Phase-3: extend absolute offset to VariableBuffers (activation write-side)

Extend _generatePromotedReferenceReset to handle VariableBuffers
(activations) in addition to ConstantBuffers.  Uses _referenceName
for baseSymbol (in scope via closure args) instead of
unraveled._instance (which may be out of scope for local activations).

Works on IC100/ML1/MMN/AD with activation promotion enabled.
CCT_2 activation promotion still fails for 32KB residual activations
that fit in L1 as 1 tile — the L2->L1 schedule has no per-tile offset
to fix, but the L3 closure would have spatially sliced the activation
per L3 tile.

Author: runwangdl

Deeploy/TilingExtension/CodeTransformationPasses/TilingCodeGeneration.py

@@ -154,17 +154,20 @@ def _generateDmaTransferCalls(self, ctxt: NetworkContext, tensorName: str, trans
 
     def _generatePromotedReferenceReset(self, ctxt: NetworkContext, tensorName: str, transfers: List[HyperRectangle],
                                         tileIdxVar: str, externalBuffer: VariableBuffer) -> Optional[CodeSnippet]:
-        """For promoted ConstantBuffers at self.externalMemory: emit an
-        absolute offset assignment BEFORE the DMA so it overrides the +0
-        initialization at the L1 closure body top (which resets every call)."""
+        """For promoted buffers (ConstantBuffer or VariableBuffer) at
+        self.externalMemory: emit an absolute offset assignment BEFORE
+        the DMA so it overrides the +0 initialization at the closure
+        body top (which resets every call).  Covers both the read side
+        (L2→L1 for promoted constants) and the write side (L1→L2 for
+        promoted activations)."""
         if len(transfers) <= 1:
             return None
-        from Deeploy.DeeployTypes import ConstantBuffer
+        from Deeploy.DeeployTypes import TransientBuffer
         try:
             unraveled = ctxt.unravelReference(externalBuffer)
         except Exception:
             unraveled = externalBuffer
-        if not isinstance(unraveled, ConstantBuffer):
+        if isinstance(unraveled, TransientBuffer):
             return None
         if getattr(unraveled, '_memoryLevel', None) != self.externalMemory:
             return None
@@ -183,12 +186,19 @@ def _generatePromotedReferenceReset(self, ctxt: NetworkContext, tensorName: str,
                 offsetBuf = ctxt.lookup(fullName)
             except Exception:
                 return None
+        # Use externalBuffer._referenceName (the buffer the ref points to)
+        # which IS available in the closure scope (passed via args or
+        # declared locally).  For ConstantBuffers this is the global static
+        # PI_L2 symbol; for VariableBuffers (activations) it's the
+        # closure-arg-passed pointer.
+        baseRef = externalBuffer._referenceName if hasattr(externalBuffer, '_referenceName') else str(
+            unraveled._instance)
         return CodeSnippet(
             self._absoluteOffsetReferenceTemplate, {
                 "reference": externalBuffer.name,
                 "tileIdxVar": tileIdxVar,
                 "typeName": externalBuffer._type.referencedType.typeName,
-                "baseSymbol": str(unraveled._instance),
+                "baseSymbol": baseRef,
                 "absoluteOffsets": offsetBuf.name,
             })