warp.tile_assign#

warp.tile_assign( dst: Tile[Any, tuple[int, ...]], src: Tile[Any, tuple[int, ...]], offset: tuple[int, ...] = ..., ) → None#

Kernel: true
Python: false
Differentiable: true

Copy a tile into a subrange of a destination tile.

dst is modified in place and requires shared storage.

When src and dst have different element types, each element is converted following C++ conversion rules. Overlapping source and destination regions assign like NumPy; t[1:] = t[:-1] shifts the tile.

In a backward pass, gradients from the overwritten region of dst are accumulated into the adjoint of src, then cleared from that region of dst.

Parameters:

dst – Destination tile, modified in place. Must have the same number of dimensions as src.
src – Source tile. Must fit inside dst at offset along every axis.
offset – Coordinate in dst at which to write src. If omitted, src is written at the origin. Must have one entry per dimension of dst and may contain runtime values.

Example

@wp.kernel
def insert_values(a: wp.array[float], out: wp.array[float]):
    dst = wp.tile_full(shape=6, value=-1.0, dtype=float)
    src = wp.tile_load(a, shape=3)

    wp.tile_assign(dst, src, offset=(2,))

    wp.tile_store(out, dst)

a = wp.array([1.0, 2.0, 3.0], dtype=float)
out = wp.zeros(6, dtype=float)

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

print(out.numpy())

[-1. -1.  1.  2.  3. -1.]