[Mlir-commits] [mlir] [mlir][XeGPU] add unroll patterns for load_matrix and store_matrix (PR #154637)

Chao Chen llvmlistbot at llvm.org
Wed Aug 27 11:00:17 PDT 2025


https://github.com/chencha3 updated https://github.com/llvm/llvm-project/pull/154637

>From 22d4193fdec06afe6f3a3518480fe6cd32e4f0dc Mon Sep 17 00:00:00 2001
From: Chao Chen <chao.chen at intel.com>
Date: Wed, 20 Aug 2025 22:09:11 +0000
Subject: [PATCH 1/4] add unroll pattern and unit test for load_matrix and
 store_matrix

---
 .../mlir/Dialect/XeGPU/Transforms/Passes.td   |  4 +-
 .../mlir/Dialect/XeGPU/Utils/XeGPUUtils.h     | 16 ++++
 .../XeGPU/Transforms/XeGPUBlocking.cpp        | 12 +--
 .../Dialect/XeGPU/Transforms/XeGPUUnroll.cpp  | 87 +++++++++++++++++--
 mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp   | 46 ++++++++++
 mlir/test/Dialect/XeGPU/xegpu-blocking.mlir   | 23 +++++
 6 files changed, 176 insertions(+), 12 deletions(-)

diff --git a/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td b/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td
index 3a88dae041dd1..ddf6b4ac85a90 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td
@@ -67,8 +67,8 @@ def XeGPUBlocking: Pass<"xegpu-blocking"> {
     to a hardware instruction.
   }];
   let dependentDialects = [
-      "memref::MemRefDialect", "xegpu::XeGPUDialect", "vector::VectorDialect"
-  ];
+      "memref::MemRefDialect", "xegpu::XeGPUDialect", "vector::VectorDialect",
+      "index::IndexDialect"];
 }
 
 #endif // MLIR_DIALECT_XEGPU_TRANSFORMS_PASSES_TD
diff --git a/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h b/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h
index db8608c6d20b8..a40dc74edb200 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h
+++ b/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h
@@ -10,6 +10,7 @@
 #define MLIR_DIALECT_XEGPU_UTILS_XEGPUUTILS_H_
 
 #include "mlir/IR/BuiltinTypes.h"
+#include "mlir/IR/OpDefinition.h"
 namespace mlir {
 
 class VectorType;
@@ -18,6 +19,7 @@ class OpResult;
 class OpBuilder;
 class ValueRange;
 class TypeConverter;
+class OpFoldResult;
 
 namespace xegpu {
 class LayoutAttr;
@@ -128,6 +130,20 @@ void doSCFStructuralTypeConversionWithTensorType(Operation *op,
 /// if no GPU module parent or XeVM target attribute exists.
 std::optional<std::string> getChipStr(Operation *op);
 
+/// Generates element-wise addition ops of two arrays with automatic alignment.
+/// When the input arrays have different sizes, the shorter array is
+/// right-aligned with the longer array, and the unmatched leading elements from
+/// the longer array are preserved unchanged. This is commonly used for offset
+/// computation where higher-dimensional offsets need to be added to
+/// lower-dimensional adjustments.
+///
+/// Example:
+///   lhs = [l1, l2, l3], rhs = [r1, r2]
+///   Result: [11, l2+r1, l3+r2]
+SmallVector<OpFoldResult> addWithRightAligned(OpBuilder &builder, Location loc,
+                                              ArrayRef<OpFoldResult> lhs,
+                                              ArrayRef<OpFoldResult> rhs);
+
 } // namespace xegpu
 
 } // namespace mlir
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
index d82c541f31359..b11f5fe87559b 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
@@ -8,6 +8,7 @@
 
 #include "mlir/Dialect/XeGPU/Transforms/Passes.h"
 
+#include "mlir/Dialect/Index/IR/IndexDialect.h"
 #include "mlir/Dialect/Vector/Transforms/VectorTransforms.h"
 #include "mlir/Dialect/XeGPU/IR/XeGPU.h"
 #include "mlir/Dialect/XeGPU/Transforms/Transforms.h"
@@ -155,10 +156,10 @@ XeGPUBlockingPass::getTileShape(const T &operandOrResult) const {
 std::optional<SmallVector<int64_t>>
 XeGPUBlockingPass::getTileShape(Operation *op) const {
   if (isa<xegpu::CreateNdDescOp, xegpu::UpdateNdOffsetOp, xegpu::CreateDescOp,
-          xegpu::UpdateOffsetOp>(op))
+          xegpu::UpdateOffsetOp, xegpu::LoadMatrixOp>(op))
     return getTileShape(op->getOpResult(0));
   if (isa<xegpu::PrefetchNdOp, xegpu::LoadNdOp, xegpu::PrefetchOp,
-          xegpu::LoadGatherOp>(op))
+          xegpu::LoadGatherOp, xegpu::StoreMatrixOp>(op))
     return getTileShape(op->getOpOperand(0));
   if (isa<xegpu::StoreNdOp, xegpu::StoreScatterOp>(op))
     return getTileShape(op->getOpOperand(1));
@@ -202,17 +203,18 @@ XeGPUBlockingPass::getTileShape(Operation *op) const {
 
 bool XeGPUBlockingPass::needsUnroll(Operation *op) const {
   // skip the op if any of its operands or results has workgroup level layouts
-  bool hasWgLayoutOperands =
+  bool hasSgLayoutOperands =
       llvm::any_of(op->getOpOperands(), [](OpOperand &opr) {
         xegpu::LayoutAttr layout = xegpu::getLayoutAttr(opr);
         return layout && layout.isWgLayout();
       });
-  bool hasWgLayoutResults =
+  bool hasSgLayoutResults =
       llvm::any_of(op->getOpResults(), [](OpResult result) {
         xegpu::LayoutAttr layout = xegpu::getLayoutAttr(result);
         return layout && layout.isWgLayout();
       });
-  if (hasWgLayoutOperands || hasWgLayoutResults) {
+
+  if (hasSgLayoutOperands || hasSgLayoutResults) {
     LDBG() << "skip unrolling for op with workgroup level layout: " << *op;
     return false;
   }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
index c793b71639e86..219e4e6f44618 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
@@ -682,13 +682,90 @@ struct UnrollUpdateOffsetOp : public UnrollPattern<xegpu::UpdateOffsetOp> {
   }
 };
 
+struct UnrollLoadMatrixOp : public UnrollPattern<xegpu::LoadMatrixOp> {
+  using UnrollPattern<xegpu::LoadMatrixOp>::UnrollPattern;
+  LogicalResult matchAndRewrite(xegpu::LoadMatrixOp op,
+                                PatternRewriter &rewriter) const override {
+    std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
+    if (!targetShape)
+      return failure();
+
+    Location loc = op.getLoc();
+    VectorType valueTy = op.getType();
+    Type elemTy = valueTy.getElementType();
+    ArrayRef<int64_t> shape = valueTy.getShape();
+    auto layout = dyn_cast<xegpu::LayoutAttr>(op.getLayoutAttr());
+
+    VectorType newValueTy = valueTy.cloneWith(*targetShape, elemTy);
+
+    SmallVector<OpFoldResult> mixedOffsets = op.getMixedOffsets();
+    SmallVector<SmallVector<OpFoldResult>> offsetsList;
+    for (SmallVector<int64_t> offsets :
+         StaticTileOffsetRange(shape, *targetShape)) {
+      auto adds = xegpu::addWithRightAligned(
+          rewriter, loc, mixedOffsets,
+          getAsIndexOpFoldResult(op.getContext(), offsets));
+      offsetsList.push_back(adds);
+    }
+
+    SmallVector<Value> newOps;
+    for (SmallVector<OpFoldResult> offsets : offsetsList) {
+      auto newOp = rewriter.create<xegpu::LoadMatrixOp>(
+          op.getLoc(), newValueTy, op.getMemDesc(), offsets,
+          layout.dropInstData());
+      newOps.push_back(newOp);
+    }
+    Value castOp = unpack(newOps, op.getType(), *targetShape, loc, rewriter);
+    rewriter.replaceOp(op, castOp);
+    return success();
+  }
+};
+
+struct UnrollStoreMatrixOp : public UnrollPattern<xegpu::StoreMatrixOp> {
+  using UnrollPattern<xegpu::StoreMatrixOp>::UnrollPattern;
+  LogicalResult matchAndRewrite(xegpu::StoreMatrixOp op,
+                                PatternRewriter &rewriter) const override {
+    std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
+    if (!targetShape)
+      return failure();
+
+    Location loc = op.getLoc();
+    VectorType valueTy = op.getData().getType();
+    ArrayRef<int64_t> shape = valueTy.getShape();
+    auto layout = dyn_cast<xegpu::LayoutAttr>(op.getLayoutAttr());
+
+    SmallVector<Type> convertedValTypes =
+        getUnrolledTypes(valueTy, *targetShape);
+    SmallVector<Value> convertedValues =
+        pack(op.getData(), convertedValTypes, *targetShape, loc, rewriter);
+
+    SmallVector<OpFoldResult> mixedOffsets = op.getMixedOffsets();
+    SmallVector<SmallVector<OpFoldResult>> offsetsList;
+    for (SmallVector<int64_t> offsets :
+         StaticTileOffsetRange(shape, *targetShape)) {
+      auto adds = xegpu::addWithRightAligned(
+          rewriter, loc, mixedOffsets,
+          getAsIndexOpFoldResult(op.getContext(), offsets));
+      offsetsList.push_back(adds);
+    }
+
+    for (auto [v, offsets] : llvm::zip_equal(convertedValues, offsetsList))
+      rewriter.create<xegpu::StoreMatrixOp>(loc, v, op.getMemDesc(), offsets,
+                                            layout.dropInstData());
+
+    rewriter.eraseOp(op);
+    return success();
+  }
+};
+
 } // namespace
 
 void mlir::xegpu::populateXeGPUUnrollPatterns(
     RewritePatternSet &patterns, const xegpu::UnrollOptions &options) {
-  patterns.add<UnrollCreateNdOp, UnrollUpdateNdOffsetOp, UnrollPrefetchNdOp,
-               UnrollLoadNdOp, UnrollStoreNdOp, UnrollDpasOp,
-               UnrollCreateDescOp, UnrollLoadGatherOp, UnrollStoreScatterOp,
-               UnrollPrefetchOp, UnrollUpdateOffsetOp>(patterns.getContext(),
-                                                       options);
+  patterns
+      .add<UnrollCreateNdOp, UnrollUpdateNdOffsetOp, UnrollPrefetchNdOp,
+           UnrollLoadNdOp, UnrollStoreNdOp, UnrollDpasOp, UnrollCreateDescOp,
+           UnrollLoadGatherOp, UnrollStoreScatterOp, UnrollPrefetchOp,
+           UnrollUpdateOffsetOp, UnrollLoadMatrixOp, UnrollStoreMatrixOp>(
+          patterns.getContext(), options);
 }
diff --git a/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp b/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
index 19eedbac0f76b..088e8a8c497d9 100644
--- a/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
+++ b/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
@@ -12,6 +12,7 @@
 
 #include "mlir/Dialect/XeGPU/Utils/XeGPUUtils.h"
 #include "mlir/Dialect/GPU/IR/GPUDialect.h"
+#include "mlir/Dialect/Index/IR/IndexOps.h"
 #include "mlir/Dialect/LLVMIR/XeVMDialect.h"
 #include "mlir/Dialect/SCF/Transforms/Patterns.h"
 #include "mlir/Dialect/Utils/IndexingUtils.h"
@@ -133,6 +134,14 @@ xegpu::LayoutAttr xegpu::getLayoutAttr(const Value value) {
     if (auto loadNd = dyn_cast<xegpu::LoadNdOp>(defOp))
       return getLayoutAttr(loadNd.getTensorDesc());
 
+    // for LoadMatrixOp, the layout is attached to the property of the op
+    if (auto loadOp = dyn_cast<xegpu::LoadMatrixOp>(defOp))
+      return dyn_cast_if_present<xegpu::LayoutAttr>(loadOp.getLayoutAttr());
+
+    // for StoreMatrixOp, the layout is attached to the property of the op
+    if (auto storeOp = dyn_cast<xegpu::StoreMatrixOp>(defOp))
+      return dyn_cast_if_present<xegpu::LayoutAttr>(storeOp.getLayoutAttr());
+
     std::string layoutName = getLayoutName(result);
     if (defOp->hasAttr(layoutName))
       return defOp->getAttrOfType<xegpu::LayoutAttr>(layoutName);
@@ -152,6 +161,13 @@ xegpu::LayoutAttr xegpu::getLayoutAttr(const Value value) {
 
 xegpu::LayoutAttr xegpu::getLayoutAttr(const OpOperand &opr) {
   Operation *op = opr.getOwner();
+
+  if (auto loadOp = dyn_cast<xegpu::LoadMatrixOp>(op))
+    return dyn_cast_if_present<xegpu::LayoutAttr>(loadOp.getLayoutAttr());
+
+  if (auto storeOp = dyn_cast<xegpu::StoreMatrixOp>(op))
+    return dyn_cast_if_present<xegpu::LayoutAttr>(storeOp.getLayoutAttr());
+
   std::string layoutName = xegpu::getLayoutName(opr);
   if (op->hasAttr(layoutName))
     return op->getAttrOfType<xegpu::LayoutAttr>(layoutName);
@@ -179,6 +195,8 @@ xegpu::setLayoutAttr<mlir::OpOperand>(const mlir::OpOperand &operand,
 void xegpu::setLayoutAttrs(Operation *op,
                            function_ref<LayoutAttr(Value)> getLayoutImpl) {
   op->walk([&](Operation *nestOp) {
+    if (isa<xegpu::LoadMatrixOp, xegpu::StoreMatrixOp>(nestOp))
+      return;
     for (OpOperand &opr : nestOp->getOpOperands()) {
       auto layout = getLayoutImpl(opr.get());
       setLayoutAttr(opr, layout);
@@ -424,3 +442,31 @@ std::optional<std::string> xegpu::getChipStr(Operation *op) {
 
   return std::nullopt;
 }
+
+/// Generates element-wise addition ops of two arrays with automatic alignment.
+/// When the input arrays have different sizes, the shorter array is
+/// right-aligned with the longer array, and the unmatched leading elements from
+/// the longer array are preserved unchanged. This is commonly used for offset
+/// computation where higher-dimensional offsets need to be added to
+/// lower-dimensional adjustments.
+///
+/// Example:
+///   lhs = [l1, l2, l3], rhs = [r1, r2]
+///   Result: [11, l2+r1, l3+r2]
+SmallVector<OpFoldResult>
+xegpu::addWithRightAligned(OpBuilder &builder, Location loc,
+                           ArrayRef<OpFoldResult> lhs,
+                           ArrayRef<OpFoldResult> rhs) {
+  // ensure a is longer than b
+  ArrayRef<OpFoldResult> a = lhs.size() >= rhs.size() ? lhs : rhs;
+  ArrayRef<OpFoldResult> b = lhs.size() >= rhs.size() ? rhs : lhs;
+  SmallVector<OpFoldResult> results(a.take_front(a.size() - b.size()));
+  a = a.slice(a.size() - b.size());
+  for (auto [l, r] : llvm::zip(a, b)) {
+    auto lval = getValueOrCreateConstantIndexOp(builder, loc, l);
+    auto rval = getValueOrCreateConstantIndexOp(builder, loc, r);
+    results.push_back(builder.createOrFold<index::AddOp>(loc, lval, rval));
+  }
+  return results;
+  return {};
+}
diff --git a/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir b/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir
index d986e5bd1cfb4..9d63c2ddd4895 100644
--- a/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir
+++ b/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir
@@ -561,3 +561,26 @@ gpu.module @test_kernel {
     gpu.return %e : vector<8x32x2xf16>
   }
 }
+
+// -----
+gpu.module @test_kernel {
+  //CHECK-LABEL: unroll_load_matrix
+  gpu.func @unroll_load_matrix(%arg0: memref<4096xi8, 3>) -> vector<32x32xf32> {
+    %0 = xegpu.create_mem_desc %arg0 : memref<4096xi8, 3> -> !xegpu.mem_desc<32x32xf32>
+    //CHECK-COUNT-8: xegpu.load_matrix {{.*}} : !xegpu.mem_desc<32x32xf32>, index, index -> vector<8x16xf32>
+    //CHECK-COUNT-8: vector.insert_strided_slice {{.*}} : vector<8x16xf32> into vector<32x32xf32>
+    %1 = xegpu.load_matrix %0[0, 0] <{layout = #xegpu.layout<inst_data = [8, 16]>}>: !xegpu.mem_desc<32x32xf32> -> vector<32x32xf32>
+    gpu.return %1: vector<32x32xf32>
+  }
+}
+
+// -----
+gpu.module @test_kernel {
+  // CHECK-LABEL: unroll_store_matrix
+  gpu.func @unroll_store_matrix(%value: vector<32x32xf32>, %arg0 : memref<32768xi8, 3>) {
+    %mdesc = xegpu.create_mem_desc %arg0 : memref<32768xi8, 3> -> !xegpu.mem_desc<64x128xf32>
+    // CHECK-COUNT-8:  xegpu.store_matrix {{.*}} : vector<8x16xf32>, !xegpu.mem_desc<64x128xf32>, index, index
+    xegpu.store_matrix %value, %mdesc[0, 0] {layout = #xegpu.layout<inst_data = [8, 16]>} : vector<32x32xf32>, !xegpu.mem_desc<64x128xf32>
+    gpu.return
+  }
+}

>From d20da858449bfd926df69de3f2b777cae4ee2f24 Mon Sep 17 00:00:00 2001
From: Chao Chen <chao.chen at intel.com>
Date: Wed, 20 Aug 2025 22:09:11 +0000
Subject: [PATCH 2/4] add unroll pattern and unit test for load_matrix and
 store_matrix

---
 .../mlir/Dialect/XeGPU/Transforms/Passes.td   |  4 +-
 .../mlir/Dialect/XeGPU/Utils/XeGPUUtils.h     |  1 +
 .../XeGPU/Transforms/XeGPUBlocking.cpp        | 12 +--
 .../Dialect/XeGPU/Transforms/XeGPUUnroll.cpp  | 87 +++++++++++++++++--
 mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp   | 17 ++++
 mlir/test/Dialect/XeGPU/xegpu-blocking.mlir   | 23 +++++
 6 files changed, 132 insertions(+), 12 deletions(-)

diff --git a/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td b/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td
index 3a88dae041dd1..ddf6b4ac85a90 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td
@@ -67,8 +67,8 @@ def XeGPUBlocking: Pass<"xegpu-blocking"> {
     to a hardware instruction.
   }];
   let dependentDialects = [
-      "memref::MemRefDialect", "xegpu::XeGPUDialect", "vector::VectorDialect"
-  ];
+      "memref::MemRefDialect", "xegpu::XeGPUDialect", "vector::VectorDialect",
+      "index::IndexDialect"];
 }
 
 #endif // MLIR_DIALECT_XEGPU_TRANSFORMS_PASSES_TD
diff --git a/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h b/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h
index b2b2d3ab85231..a40dc74edb200 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h
+++ b/mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h
@@ -19,6 +19,7 @@ class OpResult;
 class OpBuilder;
 class ValueRange;
 class TypeConverter;
+class OpFoldResult;
 
 namespace xegpu {
 class LayoutAttr;
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
index b3144e4c1e55d..fb4f00b21f2b9 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
@@ -8,6 +8,7 @@
 
 #include "mlir/Dialect/XeGPU/Transforms/Passes.h"
 
+#include "mlir/Dialect/Index/IR/IndexDialect.h"
 #include "mlir/Dialect/Vector/Transforms/VectorTransforms.h"
 #include "mlir/Dialect/XeGPU/IR/XeGPU.h"
 #include "mlir/Dialect/XeGPU/Transforms/Transforms.h"
@@ -155,10 +156,10 @@ XeGPUBlockingPass::getTileShape(const T &operandOrResult) const {
 std::optional<SmallVector<int64_t>>
 XeGPUBlockingPass::getTileShape(Operation *op) const {
   if (isa<xegpu::CreateNdDescOp, xegpu::UpdateNdOffsetOp, xegpu::CreateDescOp,
-          xegpu::UpdateOffsetOp>(op))
+          xegpu::UpdateOffsetOp, xegpu::LoadMatrixOp>(op))
     return getTileShape(op->getOpResult(0));
   if (isa<xegpu::PrefetchNdOp, xegpu::LoadNdOp, xegpu::PrefetchOp,
-          xegpu::LoadGatherOp>(op))
+          xegpu::LoadGatherOp, xegpu::StoreMatrixOp>(op))
     return getTileShape(op->getOpOperand(0));
   if (isa<xegpu::StoreNdOp, xegpu::StoreScatterOp>(op))
     return getTileShape(op->getOpOperand(1));
@@ -202,17 +203,18 @@ XeGPUBlockingPass::getTileShape(Operation *op) const {
 
 bool XeGPUBlockingPass::needsUnroll(Operation *op) const {
   // skip the op if any of its operands or results has workgroup level layouts
-  bool hasWgLayoutOperands =
+  bool hasSgLayoutOperands =
       llvm::any_of(op->getOpOperands(), [](OpOperand &opr) {
         xegpu::LayoutAttr layout = xegpu::getLayoutAttr(opr);
         return layout && layout.isForWorkgroup();
       });
-  bool hasWgLayoutResults =
+  bool hasSgLayoutResults =
       llvm::any_of(op->getOpResults(), [](OpResult result) {
         xegpu::LayoutAttr layout = xegpu::getLayoutAttr(result);
         return layout && layout.isForWorkgroup();
       });
-  if (hasWgLayoutOperands || hasWgLayoutResults) {
+
+  if (hasSgLayoutOperands || hasSgLayoutResults) {
     LDBG() << "skip unrolling for op with workgroup level layout: " << *op;
     return false;
   }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
index c793b71639e86..219e4e6f44618 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
@@ -682,13 +682,90 @@ struct UnrollUpdateOffsetOp : public UnrollPattern<xegpu::UpdateOffsetOp> {
   }
 };
 
+struct UnrollLoadMatrixOp : public UnrollPattern<xegpu::LoadMatrixOp> {
+  using UnrollPattern<xegpu::LoadMatrixOp>::UnrollPattern;
+  LogicalResult matchAndRewrite(xegpu::LoadMatrixOp op,
+                                PatternRewriter &rewriter) const override {
+    std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
+    if (!targetShape)
+      return failure();
+
+    Location loc = op.getLoc();
+    VectorType valueTy = op.getType();
+    Type elemTy = valueTy.getElementType();
+    ArrayRef<int64_t> shape = valueTy.getShape();
+    auto layout = dyn_cast<xegpu::LayoutAttr>(op.getLayoutAttr());
+
+    VectorType newValueTy = valueTy.cloneWith(*targetShape, elemTy);
+
+    SmallVector<OpFoldResult> mixedOffsets = op.getMixedOffsets();
+    SmallVector<SmallVector<OpFoldResult>> offsetsList;
+    for (SmallVector<int64_t> offsets :
+         StaticTileOffsetRange(shape, *targetShape)) {
+      auto adds = xegpu::addWithRightAligned(
+          rewriter, loc, mixedOffsets,
+          getAsIndexOpFoldResult(op.getContext(), offsets));
+      offsetsList.push_back(adds);
+    }
+
+    SmallVector<Value> newOps;
+    for (SmallVector<OpFoldResult> offsets : offsetsList) {
+      auto newOp = rewriter.create<xegpu::LoadMatrixOp>(
+          op.getLoc(), newValueTy, op.getMemDesc(), offsets,
+          layout.dropInstData());
+      newOps.push_back(newOp);
+    }
+    Value castOp = unpack(newOps, op.getType(), *targetShape, loc, rewriter);
+    rewriter.replaceOp(op, castOp);
+    return success();
+  }
+};
+
+struct UnrollStoreMatrixOp : public UnrollPattern<xegpu::StoreMatrixOp> {
+  using UnrollPattern<xegpu::StoreMatrixOp>::UnrollPattern;
+  LogicalResult matchAndRewrite(xegpu::StoreMatrixOp op,
+                                PatternRewriter &rewriter) const override {
+    std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
+    if (!targetShape)
+      return failure();
+
+    Location loc = op.getLoc();
+    VectorType valueTy = op.getData().getType();
+    ArrayRef<int64_t> shape = valueTy.getShape();
+    auto layout = dyn_cast<xegpu::LayoutAttr>(op.getLayoutAttr());
+
+    SmallVector<Type> convertedValTypes =
+        getUnrolledTypes(valueTy, *targetShape);
+    SmallVector<Value> convertedValues =
+        pack(op.getData(), convertedValTypes, *targetShape, loc, rewriter);
+
+    SmallVector<OpFoldResult> mixedOffsets = op.getMixedOffsets();
+    SmallVector<SmallVector<OpFoldResult>> offsetsList;
+    for (SmallVector<int64_t> offsets :
+         StaticTileOffsetRange(shape, *targetShape)) {
+      auto adds = xegpu::addWithRightAligned(
+          rewriter, loc, mixedOffsets,
+          getAsIndexOpFoldResult(op.getContext(), offsets));
+      offsetsList.push_back(adds);
+    }
+
+    for (auto [v, offsets] : llvm::zip_equal(convertedValues, offsetsList))
+      rewriter.create<xegpu::StoreMatrixOp>(loc, v, op.getMemDesc(), offsets,
+                                            layout.dropInstData());
+
+    rewriter.eraseOp(op);
+    return success();
+  }
+};
+
 } // namespace
 
 void mlir::xegpu::populateXeGPUUnrollPatterns(
     RewritePatternSet &patterns, const xegpu::UnrollOptions &options) {
-  patterns.add<UnrollCreateNdOp, UnrollUpdateNdOffsetOp, UnrollPrefetchNdOp,
-               UnrollLoadNdOp, UnrollStoreNdOp, UnrollDpasOp,
-               UnrollCreateDescOp, UnrollLoadGatherOp, UnrollStoreScatterOp,
-               UnrollPrefetchOp, UnrollUpdateOffsetOp>(patterns.getContext(),
-                                                       options);
+  patterns
+      .add<UnrollCreateNdOp, UnrollUpdateNdOffsetOp, UnrollPrefetchNdOp,
+           UnrollLoadNdOp, UnrollStoreNdOp, UnrollDpasOp, UnrollCreateDescOp,
+           UnrollLoadGatherOp, UnrollStoreScatterOp, UnrollPrefetchOp,
+           UnrollUpdateOffsetOp, UnrollLoadMatrixOp, UnrollStoreMatrixOp>(
+          patterns.getContext(), options);
 }
diff --git a/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp b/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
index 6835f64ad8ef7..f77749fd77831 100644
--- a/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
+++ b/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
@@ -134,6 +134,14 @@ xegpu::LayoutAttr xegpu::getLayoutAttr(const Value value) {
     if (auto loadNd = dyn_cast<xegpu::LoadNdOp>(defOp))
       return getLayoutAttr(loadNd.getTensorDesc());
 
+    // for LoadMatrixOp, the layout is attached to the property of the op
+    if (auto loadOp = dyn_cast<xegpu::LoadMatrixOp>(defOp))
+      return dyn_cast_if_present<xegpu::LayoutAttr>(loadOp.getLayoutAttr());
+
+    // for StoreMatrixOp, the layout is attached to the property of the op
+    if (auto storeOp = dyn_cast<xegpu::StoreMatrixOp>(defOp))
+      return dyn_cast_if_present<xegpu::LayoutAttr>(storeOp.getLayoutAttr());
+
     std::string layoutName = getLayoutName(result);
     if (defOp->hasAttr(layoutName))
       return defOp->getAttrOfType<xegpu::LayoutAttr>(layoutName);
@@ -153,6 +161,13 @@ xegpu::LayoutAttr xegpu::getLayoutAttr(const Value value) {
 
 xegpu::LayoutAttr xegpu::getLayoutAttr(const OpOperand &opr) {
   Operation *op = opr.getOwner();
+
+  if (auto loadOp = dyn_cast<xegpu::LoadMatrixOp>(op))
+    return dyn_cast_if_present<xegpu::LayoutAttr>(loadOp.getLayoutAttr());
+
+  if (auto storeOp = dyn_cast<xegpu::StoreMatrixOp>(op))
+    return dyn_cast_if_present<xegpu::LayoutAttr>(storeOp.getLayoutAttr());
+
   std::string layoutName = xegpu::getLayoutName(opr);
   if (op->hasAttr(layoutName))
     return op->getAttrOfType<xegpu::LayoutAttr>(layoutName);
@@ -180,6 +195,8 @@ xegpu::setLayoutAttr<mlir::OpOperand>(const mlir::OpOperand &operand,
 void xegpu::setLayoutAttrs(Operation *op,
                            function_ref<LayoutAttr(Value)> getLayoutImpl) {
   op->walk([&](Operation *nestOp) {
+    if (isa<xegpu::LoadMatrixOp, xegpu::StoreMatrixOp>(nestOp))
+      return;
     for (OpOperand &opr : nestOp->getOpOperands()) {
       auto layout = getLayoutImpl(opr.get());
       setLayoutAttr(opr, layout);
diff --git a/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir b/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir
index d986e5bd1cfb4..9d63c2ddd4895 100644
--- a/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir
+++ b/mlir/test/Dialect/XeGPU/xegpu-blocking.mlir
@@ -561,3 +561,26 @@ gpu.module @test_kernel {
     gpu.return %e : vector<8x32x2xf16>
   }
 }
+
+// -----
+gpu.module @test_kernel {
+  //CHECK-LABEL: unroll_load_matrix
+  gpu.func @unroll_load_matrix(%arg0: memref<4096xi8, 3>) -> vector<32x32xf32> {
+    %0 = xegpu.create_mem_desc %arg0 : memref<4096xi8, 3> -> !xegpu.mem_desc<32x32xf32>
+    //CHECK-COUNT-8: xegpu.load_matrix {{.*}} : !xegpu.mem_desc<32x32xf32>, index, index -> vector<8x16xf32>
+    //CHECK-COUNT-8: vector.insert_strided_slice {{.*}} : vector<8x16xf32> into vector<32x32xf32>
+    %1 = xegpu.load_matrix %0[0, 0] <{layout = #xegpu.layout<inst_data = [8, 16]>}>: !xegpu.mem_desc<32x32xf32> -> vector<32x32xf32>
+    gpu.return %1: vector<32x32xf32>
+  }
+}
+
+// -----
+gpu.module @test_kernel {
+  // CHECK-LABEL: unroll_store_matrix
+  gpu.func @unroll_store_matrix(%value: vector<32x32xf32>, %arg0 : memref<32768xi8, 3>) {
+    %mdesc = xegpu.create_mem_desc %arg0 : memref<32768xi8, 3> -> !xegpu.mem_desc<64x128xf32>
+    // CHECK-COUNT-8:  xegpu.store_matrix {{.*}} : vector<8x16xf32>, !xegpu.mem_desc<64x128xf32>, index, index
+    xegpu.store_matrix %value, %mdesc[0, 0] {layout = #xegpu.layout<inst_data = [8, 16]>} : vector<32x32xf32>, !xegpu.mem_desc<64x128xf32>
+    gpu.return
+  }
+}

>From 442c18aed86ab2423e20401d8c17f9c5b73543b3 Mon Sep 17 00:00:00 2001
From: Chao Chen <chao.chen at intel.com>
Date: Wed, 27 Aug 2025 17:54:16 +0000
Subject: [PATCH 3/4] merge

---
 mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp b/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
index 95aa1fc58f4f6..2e17e559fdd2d 100644
--- a/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
+++ b/mlir/lib/Dialect/XeGPU/Utils/XeGPUUtils.cpp
@@ -136,11 +136,11 @@ xegpu::DistributeLayoutAttr xegpu::getDistributeLayoutAttr(const Value value) {
 
     // for LoadMatrixOp, the layout is attached to the property of the op
     if (auto loadOp = dyn_cast<xegpu::LoadMatrixOp>(defOp))
-      return dyn_cast_if_present<xegpu::LayoutAttr>(loadOp.getLayoutAttr());
+      return loadOp.getLayoutAttr();
 
     // for StoreMatrixOp, the layout is attached to the property of the op
     if (auto storeOp = dyn_cast<xegpu::StoreMatrixOp>(defOp))
-      return dyn_cast_if_present<xegpu::LayoutAttr>(storeOp.getLayoutAttr());
+      return storeOp.getLayoutAttr();
 
     std::string layoutName = getLayoutName(result);
     if (defOp->hasAttr(layoutName))
@@ -164,10 +164,10 @@ xegpu::getDistributeLayoutAttr(const OpOperand &opr) {
   Operation *op = opr.getOwner();
 
   if (auto loadOp = dyn_cast<xegpu::LoadMatrixOp>(op))
-    return dyn_cast_if_present<xegpu::LayoutAttr>(loadOp.getLayoutAttr());
+    return loadOp.getLayoutAttr();
 
   if (auto storeOp = dyn_cast<xegpu::StoreMatrixOp>(op))
-    return dyn_cast_if_present<xegpu::LayoutAttr>(storeOp.getLayoutAttr());
+    return storeOp.getLayoutAttr();
 
   std::string layoutName = xegpu::getLayoutName(opr);
   if (op->hasAttr(layoutName))
@@ -199,6 +199,7 @@ void xegpu::setDistributeLayoutAttrs(
   op->walk([&](Operation *nestOp) {
     if (isa<xegpu::LoadMatrixOp, xegpu::StoreMatrixOp>(nestOp))
       return;
+
     for (OpOperand &opr : nestOp->getOpOperands()) {
       auto layout = getLayoutImpl(opr.get());
       setDistributeLayoutAttr(opr, layout);
@@ -471,5 +472,4 @@ xegpu::addWithRightAligned(OpBuilder &builder, Location loc,
     results.push_back(builder.createOrFold<index::AddOp>(loc, lval, rval));
   }
   return results;
-  return {};
 }

>From a368430b6636f3285062c45b7657d9d9103485f7 Mon Sep 17 00:00:00 2001
From: Chao Chen <chao.chen at intel.com>
Date: Wed, 27 Aug 2025 17:59:56 +0000
Subject: [PATCH 4/4] roll back unnecessary change

---
 mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
index 1ccb4e89fd6a5..5d5ff69e06886 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUBlocking.cpp
@@ -205,20 +205,19 @@ XeGPUBlockingPass::getTileShape(Operation *op) const {
 
 bool XeGPUBlockingPass::needsUnroll(Operation *op) const {
   // skip the op if any of its operands or results has workgroup level layouts
-  bool hasSgLayoutOperands =
+  bool hasWgLayoutOperands =
       llvm::any_of(op->getOpOperands(), [](OpOperand &opr) {
         xegpu::DistributeLayoutAttr layout =
             xegpu::getDistributeLayoutAttr(opr);
         return layout && layout.isForWorkgroup();
       });
-  bool hasSgLayoutResults =
+  bool hasWgLayoutResults =
       llvm::any_of(op->getOpResults(), [](OpResult result) {
         xegpu::DistributeLayoutAttr layout =
             xegpu::getDistributeLayoutAttr(result);
         return layout && layout.isForWorkgroup();
       });
-
-  if (hasSgLayoutOperands || hasSgLayoutResults) {
+  if (hasWgLayoutOperands || hasWgLayoutResults) {
     LDBG() << "skip unrolling for op with workgroup level layout: " << *op;
     return false;
   }



More information about the Mlir-commits mailing list