[llvm-branch-commits] [mlir] 20f7777 - Revert "[MLIR][XeGPU] Updates XeGPU TensorDescAttr and Refine Gather/Scatter …"

via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Mon Sep 23 07:03:56 PDT 2024


Author: Chao Chen
Date: 2024-09-23T09:03:53-05:00
New Revision: 20f77775cd2516c1a0ef12d5f8a625b31b2448d5

URL: https://github.com/llvm/llvm-project/commit/20f77775cd2516c1a0ef12d5f8a625b31b2448d5
DIFF: https://github.com/llvm/llvm-project/commit/20f77775cd2516c1a0ef12d5f8a625b31b2448d5.diff

LOG: Revert "[MLIR][XeGPU] Updates XeGPU TensorDescAttr and Refine Gather/Scatter …"

This reverts commit 21627236363d629f6a5b820f45a6071371e4b8db.

Added: 
    

Modified: 
    mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
    mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
    mlir/include/mlir/Dialect/XeGPU/IR/XeGPUTypes.td
    mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
    mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
    mlir/test/Dialect/XeGPU/XeGPUOps.mlir
    mlir/test/Dialect/XeGPU/invalid.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
index 26eec0d4f2082a..f3ca09a6a68ea8 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
@@ -19,18 +19,12 @@ class XeGPUAttr<string name, string attrMnemonic, list<Trait> traits = [],
   let mnemonic = attrMnemonic;
 }
 
-class XeGPU_TensorDescAttr<string name, string attrMnemonic, list<Trait> traits = [],
-                         string baseCppClass = "::mlir::Attribute">
-    : XeGPUAttr<name, attrMnemonic, traits, baseCppClass> {
-  let assemblyFormat = "`<` struct(params) `>`";
-}
-
-def XeGPU_BlockTensorDescAttr: XeGPU_TensorDescAttr<"BlockTensorDesc", "block_tdesc_attr"> {
+def XeGPU_TensorDescAttr: XeGPUAttr<"TensorDesc", "tdesc_attr"> {
   let summary = [{a composite attribute for `TensorDescType`}];
-  let description = [{`BlockTensorDesc` (or `block_tdesc_attr`) is a composite
+  let description = [{`TensorDescAttr` (or `tdesc_attr`) is a composite
     attribute defined for `TensorDescType` for describing following
     properties of a `TensorDesc`.
-    1. `memory_space`: It describes where the data block described by the
+    1. `memory_scope`: It describes where the data block described by the
         TensorDesc is located, `Global` device memory or `Shared` local memory.
         It is default to `Global`.
     2. `array_length`: It describes how many horizontally consecutive blocks
@@ -39,63 +33,43 @@ def XeGPU_BlockTensorDescAttr: XeGPU_TensorDescAttr<"BlockTensorDesc", "block_td
         8x32. Its default value is 1.
     3. `boundary_check`: It is used to indicates the hardware whether to do
         out-of-boundary check. The default value is true.
+    4. `scattered`: It is used to 
diff erenciate TensorDescs created from
+       `create_nd_tdesc` vs from `create_tdesc`.
   }];
 
   let parameters = (ins
-    OptionalParameter<"MemorySpaceAttr">: $memory_space,
+    OptionalParameter<"MemoryScopeAttr">: $memory_scope,
     OptionalParameter<"IntegerAttr", "1">: $array_length,
-    OptionalParameter<"BoolAttr", "true">: $boundary_check
+    OptionalParameter<"BoolAttr", "true">: $boundary_check,
+    OptionalParameter<"BoolAttr", "false">: $scattered
   );
 
   let builders = [
     AttrBuilder<(ins
-      CArg<"xegpu::MemorySpace", "xegpu::MemorySpace::Global">:$memory_space,
+      CArg<"xegpu::MemoryScope", "xegpu::MemoryScope::Global">:$memory_scope,
       CArg<"int", "1">:$array_length,
-      CArg<"bool", "true">: $boundary_check
+      CArg<"bool", "true">: $boundary_check,
+      CArg<"bool", "false">: $scattered
     )>
   ];
 
+  let assemblyFormat = "`<` struct(params) `>`";
 }
 
-def XeGPU_ScatterTensorDescAttr: XeGPU_TensorDescAttr<"ScatterTensorDesc", "scatter_tdesc_attr"> {
-  let summary = [{a composite attribute for `TensorDescType`}];
-  let description = [{`ScatterTensorDesc` (or `scatter_tdesc_attr`) is a composite
-    attribute defined for `TensorDescType` for describing following
-    properties of a `TensorDesc`.
-    1. `memory_space`: It describes where the data block described by the
-        TensorDesc is located, `Global` device memory or `Shared` local memory.
-        It is default to `Global`.
-    2.  `chunk_size`: indicates number of continious elements accessed for each
-        offset, default is 1. It is used with `scattered` attr only.
-  }];
-
-  let parameters = (ins
-    OptionalParameter<"MemorySpaceAttr">: $memory_space,
-    OptionalParameter<"IntegerAttr", "1">: $chunk_size
-  );
-
-  let builders = [
-    AttrBuilder<(ins
-      CArg<"xegpu::MemorySpace", "xegpu::MemorySpace::Global">:$memory_space,
-      CArg<"int", "1">: $chunk_size
-    )>
-  ];
- }
-
 //===----------------------------------------------------------------------===//
 // XeGPU Memory Scope Enums.
 //===----------------------------------------------------------------------===//
-def XeGPU_MemorySpaceGlobal: I32EnumAttrCase<"Global", 0, "global">;
-def XeGPU_MemorySpaceShared: I32EnumAttrCase<"SLM", 3, "slm">;
-def XeGPU_MemorySpace: I32EnumAttr<"MemorySpace",
+def XeGPU_MemoryScopeGlobal: I32EnumAttrCase<"Global", 0, "global">;
+def XeGPU_MemoryScopeShared: I32EnumAttrCase<"SLM", 1, "slm">;
+def XeGPU_MemoryScope: I32EnumAttr<"MemoryScope",
       "The address space of the memory the tensor descritor is created for",
-      [XeGPU_MemorySpaceGlobal, XeGPU_MemorySpaceShared]> {
+      [XeGPU_MemoryScopeGlobal, XeGPU_MemoryScopeShared]> {
   let genSpecializedAttr = 0;
   let cppNamespace = "::mlir::xegpu";
 }
 
-def XeGPU_MemorySpaceAttr:
-  EnumAttr<XeGPU_Dialect, XeGPU_MemorySpace, "memory_space"> {
+def XeGPU_MemoryScopeAttr:
+  EnumAttr<XeGPU_Dialect, XeGPU_MemoryScope, "memory_scope"> {
     let summary = [{Describe the location of data described by a `TensorDesc`:
                  Global device memory (`Global`) or Shared local memory (`SLM`).}];
     let assemblyFormat = "$value";
@@ -142,4 +116,4 @@ def XeGPU_FenceScopeAttr:
     let assemblyFormat = "$value";
 }
 
-#endif // MLIR_DIALECT_XEGPU_IR_XEGPUATTRS_TD
+#endif // MLIR_DIALECT_XEGPU_IR_XEGPUATTRS_TD
\ No newline at end of file

diff  --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
index e24a056de2caf3..c32c7541c39791 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
@@ -218,23 +218,6 @@ def XeGPU_CreateNdDescOp: XeGPU_Op<"create_nd_tdesc", [Pure, ViewLikeOpInterface
     static unsigned getOffsetSizeAndStrideStartOperandIndex() { return 1; }
 
     mlir::Value getViewSource() { return getSource(); }
-
-    unsigned getSourceMemorySpace() {
-      auto srcTy = getSourceType();
-      if (auto memrefTy = llvm::dyn_cast<mlir::MemRefType>(srcTy)) {
-        auto attr = memrefTy.getMemorySpace();
-        if (attr) {
-          if (auto intAttr = llvm::dyn_cast<mlir::IntegerAttr>(attr)) {
-            return static_cast<unsigned>(intAttr.getInt());
-          }
-          if (auto memSpaceAttr = llvm::dyn_cast<MemorySpaceAttr>(attr))
-            return static_cast<unsigned>(memSpaceAttr.getValue());
-        }
-      }
-      // take global as default memory scope.
-      return static_cast<unsigned>(MemorySpace::Global);
-    }
-
   }];
 }
 
@@ -428,10 +411,8 @@ def XeGPU_CreateDescOp: XeGPU_Op<"create_tdesc", [Pure, ViewLikeOpInterface]> {
       is fixed to the hardware supportted subgroup size, e.g., 16 on PVC,
       implying each element in the array corresponds to a work-item (SIMT lane)
       in the subgroup.
-
-    The first dimension of the result TensorDesc corresponds to work-items, so it should
-    match the dimension of offsets. It may also has a second dimension corresponding to
-    the chunk_size if the chunk size is larger than 1.
+    * chunk_size: [optional attribute] indicates number of continious
+      elements accessed for each offset, default is 1.
 
     Example 1. It assumes subgroup size is 4, and accesses a[0], a[16], a[32], a[64]
     ```mlir
@@ -443,22 +424,29 @@ def XeGPU_CreateDescOp: XeGPU_Op<"create_tdesc", [Pure, ViewLikeOpInterface]> {
                It will access totally 32 data elements: a[0:7], a[16:23], a[32:39], a[64:71]
     ```mlir
     %0 = memref.alloc() : memref<1024xf32>
-    %1 = xegpu.create_tdesc %0[0, 16, 32, 64] : memref<1024xf32> -> TensorDesc<4x8xf32, chunk_size = 8>
+    %1 = xegpu.create_tdesc %0[0, 16, 32, 64] {chunk_size = 8}: memref<1024xf32> -> TensorDesc<4x8xf32>
     ```
 
     Example 3. It is similar to Example 2, but there is some overlaps among workitems.
                It accesses: a[0:7], a[4:11], a[8:15], a[12:19]
     ```mlir
     %0 = memref.alloc() : memref<1024xf32>
-    %1 = xegpu.create_tdesc %0[0, 4, 8, 12] : memref<1024xf32> -> TensorDesc<4x8xf32, chunk_size = 8>>
+    %1 = xegpu.create_tdesc %0[0, 4, 8, 12] {chunk_size = 8}: memref<1024xf32> -> TensorDesc<4x8xf32>
     ```
   }];
 
   let arguments = (ins XeGPU_BaseAddrType: $source,
                        Variadic<Index>: $offsets,
-                       DenseI64ArrayAttr: $const_offsets);
+                       DenseI64ArrayAttr: $const_offsets,
+                       DefaultValuedAttr<I64Attr, "1">: $chunk_size);
   let results = (outs XeGPU_TensorDesc:$TensorDesc);
 
+  let builders = [
+    OpBuilder<(ins "xegpu::TensorDescType": $TensorDesc, "Value": $source,
+                   "llvm::ArrayRef<OpFoldResult>": $offsets,
+                   CArg<"uint32_t", "1"> : $chunk_size)>,
+  ];
+
   let assemblyFormat = [{
     $source
     custom<DynamicIndexList>($offsets, $const_offsets)
@@ -485,22 +473,6 @@ def XeGPU_CreateDescOp: XeGPU_Op<"create_tdesc", [Pure, ViewLikeOpInterface]> {
       assert(idx < getNumOffsets() && "Invalid out of bound access.");
       return getMixedOffsets()[idx];
     }
-
-    unsigned getSourceMemorySpace() {
-      auto srcTy = getSource().getType();
-      if (auto memrefTy = llvm::dyn_cast<mlir::MemRefType>(srcTy)) {
-        auto attr = memrefTy.getMemorySpace();
-        if (attr) {
-          if (auto intAttr = llvm::dyn_cast<mlir::IntegerAttr>(attr))
-            return static_cast<unsigned>(intAttr.getInt());
-          if (auto memSpaceAttr = llvm::dyn_cast<MemorySpaceAttr>(attr))
-            return static_cast<unsigned>(memSpaceAttr.getValue());
-        }
-      }
-      // take global as default memory scope.
-      return static_cast<unsigned>(MemorySpace::Global);
-    }
-
   }];
 
   let hasVerifier = 1;
@@ -548,31 +520,28 @@ def XeGPU_LoadGatherOp : XeGPU_Op<"load", [AllRanksMatch<["value", "TensorDesc"]
 
   let description = [{ It (aka. load) load data per each work-item. The output
     describes the data being loaded at the subgroup level, so its size is
-    consistent with the number of work-items in a subgroup. When the chunk size
-    is larger than 2, the output vector is a 2D vector, with dim-1 correspoding
-    to work-items, and dim-0 corresponding to the chunk_size loaded by each work-item.
-    Specially, there is a transpose effect on the result (as compared to the TensorDesc)
-    due to the hardware implementation. Therefore, a transpose attribute is introduced
-    on purpose, making sure users are aware of this implicit transformation.
+    consistent with the number of work-items in a subgroup. When `chunk_size_per_lane`
+    attribute is larger than 1 in TensorDesc, the output vector will be 2D vector,
+    with dim-1 correspoding to the chunk size.
 
     The mask operand masks out memory access so that it is safe to pass out-of-boundary
     addresses/offsets as long as they are masked. It applies to slots of SIMD lanes.
 
   Example:
   ```mlir
-    %2 = xegpu.load %1, %0 {transpose,
+    %2 = xegpu.load %1, %0 {transpose = [1, 0],
                             l1_hint = #xegpu.cache_hint<cached>,
                             l2_hint = #xegpu.cache_hint<uncached>,
                             l3_hint = #xegpu.cache_hint<uncached>}
-          : !xegpu.tensor_desc<16xf32, #xegpu.scatter_tdesc_attr<memory_space=global>>,
-            vector<16xi1> -> vector<16xf32>
+          : !xegpu.tensor_desc<16xf32, #xegpu.tdesc_attr<scattered=true>>, vector<16xi1>
+            -> vector<16xf32>
   ```
 
   }];
 
   let arguments = (ins XeGPU_TensorDesc: $TensorDesc,
                        XeGPU_MaskType: $mask,
-                       OptionalAttr<UnitAttr>: $transpose,
+                       OptionalAttr<DenseI64ArrayAttr>: $transpose,
                        OptionalAttr<XeGPU_CacheHintAttr>: $l1_hint,
                        OptionalAttr<XeGPU_CacheHintAttr>: $l2_hint,
                        OptionalAttr<XeGPU_CacheHintAttr>: $l3_hint);
@@ -604,15 +573,11 @@ def XeGPU_LoadGatherOp : XeGPU_Op<"load", [AllRanksMatch<["value", "TensorDesc"]
   let hasVerifier = 1;
 }
 
-def XeGPU_StoreScatterOp : XeGPU_Op<"store", [AllElementCountsMatch<["value", "TensorDesc"]>,
-                                              AllElementTypesMatch<["value", "TensorDesc"]>]> {
+def XeGPU_StoreScatterOp : XeGPU_Op<"store", [AllShapesMatch<["value", "TensorDesc"]>,
+                                        AllElementTypesMatch<["value", "TensorDesc"]>]> {
   let summary = "store data to scattered memory locations.";
-  let description = [{ It (aka. store) stores data to scattered memory locations. The value is
-  typically a 1D vector. But when the chunk size of the TensorDesc is larger than 1, it will be
-  a 2D vector instead. For the later case, dim-1 of the value correspods to the simd lanes
-  and the dim-0 of the value corresponds to the chunk_size stored per lane. So `store_scatter`
-  has transpose effect, which is similar to `load_gather`. Therefore, a transpose attribute is
-  introduced on purpose, making sure users are aware of this implicit transformation.
+  let description = [{ It (aka. store) stores data to scattered memory locations.
+  It has similar semantic to `load_gather`.
 
   Example:
   ```mlir
@@ -627,7 +592,6 @@ def XeGPU_StoreScatterOp : XeGPU_Op<"store", [AllElementCountsMatch<["value", "T
     XeGPU_ValueType: $value,
     XeGPU_TensorDesc: $TensorDesc,
     XeGPU_MaskType: $mask,
-    OptionalAttr<UnitAttr>: $transpose,
     OptionalAttr<XeGPU_CacheHintAttr>: $l1_hint,
     OptionalAttr<XeGPU_CacheHintAttr>: $l2_hint,
     OptionalAttr<XeGPU_CacheHintAttr>: $l3_hint);
@@ -759,7 +723,7 @@ def XeGPU_DpasOp : XeGPU_Op<"dpas", [Pure, AllElementTypesMatch<["lhs", "rhs"]>]
 
 def XeGPU_AtomicRMWOp: XeGPU_Op<"atomic_rmw", [Pure,
       AllElementTypesMatch<["tensorDesc", "value", "result"]>,
-      AllShapesMatch<["tensorDesc", "value", "result"]>]> {
+      AllShapesMatch<["tensorDesc", "mask", "value", "result"]>]> {
   let summary = "Atomic ready-modify-write operation on the TensorDesc. ";
 
   let description = [{
@@ -844,7 +808,7 @@ def XeGPU_FenceOp: XeGPU_Op<"fence", []> {
     2. `Fence_scope` describes the scope of fence. "Workgroup" means that the scope would be
         within each workgroup. "GPU" means the scope would be across workgroups within the GPU.
   }];
-  let arguments = (ins XeGPU_MemorySpaceAttr: $memory_kind,
+  let arguments = (ins XeGPU_MemoryScopeAttr: $memory_kind,
                        XeGPU_FenceScopeAttr: $fence_scope);
   let assemblyFormat = [{`memory_kind` `=` `` $memory_kind `,` `fence_scope` `=` `` $fence_scope attr-dict}];
   let extraClassDeclaration = extraBaseClassDeclaration;

diff  --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUTypes.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUTypes.td
index 0ce1211664b5ba..9f101a71697b56 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUTypes.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUTypes.td
@@ -48,7 +48,7 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
 
     Similar to the builtin tensor, it also provides an optinal attribute to encoding
     the following information via the TensorDescAttr object:
-    * memory_space (xegpu::MemorySpace): [optional] where the data is located,
+    * memory_scope (xegpu::MemoryScope): [optional] where the data is located,
                 global memory or shared memory. It is default to Global.
     * array_length (int): [optional] The number of contiguous blocks with size as `shape`,
                that will be loaded by block load at a time. It is default to 1.
@@ -63,7 +63,7 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
     element-type ::= float-type | integer-type | index-type
     dim-list := (static-dim-list `x`)?
     static-dim-list ::= decimal-literal `x` decimal-literal
-    attr-list = (, memory_space = value)? (, arr_len = value)? (, boundary_check = value)? (, scattered = value)?
+    attr-list = (, memory_scope = value)? (, arr_len = value)? (, boundary_check = value)? (, scattered = value)?
     ```
 
     Examples:
@@ -76,7 +76,7 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
     xegpu.tensor_desc<8x16xf32>
 
     // A TensorDesc with 8x16 f32 elements for a memory region in shared memory space.
-    xegpu.tensor_desc<8x16xf32, #xegpu.tdesc_attr<memory_space = slm>>
+    xegpu.tensor_desc<8x16xf32, #xegpu.tdesc_attr<memory_scope = slm>>
     ```
   }];
 
@@ -88,14 +88,11 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
     TypeBuilderWithInferredContext<(ins
       "llvm::ArrayRef<int64_t>": $shape,
       "mlir::Type": $elementType,
+      CArg<"bool", "false">: $scattered,
       CArg<"int", "1">: $array_length,
-      CArg<"bool", "true">: $boundary_check,
-      CArg<"xegpu::MemorySpace", "xegpu::MemorySpace::Global">:$memory_space)>,
-    TypeBuilderWithInferredContext<(ins
-      "llvm::ArrayRef<int64_t>": $shape,
-      "mlir::Type": $elementType,
-      CArg<"int", "1">: $chunk_size,
-      CArg<"xegpu::MemorySpace", "xegpu::MemorySpace::Global">:$memory_space)>
+      CArg<"xegpu::MemoryScope", "xegpu::MemoryScope::Global">:$memory_scope,
+      CArg<"bool", "true">: $boundary_check
+    )>
   ];
 
   let extraClassDeclaration = [{
@@ -113,58 +110,40 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
       return llvm::cast<TensorDescType>(cloneWith(getShape(), elementType));
     }
 
-    BlockTensorDescAttr getEncodingAsBlockTensorDescAttr() const {
-      return llvm::dyn_cast_if_present<BlockTensorDescAttr>(getEncoding());
+    TensorDescAttr getEncodingAsTensorDescAttr() const {
+      return llvm::dyn_cast_if_present<TensorDescAttr>(getEncoding());
     }
 
-    ScatterTensorDescAttr getEncodingAsScatterTensorDescAttr() const {
-      return llvm::dyn_cast_if_present<ScatterTensorDescAttr>(getEncoding());
-    }
-
-    xegpu::MemorySpace getMemorySpace() const {
-      auto block_attr = getEncodingAsBlockTensorDescAttr();
-      if (block_attr && block_attr.getMemorySpace())
-        return block_attr.getMemorySpace().getValue();
-
-      auto scatter_attr = getEncodingAsScatterTensorDescAttr();
-      if (scatter_attr && scatter_attr.getMemorySpace())
-        return scatter_attr.getMemorySpace().getValue();
-
+    xegpu::MemoryScope getMemoryScope() const {
+      auto attr = getEncodingAsTensorDescAttr();
+      if (attr && attr.getMemoryScope())
+        return attr.getMemoryScope().getValue();
       // return default value
-      return MemorySpace::Global;
+      return MemoryScope::Global;
     }
 
     int getArrayLength() {
-      auto attr = getEncoding();
-      auto block_attr = mlir::dyn_cast_if_present<BlockTensorDescAttr>(attr);
-      assert((!attr || block_attr) && "invalid on non BlockTensorDescAttr.");
-      if (block_attr && block_attr.getArrayLength())
-        return block_attr.getArrayLength().getInt();
+      auto attr = getEncodingAsTensorDescAttr();
+      if (attr && attr.getArrayLength())
+        return attr.getArrayLength().getInt();
       // return default value
       return 1;
     }
 
     bool getBoundaryCheck() {
-      auto attr = getEncoding();
-      auto block_attr = mlir::dyn_cast_if_present<BlockTensorDescAttr>(attr);
-      assert((!attr || block_attr) && "invalid on non BlockTensorDescAttr.");
-      if (block_attr && block_attr.getBoundaryCheck())
-        return block_attr.getBoundaryCheck().getValue();
+      auto attr = getEncodingAsTensorDescAttr();
+      if (attr && attr.getBoundaryCheck())
+        return attr.getBoundaryCheck().getValue();
       // return default value
       return true;
     }
 
-    bool isScattered() {
-      return bool(getEncodingAsScatterTensorDescAttr());
-    }
-
-    int getChunkSize() {
-      auto attr = getEncoding();
-      auto scatter_attr = mlir::dyn_cast_if_present<ScatterTensorDescAttr>(attr);
-      assert((!attr || scatter_attr) && "invalid on non ScatterTensorDescAttr.");
-      if (scatter_attr && scatter_attr.getChunkSize())
-        return scatter_attr.getChunkSize().getInt();
-      return 1;
+    bool getScattered() {
+      auto attr = getEncodingAsTensorDescAttr();
+      if (attr && attr.getScattered())
+        return attr.getScattered().getValue();
+      // return default value
+      return false;
     }
   }];
 

diff  --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
index 1dfbaed454c193..24719fe748fe4f 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
@@ -30,35 +30,23 @@ void XeGPUDialect::initialize() {
 }
 
 //===----------------------------------------------------------------------===//
-// XeGPU_BlockTensorDescAttr
+// XeGPU_TensorDescAttr
 //===----------------------------------------------------------------------===//
-BlockTensorDescAttr BlockTensorDescAttr::get(mlir::MLIRContext *context,
-                                             xegpu::MemorySpace memory_space,
-                                             int array_length,
-                                             bool boundary_check) {
-  auto scopeAttr = MemorySpaceAttr::get(context, memory_space);
+TensorDescAttr TensorDescAttr::get(mlir::MLIRContext *context,
+                                   xegpu::MemoryScope memory_scope,
+                                   int array_length, bool boundary_check,
+                                   bool scattered) {
+  auto scopeAttr = MemoryScopeAttr::get(context, memory_scope);
   auto lengthAttr =
       IntegerAttr::get(IntegerType::get(context, 64), array_length);
   auto boundaryAttr = BoolAttr::get(context, boundary_check);
-  return Base::get(context, scopeAttr, lengthAttr, boundaryAttr);
-}
-
-//===----------------------------------------------------------------------===//
-// XeGPU_ScatterTensorDescAttr
-//===----------------------------------------------------------------------===//
-ScatterTensorDescAttr
-ScatterTensorDescAttr::get(mlir::MLIRContext *context,
-                           xegpu::MemorySpace memory_space, int chunk_size) {
-  auto scopeAttr = MemorySpaceAttr::get(context, memory_space);
-  auto chunkSizeAttr =
-      IntegerAttr::get(IntegerType::get(context, 64), chunk_size);
-  return Base::get(context, scopeAttr, chunkSizeAttr);
+  auto scatteredAttr = BoolAttr::get(context, scattered);
+  return Base::get(context, scopeAttr, lengthAttr, boundaryAttr, scatteredAttr);
 }
 
 //===----------------------------------------------------------------------===//
 // XeGPU_TensorDescType
 //===----------------------------------------------------------------------===//
-
 mlir::Type TensorDescType::parse(::mlir::AsmParser &parser) {
   llvm::SmallVector<int64_t> shape;
   mlir::Type elementType;
@@ -120,20 +108,12 @@ void TensorDescType::print(::mlir::AsmPrinter &printer) const {
 }
 
 TensorDescType TensorDescType::get(llvm::ArrayRef<int64_t> shape,
-                                   mlir::Type elementType, int array_length,
-                                   bool boundary_check,
-                                   MemorySpace memory_space) {
-  auto context = elementType.getContext();
-  auto attr = BlockTensorDescAttr::get(context, memory_space, array_length,
-                                       boundary_check);
-  return Base::get(context, shape, elementType, attr);
-}
-
-TensorDescType TensorDescType::get(llvm::ArrayRef<int64_t> shape,
-                                   mlir::Type elementType, int chunk_size,
-                                   MemorySpace memory_space) {
+                                   mlir::Type elementType, bool scattered,
+                                   int array_length, MemoryScope memory_scope,
+                                   bool boundary_check) {
   auto context = elementType.getContext();
-  auto attr = ScatterTensorDescAttr::get(context, memory_space, chunk_size);
+  auto attr = TensorDescAttr::get(context, memory_scope, array_length,
+                                  boundary_check, scattered);
   return Base::get(context, shape, elementType, attr);
 }
 

diff  --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
index 7f0c8b72b3c252..9c517337a3aa57 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
@@ -124,17 +124,6 @@ LogicalResult CreateNdDescOp::verify() {
   bool invalidRank = false;
   bool invalidElemTy = false;
 
-  // Memory space of created TensorDesc should match with the source.
-  // Both source and TensorDesc are considered for global memory by default,
-  // if the memory scope attr is not specified. If source is an integer,
-  // it is considered as ptr to global memory.
-  auto srcMemorySpace = getSourceMemorySpace();
-  auto tdescMemorySpace = static_cast<unsigned>(getType().getMemorySpace());
-  if (srcMemorySpace != tdescMemorySpace)
-    return emitOpError("Memory space mismatch.")
-           << " Source: " << srcMemorySpace
-           << ", TensorDesc: " << tdescMemorySpace;
-
   // check source type matches the rank if it is a memref.
   // It also should have the same ElementType as TensorDesc.
   auto memrefTy = dyn_cast<MemRefType>(getSourceType());
@@ -163,13 +152,9 @@ LogicalResult CreateNdDescOp::verify() {
     return emitOpError("TensorDesc should have the same element "
                        "type with the source if it is a memref.\n");
 
-  if (getType().isScattered())
+  if (getType().getScattered())
     return emitOpError("Expects a non-scattered TensorDesc.\n");
 
-  if (getType().getRank() == 2 &&
-      tdescMemorySpace == static_cast<unsigned>(MemorySpace::SLM))
-    return emitOpError("SLM is not supported for 2D Block TensorDesc.\n");
-
   return success();
 }
 
@@ -178,7 +163,7 @@ LogicalResult CreateNdDescOp::verify() {
 //===----------------------------------------------------------------------===//
 LogicalResult PrefetchNdOp::verify() {
   auto tdescTy = getTensorDescType();
-  if (tdescTy.isScattered())
+  if (tdescTy.getScattered())
     return emitOpError("Expects a non-scattered TensorDesc.\n");
 
   if (!isReadHintOrNone(getL1HintAttr()))
@@ -203,7 +188,7 @@ LogicalResult LoadNdOp::verify() {
   if (tdescTy.getRank() > 2)
     return emitOpError("Expecting a 1D/2D TensorDesc.\n");
 
-  if (tdescTy.isScattered())
+  if (tdescTy.getScattered())
     return emitOpError("Expects a non-scattered TensorDesc.\n");
 
   if (!valueTy)
@@ -243,8 +228,8 @@ LogicalResult LoadNdOp::verify() {
       tdescShape[axis] /= vnni_factor;
       tdescShape.push_back(vnni_factor);
     } else {
-      emitWarning("Invalid Packed Attr. It is ignored (available for 2D "
-                  "TensorDesc only).");
+      return emitWarning("Invalid Packed Attr. It is ignored (available for 2D "
+                         "TensorDesc only).");
     }
   }
 
@@ -271,7 +256,7 @@ LogicalResult StoreNdOp::verify() {
   if (dstTy.getRank() > 2)
     return emitOpError("Expecting a 1D/2D TensorDesc.\n");
 
-  if (dstTy.isScattered())
+  if (dstTy.getScattered())
     return emitOpError("Expects a non-scattered TensorDesc.\n");
 
   if (!valTy)
@@ -294,7 +279,7 @@ LogicalResult StoreNdOp::verify() {
 //===----------------------------------------------------------------------===//
 LogicalResult UpdateNdOffsetOp::verify() {
   auto ty = getTensorDescType();
-  if (ty.isScattered())
+  if (ty.getScattered())
     return emitOpError("Expects a non-scattered TensorDesc.\n");
 
   // number of offsets specified must match the rank of the tensor descriptor
@@ -307,55 +292,28 @@ LogicalResult UpdateNdOffsetOp::verify() {
 //===----------------------------------------------------------------------===//
 // XeGPU_CreateDescOp
 //===----------------------------------------------------------------------===//
+void CreateDescOp::build(OpBuilder &builder, OperationState &state,
+                         TensorDescType TensorDesc, Value source,
+                         llvm::ArrayRef<OpFoldResult> offsets,
+                         uint32_t chunk_size) {
+  llvm::SmallVector<int64_t> staticOffsets;
+  llvm::SmallVector<Value> dynamicOffsets;
+  dispatchIndexOpFoldResults(offsets, dynamicOffsets, staticOffsets);
+  build(builder, state, TensorDesc, source, dynamicOffsets, staticOffsets,
+        chunk_size);
+}
 
 LogicalResult CreateDescOp::verify() {
   auto tdescTy = getTensorDescType();
+  auto chunkSize = getChunkSize();
 
   if (getRankOf(getSource()) > 1)
     return emitOpError(
         "Expecting the source is a 1D memref or pointer (uint64_t).");
 
-  if (!tdescTy.isScattered())
+  if (!tdescTy.getScattered())
     return emitOpError("Expects a scattered TensorDesc.\n");
 
-  // Memory space of created TensorDesc should match with the source.
-  // Both source and TensorDesc are considered for global memory by default,
-  // if the memory scope attr is not specified. If source is an integer,
-  // it is considered as ptr to global memory.
-  auto srcMemorySpace = getSourceMemorySpace();
-  auto tdescMemorySpace = static_cast<unsigned>(tdescTy.getMemorySpace());
-  if (srcMemorySpace != tdescMemorySpace)
-    return emitOpError("Memory space mismatch.")
-           << " Source: " << srcMemorySpace
-           << ", TensorDesc: " << tdescMemorySpace;
-
-  auto chunkSize = tdescTy.getChunkSize();
-
-  // check chunk_size
-  llvm::SmallVector<int64_t> supportedChunkSizes = {1,  2,  3,  4,   8,
-                                                    16, 32, 64, 128, 256};
-  if (!llvm::is_contained(supportedChunkSizes, chunkSize))
-    return emitOpError("Invalid chunk_size. Supported values are 1, 2, 3, 4, "
-                       "8, 16, 32, 64, 128, or 256.");
-
-  // check total size
-  auto elemBits = tdescTy.getElementType().getIntOrFloatBitWidth();
-  auto bitsPerLane = elemBits * chunkSize;
-  if (chunkSize > 1 && bitsPerLane % 32) {
-    // For 8-bit and 16-bit data, the hardware only supports chunk size of 1.
-    // For 32-bit data, the hardware can support larger chunk size. So
-    // we can bitcast 8-bit/16-bit data to 32-bit data for better performance.
-    // But this requires the total size is 32 bit aligned to make the
-    // optimization work.
-    return emitOpError(
-        "access size (chunk_size * sizeof(elemTy)) should be 32-bit aligned.");
-  }
-
-  auto lscConstraints = 512 * 8; // each access is upto 512 bytes.
-  if (elemBits * tdescTy.getNumElements() > lscConstraints)
-    return emitOpError("total access size (simd_lanes * chunk_size * "
-                       "sizeof(elemTy)) is upto 512 bytes.");
-
   SmallVector<int64_t> shape({(int64_t)getNumOffsets()});
   if (chunkSize != 1)
     shape.push_back(chunkSize);
@@ -373,7 +331,7 @@ LogicalResult CreateDescOp::verify() {
 //===----------------------------------------------------------------------===//
 LogicalResult PrefetchOp::verify() {
   auto tdescTy = getTensorDescType();
-  if (!tdescTy.isScattered())
+  if (!tdescTy.getScattered())
     return emitOpError("Expects a scattered TensorDesc.\n");
 
   if (!isReadHintOrNone(getL1HintAttr()))
@@ -396,7 +354,7 @@ LogicalResult LoadGatherOp::verify() {
   auto maskTy = getMaskType();
   auto valueTy = getValueType();
 
-  if (!tdescTy.isScattered())
+  if (!tdescTy.getScattered())
     return emitOpError("Expects a scattered TensorDesc.\n");
 
   if (!isReadHintOrNone(getL1HintAttr()))
@@ -421,10 +379,12 @@ LogicalResult LoadGatherOp::verify() {
   if (tdescShape[0] != maskShape[0])
     return emitOpError("dim-0 of the Mask and TensorDesc should be the same.");
 
-  if (tdescTy.getRank() == 2) {
-    if (!getTransposeAttr())
-      return emitOpError("load_gather has to be transposed.");
-    transpose({1, 0}, tdescShape);
+  if (getTransposeAttr()) {
+    auto trans = getTranspose().value();
+    if (tdescShape.size() < trans.size())
+      emitWarning("Invalid transpose attr. It is ignored.");
+    else
+      transpose(trans, tdescShape);
   }
 
   if (valueShape != tdescShape)
@@ -440,7 +400,7 @@ LogicalResult LoadGatherOp::verify() {
 //===----------------------------------------------------------------------===//
 LogicalResult StoreScatterOp::verify() {
   auto tdescTy = getTensorDescType();
-  if (!tdescTy.isScattered())
+  if (!tdescTy.getScattered())
     return emitOpError("Expects a scattered TensorDesc.\n");
 
   if (!isWriteHintOrNone(getL1HintAttr()))
@@ -453,24 +413,11 @@ LogicalResult StoreScatterOp::verify() {
     return emitOpError("invlid l3_hint: ") << getL3HintAttr();
 
   auto maskTy = getMaskType();
-  auto valueTy = getValueType();
   auto maskShape = getShapeOf(maskTy);
   auto tdescShape = getShapeOf(tdescTy);
-  auto valueShape = getShapeOf(valueTy);
   if (tdescShape[0] != maskShape[0])
     return emitOpError("dim-0 of the Mask and TensorDesc should be the same.");
 
-  if (tdescTy.getRank() == 2) {
-    if (!getTransposeAttr())
-      return emitOpError("load_gather has to be transposed.");
-    transpose({1, 0}, tdescShape);
-  }
-
-  if (valueShape != tdescShape)
-    return emitOpError("Unexpected value shape")
-           << "(Expected shape: " << makeString(tdescShape)
-           << ", Given shape: " << makeString(valueShape) << ").\n";
-
   return success();
 }
 //===----------------------------------------------------------------------===//

diff  --git a/mlir/test/Dialect/XeGPU/XeGPUOps.mlir b/mlir/test/Dialect/XeGPU/XeGPUOps.mlir
index c1126efb6046dc..35d44cf56a239b 100644
--- a/mlir/test/Dialect/XeGPU/XeGPUOps.mlir
+++ b/mlir/test/Dialect/XeGPU/XeGPUOps.mlir
@@ -24,8 +24,8 @@ gpu.func @test_create_nd_tdesc_vc_2(%src: ui64, %w : index, %h : index, %x : ind
 
 // CHECK: gpu.func @test_create_nd_tdesc_vc_3(%[[arg0:.*]]: memref<24x32xf32>) {
 gpu.func @test_create_nd_tdesc_vc_3(%src: memref<24x32xf32>) {
-  // CHECK: %[[REG:.*]] = xegpu.create_nd_tdesc %[[arg0]][0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x16xf32, #xegpu.block_tdesc_attr<array_length = 2 : i64>
-  %1 = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x16xf32, #xegpu.block_tdesc_attr<array_length = 2>>
+  // CHECK: %[[REG:.*]] = xegpu.create_nd_tdesc %[[arg0]][0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x16xf32, #xegpu.tdesc_attr<array_length = 2 : i64>
+  %1 = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x16xf32, #xegpu.tdesc_attr<array_length = 2>>
   gpu.return
 }
 
@@ -36,13 +36,6 @@ gpu.func @test_create_nd_tdesc_vc_4(%src: memref<2x24x32xf32>) {
   gpu.return
 }
 
-// CHECK: gpu.func @test_create_nd_tdesc_vc_5(%[[arg0:.*]]: memref<2x24x32xf32, 3>) {
-gpu.func @test_create_nd_tdesc_vc_5(%src: memref<2x24x32xf32, 3>) {
-  // CHECK: %[[REG:.*]] = xegpu.create_nd_tdesc %arg0[0, 0, 0] : memref<2x24x32xf32, 3> -> !xegpu.tensor_desc<16xf32, #xegpu.block_tdesc_attr<memory_space = slm>>
-  %1 = xegpu.create_nd_tdesc %src[0, 0, 0] : memref<2x24x32xf32, 3> -> !xegpu.tensor_desc<16xf32, #xegpu.block_tdesc_attr<memory_space = slm>>
-  gpu.return
-}
-
 // CHECK: gpu.func @test_prefetch_nd_vc(%[[arg0:.*]]: memref<24x32xf16>) {
 gpu.func @test_prefetch_nd_vc(%src: memref<24x32xf16>) {
   // CHECK: %[[R0:.*]] = xegpu.create_nd_tdesc %[[arg0]][0, 0] : memref<24x32xf16> -> !xegpu.tensor_desc<8x16xf16>
@@ -104,24 +97,17 @@ gpu.func @test_create_update_nd_tdesc_vc(%src: memref<24x32xf32>) {
 
 // CHECK: gpu.func @test_create_tdesc_vc(%[[arg0:.*]]: ui64) {
 gpu.func @test_create_tdesc_vc(%src: ui64) {
-  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
-  gpu.return
-}
-
-// CHECK: gpu.func @test_create_tdesc_vc_1(%[[arg0:.*]]: memref<?xf32, 3>) {
-gpu.func @test_create_tdesc_vc_1(%src: memref<?xf32, 3>) {
-  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] : memref<?xf32, 3> -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<memory_space =  slm, chunk_size = 2 : i64>>
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] : memref<?xf32, 3>  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<memory_space = slm, chunk_size = 2>>
+  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] {chunk_size = 2 : i64} : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2} : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
   gpu.return
 }
 
 // CHECK: gpu.func @test_prefetch_vc(%[[arg0:.*]]: ui64) {
 gpu.func @test_prefetch_vc(%src: ui64) {
-  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
-  // CHECK: xegpu.prefetch %[[R0]] <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>}> : !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>
-  xegpu.prefetch %1 <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>}>: !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] {chunk_size = 2 : i64} : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2} : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  // CHECK: xegpu.prefetch %[[R0]] <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>}> : !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  xegpu.prefetch %1 <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>}>: !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
   gpu.return
 }
 
@@ -129,12 +115,12 @@ gpu.func @test_prefetch_vc(%src: ui64) {
 gpu.func @test_load_gather_vc(%src: ui64) {
   //CHECK: %[[cst:.*]] = arith.constant dense<true> : vector<4xi1>
   %0 = arith.constant dense<1>: vector<4xi1>
-  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
-  //CHECK: %[[R1:.*]] = xegpu.load %[[R0]], %[[cst]] <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>, transpose}>
-  //CHECK-SAME: !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>, vector<4xi1> -> vector<2x4xf32>
-  %2 = xegpu.load %1, %0 <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>, transpose}>
-        : !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>, vector<4xi1> -> vector<2x4xf32>
+  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] {chunk_size = 2 : i64} : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2} : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  //CHECK: %[[R1:.*]] = xegpu.load %[[R0]], %[[cst]] <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>}>
+  //CHECK-SAME: !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>, vector<4xi1> -> vector<4x2xf32>
+  %2 = xegpu.load %1, %0 <{l1_hint = #xegpu.cache_hint<cached>, l2_hint = #xegpu.cache_hint<uncached>}>
+        : !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>, vector<4xi1> -> vector<4x2xf32>
   gpu.return
 }
 
@@ -142,23 +128,23 @@ gpu.func @test_load_gather_vc(%src: ui64) {
 gpu.func @test_store_scatter_vc(%src: ui64) {
   //CHECK: %[[c0:.*]] = arith.constant dense<true> : vector<4xi1>
   %0 = arith.constant dense<1>: vector<4xi1>
-  //CHECK: %[[c1:.*]] = arith.constant dense<2.900000e+00> : vector<2x4xf32>
-  %1 = arith.constant dense<2.9>: vector<2x4xf32>
-  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>
-  %2 = xegpu.create_tdesc %src[0, 8, 16, 24] : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
-  //CHECK: xegpu.store %[[c1]], %[[R0]], %[[c0]] <{l1_hint = #xegpu.cache_hint<write_back>, l2_hint = #xegpu.cache_hint<uncached>, transpose}>
-  //CHECK-SAME: vector<2x4xf32>, !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>, vector<4xi1>
-  xegpu.store %1, %2, %0 <{l1_hint = #xegpu.cache_hint<write_back>, l2_hint = #xegpu.cache_hint<uncached>, transpose}>
-        : vector<2x4xf32>, !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>, vector<4xi1>
+  //CHECK: %[[c1:.*]] = arith.constant dense<2.900000e+00> : vector<4x2xf32>
+  %1 = arith.constant dense<2.9>: vector<4x2xf32>
+  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] {chunk_size = 2 : i64} : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  %2 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2} : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  //CHECK: xegpu.store %[[c1]], %[[R0]], %[[c0]] <{l1_hint = #xegpu.cache_hint<write_back>, l2_hint = #xegpu.cache_hint<uncached>}>
+  //CHECK-SAME: vector<4x2xf32>, !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>, vector<4xi1>
+  xegpu.store %1, %2, %0 <{l1_hint = #xegpu.cache_hint<write_back>, l2_hint = #xegpu.cache_hint<uncached>}>
+        : vector<4x2xf32>, !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>, vector<4xi1>
   gpu.return
 }
 
 // CHECK: gpu.func @test_create_update_tdesc_vc(%[[arg0:.*]]: ui64) {
 gpu.func @test_create_update_tdesc_vc(%src: ui64) {
-  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24]: ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
-  //CHECK: %[[R1:.*]] = xegpu.update_offset %[[R0]], [32, 32, 32, 32] : !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2 : i64>>
-  %2 = xegpu.update_offset %1, [32, 32, 32, 32] : !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %arg0 [0, 8, 16, 24] {chunk_size = 2 : i64} : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2} : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  //CHECK: %[[R1:.*]] = xegpu.update_offset %[[R0]], [32, 32, 32, 32] : !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
+  %2 = xegpu.update_offset %1, [32, 32, 32, 32] : !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
   gpu.return
 }
 
@@ -179,10 +165,10 @@ gpu.func @test_dpas_vc_with_packed_b(%a : vector<8x16xf16>, %b: vector<8x16x2xf1
 
 // CHECK: gpu.func @test_atomic_rmw(%[[arg0:.*]]: ui64, %[[arg1:.*]]: vector<16xf32>, %[[arg2:.*]]: vector<16xi1>)
 gpu.func @test_atomic_rmw(%src: ui64, %value : vector<16xf32>, %mask : vector<16xi1>) {
-  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %[[arg0]] [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] : ui64 -> !xegpu.tensor_desc<16xf32, #xegpu.scatter_tdesc_attr<>>
-  %1 = xegpu.create_tdesc %src[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]: ui64 -> !xegpu.tensor_desc<16xf32, #xegpu.scatter_tdesc_attr<>>
-  //CHECK: %[[R1:.*]] = xegpu.atomic_rmw addf %[[R0]], %[[arg2]], %[[arg1]] : !xegpu.tensor_desc<16xf32, #xegpu.scatter_tdesc_attr<>>, vector<16xi1>, vector<16xf32> -> vector<16xf32>
-  xegpu.atomic_rmw addf %1, %mask, %value: !xegpu.tensor_desc<16xf32, #xegpu.scatter_tdesc_attr<>>, vector<16xi1>, vector<16xf32> -> vector<16xf32>
+  //CHECK: %[[R0:.*]] = xegpu.create_tdesc %[[arg0]] [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] : ui64 -> !xegpu.tensor_desc<16xf32, #xegpu.tdesc_attr<scattered = true>>
+  %1 = xegpu.create_tdesc %src[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]: ui64 -> !xegpu.tensor_desc<16xf32, #xegpu.tdesc_attr<scattered = true>>
+  //CHECK: %[[R1:.*]] = xegpu.atomic_rmw addf %[[R0]], %[[arg2]], %[[arg1]] : !xegpu.tensor_desc<16xf32, #xegpu.tdesc_attr<scattered = true>>, vector<16xi1>, vector<16xf32> -> vector<16xf32>
+  xegpu.atomic_rmw addf %1, %mask, %value: !xegpu.tensor_desc<16xf32, #xegpu.tdesc_attr<scattered = true>>, vector<16xi1>, vector<16xf32> -> vector<16xf32>
   gpu.return
 }
 

diff  --git a/mlir/test/Dialect/XeGPU/invalid.mlir b/mlir/test/Dialect/XeGPU/invalid.mlir
index 193dae352e3707..7ef50bb2b5fadf 100644
--- a/mlir/test/Dialect/XeGPU/invalid.mlir
+++ b/mlir/test/Dialect/XeGPU/invalid.mlir
@@ -15,20 +15,6 @@ func.func @test_create_nd_tdesc_vc_2(%src: memref<24x32xf32>) {
   return
 }
 
-// -----
-func.func @test_create_nd_tdesc_vc_3(%src: memref<2x24x32xf32, 3>) {
-  // expected-error at +1 {{SLM is not supported for 2D Block TensorDesc}}
-  %1 = xegpu.create_nd_tdesc %src[0, 0, 0] : memref<2x24x32xf32, 3> -> !xegpu.tensor_desc<8x16xf32, #xegpu.block_tdesc_attr<memory_space = slm>>
-  return
-}
-
-// -----
-func.func @test_create_nd_tdesc_vc_4(%src: memref<2x24x32xf32, 3>) {
-  // expected-error at +1 {{Memory space mismatch}}
-  %1 = xegpu.create_nd_tdesc %src[0, 0, 0] : memref<2x24x32xf32, 3> -> !xegpu.tensor_desc<16xf32>
-  return
-}
-
 // -----
 func.func @test_prefetch_nd_vc_1(%src: memref<24x32xf16>) {
   %1 = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf16> -> !xegpu.tensor_desc<8x16xf16>
@@ -40,10 +26,10 @@ func.func @test_prefetch_nd_vc_1(%src: memref<24x32xf16>) {
 // -----
 func.func @test_prefetch_nd_vc_2(%src: memref<24xf16>) {
   %1 = xegpu.create_tdesc %src[0, 1, 2, 3, 4, 5, 6, 7]
-        : memref<24xf16> -> !xegpu.tensor_desc<8xf16, #xegpu.scatter_tdesc_attr<>>
+        : memref<24xf16> -> !xegpu.tensor_desc<8xf16, #xegpu.tdesc_attr<scattered=true>>
   // expected-error at +1 {{Expects a non-scattered TensorDesc}}
   xegpu.prefetch_nd %1 <{l1_hint = #xegpu.cache_hint<cached>}>
-        : !xegpu.tensor_desc<8xf16, #xegpu.scatter_tdesc_attr<>>
+        : !xegpu.tensor_desc<8xf16, #xegpu.tdesc_attr<scattered=true>>
   return
 }
 
@@ -58,11 +44,11 @@ func.func @test_load_nd_vc_1(%src: memref<8x16xf16>) {
 
 // -----
 func.func @test_load_nd_vc_2(%src: memref<16xf16>) {
-  %1 = xegpu.create_tdesc %src[0, 2, 4, 6, 8, 10, 12, 14]
-        : memref<16xf16> -> !xegpu.tensor_desc<8x2xf16, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  %1 = xegpu.create_tdesc %src[0, 2, 4, 6, 8, 10, 12, 14] {chunk_size = 2}
+        : memref<16xf16> -> !xegpu.tensor_desc<8x2xf16, #xegpu.tdesc_attr<scattered=true>>
   // expected-error at +1 {{Expects a non-scattered TensorDesc.}}
   %2 = xegpu.load_nd %1 <{l1_hint = #xegpu.cache_hint<cached>}>
-      : !xegpu.tensor_desc<8x2xf16, #xegpu.scatter_tdesc_attr<chunk_size = 2>> -> vector<8x2xf16>
+      : !xegpu.tensor_desc<8x2xf16, #xegpu.tdesc_attr<scattered=true>> -> vector<8x2xf16>
   return
 }
 
@@ -87,28 +73,28 @@ func.func @test_store_nd_vc_1(%dst: memref<24x32xf16>) {
 // -----
 func.func @test_store_nd_vc_2(%dst: memref<16xf16>) {
   %1 = arith.constant dense<1.0>: vector<8x2xf16>
-  %2 = xegpu.create_tdesc %dst[0, 2, 4, 6, 8, 10, 12, 14]
-        : memref<16xf16> -> !xegpu.tensor_desc<8x2xf16, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  %2 = xegpu.create_tdesc %dst[0, 2, 4, 6, 8, 10, 12, 14] {chunk_size = 2}
+        : memref<16xf16> -> !xegpu.tensor_desc<8x2xf16, #xegpu.tdesc_attr<scattered=true>>
   // expected-error at +1 {{Expects a non-scattered TensorDesc}}
   xegpu.store_nd %1, %2 <{l1_hint = #xegpu.cache_hint<streaming>}>
-        : vector<8x2xf16>, !xegpu.tensor_desc<8x2xf16, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+        : vector<8x2xf16>, !xegpu.tensor_desc<8x2xf16, #xegpu.tdesc_attr<scattered=true>>
   return
 }
 
 // -----
 func.func @test_update_nd_offset_1(%dst: memref<16xf16>) {
-  %1 = xegpu.create_tdesc %dst[0, 2, 4, 6, 8, 10, 12, 14]
-        : memref<16xf16> -> !xegpu.tensor_desc<8x2xf16, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  %1 = xegpu.create_tdesc %dst[0, 2, 4, 6, 8, 10, 12, 14] {chunk_size = 2}
+        : memref<16xf16> -> !xegpu.tensor_desc<8x2xf16, #xegpu.tdesc_attr<scattered=true>>
   // expected-error at +1 {{Expects a non-scattered TensorDesc}}
-  xegpu.update_nd_offset %1, [0, 2] : !xegpu.tensor_desc<8x2xf16, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  xegpu.update_nd_offset %1, [0, 2] : !xegpu.tensor_desc<8x2xf16, #xegpu.tdesc_attr<scattered=true>>
   return
 }
 
 // -----
 func.func @test_create_tdesc_vc_1(%src: ui64) {
   // expected-error at +1 {{Expects a scattered TensorDesc}}
-  %1 = xegpu.create_tdesc %src[0, 2, 4, 6, 8, 10, 12, 14]
-        : ui64 -> !xegpu.tensor_desc<8xf16>
+  %1 = xegpu.create_tdesc %src[0, 2, 4, 6, 8, 10, 12, 14] {chunk_size = 2}
+        : ui64 -> !xegpu.tensor_desc<8x2xf16>
   return
 }
 
@@ -116,14 +102,7 @@ func.func @test_create_tdesc_vc_1(%src: ui64) {
 func.func @test_create_tdesc_vc_2(%src: ui64) {
   // expected-error at +1 {{Incorrect TensorDesc shape}}
   %1 = xegpu.create_tdesc %src[0, 2, 4, 6, 8, 10, 12, 14] {chunk_size = 2}
-        : ui64 -> !xegpu.tensor_desc<8x4xf16, #xegpu.scatter_tdesc_attr<>>
-  return
-}
-
-// -----
-func.func @test_create_tdesc_vc_1(%src: memref<?xf32>) {
-  // expected-error at +1 {{Memory space mismatch}}
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] : memref<?xf32>  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<memory_space = slm, chunk_size = 2>>
+        : ui64 -> !xegpu.tensor_desc<8x4xf16, #xegpu.tdesc_attr<scattered = true>>
   return
 }
 
@@ -137,9 +116,9 @@ func.func @test_prefetch_vc_1(%src: memref<24x32xf16>) {
 
 // -----
 func.func @test_prefetch_vc_2(%src: ui64) {
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2} : ui64  -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
   // expected-error at +1 {{invlid l1_hint: #xegpu.cache_hint<write_back>}}
-  xegpu.prefetch %1 <{l1_hint = #xegpu.cache_hint<write_back>}>: !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  xegpu.prefetch %1 <{l1_hint = #xegpu.cache_hint<write_back>}>: !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
   return
 }
 
@@ -156,11 +135,11 @@ func.func @test_load_gather_vc_1(%src: memref<24x32xf16>) {
 // -----
 func.func @test_load_gather_vc_2(%src: ui64) {
   %0 = arith.constant dense<1>: vector<4xi1>
-  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] : ui64
-        -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  %1 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2} : ui64
+        -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
   // expected-error at +1 {{invlid l1_hint: #xegpu.cache_hint<write_back>}}
   %2 = xegpu.load %1, %0 <{l1_hint = #xegpu.cache_hint<write_back>}>
-        : !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>, vector<4xi1>
+        : !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>, vector<4xi1>
           -> vector<4x2xf32>
   return
 }
@@ -180,11 +159,11 @@ func.func @test_store_scatter_vc_1(%src: memref<24x32xf32>) {
 func.func @test_store_scatter_vc_2(%src: ui64) {
   %0 = arith.constant dense<1>: vector<4xi1>
   %1 = arith.constant dense<2.9>: vector<4x2xf32>
-  %2 = xegpu.create_tdesc %src[0, 8, 16, 24]
-          : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>
+  %2 = xegpu.create_tdesc %src[0, 8, 16, 24] {chunk_size = 2}
+          : ui64 -> !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>
   // expected-error at +1 {{invlid l1_hint: #xegpu.cache_hint<streaming>}}
   xegpu.store %1, %2, %0 <{l1_hint = #xegpu.cache_hint<streaming>}> : vector<4x2xf32>,
-          !xegpu.tensor_desc<4x2xf32, #xegpu.scatter_tdesc_attr<chunk_size = 2>>, vector<4xi1>
+          !xegpu.tensor_desc<4x2xf32, #xegpu.tdesc_attr<scattered = true>>, vector<4xi1>
   return
 }
 
@@ -203,9 +182,9 @@ func.func @test_dpas_vc_2(%a : vector<8x8x2xf16>, %b: vector<8x16x2xf16>) {
 }
 
 // -----
-func.func @test_atomic_rmw(%src: ui64, %value : vector<16x4xf32>, %mask : vector<16xi1>) {
-  %1 = xegpu.create_tdesc %src[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] : ui64 -> !xegpu.tensor_desc<16x8xf32, #xegpu.scatter_tdesc_attr<chunk_size = 8>>
-  // expected-error at +1 {{failed to verify that all of {tensorDesc, value, result} have same shape}}
-  xegpu.atomic_rmw addf %1, %mask, %value: !xegpu.tensor_desc<16x8xf32, #xegpu.scatter_tdesc_attr<chunk_size = 8>>, vector<16xi1>, vector<16x4xf32> -> vector<16x8xf32>
-  return
+func.func @test_atomic_rmw(%src: ui64, %value : vector<16x8xf32>, %mask : vector<16xi1>) {
+  %1 = xegpu.create_tdesc %src[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] {chunk_size = 8}: ui64 -> !xegpu.tensor_desc<16x8xf32, #xegpu.tdesc_attr<scattered = true>>
+  // expected-error at +1 {{failed to verify that all of {tensorDesc, mask, value, result} have same shape}}
+  xegpu.atomic_rmw addf %1, %mask, %value: !xegpu.tensor_desc<16x8xf32, #xegpu.tdesc_attr<scattered = true>>, vector<16xi1>, vector<16x8xf32> -> vector<16x8xf32>
+  gpu.return
 }
\ No newline at end of file


        


More information about the llvm-branch-commits mailing list