; it is an address, not the tensor itself. Triton converts a PyTorch tensor to a pointer when passing it to the kernel. Loads and stores combine that base address with computed offsets to reach individual elements. section 2.2.2
def add_kernel(
x_ptr, y_ptr, z_ptr,
n_elements,
BLOCK_SIZE: tl.constexpr,
):
The kernel signature names three pointers for its input and output data. section 2.2.2



