[mlir][linalg] Add mixed precision folding pattern in vectorize_children_and_apply_patterns TD Op (#148684)

In case of mixed precision inputs, the inputs are generally casted to
match output type thereby introduces arith.extFOp/extIOp instructions.

Folding such pattern into vector.contract is desirable for HW having
mixed precision ISA support.

This patch adds folding of mixed precision pattern into vector.contract
optionaly which can be enabled using attribute
`fold_type_extensions_into_contract`.

Author: shahidact

mlir/include/mlir/Dialect/Linalg/TransformOps/LinalgTransformOps.td

@@ -2348,6 +2348,9 @@ def VectorizeChildrenAndApplyPatternsOp :
     operation that is contained inside the vectorization target.
 
     This transformation supports the following attributes:
+    - `fold_type_extensions_into_contract`: a `UnitAttr` to enable the folding of
+      type extension operations into `vector.contract` to create a mixed precision
+      operation.
     - `vectorize_padding`: a `UnitAttr` to activate the vectorization of
       `tensor.pad` ops. Different pipelines may prefer to lower such ops to
       loops.
@@ -2368,6 +2371,7 @@ def VectorizeChildrenAndApplyPatternsOp :
   }];
 
   let arguments = (ins TransformHandleTypeInterface:$target,
+                   UnitAttr:$fold_type_extensions_into_contract,
                    UnitAttr:$vectorize_padding,
                    UnitAttr:$vectorize_nd_extract,
                    UnitAttr:$flatten_1d_depthwise_conv,
@@ -2381,6 +2385,7 @@ def VectorizeChildrenAndApplyPatternsOp :
 
   let builders = [
     OpBuilder<(ins "Value":$target,
+               CArg<"bool", "false">:$foldTypeExtensionsIntoContract,
                CArg<"bool", "false">:$vectorizePadding,
                CArg<"bool", "false">:$vectorizeNDExtract,
                CArg<"bool", "false">:$flatten1DDepthwise)>

mlir/lib/Dialect/Linalg/TransformOps/LinalgTransformOps.cpp

@@ -3783,8 +3783,15 @@ LogicalResult TileUsingForallOp::verify() {
 
 void transform::VectorizeChildrenAndApplyPatternsOp::build(
     OpBuilder &builder, OperationState &result, Value target,
-    bool vectorizePadding, bool vectorizeExtract, bool flatten1DDepthwiseConv) {
+    bool foldTypeExtensionsIntoContract, bool vectorizePadding,
+    bool vectorizeExtract, bool flatten1DDepthwiseConv) {
   result.addOperands(target);
+  if (foldTypeExtensionsIntoContract) {
+    result.addAttribute(
+        VectorizeChildrenAndApplyPatternsOp::
+            getFoldTypeExtensionsIntoContractAttrName(result.name),
+        builder.getUnitAttr());
+  }
   if (vectorizePadding) {
     result.addAttribute(
         VectorizeChildrenAndApplyPatternsOp::getVectorizePaddingAttrName(
@@ -3875,6 +3882,9 @@ transform::VectorizeChildrenAndApplyPatternsOp::applyToOne(
 
   patterns.add<CopyVectorizationPattern>(ctx);
 
+  if (getFoldTypeExtensionsIntoContract())
+    vector::populateFoldArithExtensionPatterns(patterns);
+
   if (getVectorizePadding()) {
     linalg::populatePadOpVectorizationPatterns(patterns);
     // This creates an alternative path for lowering tensor.pad - by