iron: accept tile=None in Worker, ObjectFifoLink, RuntimeEndpoint

Worker, ObjectFifoLink (used by split/forward/join), and RuntimeEndpoint
(used by rt.fill/rt.drain) now silently accept tile=None, treating it
the same as the default (AnyComputeTile, AnyMemTile, or AnyShimTile
respectively). This allows callers to pass None to mean "unplaced"
without needing conditional keyword splat patterns.

Signatures and defaults are unchanged — existing code is unaffected.

Refactor mobilenet builders to use direct tile=value instead of the
**tile_kw() splat helper, which is now removed.

Co-Authored-By: Claude Opus 4 (1M context) <noreply@anthropic.com>

Author: yenjames

programming_examples/ml/mobilenet/aie2_mobilenet_iron.py

@@ -38,7 +38,6 @@
 
 sys.path.insert(0, str(pathlib.Path(__file__).parent))
 
-from bottleneck._common import tile_kw
 from bottleneck.init import init_conv
 from bottleneck.regular import regular_bottlenecks
 from bottleneck.pipeline import pipeline_bottlenecks
@@ -168,7 +167,7 @@ def _wts_tap(byte_offset, byte_size):
         rt.fill(
             act_in.prod(depth=1),
             inp,
-            **tile_kw(shim.get("input")),
+            tile=shim.get("input"),
             task_group=tg1,
         )
         # bn13/14 L1+L3 weight chunks from the combined cascade buffer
@@ -180,7 +179,7 @@ def _wts_tap(byte_offset, byte_size):
                 fifo.prod(),
                 cascade_wts,
                 _wts_tap(off, sz),
-                **tile_kw(s),
+                tile=s,
                 task_group=tg1,
             )
         # Round-trip avgpool output through L3 (shim 30/40 hop). Reuse `inp`
@@ -203,7 +202,7 @@ def _wts_tap(byte_offset, byte_size):
             tap=_post_l1_scratch_tap,
             wait=True,
             task_group=tg1,
-            **tile_kw(shim.get("scratch_drain")),
+            tile=shim.get("scratch_drain"),
         )
         rt.finish_task_group(tg1)
 
@@ -218,7 +217,7 @@ def _wts_tap(byte_offset, byte_size):
             inp,
             tap=_post_l1_scratch_tap,
             task_group=tg2,
-            **tile_kw(shim.get("fc_fill")),
+            tile=shim.get("fc_fill"),
         )
         _post_fc_out_tap = TensorAccessPattern(
             (_inp_sz_i32,),
@@ -232,7 +231,7 @@ def _wts_tap(byte_offset, byte_size):
             tap=_post_fc_out_tap,
             wait=True,
             task_group=tg2,
-            **tile_kw(shim.get("fc_drain")),
+            tile=shim.get("fc_drain"),
         )
         rt.finish_task_group(tg2)
 
@@ -243,14 +242,14 @@ def _wts_tap(byte_offset, byte_size):
             inp,
             tap=_post_fc_out_tap,
             task_group=tg3,
-            **tile_kw(shim.get("fc_fill")),
+            tile=shim.get("fc_fill"),
         )
         rt.drain(
             act_out_of.cons(),
             out,
             wait=True,
             task_group=tg3,
-            **tile_kw(shim.get("fc_drain")),
+            tile=shim.get("fc_drain"),
         )
         rt.finish_task_group(tg3)
 

programming_examples/ml/mobilenet/bottleneck/_common.py

@@ -63,21 +63,3 @@ def layer_sf(blk, sf, idx):
 def skip_sf(blk, sf):
     """Scale factor for the skip-add (only valid when blk.skip is True)."""
     return sf[sf_key(blk.name)][blk.skip_sf_key]
-
-
-def tile_kw(tile_or_dict, dict_key=None, *, kw="tile"):
-    """Build a keyword-arg dict for optional tile placement.
-
-    Returns ``{kw: tile}`` when a tile is provided, or ``{}`` when ``None``.
-    Intended for ``**tile_kw(...)`` splat into Worker / ObjectFifo calls so
-    that placement is only passed when explicitly requested.
-
-    Args:
-        tile_or_dict: A Tile, a dict of tiles, or None.
-        dict_key: If *tile_or_dict* is a dict, extract this key.
-        kw: Keyword name to use (default ``"tile"``).
-    """
-    if tile_or_dict is None:
-        return {}
-    tile = tile_or_dict[dict_key] if dict_key is not None else tile_or_dict
-    return {kw: tile}

programming_examples/ml/mobilenet/bottleneck/cascade.py

@@ -34,7 +34,6 @@
     load_wts,
     layer_sf as _layer_sf,
     skip_sf as _skip_sf,
-    tile_kw,
 )
 from network_spec import block as nsblock
 
@@ -332,21 +331,22 @@ def build_cascade(blk, l3_get_sym, act_in, skip_in, sf, *, data_dir, tiles=None)
     )
 
     # Streaming weight fifos (Shim → MemTile → split → put/get tiles)
+    t = tiles.get if tiles else lambda k: None
     wts_l1_full = ObjectFifo(_ty_l1_full_wts, depth=1)
     wts_l1_put_h, wts_l1_get_h = wts_l1_full.cons().split(
         offsets=[0, _l1_full_wts_sz // 2],
         depths=[1, 1],
         obj_types=[_ty_l1_split_wts, _ty_l1_split_wts],
         repeat_counts=[_InH, _InH],
-        **tile_kw(tiles, "mem_l1"),
+        tile=t("mem_l1"),
     )
     wts_l3_full = ObjectFifo(_ty_l3_full_wts, depth=1)
     wts_l3_put_h, wts_l3_get_h = wts_l3_full.cons().split(
         offsets=[0, _l3_full_wts_sz // 2],
         depths=[1, 1],
         obj_types=[_ty_l3_split_wts, _ty_l3_split_wts],
         repeat_counts=[_InH, _InH],
-        **tile_kw(tiles, "mem_l3"),
+        tile=t("mem_l3"),
     )
 
     # L2 DW weights are static (compile-time bake-in)
@@ -377,7 +377,7 @@ def build_cascade(blk, l3_get_sym, act_in, skip_in, sf, *, data_dir, tiles=None)
     l1_get_cons = act_in.cons()
     # depth=6 on the cons handle lets the skip path buffer enough rows to
     # outlive the 5-tile cascade pipeline lag (l1_put → l1_get → l2 → l3_put → l3_get).
-    skip_fifo = skip_in.cons(depth=6).forward(depth=2, **tile_kw(tiles, "mem_skip"))
+    skip_fifo = skip_in.cons(depth=6).forward(depth=2, tile=t("mem_skip"))
 
     bws = [
         Worker(
@@ -394,7 +394,7 @@ def build_cascade(blk, l3_get_sym, act_in, skip_in, sf, *, data_dir, tiles=None)
                 _OC8,
                 s1,
             ],
-            **tile_kw(tiles, "l1_put"),
+            tile=t("l1_put"),
         ),
         Worker(
             _l1_get_fn,
@@ -411,7 +411,7 @@ def build_cascade(blk, l3_get_sym, act_in, skip_in, sf, *, data_dir, tiles=None)
                 _OC8,
                 s1,
             ],
-            **tile_kw(tiles, "l1_get"),
+            tile=t("l1_get"),
         ),
         Worker(
             _l2_fn,
@@ -425,7 +425,7 @@ def build_cascade(blk, l3_get_sym, act_in, skip_in, sf, *, data_dir, tiles=None)
                 _L1_OutC,
                 s2,
             ],
-            **tile_kw(tiles, "l2"),
+            tile=t("l2"),
         ),
         Worker(
             _l3_put_fn,
@@ -441,7 +441,7 @@ def build_cascade(blk, l3_get_sym, act_in, skip_in, sf, *, data_dir, tiles=None)
                 _OC8_out,
                 s3,
             ],
-            **tile_kw(tiles, "l3_put"),
+            tile=t("l3_put"),
         ),
         Worker(
             _l3_get_fn,
@@ -460,7 +460,7 @@ def build_cascade(blk, l3_get_sym, act_in, skip_in, sf, *, data_dir, tiles=None)
                 s3,
                 s_add,
             ],
-            **tile_kw(tiles, "l3_get"),
+            tile=t("l3_get"),
         ),
     ]
     # Cascade flows: L1 put→get and L3 put→get share streams between adjacent tiles.

programming_examples/ml/mobilenet/bottleneck/init.py

@@ -19,7 +19,7 @@
 from aie.iron import Buffer, Kernel, ObjectFifo, Worker
 from aie.iron.controlflow import range_
 
-from bottleneck._common import i8, u8, load_wts, tile_kw
+from bottleneck._common import i8, u8, load_wts
 from network_spec import block as nsblock
 
 
@@ -135,7 +135,7 @@ def init_fn(act_in, act_out, wts, k, inW, inH, inC, outW, outH, outC, sf):
             init_OutC,
             init_scaleFactor,
         ],
-        **tile_kw(placement),
+        tile=placement,
     )
 
     return [w_init], act_in, act_init_out

programming_examples/ml/mobilenet/bottleneck/pipeline.py

@@ -26,7 +26,6 @@
     load_wts as _load_weights,
     layer_sf as _layer_sf,
     skip_sf as _skip_sf,
-    tile_kw,
     wts_buffer as _wts_buf,
 )
 from network_spec import block as nsblock
@@ -129,26 +128,25 @@ def call(r_in, r_out, _):
         # L3 reads a forwarded skip path — caller passes skip_in already; we
         # only need its .cons() handle for this Worker.
         skip_h = skip_in.cons()
+    t = tiles.get if tiles else lambda k: None
     workers = [
-        Worker(l1_fn, [l1_in_h, of_12.prod(), l1_wts, k_l1], **tile_kw(tiles, "l1")),
-        Worker(
-            l2_fn, [of_12.cons(), of_23.prod(), l2_wts, k_l2], **tile_kw(tiles, "l2")
-        ),
+        Worker(l1_fn, [l1_in_h, of_12.prod(), l1_wts, k_l1], tile=t("l1")),
+        Worker(l2_fn, [of_12.cons(), of_23.prod(), l2_wts, k_l2], tile=t("l2")),
     ]
     if has_skip:
         workers.append(
             Worker(
                 l3_fn,
                 [of_23.cons(), skip_h, out_fifo.prod(), l3_wts, k_l3],
-                **tile_kw(tiles, "l3"),
+                tile=t("l3"),
             )
         )
     else:
         workers.append(
             Worker(
                 l3_fn,
                 [of_23.cons(), out_fifo.prod(), l3_wts, k_l3],
-                **tile_kw(tiles, "l3"),
+                tile=t("l3"),
             )
         )
     return out_fifo, workers
@@ -305,11 +303,12 @@ def _pw():
         dw_tmp_prod.release(1)
         _pw()
 
+    t = tiles.get if tiles else lambda k: None
     workers = [
         Worker(
             bn12_l1_fn,
             [act_in.cons(), bn12_of_12.prod(), bn12_l1_wts, k_bn12_l1],
-            **tile_kw(tiles, "l1"),
+            tile=t("l1"),
         ),
         Worker(
             bn12_l23_fn,
@@ -322,7 +321,7 @@ def _pw():
                 k_bn12_dw,
                 k_bn12_pw,
             ],
-            **tile_kw(tiles, "l23"),
+            tile=t("l23"),
         ),
     ]
     return act_bn12_out, workers
@@ -354,7 +353,9 @@ def pipeline_bottlenecks(
         tiles = p.get(name)
         skip_in = None
         if has_skip:
-            skip_in = act.cons(depth=6).forward(depth=2, **tile_kw(tiles, "mem_skip"))
+            skip_in = act.cons(depth=6).forward(
+                depth=2, tile=tiles.get("mem_skip") if tiles else None
+            )
             # Strip mem_skip so only l1/l2/l3 are passed to the builder.
             if tiles is not None:
                 tiles = {k: tiles[k] for k in ("l1", "l2", "l3")}

programming_examples/ml/mobilenet/bottleneck/post_l1.py

@@ -22,7 +22,7 @@
 from aie.iron.dataflow.endpoint import ObjectFifoEndpoint
 from aie.iron.device import AnyMemTile
 
-from bottleneck._common import i8, load_wts, tile_kw
+from bottleneck._common import i8, load_wts
 from network_spec import block as nsblock
 
 
@@ -154,7 +154,7 @@ def post_l1_fn(
         # loop intact with 1 call site). The dynamic lowering uses runtime
         # modulo indexing, preserving the loop structure.
         dynamic_objfifo_lowering=True,
-        **tile_kw(placement, "compute"),
+        tile=placement["compute"] if placement else None,
     )
 
     return [w_post_l1], act_out_post_avgpool_shim

programming_examples/ml/mobilenet/bottleneck/post_l2.py

@@ -21,7 +21,7 @@
 from aie.iron.dataflow.endpoint import ObjectFifoEndpoint
 from aie.iron.device import AnyMemTile
 
-from bottleneck._common import i8, load_wts, tile_kw
+from bottleneck._common import i8, load_wts
 from network_spec import block as nsblock
 
 
@@ -76,7 +76,7 @@ def post_l2(act_in, sf, *, placement=None, data_dir):
         offsets=[i * fc_out_per_tile for i in range(n_fc_tiles)],
         depths=[2] * n_fc_tiles,
         obj_types=[np.ndarray[(co,), np.dtype[np.uint16]]] * n_fc_tiles,
-        **tile_kw(placement, "join_memtile"),
+        tile=placement["join_memtile"] if placement else None,
     )
 
     def _u16(shape):

programming_examples/ml/mobilenet/bottleneck/regular.py

@@ -30,7 +30,6 @@
     u8 as _u8,
     layer_sf as _layer_sf,
     skip_sf as _skip_sf,
-    tile_kw,
     wts_buffer as _wts_buffer,
 )
 from network_spec import block as nsblock
@@ -270,7 +269,7 @@ def _pw_pair(rows_in):
             k_l3,
         ],
         while_true=False,
-        **tile_kw(tile),
+        tile=tile,
     )
     return out_fifo, worker
 
@@ -386,7 +385,7 @@ def _skip(skip_row):
             k_skip,
         ],
         while_true=False,
-        **tile_kw(tile),
+        tile=tile,
     )
     return out_fifo, worker
 
@@ -473,14 +472,12 @@ def _kernels(name, dw_ch, in_c_local, out_c_local, l1_sz, l2_sz, l3_sz):
     out_fifo = ObjectFifo(_i8((in_w, 1, b_out_c)), depth=out_depth)
 
     # Self-loop fifos (no synchronization — single core).
-    _dt = tile_kw(alloc_tile, kw="delegate_tile")
-
     def _of(ch, depth):
         return ObjectFifo(
             _u8((in_w, 1, ch)),
             depth=depth,
             disable_synchronization=True,
-            **_dt,
+            delegate_tile=alloc_tile,
         )
 
     f_a12 = _of(a_dw_ch, 3)
@@ -489,7 +486,7 @@ def _of(ch, depth):
         _i8((in_w, 1, a_out_c)),
         depth=2,
         disable_synchronization=True,
-        **_dt,
+        delegate_tile=alloc_tile,
     )
     f_b12 = _of(b_dw_ch, 3)
     f_b23 = _of(b_dw_ch, 1)
@@ -666,7 +663,7 @@ def b_sliding_iter():
             kb_skip,
         ],
         while_true=False,
-        **tile_kw(compute_tile),
+        tile=compute_tile,
     )
     return out_fifo, worker
 

python/iron/dataflow/objectfifo.py

@@ -779,7 +779,7 @@ def __init__(
         Args:
             srcs (list[ObjectFifoHandle] | ObjectFifoHandle): A list of consumer ObjectFifoHandles to link.
             dsts (list[ObjectFifoHandle] | ObjectFifoHandle): A list of producer ObjectFifoHandles to link.
-            tile (Tile, optional): The tile where the link occurs. Defaults to AnyMemTile.
+            tile (Tile, optional): The tile where the link occurs. Also accepts None (treated as AnyMemTile). Defaults to AnyMemTile.
             src_offsets (list[int], optional): If many sources, one offset per source is required to split the destination. Defaults to [].
             dst_offsets (list[int], optional): If many destinations, one offset per destination is required to split the source. Defaults to [].
 
@@ -813,6 +813,8 @@ def __init__(
             s.endpoint = self
         for d in self._dsts:
             d.endpoint = self
+        if tile is None:
+            tile = AnyMemTile
         tile = tile.copy()
         if tile.tile_type is None:
             tile.tile_type = AIETileType.MemTile

python/iron/runtime/endpoint.py

@@ -19,6 +19,8 @@ class RuntimeEndpoint(ObjectFifoEndpoint):
     """
 
     def __init__(self, tile: Tile = AnyShimTile) -> None:
+        if tile is None:
+            tile = AnyShimTile
         tile = tile.copy()
         if tile.tile_type is not None and tile.tile_type != AIETileType.ShimNOCTile:
             raise ValueError(