add optional offsets to load_gather

Author: Jianhui-Li

mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td

@@ -655,7 +655,7 @@ def XeGPU_PrefetchOp : XeGPU_Op<"prefetch", []> {
 }
 
 def XeGPU_LoadGatherOp : XeGPU_Op<"load", [
-  AllElementTypesMatch<["value", "TensorDesc"]>, MemoryEffects<[MemRead]>
+  AllElementTypesMatch<["value", "source"]>, MemoryEffects<[MemRead]>
   ]> {
   let summary = "load a set of scattered data points from memory.";
 
@@ -698,16 +698,21 @@ def XeGPU_LoadGatherOp : XeGPU_Op<"load", [
 
   }];
 
-  let arguments = (ins XeGPU_TensorDesc: $TensorDesc,
+  let arguments = (ins XeGPU_TensorDesc_or_MemRef: $source,
                        XeGPU_MaskType: $mask,
                        OptionalAttr<XeGPU_CacheHintAttr>: $l1_hint,
                        OptionalAttr<XeGPU_CacheHintAttr>: $l2_hint,
                        OptionalAttr<XeGPU_CacheHintAttr>: $l3_hint);
   let results = (outs XeGPU_ValueType: $value);
 
   let extraClassDeclaration = extraBaseClassDeclaration # [{
+
+    Type getSourceType() {
+      return getSource().getType();
+    }
+
     xegpu::TensorDescType getTensorDescType() {
-      return getTensorDesc().getType();
+       return dyn_cast<xegpu::TensorDescType>(getSourceType());
     }
 
     mlir::Type getElementType() {
@@ -725,8 +730,8 @@ def XeGPU_LoadGatherOp : XeGPU_Op<"load", [
 
   }];
 
-  let assemblyFormat = [{$TensorDesc `,` $mask prop-dict attr-dict
-      `:` qualified(type($TensorDesc)) `,` type($mask) `->` type($value)}];
+  let assemblyFormat = [{$source `,` $mask prop-dict attr-dict
+      `:` qualified(type($source)) `,` type($mask) `->` type($value)}];
 
   let hasVerifier = 1;
 }

mlir/include/mlir/Dialect/XeGPU/IR/XeGPUTypes.td

@@ -189,6 +189,7 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
   let genVerifyDecl = 1;
 }
 
+def XeGPU_TensorDesc_or_MemRef : AnyTypeOf<[XeGPU_TensorDesc,Non0RankedMemRefOf<[XeGPU_ScalarType]>]>;
 
 def XeGPU_Nbarrier: XeGPUTypeDef<"Nbarrier", "nbarrier", [], "mlir::Type"> {
   let summary = "!xegpu.nbarrier a custom XeGPU type representing a barrier.";

mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp

@@ -502,7 +502,7 @@ struct UnrollLoadGatherOp : public UnrollPattern<xegpu::LoadGatherOp> {
     SmallVector<Type> convertedTdescTypes =
         getUnrolledTypes(tdescTy, *targetShape);
     SmallVector<Value> convertedTdescs = pack(
-        op.getTensorDesc(), convertedTdescTypes, *targetShape, loc, rewriter);
+        op.getSource(), convertedTdescTypes, *targetShape, loc, rewriter);
 
     SmallVector<Type> convertedMaskTypes;
     SmallVector<Value> convertedMasks;