[Mlir-commits] [mlir] [MLIR][XeGPU] Lowering 2-Dimensional Reductions of N-D Tensors into Chained 1-D Reductions (PR #186034)

Jianhui Li llvmlistbot at llvm.org
Wed Mar 11 21:23:08 PDT 2026


https://github.com/Jianhui-Li created https://github.com/llvm/llvm-project/pull/186034

This PR relaxes the 2d reduction lowering in the peephole optimization pass to allow source tensor to have n-d shape. 
It also fixes a minor bug of accumulator lowering in the current implementation. 

>From d6db28591c8514a03fc85accbf5f81829366b664 Mon Sep 17 00:00:00 2001
From: Jianhui Li <jian.hui.li at intel.com>
Date: Wed, 11 Mar 2026 05:33:32 +0000
Subject: [PATCH 1/3] passing 2d reduction test with layouts

---
 .../XeGPU/Transforms/XeGPULayoutImpl.cpp      |  6 ++-
 .../Transforms/XeGPUPeepHoleOptimizer.cpp     | 46 ++++++++-----------
 .../XeGPU/Transforms/XeGPUPropagateLayout.cpp |  2 -
 .../test/Dialect/XeGPU/peephole-optimize.mlir | 20 +++++---
 4 files changed, 35 insertions(+), 39 deletions(-)

diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp
index feefeb727a732..bbe2581aea99e 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp
@@ -489,13 +489,15 @@ xegpu::SliceAttr xegpu::setupMultiReductionResultLayout(
     SmallVector<int64_t> instData(srcRank, 1);
     instData[srcRank - 2] =
         std::min(maxReduceVectorSize, srcShape[srcRank - 2]);
-    instData[srcRank - 1] = subgroupSize;
+    instData[srcRank - 1] =
+        std::min(static_cast<int64_t>(subgroupSize), srcShape[srcRank - 1]);
     srcLayout = xegpu::LayoutAttr::get(context, toInt32Attr(instData));
 
   } else if (layoutKind == xegpu::LayoutKind::Lane) {
 
     SmallVector<int64_t> laneLayout(srcRank, 1), laneData(srcRank, 1);
-    laneLayout[srcRank - 1] = subgroupSize;
+    laneLayout[srcRank - 1] =
+        std::min(static_cast<int64_t>(subgroupSize), srcShape[srcRank - 1]);
     laneData[srcRank - 2] =
         std::min(maxReduceVectorSize, srcShape[srcRank - 2]);
     srcLayout = xegpu::LayoutAttr::get(context, toInt32Attr(laneLayout),
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPeepHoleOptimizer.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPeepHoleOptimizer.cpp
index d7a9b7ba377f9..9b55ae1aa319b 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPeepHoleOptimizer.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPeepHoleOptimizer.cpp
@@ -428,10 +428,6 @@ class MultiRed2dOpPattern
   matchAndRewrite(vector::MultiDimReductionOp reductionOp, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
     auto sourceVecType = reductionOp.getSourceVectorType();
-    if (reductionOp.getReductionDims().size() != 2 ||
-        sourceVecType.getRank() != 2)
-      return rewriter.notifyMatchFailure(
-          reductionOp, "Expected 2D multi reduction of a 2D source");
     auto resLayout = xegpu::getDistributeLayoutAttr(reductionOp.getResult());
     // Retrieve and order dims for 1D decomposition (prefer intra-lane first).
     auto dims = llvm::to_vector(reductionOp.getReductionDims());
@@ -444,33 +440,27 @@ class MultiRed2dOpPattern
     auto loc = reductionOp.getLoc();
     auto acc = reductionOp.getAcc();
 
-    // The first reduction's dist attribute does not have the cross lane dim.
-    auto resSliceLayoutAttr = cast<xegpu::SliceAttr>(resLayout);
-    SmallVector<int64_t> dropDims{crossLaneDim};
-    auto intraLaneRedResLayout = resSliceLayoutAttr.dropSliceDims(dropDims);
-
     SmallVector<int64_t> accShape(sourceVecType.getShape());
     accShape.erase(accShape.begin() + intraLaneDim);
-    if (acc) {
-      acc = vector::BroadcastOp::create(
-          rewriter, loc,
-          VectorType::get(accShape, sourceVecType.getElementType()), acc);
-      xegpu::setDistributeLayoutAttr(
-          llvm::dyn_cast<OpResult>(acc),
-          cast<xegpu::DistributeLayoutAttr>(intraLaneRedResLayout));
-    }
+    Type eTy = sourceVecType.getElementType();
+    Attribute eVal;
+    if (eTy.isFloat())
+      eVal = FloatAttr::get(eTy, 0.0);
+    else
+      eVal = IntegerAttr::get(eTy, 0);
+    Value const_zero = arith::ConstantOp::create(
+        rewriter, loc,
+        DenseElementsAttr::get(VectorType::get(accShape, eTy), eVal));
     Value intraLaneReduced = vector::MultiDimReductionOp::create(
-        rewriter, loc, reductionOp.getKind(), reductionOp.getSource(), acc,
-        ArrayRef<int64_t>(intraLaneDim));
-    xegpu::setDistributeLayoutAttr(
-        llvm::dyn_cast<OpResult>(intraLaneReduced),
-        cast<xegpu::DistributeLayoutAttr>(intraLaneRedResLayout));
-
-    Value crossLaneReduced = vector::ReductionOp::create(
-        rewriter, loc, reductionOp.getKind(), intraLaneReduced, nullptr);
-    xegpu::setDistributeLayoutAttr(
-        llvm::dyn_cast<OpResult>(crossLaneReduced),
-        cast<xegpu::DistributeLayoutAttr>(resLayout));
+        rewriter, loc, reductionOp.getKind(), reductionOp.getSource(),
+        const_zero, ArrayRef<int64_t>(intraLaneDim));
+
+    // Adjust crossLaneDim after the first reduction.
+    if (crossLaneDim > intraLaneDim)
+      crossLaneDim -= 1;
+    Value crossLaneReduced = vector::MultiDimReductionOp::create(
+        rewriter, loc, reductionOp.getKind(), intraLaneReduced, acc,
+        ArrayRef<int64_t>(crossLaneDim));
     assert(crossLaneReduced.getType() == reductionOp.getResult().getType() &&
            "Type mismatch");
     rewriter.replaceOp(reductionOp, crossLaneReduced);
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
index b3fb00f35b167..62ba6323134a9 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
@@ -1135,7 +1135,6 @@ void LayoutInfoPropagation::visitLoadMatrixOp(
   if (!hasParamsOfLayoutKind(anchorLayout)) {
     VectorType resVecTy =
         llvm::cast<VectorType>(loadMatrixOp.getRes().getType());
-    assert(resVecTy.getRank() == 2 && "Expecting 2D vector for store matrix.");
     const uArch *uArch = getUArch(getChipStr(loadMatrixOp).value_or(""));
     if (!uArch)
       return;
@@ -1156,7 +1155,6 @@ void LayoutInfoPropagation::visitStoreMatrixOp(
   } else {
     VectorType srcVecTy =
         llvm::cast<VectorType>(storeMatrix.getData().getType());
-    assert(srcVecTy.getRank() == 2 && "Expecting 2D vector for store matrix.");
     const uArch *uArch = getUArch(getChipStr(storeMatrix).value_or(""));
     if (!uArch)
       return;
diff --git a/mlir/test/Dialect/XeGPU/peephole-optimize.mlir b/mlir/test/Dialect/XeGPU/peephole-optimize.mlir
index 83fec045b9973..04f4bf9430801 100644
--- a/mlir/test/Dialect/XeGPU/peephole-optimize.mlir
+++ b/mlir/test/Dialect/XeGPU/peephole-optimize.mlir
@@ -293,14 +293,20 @@ gpu.func @array_length(%arg0: vector<8x16xf16>, %arg1: memref<256x256xf16>, %arg
 
 // -----
 // CHECK-LABEL: gpu.func @vector_reduce_2d(
-// CHECK-SAME: %[[ARG0:[0-9a-zA-Z]+]]: memref<4x16xf32>, %[[ARG2:[0-9a-zA-Z]+]]: memref<256xf32>) {
-// CHECK:      %[[ACC_VEC:.*]] = arith.constant dense<1.000000e+00> : vector<16xf32>
+// CHECK-SAME: %[[ARG0:[0-9a-zA-Z]+]]: memref<4x16xf32>, %[[ARG1:[0-9a-zA-Z]+]]: memref<256xf32>) {
+// CHECK:      %[[MASK:.*]] = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<true> : vector<16xi1>
+// CHECK:      %[[OFFSET:.*]] = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<0> : vector<16xindex>
+// CHECK:      %[[ACC_VEC:.*]] = arith.constant dense<0.000000e+00> : vector<16xf32>
+// CHECK:      %[[ACC_SCALAR:.*]] = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>, dims = [0, 1]>} 1.000000e+00 : f32
 // CHECK:      %[[TDESC:.*]] = xegpu.create_nd_tdesc %[[ARG0]] : memref<4x16xf32> -> !xegpu.tensor_desc<4x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
-// CHECK:      %[[LOADED:.*]] = xegpu.load_nd %[[TDESC]][0, 0]  : !xegpu.tensor_desc<4x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>> -> vector<4x16xf32>
-// CHECK:      %[[LOADED_REDUCED:.*]] = vector.multi_reduction <add>, %[[LOADED]], %[[ACC_VEC]]
-// CHECK-SAME: {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>, dims = [0]>} [0] : vector<4x16xf32> to vector<16xf32>
-// CHECK:      %[[LOADED_REDUCED_FOR_CROSS:.*]] = vector.reduction <add>, %[[LOADED_REDUCED]]
-// CHECK-SAME: {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>, dims = [0, 1]>} : vector<16xf32> into f32
+// CHECK:      %[[LOADED:.*]] = xegpu.load_nd %[[TDESC]][0, 0] : !xegpu.tensor_desc<4x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>> -> vector<4x16xf32>
+// CHECK:      %[[REDUCE_1:.*]] = vector.multi_reduction <add>, %[[LOADED]], %[[ACC_VEC]] [0] : vector<4x16xf32> to vector<16xf32>
+// CHECK:      %[[REDUCE_2:.*]] = vector.multi_reduction <add>, %[[REDUCE_1]], %[[ACC_SCALAR]] [0] : vector<16xf32> to f32
+// CHECK:      %[[BCAST:.*]] = vector.broadcast %[[REDUCE_2]]
+// CHECK-SAME: {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>, dims = [0]>} : f32 to vector<16xf32>
+// CHECK:      xegpu.store %[[BCAST]], %[[ARG1]][%[[OFFSET]]], %[[MASK]]
+// CHECK-SAME: <{layout = #xegpu.slice<#xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>, dims = [0]>}>
+// CHECK-SAME: : vector<16xf32>, memref<256xf32>, vector<16xindex>, vector<16xi1>
 gpu.module @xevm_test {
   gpu.func @vector_reduce_2d(%src: memref<4x16xf32>, %dst: memref<256xf32>) {
     %cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>, dims = [0, 1]>} 1.0 : f32

>From 16ffe18058529689baa8adefd4df564138f4678e Mon Sep 17 00:00:00 2001
From: Jianhui Li <jian.hui.li at intel.com>
Date: Thu, 12 Mar 2026 03:51:43 +0000
Subject: [PATCH 2/3] clean up

---
 mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp      | 6 ++----
 mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp | 2 ++
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp
index bbe2581aea99e..feefeb727a732 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp
@@ -489,15 +489,13 @@ xegpu::SliceAttr xegpu::setupMultiReductionResultLayout(
     SmallVector<int64_t> instData(srcRank, 1);
     instData[srcRank - 2] =
         std::min(maxReduceVectorSize, srcShape[srcRank - 2]);
-    instData[srcRank - 1] =
-        std::min(static_cast<int64_t>(subgroupSize), srcShape[srcRank - 1]);
+    instData[srcRank - 1] = subgroupSize;
     srcLayout = xegpu::LayoutAttr::get(context, toInt32Attr(instData));
 
   } else if (layoutKind == xegpu::LayoutKind::Lane) {
 
     SmallVector<int64_t> laneLayout(srcRank, 1), laneData(srcRank, 1);
-    laneLayout[srcRank - 1] =
-        std::min(static_cast<int64_t>(subgroupSize), srcShape[srcRank - 1]);
+    laneLayout[srcRank - 1] = subgroupSize;
     laneData[srcRank - 2] =
         std::min(maxReduceVectorSize, srcShape[srcRank - 2]);
     srcLayout = xegpu::LayoutAttr::get(context, toInt32Attr(laneLayout),
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
index 62ba6323134a9..b3fb00f35b167 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
@@ -1135,6 +1135,7 @@ void LayoutInfoPropagation::visitLoadMatrixOp(
   if (!hasParamsOfLayoutKind(anchorLayout)) {
     VectorType resVecTy =
         llvm::cast<VectorType>(loadMatrixOp.getRes().getType());
+    assert(resVecTy.getRank() == 2 && "Expecting 2D vector for store matrix.");
     const uArch *uArch = getUArch(getChipStr(loadMatrixOp).value_or(""));
     if (!uArch)
       return;
@@ -1155,6 +1156,7 @@ void LayoutInfoPropagation::visitStoreMatrixOp(
   } else {
     VectorType srcVecTy =
         llvm::cast<VectorType>(storeMatrix.getData().getType());
+    assert(srcVecTy.getRank() == 2 && "Expecting 2D vector for store matrix.");
     const uArch *uArch = getUArch(getChipStr(storeMatrix).value_or(""));
     if (!uArch)
       return;

>From b093272e0483e96dcef6ecccf80d73cf0d78b150 Mon Sep 17 00:00:00 2001
From: Jianhui Li <jian.hui.li at intel.com>
Date: Thu, 12 Mar 2026 04:16:57 +0000
Subject: [PATCH 3/3] adding test

---
 .../test/Dialect/XeGPU/peephole-optimize.mlir | 41 +++++++++++++++++++
 1 file changed, 41 insertions(+)

diff --git a/mlir/test/Dialect/XeGPU/peephole-optimize.mlir b/mlir/test/Dialect/XeGPU/peephole-optimize.mlir
index 04f4bf9430801..06008ccafbccd 100644
--- a/mlir/test/Dialect/XeGPU/peephole-optimize.mlir
+++ b/mlir/test/Dialect/XeGPU/peephole-optimize.mlir
@@ -329,3 +329,44 @@ gpu.module @xevm_test {
     gpu.return
   }
 }
+
+// -----
+// CHECK-LABEL: gpu.func @vector_reduce_2d_with_leading_unit_dims(
+// CHECK-SAME: %[[ARG0:[0-9a-zA-Z]+]]: memref<4x16xf32>, %[[ARG1:[0-9a-zA-Z]+]]: memref<256xf32>) {
+// CHECK:      %[[MASK:.*]] = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<true> : vector<16xi1>
+// CHECK:      %[[OFFSET:.*]] = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<0> : vector<16xindex>
+// CHECK:      %[[ACC_2D:.*]] = arith.constant dense<0.000000e+00> : vector<1x16xf32>
+// CHECK:      %[[ACC_1D:.*]] = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 1, 16], lane_data = [1, 1, 1]>, dims = [1, 2]>} dense<1.000000e+00> : vector<1xf32>
+// CHECK:      %[[TDESC:.*]] = xegpu.create_nd_tdesc %[[ARG0]] : memref<4x16xf32> -> !xegpu.tensor_desc<4x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
+// CHECK:      %[[LOADED:.*]] = xegpu.load_nd %[[TDESC]][0, 0] : !xegpu.tensor_desc<4x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>> -> vector<4x16xf32>
+// CHECK:      %[[SHAPED:.*]] = vector.shape_cast %[[LOADED]] : vector<4x16xf32> to vector<1x4x16xf32>
+// CHECK:      %[[REDUCE_1:.*]] = vector.multi_reduction <add>, %[[SHAPED]], %[[ACC_2D]] [1] : vector<1x4x16xf32> to vector<1x16xf32>
+// CHECK:      %[[REDUCE_2:.*]] = vector.multi_reduction <add>, %[[REDUCE_1]], %[[ACC_1D]] [1] : vector<1x16xf32> to vector<1xf32>
+// CHECK:      %[[BCAST:.*]] = vector.broadcast %[[REDUCE_2]]
+// CHECK-SAME: {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} : vector<1xf32> to vector<16xf32>
+// CHECK:      xegpu.store %[[BCAST]], %[[ARG1]][%[[OFFSET]]], %[[MASK]]
+// CHECK-SAME: <{layout = #xegpu.layout<lane_layout = [16], lane_data = [1]>}>
+// CHECK-SAME: : vector<16xf32>, memref<256xf32>, vector<16xindex>, vector<16xi1>
+gpu.module @xevm_test {
+  gpu.func @vector_reduce_2d_with_leading_unit_dims(%src: memref<4x16xf32>, %dst: memref<256xf32>) {
+    %cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 1, 16], lane_data = [1, 1, 1]>, dims = [1, 2]>} dense<1.000000e+00> : vector<1xf32>
+    %tdesc = xegpu.create_nd_tdesc %src : memref<4x16xf32>
+      -> !xegpu.tensor_desc<4x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
+    %load =  xegpu.load_nd %tdesc[0, 0]
+      : !xegpu.tensor_desc<4x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
+      -> vector<4x16xf32>
+    %load1 = vector.broadcast %load {layout_result_0 = #xegpu.layout<lane_layout = [1, 1, 16], lane_data = [1, 1, 1]>}: vector<4x16xf32> to vector<1x4x16xf32>
+    %reduce = vector.multi_reduction <add>, %load1, %cst
+     {layout_result_0 = #xegpu.slice<#xegpu.layout<lane_layout = [1, 1, 16], lane_data = [1, 1, 1]>, dims = [1, 2]>}
+     [1, 2] : vector<1x4x16xf32> to vector<1xf32>
+    %reduce_bcast = vector.broadcast %reduce
+     {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>}
+     : vector<1xf32> to vector<16xf32>
+
+    %offset = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<0> : vector<16xindex>
+    %mask = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<1> : vector<16xi1>
+
+    xegpu.store %reduce_bcast, %dst[%offset], %mask {layout = #xegpu.layout<lane_layout = [16], lane_data = [1]>} : vector<16xf32>, memref<256xf32>, vector<16xindex>, vector<16xi1>
+    gpu.return
+  }
+}



More information about the Mlir-commits mailing list