warp.tile_empty#

warp.tile_empty(
shape: tuple[int, ...],
dtype: Any = float,
storage: str = 'register',
) → Tile[Any, tuple[int, ...]]#
  • Kernel

Allocate a tile of uninitialized items.

The tile’s contents are undefined; overwrite every element before any read. This matches the semantics of numpy.empty.

Because it skips initialization, tile_empty can avoid unnecessary stores when every element will be overwritten, especially for "shared" tiles. For accumulator patterns (a += ...), use tile_zeros() instead.

Parameters:

  • shape – Shape of the output tile. Must be a compile-time constant.

  • dtype – Data type of output tile’s elements. Must be a compile-time constant.

  • storage – The storage location for the tile: "register" for registers or "shared" for shared memory. Must be a compile-time constant.

Returns:

An uninitialized tile with the requested shape and data type.

Example

@wp.kernel
def concatenate(a: wp.array[float], b: wp.array[float], out: wp.array[float]):
    # every element is written below, so skipping initialization is safe
    t = wp.tile_empty(shape=(8,), dtype=float, storage="shared")
    wp.tile_assign(t, wp.tile_load(a, shape=(4,)), offset=(0,))
    wp.tile_assign(t, wp.tile_load(b, shape=(4,)), offset=(4,))
    wp.tile_store(out, t)

a = wp.array([1.0, 2.0, 3.0, 4.0], dtype=float)
b = wp.array([10.0, 20.0, 30.0, 40.0], dtype=float)
out = wp.zeros(8, dtype=float)

wp.launch_tiled(concatenate, dim=1, inputs=[a, b], outputs=[out], block_dim=4)

print(out.numpy())

[ 1.  2.  3.  4. 10. 20. 30. 40.]
warp.tile_empty(
shape: int32,
dtype: Any = float,
storage: str = 'register',
) → Tile[Any, tuple[int, ...]]
  • Kernel

Allocate a tile of uninitialized items.

Overload for 1D tiles: shape is the number of elements, equivalent to passing (shape,). See the overload taking a tuple-valued shape argument for usage details and an example.