Add tile-count performance hint to TransposeTileConstraint

Add a soft PerformanceHint constraining the output to at least 4 tiles.
Transpose is bandwidth-bound, so a higher tile count keeps DMA and
compute pipelined under double buffering.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Xeratec

Deeploy/Targets/Generic/TileConstraints/TransposeTileConstraint.py

@@ -4,14 +4,16 @@
 
 from typing import Dict, List, Tuple
 
+import numpy as np
+
 from Deeploy.AbstractDataTypes import PointerClass
 from Deeploy.CommonExtensions.DataTypes import uint16_t
 from Deeploy.CommonExtensions.OptimizationPasses.TopologyOptimizationPasses.LoweringOptimizationPasses import \
     _invertPermutation, _permuteHyperRectangle
 from Deeploy.DeeployTypes import NetworkContext, OperatorRepresentation
 from Deeploy.TilingExtension.MemoryConstraints import NodeMemoryConstraint
 from Deeploy.TilingExtension.TileConstraint import TileConstraint
-from Deeploy.TilingExtension.TilerModel import TilerModel
+from Deeploy.TilingExtension.TilerModel import PerformanceHint, TilerModel
 from Deeploy.TilingExtension.TilingCodegen import AbsoluteHyperRectangle, TilingSchedule, VariableReplacementScheme
 
 
@@ -35,6 +37,24 @@ def addGeometricalConstraint(tilerModel: TilerModel, parseDict: Dict, ctxt: Netw
 
         return tilerModel
 
+    @staticmethod
+    def addPolicyConstraint(tilerModel: TilerModel, parseDict: Dict, ctxt: NetworkContext) -> TilerModel:
+        # Soft hint: split the output across at least 4 tiles. Transpose is
+        # bandwidth-bound, so a higher tile count keeps DMA and compute pipelined
+        # under double buffering.
+        outputBufferName = parseDict['data_out']
+        outputBuffer = ctxt.lookup(outputBufferName)
+
+        totalElements = int(np.prod(outputBuffer.shape))
+        if totalElements < 4:
+            return tilerModel
+
+        tilerModel.addTensorNumOfEltToModel(ctxt, outputBufferName)
+        numEltVar = tilerModel.getTensorNumberOfEltVar(outputBufferName)
+        tilerModel.addConstraint(numEltVar * 4 <= totalElements, strategy = PerformanceHint(priority = 0))
+
+        return tilerModel
+
     @classmethod
     def serializeTilingSolution(
             cls, tilingSolution: NodeMemoryConstraint, absoluteOutputCubes: List[AbsoluteHyperRectangle],