[mlir] Implement Mesh's ShardingInterface for Linalg ops

Allows linalg structured operations to be handled during spmdization and
sharding propagation.

There is only support for projected permutation indexing maps.

Author: sogartar

mlir/include/mlir/Dialect/Linalg/Transforms/AllInterfaces.h

@@ -0,0 +1,26 @@
+//===- AllInterfaces.h - ----------------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines a common entry point for registering all external
+// interface implementations to the linalg dialect.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_DIALECT_LINALG_TRANSFORMS_ALLINTERFACES_H
+#define MLIR_DIALECT_LINALG_TRANSFORMS_ALLINTERFACES_H
+
+namespace mlir {
+class DialectRegistry;
+
+namespace linalg {
+void registerAllDialectInterfaceImplementations(DialectRegistry &registry);
+} // namespace linalg
+
+} // namespace mlir
+
+#endif // MLIR_DIALECT_LINALG_TRANSFORMS_ALLINTERFACES_H

mlir/include/mlir/Dialect/Linalg/Transforms/MeshShardingInterfaceImpl.h

@@ -0,0 +1,20 @@
+//===- MeshShardingInterfaceImpl.h ----------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_DIALECT_LINALG_MESHSHARDINGINTERFACEIMPL_H
+#define MLIR_DIALECT_LINALG_MESHSHARDINGINTERFACEIMPL_H
+
+namespace mlir {
+class DialectRegistry;
+
+namespace linalg {
+void registerMeshShardingInterfaceExternalModels(DialectRegistry &registry);
+} // namespace linalg
+} // namespace mlir
+
+#endif // MLIR_DIALECT_LINALG_MESHSHARDINGINTERFACEIMPL_H

mlir/include/mlir/Dialect/Mesh/IR/MeshBase.td

@@ -46,6 +46,12 @@ def Mesh_ReductionKind : I32EnumAttr<"ReductionKind",
   I32EnumAttrCase<"Sum", 1, "sum">,
   I32EnumAttrCase<"Max", 2, "max">,
   I32EnumAttrCase<"Min", 3, "min">,
+  I32EnumAttrCase<"Product", 4, "product">,
+  // Arithmetic mean.
+  I32EnumAttrCase<"Average", 5, "average">,
+  I32EnumAttrCase<"BitwiseAnd", 6, "bitwise_and">,
+  I32EnumAttrCase<"BitwiseOr", 7, "bitwise_or">,
+  I32EnumAttrCase<"BitwiseXor", 8, "bitwise_xor">,
   I32EnumAttrCase<"Generic", 100, "generic">
 ]> {
   let genSpecializedAttr = 0;

mlir/include/mlir/Dialect/Mesh/IR/MeshOps.td

@@ -353,6 +353,10 @@ def Mesh_AllReduceOp : Mesh_CollectiveCommunicationOpBase<"all_reduce", [
     attr-dict `:` type($input) `->` type($result)
   }];
   let hasCanonicalizer = 1;
+    let builders = [
+    OpBuilder<(ins "Value":$input, "StringRef":$mesh,
+      "ArrayRef<MeshAxis>":$meshAxes, "ReductionKind":$reduction)>
+  ];
 }
 
 def Mesh_AllSliceOp : Mesh_CollectiveCommunicationOpBase<"all_slice", [

mlir/include/mlir/Dialect/Mesh/Interfaces/ShardingInterfaceImpl.h

@@ -22,6 +22,24 @@ class SymbolTableCollection;
 
 namespace mesh {
 
+// Retrieve the mesh axes corresponding to each operation loop iterator based
+// on the provided shardings for the op's operands and results.
+// Assumes that the indexingMaps are projected permutations.
+ShardingArray getMeshAxisAssignmentForLoopIterators(
+    ArrayRef<MeshShardingAttr> operandShardings,
+    ArrayRef<MeshShardingAttr> resultShardings,
+    ArrayRef<utils::IteratorType> loopIteratorTypes,
+    ArrayRef<AffineMap> indexingMaps);
+
+bool isAtLeastOneReductionIteratorSharded(
+    ArrayRef<utils::IteratorType> loopIteratorTypes,
+    ArrayRef<SmallVector<MeshAxis>> meshAxisAssignmentForLoopIterators);
+
+// Get the set of mesh axes that correspond to reduction loop iterators.
+SmallVector<MeshAxis> getReductionMeshAxes(
+    ArrayRef<utils::IteratorType> loopIteratorTypes,
+    ArrayRef<SmallVector<MeshAxis>> meshAxisAssignmentForLoopIterators);
+
 // Inserts a clone of the operation that has all ranked tensor
 // arguments/results sharded.
 void spmdizeTriviallyShardableOperation(

mlir/include/mlir/Dialect/Mesh/Transforms/Transforms.h

@@ -13,6 +13,7 @@
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Value.h"
 #include "mlir/Support/LLVM.h"
+#include "llvm/ADT/ArrayRef.h"
 
 namespace mlir {
 class RewritePatternSet;
@@ -37,6 +38,11 @@ TypedValue<IndexType>
 createCollectiveProcessGroupSize(MeshOp mesh, ArrayRef<MeshAxis> axes,
                                  ImplicitLocOpBuilder &builder);
 
+// Get process linear index along the given mesh axes.
+TypedValue<IndexType> createProcessLinearIndex(StringRef mesh,
+                                               ArrayRef<MeshAxis> meshAxes,
+                                               ImplicitLocOpBuilder &builder);
+
 } // namespace mesh
 } // namespace mlir
 

mlir/include/mlir/InitAllDialects.h

@@ -43,10 +43,7 @@
 #include "mlir/Dialect/LLVMIR/NVVMDialect.h"
 #include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
 #include "mlir/Dialect/Linalg/IR/Linalg.h"
-#include "mlir/Dialect/Linalg/IR/ValueBoundsOpInterfaceImpl.h"
-#include "mlir/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.h"
-#include "mlir/Dialect/Linalg/Transforms/SubsetInsertionOpInterfaceImpl.h"
-#include "mlir/Dialect/Linalg/Transforms/TilingInterfaceImpl.h"
+#include "mlir/Dialect/Linalg/Transforms/AllInterfaces.h"
 #include "mlir/Dialect/MLProgram/IR/MLProgram.h"
 #include "mlir/Dialect/MLProgram/Transforms/BufferizableOpInterfaceImpl.h"
 #include "mlir/Dialect/MPI/IR/MPI.h"
@@ -155,10 +152,7 @@ inline void registerAllDialects(DialectRegistry &registry) {
   cf::registerBufferizableOpInterfaceExternalModels(registry);
   cf::registerBufferDeallocationOpInterfaceExternalModels(registry);
   gpu::registerBufferDeallocationOpInterfaceExternalModels(registry);
-  linalg::registerBufferizableOpInterfaceExternalModels(registry);
-  linalg::registerSubsetOpInterfaceExternalModels(registry);
-  linalg::registerTilingInterfaceExternalModels(registry);
-  linalg::registerValueBoundsOpInterfaceExternalModels(registry);
+  linalg::registerAllDialectInterfaceImplementations(registry);
   memref::registerAllocationOpInterfaceExternalModels(registry);
   memref::registerRuntimeVerifiableOpInterfaceExternalModels(registry);
   memref::registerValueBoundsOpInterfaceExternalModels(registry);

mlir/lib/Dialect/Linalg/Transforms/AllInterfaces.cpp

@@ -0,0 +1,24 @@
+//===- AllInterfaces.cpp - --------------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Linalg/Transforms/AllInterfaces.h"
+
+#include "mlir/Dialect/Linalg/IR/ValueBoundsOpInterfaceImpl.h"
+#include "mlir/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.h"
+#include "mlir/Dialect/Linalg/Transforms/MeshShardingInterfaceImpl.h"
+#include "mlir/Dialect/Linalg/Transforms/SubsetInsertionOpInterfaceImpl.h"
+#include "mlir/Dialect/Linalg/Transforms/TilingInterfaceImpl.h"
+
+void mlir::linalg::registerAllDialectInterfaceImplementations(
+    DialectRegistry &registry) {
+  registerBufferizableOpInterfaceExternalModels(registry);
+  registerMeshShardingInterfaceExternalModels(registry);
+  registerSubsetOpInterfaceExternalModels(registry);
+  registerTilingInterfaceExternalModels(registry);
+  registerValueBoundsOpInterfaceExternalModels(registry);
+}

mlir/lib/Dialect/Linalg/Transforms/CMakeLists.txt

@@ -1,4 +1,5 @@
 add_mlir_dialect_library(MLIRLinalgTransforms
+  AllInterfaces.cpp
   BubbleUpExtractSlice.cpp
   BufferizableOpInterfaceImpl.cpp
   Bufferize.cpp
@@ -21,6 +22,7 @@ add_mlir_dialect_library(MLIRLinalgTransforms
   InlineScalarOperands.cpp
   Interchange.cpp
   Loops.cpp
+  MeshShardingInterfaceImpl.cpp
   NamedOpConversions.cpp
   Padding.cpp
   Promotion.cpp
@@ -61,12 +63,15 @@ add_mlir_dialect_library(MLIRLinalgTransforms
   MLIRIR
   MLIRMemRefDialect
   MLIRMemRefTransforms
+  MLIRMeshDialect
+  MLIRMeshTransforms
   MLIRLinalgDialect
   MLIRLinalgUtils
   MLIRSCFDialect
   MLIRSCFTransforms
   MLIRSCFUtils
   MLIRPass
+  MLIRShardingInterface
   MLIRSubsetOpInterface
   MLIRSparseTensorDialect
   MLIRTensorDialect