[Mlir-commits] [mlir] [MLIR][XeGPU] Add uArch limitation to scatter load store (PR #172845)

Artem Kroviakov llvmlistbot at llvm.org
Thu Jan 8 07:19:19 PST 2026


https://github.com/akroviakov updated https://github.com/llvm/llvm-project/pull/172845

>From 2970ef37c677866d693651082fbce8bede731089 Mon Sep 17 00:00:00 2001
From: Artem Kroviakov <artem.kroviakov at intel.com>
Date: Thu, 18 Dec 2025 12:30:40 +0000
Subject: [PATCH 1/2] [MLIR][XeGPU] Add uArch limitation to scatter load store

---
 .../mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h    | 34 +++++++++++++--
 .../mlir/Dialect/XeGPU/uArch/uArchBase.h      |  8 +++-
 .../XeGPU/Transforms/XeGPUPropagateLayout.cpp | 41 +++++++++++++------
 .../XeGPU/propagate-layout-inst-data.mlir     | 22 ++++++++++
 4 files changed, 87 insertions(+), 18 deletions(-)

diff --git a/mlir/include/mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h b/mlir/include/mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h
index b3231a173f33a..055c10ab50652 100644
--- a/mlir/include/mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h
+++ b/mlir/include/mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h
@@ -215,6 +215,26 @@ struct SubgroupMatrixMultiplyAcc : public Instruction,
   const unsigned packedFormatBitSizeB;
 };
 
+struct StoreScatterInstruction : public Instruction {
+  StoreScatterInstruction()
+      : Instruction(InstructionKind::StoreScatter, InstructionScope::Lane) {}
+  static bool classof(const Instruction *B) {
+    return B->getInstructionKind() == InstructionKind::StoreScatter;
+  }
+
+  int32_t getMaxBitSize() const { return 128; }
+};
+
+struct LoadGatherInstruction : public Instruction {
+  LoadGatherInstruction()
+      : Instruction(InstructionKind::LoadGather, InstructionScope::Lane) {}
+  static bool classof(const Instruction *B) {
+    return B->getInstructionKind() == InstructionKind::LoadGather;
+  }
+
+  int32_t getMaxBitSize() const { return 128; }
+};
+
 //===----------------------------------------------------------------------===//
 // uArch instances
 //===----------------------------------------------------------------------===//
@@ -225,8 +245,11 @@ struct PVCuArch final : public Xe2Plus {
     static const Subgroup2DBlockLoadInstruction loadNdInst;
     static const Subgroup2DBlockStoreInstruction storeNdInst;
     static const Subgroup2DBlockPrefetchInstruction prefetchNdInst;
-    static const Instruction *arr[] = {&dpasInst, &loadNdInst, &storeNdInst,
-                                       &prefetchNdInst};
+    static const StoreScatterInstruction storeScatterInst;
+    static const LoadGatherInstruction loadGatherInst;
+    static const Instruction *arr[] = {&dpasInst,         &loadNdInst,
+                                       &storeNdInst,      &prefetchNdInst,
+                                       &storeScatterInst, &loadGatherInst};
     return arr;
   }
 
@@ -248,8 +271,11 @@ struct BMGuArch : public Xe2Plus {
     static const Subgroup2DBlockLoadInstruction loadNdInst;
     static const Subgroup2DBlockStoreInstruction storeNdInst;
     static const Subgroup2DBlockPrefetchInstruction prefetchNdInst;
-    static const Instruction *arr[] = {&dpasInst, &loadNdInst, &storeNdInst,
-                                       &prefetchNdInst};
+    static const StoreScatterInstruction storeScatterInst;
+    static const LoadGatherInstruction loadGatherInst;
+    static const Instruction *arr[] = {&dpasInst,         &loadNdInst,
+                                       &storeNdInst,      &prefetchNdInst,
+                                       &storeScatterInst, &loadGatherInst};
     return arr;
   }
 
diff --git a/mlir/include/mlir/Dialect/XeGPU/uArch/uArchBase.h b/mlir/include/mlir/Dialect/XeGPU/uArch/uArchBase.h
index 8f23b89134773..db1984b2edb1d 100644
--- a/mlir/include/mlir/Dialect/XeGPU/uArch/uArchBase.h
+++ b/mlir/include/mlir/Dialect/XeGPU/uArch/uArchBase.h
@@ -38,7 +38,9 @@ enum class InstructionKind {
                              // matrix multiply-add operation
   Subgroup2DBlockStore,      // Subgroup-level 2D block write instruction
   Subgroup2DBlockLoad,       // Subgroup-level 2D block load instruction
-  Subgroup2DBlockPrefetch    // Subgroup-level 2D block prefetch instruction
+  Subgroup2DBlockPrefetch,   // Subgroup-level 2D block prefetch instruction
+  StoreScatter,              // Lane-level store (scalar, vector)
+  LoadGather                 // Lane-level load (scalar, vector)
   // @TODO: Add more instructions as needed
 };
 
@@ -65,6 +67,10 @@ struct Instruction {
       return "load_nd";
     case InstructionKind::Subgroup2DBlockPrefetch:
       return "prefetch_nd";
+    case InstructionKind::StoreScatter:
+      return "store";
+    case InstructionKind::LoadGather:
+      return "load";
     }
     llvm_unreachable("Unknown InstructionKind");
   }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
index 3ac23f348f8a9..76fa7c0e2698d 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
@@ -986,15 +986,21 @@ void LayoutInfoPropagation::visitLoadGatherOp(
       return;
     }
     auto uArch = getUArch(getChipStr(load).value_or(""));
+    const auto *uArchInstruction =
+        dyn_cast<xegpu::uArch::LoadGatherInstruction>(
+            uArch->getInstruction(xegpu::uArch::InstructionKind::LoadGather));
+    const int maxElemsPerInst =
+        uArchInstruction->getMaxBitSize() /
+        payloadTy.getElementType().getIntOrFloatBitWidth();
+
     const int subgroupSize = uArch->getSubgroupSize();
     SmallVector<int> instData{subgroupSize};
-    if (auto chunkSize = load.getChunkSize().value_or(0); chunkSize > 1)
-      instData.push_back(chunkSize);
-    else if (auto srcTdescTy =
-                 dyn_cast<xegpu::TensorDescType>(load.getSourceType())) {
-      if (srcTdescTy.getChunkSizeAsInt() > 1)
-        instData.push_back(chunkSize);
+    auto chunkSize = load.getChunkSize().value_or(0);
+    if (auto srcTdescTy = dyn_cast<xegpu::TensorDescType>(load.getSourceType());
+        !chunkSize && srcTdescTy) {
+      chunkSize = srcTdescTy.getChunkSizeAsInt();
     }
+    instData.push_back(std::min(static_cast<int>(chunkSize), maxElemsPerInst));
 
     if (layoutKind == LayoutKind::InstData)
       loadLayout =
@@ -1060,15 +1066,24 @@ void LayoutInfoPropagation::visitStoreScatterOp(
     const int subgroupSize = uArch->getSubgroupSize();
 
     if (layoutKind == LayoutKind::InstData) {
+      const auto *uArchInstruction =
+          dyn_cast<xegpu::uArch::LoadGatherInstruction>(
+              uArch->getInstruction(xegpu::uArch::InstructionKind::LoadGather));
+      const int maxElemsPerInst =
+          uArchInstruction->getMaxBitSize() /
+          payloadTy.getElementType().getIntOrFloatBitWidth();
+
+      const int subgroupSize = uArch->getSubgroupSize();
       SmallVector<int> instData{subgroupSize};
-      if (auto chunkSize = storeScatter.getChunkSize().value_or(0);
-          chunkSize > 1)
-        instData.push_back(chunkSize);
-      else if (auto dstTdescTy = dyn_cast<xegpu::TensorDescType>(
-                   storeScatter.getDestType())) {
-        if (dstTdescTy.getChunkSizeAsInt() > 1)
-          instData.push_back(chunkSize);
+      auto chunkSize = storeScatter.getChunkSize().value_or(0);
+      if (auto srcTdescTy =
+              dyn_cast<xegpu::TensorDescType>(storeScatter.getDestType());
+          !chunkSize && srcTdescTy) {
+        chunkSize = srcTdescTy.getChunkSizeAsInt();
       }
+      instData.push_back(
+          std::min(static_cast<int>(chunkSize), maxElemsPerInst));
+
       payloadLayout = LayoutInfo(
           xegpu::LayoutAttr::get(storeScatter.getContext(), instData));
     } else {
diff --git a/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir b/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir
index 5f70831f45e97..f2b270ab16218 100644
--- a/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir
+++ b/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir
@@ -153,3 +153,25 @@ func.func @scatter_ops_chunksize(%src: memref<256xf16>) {
   return
 }
 }
+
+// -----
+
+gpu.module @test {
+// CHECK-LABEL: func.func @scatter_ops_chunksize_excessive(
+// CHECK-SAME: %[[ARG0:[0-9a-zA-Z]+]]: memref<1024xf32>) {
+// CHECK: %{{.*}} = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<true> : vector<16xi1>
+// CHECK: %{{.*}} = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<12> : vector<16xindex>
+// CHECK: %{{.*}} = xegpu.load %[[ARG0]][%{{.*}}], %{{.*}} <{chunk_size = 16 : i64, layout = #xegpu.layout<inst_data = [16, 4]>}>
+// CHECK-SAME: {layout_result_0 = #xegpu.layout<inst_data = [16, 4]>} : memref<1024xf32>, vector<16xindex>, vector<16xi1> -> vector<16x16xf32>
+// CHECK: xegpu.store %0, %[[ARG0]][%{{.*}}], %{{.*}} <{chunk_size = 16 : i64, layout = #xegpu.layout<inst_data = [16, 4]>}> :
+// CHECK-SAME: vector<16x16xf32>, memref<1024xf32>, vector<16xindex>, vector<16xi1>
+func.func @scatter_ops_chunksize_excessive(%src: memref<1024xf32>) {
+  %1 = arith.constant dense<1>: vector<16xi1>
+  %offset = arith.constant dense<12> : vector<16xindex>
+  %3 = xegpu.load %src[%offset], %1 <{chunk_size=16}>
+      : memref<1024xf32>, vector<16xindex>, vector<16xi1> -> vector<16x16xf32>
+  xegpu.store %3, %src[%offset], %1 <{chunk_size=16}>
+      : vector<16x16xf32>, memref<1024xf32>, vector<16xindex>, vector<16xi1>
+  return
+}
+}

>From e80cad1faec06e5dabf39388343248d3047211a0 Mon Sep 17 00:00:00 2001
From: Artem Kroviakov <artem.kroviakov at intel.com>
Date: Fri, 19 Dec 2025 12:30:18 +0000
Subject: [PATCH 2/2] LoadGather to modify its inst_data

---
 .../XeGPU/Transforms/XeGPUPropagateLayout.cpp | 72 +++++++++++++------
 .../XeGPU/propagate-layout-inst-data.mlir     | 26 ++++++-
 2 files changed, 75 insertions(+), 23 deletions(-)

diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
index 76fa7c0e2698d..ef57278662cb0 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
@@ -978,8 +978,19 @@ void LayoutInfoPropagation::visitLoadGatherOp(
     loadLayout = LayoutInfo(anchorLayout);
     maskLayout = loadLayout;
   } else {
+    LayoutInfo valueLayout = results[0]->getValue();
+    // Need the layout of the value to propagate to the tensor descriptor.
+    if (!valueLayout.isAssigned())
+      return;
+
+    auto resAttr = dyn_cast<xegpu::DistributeLayoutAttr>(valueLayout.get());
+    auto instDataIncoming = resAttr.getEffectiveInstDataAsInt();
+    if (auto sliceAttr = dyn_cast<xegpu::SliceAttr>(resAttr))
+      instDataIncoming = SmallVector<int64_t>(
+          cast<xegpu::LayoutAttr>(sliceAttr.flatten().getParent())
+              .getInstData()
+              .asArrayRef());
 
-    // The layout is strictly determined by the payload type.
     VectorType payloadTy = load.getValueType();
     if (!payloadTy) {
       load.emitWarning("Not propagating, non-vector payload supplied.");
@@ -989,30 +1000,48 @@ void LayoutInfoPropagation::visitLoadGatherOp(
     const auto *uArchInstruction =
         dyn_cast<xegpu::uArch::LoadGatherInstruction>(
             uArch->getInstruction(xegpu::uArch::InstructionKind::LoadGather));
-    const int maxElemsPerInst =
-        uArchInstruction->getMaxBitSize() /
-        payloadTy.getElementType().getIntOrFloatBitWidth();
 
     const int subgroupSize = uArch->getSubgroupSize();
-    SmallVector<int> instData{subgroupSize};
-    auto chunkSize = load.getChunkSize().value_or(0);
-    if (auto srcTdescTy = dyn_cast<xegpu::TensorDescType>(load.getSourceType());
-        !chunkSize && srcTdescTy) {
-      chunkSize = srcTdescTy.getChunkSizeAsInt();
-    }
-    instData.push_back(std::min(static_cast<int>(chunkSize), maxElemsPerInst));
-
-    if (layoutKind == LayoutKind::InstData)
-      loadLayout =
-          LayoutInfo(xegpu::LayoutAttr::get(load.getContext(), instData));
-    else
-      loadLayout = getDefaultSIMTLayoutInfo(
-          payloadTy, uArch, uArch->getGeneralPackedFormatBitSize(),
-          /*scattered*/ true);
-
     // Mask operand should have 1D default layout.
     maskLayout = getDefaultSIMTLayoutInfo(load->getContext(), 1, subgroupSize);
 
+    // Check if value inst_data complies with uArch
+    if (!instDataIncoming.empty()) {
+      const int maxElemsPerInst =
+          uArchInstruction->getMaxBitSize() /
+          payloadTy.getElementType().getIntOrFloatBitWidth();
+
+      xegpu::LayoutAttr sourceAttr;
+      // Each lane loads either one element
+      SmallVector<int> instDataUarch(instDataIncoming.size(), 1);
+      // Or multiple elements as 2D with lane's elements in the inner dimension
+      if (payloadTy.getRank() == 1) {
+        instDataUarch.back() = subgroupSize;
+      } else {
+        *std::prev(instDataUarch.end(), 2) = subgroupSize;
+        instDataUarch.back() = (std::min(
+            static_cast<int>(payloadTy.getShape().back()), maxElemsPerInst));
+      }
+      // If inst data does not match, enforce the uArch-based one
+      if (!llvm::equal(instDataIncoming, instDataUarch)) {
+        xegpu::LayoutAttr sourceAttr = dyn_cast<xegpu::LayoutAttr>(resAttr);
+        if (auto sliceAttr = dyn_cast<xegpu::SliceAttr>(resAttr)) {
+          sourceAttr = cast<xegpu::LayoutAttr>(sliceAttr.flatten().getParent());
+        }
+        assert(sourceAttr);
+        xegpu::DistributeLayoutAttr updatedLayoutAttr = xegpu::LayoutAttr::get(
+            load.getContext(), sourceAttr.getSgLayout(), sourceAttr.getSgData(),
+            DenseI32ArrayAttr::get(load.getContext(), instDataUarch),
+            sourceAttr.getLaneLayout(), sourceAttr.getLaneData(),
+            sourceAttr.getOrder());
+
+        if (auto sliceAttr = dyn_cast<xegpu::SliceAttr>(resAttr))
+          updatedLayoutAttr = xegpu::SliceAttr::get(
+              load.getContext(), updatedLayoutAttr, sliceAttr.getDims());
+        valueLayout = LayoutInfo(updatedLayoutAttr);
+      }
+    }
+    loadLayout = valueLayout;
     load.setLayoutAttr(dyn_cast<xegpu::DistributeLayoutAttr>(loadLayout.get()));
   }
   // Propagate the new layout to the tensor descriptor operand.
@@ -1235,7 +1264,8 @@ static LogicalResult updateOp(mlir::OpBuilder &builder, mlir::Operation *op,
     }
     // If the result is a vector type, add a temporary layout attribute to the
     // op.
-    xegpu::setDistributeLayoutAttr(result, layout);
+    if (!isa<xegpu::LoadGatherOp>(op))
+      xegpu::setDistributeLayoutAttr(result, layout);
   }
   return success();
 }
diff --git a/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir b/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir
index f2b270ab16218..1fe32d380eafc 100644
--- a/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir
+++ b/mlir/test/Dialect/XeGPU/propagate-layout-inst-data.mlir
@@ -161,8 +161,8 @@ gpu.module @test {
 // CHECK-SAME: %[[ARG0:[0-9a-zA-Z]+]]: memref<1024xf32>) {
 // CHECK: %{{.*}} = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<true> : vector<16xi1>
 // CHECK: %{{.*}} = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<12> : vector<16xindex>
-// CHECK: %{{.*}} = xegpu.load %[[ARG0]][%{{.*}}], %{{.*}} <{chunk_size = 16 : i64, layout = #xegpu.layout<inst_data = [16, 4]>}>
-// CHECK-SAME: {layout_result_0 = #xegpu.layout<inst_data = [16, 4]>} : memref<1024xf32>, vector<16xindex>, vector<16xi1> -> vector<16x16xf32>
+// CHECK: %{{.*}} = xegpu.load %[[ARG0]][%{{.*}}], %{{.*}} <{chunk_size = 16 : i64, layout = #xegpu.layout<inst_data = [16, 4]>}> :
+// CHECK-SAME: memref<1024xf32>, vector<16xindex>, vector<16xi1> -> vector<16x16xf32>
 // CHECK: xegpu.store %0, %[[ARG0]][%{{.*}}], %{{.*}} <{chunk_size = 16 : i64, layout = #xegpu.layout<inst_data = [16, 4]>}> :
 // CHECK-SAME: vector<16x16xf32>, memref<1024xf32>, vector<16xindex>, vector<16xi1>
 func.func @scatter_ops_chunksize_excessive(%src: memref<1024xf32>) {
@@ -175,3 +175,25 @@ func.func @scatter_ops_chunksize_excessive(%src: memref<1024xf32>) {
   return
 }
 }
+
+// -----
+
+gpu.module @test {
+// CHECK-LABEL: func.func @scatter_ops_chunksize_excessive_slice(
+// CHECK-SAME: %[[ARG0:[0-9a-zA-Z]+]]: memref<1024xf32>) {
+// CHECK: %{{.*}} = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<true> : vector<16xi1>
+// CHECK: %{{.*}} = arith.constant {layout_result_0 = #xegpu.layout<lane_layout = [16], lane_data = [1]>} dense<12> : vector<16xindex>
+// CHECK: %{{.*}} = xegpu.load %[[ARG0]][%{{.*}}], %{{.*}} <{chunk_size = 16 : i64, layout = #xegpu.slice<#xegpu.layout<inst_data = [1, 16, 4]>, dims = [0]>}> :
+// CHECK-SAME: memref<1024xf32>, vector<16xindex>, vector<16xi1> -> vector<16x16xf32>
+// CHECK: xegpu.store %0, %[[ARG0]][%{{.*}}], %{{.*}} <{chunk_size = 16 : i64, layout = #xegpu.slice<#xegpu.layout<inst_data = [1, 16, 16]>, dims = [0]>}> :
+// CHECK-SAME: vector<16x16xf32>, memref<1024xf32>, vector<16xindex>, vector<16xi1>
+func.func @scatter_ops_chunksize_excessive_slice(%src: memref<1024xf32>) {
+  %1 = arith.constant dense<1>: vector<16xi1>
+  %offset = arith.constant dense<12> : vector<16xindex>
+  %3 = xegpu.load %src[%offset], %1 <{chunk_size=16}>
+      : memref<1024xf32>, vector<16xindex>, vector<16xi1> -> vector<16x16xf32>
+  xegpu.store %3, %src[%offset], %1 <{chunk_size=16, layout = #xegpu.slice<#xegpu.layout<inst_data = [1, 16, 16]>, dims = [0]>}>
+      : vector<16x16xf32>, memref<1024xf32>, vector<16xindex>, vector<16xi1>
+  return
+}
+}



More information about the Mlir-commits mailing list