fix(tiling): emit cumulative numTiles for L1 when only one tiling schedule

When there is a single tiling schedule (len == 1), L1 is the outermost
tile loop and is called exactly once.  The prior code always emitted
per-tile numTiles {0,1,...,N} at L1, so only tile 0 ran — causing
failures when a node tiles into multiple L1 slices with no L2 spatial loop
wrapping the L1 closure.

Switch to cumulative {0, total} when len(tilingSchedules) == 1 so the
single call walks all tiles.  Keep per-tile for len > 1 (L2 spatial loop
drives L1 once per step).

Also add profilingNodes field to CodeGenVerbosity and honour it in
PULPClusterTiling / PULPL3Tiling to restrict cycle-counter instrumentation
to a named subset of nodes.

Author: runwangdl

Deeploy/DeeployTypes.py

@@ -53,6 +53,8 @@ class CodeGenVerbosity:
 
     tilingProfiling: Optional[bool] = False  # Specifies if we should profile the tiling code
     untiledProfiling: Optional[bool] = None  #  Specifies if we should profile the untilied code
+    profilingNodes: Optional[
+        list] = None  # Restrict tiling profiling to nodes whose name contains any of these substrings; None means all nodes
 
 
 _NoVerbosity = CodeGenVerbosity(None)

Deeploy/Targets/PULPOpen/CodeTransformationPasses/PULPClusterTiling.py

@@ -42,7 +42,9 @@ def apply(self,
               name: str,
               verbose: CodeGenVerbosity = _NoVerbosity) -> Tuple[NetworkContext, ExecutionBlock]:
 
-        if verbose.tilingProfiling:
+        nodes = verbose.profilingNodes
+        profileThis = verbose.tilingProfiling and (nodes is None or any(s in name for s in nodes))
+        if profileThis:
             ctxt, executionBlock = self.profilingSB.apply(ctxt, executionBlock, name)
             ctxt, executionBlock = self.profilingDB.apply(ctxt, executionBlock, name)
         else:

Deeploy/Targets/PULPOpen/CodeTransformationPasses/PULPL3Tiling.py

@@ -42,7 +42,9 @@ def apply(self,
               name: str,
               verbose: CodeGenVerbosity = _NoVerbosity) -> Tuple[NetworkContext, ExecutionBlock]:
 
-        if verbose.tilingProfiling:
+        nodes = verbose.profilingNodes
+        profileThis = verbose.tilingProfiling and (nodes is None or any(s in name for s in nodes))
+        if profileThis:
             ctxt, executionBlock = self.profilingSB.apply(ctxt, executionBlock, name)
             ctxt, executionBlock = self.profilingDB.apply(ctxt, executionBlock, name)
         else:

Deeploy/TilingExtension/CodeTransformationPasses/TilingHoistingMixIn.py

@@ -80,14 +80,21 @@ def _hoistTileNumAndIdxPtr(self, ctxt: NetworkContext,
 
         # Core extension: at the innermost memory level (L1), emit a per-tile
         # boundary so each invocation of the inner closure processes exactly one
-        # tile. The outer-level (L3→L2) closure iterates N_outer times and calls
-        # inner once per iter; with the baseline cumulative layout
+        # tile. The outer-level (L2 spatial) closure iterates N_outer times and
+        # calls inner once per iter; with the baseline cumulative layout
         # `{0, N1, N1+N2, ...}` inner would process many tiles per call and only
         # tolerate `len(tilingSchedules)` outer iters before reading numTiles
         # OOB. Per-tile layout `{0,1,2,...,total}` keeps
         # outer_iters == inner_calls == total_tiles. Outer memory levels keep
         # cumulative layout to iterate per L2 tile.
-        if self.memory == "L1":
+        #
+        # Detect whether an outer driver will call L1 multiple times:
+        # `len(tilingSchedules) > 1` means there are multiple L2 spatial steps,
+        # so an L2 outer loop drives L1 once per step → per-tile.
+        # When len == 1 (L1 is the single outermost tile loop, called once from
+        # RunNetwork / the L3 closure), emit cumulative {0, total} so the single
+        # call walks all tiles.
+        if self.memory == "L1" and len(tilingSchedules) > 1:
             total = sum(stepsNumTiles)
             cumulativeNumTiles = list(range(total + 1))
         else: