only releveant change

compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/Passes.cpp

@@ -481,15 +481,15 @@ void addConvDecomposePassPipeline(OpPassManager &funcPassManager,
   // Pad the linalg operation
   {
     AMDAIEPadOptions padOptions;
-    padOptions.paddingLevel = 0;
+    padOptions.paddingLevel = 1;
     funcPassManager.addPass(createAMDAIEPadPass(padOptions));
   }
 
   // Only promote the result to local memory
   {
     AMDAIEBufferizeToAllocationOptions bufferizeOptions;
     bufferizeOptions.memorySpace = 2;
-    bufferizeOptions.bufferizeOperand = BufferizeOperand::InputOutput;
+    bufferizeOptions.bufferizeOperand = BufferizeOperand::Output;
     funcPassManager.addPass(
         createAMDAIEBufferizeToAllocationPass(bufferizeOptions));
   }
@@ -505,6 +505,22 @@ void addConvDecomposePassPipeline(OpPassManager &funcPassManager,
   funcPassManager.addPass(createCanonicalizerPass());
   funcPassManager.addPass(createCSEPass());
 
+  // Pad the linalg operation
+  {
+    AMDAIEPadOptions padOptions;
+    padOptions.paddingLevel = 2;
+    funcPassManager.addPass(createAMDAIEPadPass(padOptions));
+  }
+
+  // Promote the inputs to local memory
+  {
+    AMDAIEBufferizeToAllocationOptions bufferizeOptions;
+    bufferizeOptions.memorySpace = 2;
+    bufferizeOptions.bufferizeOperand = BufferizeOperand::Input;
+    funcPassManager.addPass(
+        createAMDAIEBufferizeToAllocationPass(bufferizeOptions));
+  }
+
   // Decompose Conv2d ops to Conv1d ops
   funcPassManager.addPass(createDecomposeConvolutionToLowerDimOpsPass());
 
@@ -628,23 +644,6 @@ void addMLIRAIRLoweringPasses(OpPassManager &passManager, AMDAIEDevice device) {
   passManager.addPass(memref::createFoldMemRefAliasOpsPass());
   passManager.addPass(createAMDAIEBridgeToAIRPass());
 
-  // Running canonicalization for all pipelines here results in failures.
-  // Example
-  // ```
-  // 'memref.cast' op is an unsupported operation. This pass currently only
-  // supports AllocOp and SubViewOp as inputs.
-  // ```
-  // It is currently required for the convolution pipeline though, to remove the
-  // extra (size-1) thread- and group- dimensions.
-  //
-  // TODO(newling) there are better solutions like:
-  // 1) make canonicalization work for scf.forall
-  // 2) pass to collapse rank-4 scf.foralls to rank-2 scf.foralls.
-  // 3) resolve above 'unsupproted operation' error.
-  if (clUseTilePipeline == TilePassPipeline::ConvDecomposePipeline) {
-    passManager.addPass(createCanonicalizerPass());
-  }
-
   // TODO (Erwei): Figure out a way to work with AMDAIEPackToDmaPass.
   if (clUseTilePipeline == TilePassPipeline::PackPeelPipeline)
     passManager.addPass(createAMDAIEDecomposeLinalgExtPackUnPackToAIRPass());