Bug Description

Passing shared-memory tiles to @wp.func fails to compile

Problem

When a @wp.func has a tile parameter (e.g., wp.tile(dtype=float, shape=(M, N))), the tile annotation defaults to register storage. If the function is called with a shared-memory tile, the generated C++ fails to compile because the function signature uses tile_register_t but the caller passes tile_shared_t.

Reproduction

import warp as wp
import numpy as np

BLOCK_DIM = 64
TILE_M = 4
TILE_N = 4

@wp.func
def scale_tile(t: wp.tile(dtype=float, shape=(TILE_M, TILE_N))):
    return t * 2.0

@wp.kernel(enable_backward=False)
def compute(input: wp.array2d(dtype=float), out: wp.array2d(dtype=float)):
    i = wp.tid()
    t = wp.tile_load(input, shape=(TILE_M, TILE_N), offset=(0, 0), storage="shared")
    r = scale_tile(t)  # fails here
    wp.tile_store(out, r)

inp = wp.array(np.ones((TILE_M, TILE_N), dtype=np.float32) * 3.0, device="cpu")
out = wp.zeros((TILE_M, TILE_N), dtype=float, device="cpu")
wp.launch_tiled(compute, dim=[1], inputs=[inp, out], block_dim=BLOCK_DIM, device="cpu")

Error:

error: no matching function for call to 'scale_tile_0'
note: candidate function not viable: no known conversion from
  'tile_shared_t<float, tile_layout_strided_t<...>>'
  to 'tile_register_t<float, tile_layout_register_t<...>>'

Impact

Users cannot pass shared-memory tiles to @wp.func functions. This is a common pattern — tile_load with storage="shared" and tile_matmul both produce shared tiles, and users often want to wrap tile operations in helper functions.

The workaround is to avoid @wp.func and inline the tile operations directly in the kernel, which hurts code reuse and readability.

System Information

No response