[mlir][mesh, MPI] Mesh2mpi

mlir/include/mlir/Conversion/MeshToMPI/MeshToMPI.h

@@ -0,0 +1,27 @@
+//===- MeshToMPI.h - Convert Mesh to MPI dialect ----------------*- 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
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_CONVERSION_MESHTOMPI_MESHTOMPI_H
+#define MLIR_CONVERSION_MESHTOMPI_MESHTOMPI_H
+
+#include "mlir/Pass/Pass.h"
+#include "mlir/Support/LLVM.h"
+
+namespace mlir {
+class Pass;
+
+#define GEN_PASS_DECL_CONVERTMESHTOMPIPASS
+#include "mlir/Conversion/Passes.h.inc"
+
+/// Lowers Mesh communication operations (updateHalo, AllGater, ...)
+/// to MPI primitives.
+std::unique_ptr<::mlir::Pass> createConvertMeshToMPIPass();
+
+} // namespace mlir
+
+#endif // MLIR_CONVERSION_MESHTOMPI_MESHTOMPI_H

mlir/include/mlir/Conversion/Passes.h

@@ -51,6 +51,7 @@
 #include "mlir/Conversion/MemRefToEmitC/MemRefToEmitCPass.h"
 #include "mlir/Conversion/MemRefToLLVM/MemRefToLLVM.h"
 #include "mlir/Conversion/MemRefToSPIRV/MemRefToSPIRVPass.h"
+#include "mlir/Conversion/MeshToMPI/MeshToMPI.h"
 #include "mlir/Conversion/NVGPUToNVVM/NVGPUToNVVM.h"
 #include "mlir/Conversion/NVVMToLLVM/NVVMToLLVM.h"
 #include "mlir/Conversion/OpenACCToSCF/ConvertOpenACCToSCF.h"

mlir/include/mlir/Conversion/Passes.td

@@ -878,6 +878,29 @@ def ConvertMemRefToSPIRV : Pass<"convert-memref-to-spirv"> {
   ];
 }
 
+//===----------------------------------------------------------------------===//
+// MeshToMPI
+//===----------------------------------------------------------------------===//
+
+def ConvertMeshToMPIPass : Pass<"convert-mesh-to-mpi"> {
+  let summary = "Convert Mesh dialect to MPI dialect.";
+  let constructor = "mlir::createConvertMeshToMPIPass()";
+  let description = [{
+    This pass converts communication operations from the Mesh dialect to the
+    MPI dialect.
+    If it finds a global named "static_mpi_rank" it will use that splat value
+    instead of calling MPI_Comm_rank. This allows optimizations like constant
+    shape propagation and fusion because shard/partition sizes depend on the
+    rank.
+  }];
+  let dependentDialects = [
+    "memref::MemRefDialect",
+    "mpi::MPIDialect",
+    "scf::SCFDialect",
+    "bufferization::BufferizationDialect"
+  ];
+}
+
 //===----------------------------------------------------------------------===//
 // NVVMToLLVM
 //===----------------------------------------------------------------------===//

mlir/include/mlir/Dialect/MPI/IR/MPIOps.td

@@ -84,6 +84,7 @@ def MPI_SendOp : MPI_Op<"send", []> {
   let assemblyFormat = "`(` $ref `,` $tag `,` $rank `)` attr-dict `:` "
                        "type($ref) `,` type($tag) `,` type($rank)"
                        "(`->` type($retval)^)?";
+  let hasCanonicalizer = 1;
 }
 
 //===----------------------------------------------------------------------===//
@@ -114,6 +115,7 @@ def MPI_RecvOp : MPI_Op<"recv", []> {
   let assemblyFormat = "`(` $ref `,` $tag `,` $rank `)` attr-dict `:` "
                        "type($ref) `,` type($tag) `,` type($rank)"
                        "(`->` type($retval)^)?";
+  let hasCanonicalizer = 1;
 }
 
 

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

@@ -156,6 +156,40 @@ def Mesh_ProcessLinearIndexOp : Mesh_Op<"process_linear_index", [
   ];
 }
 
+def Mesh_NeighborsLinearIndicesOp : Mesh_Op<"neighbors_linear_indices", [
+  Pure,
+  DeclareOpInterfaceMethods<SymbolUserOpInterface>,
+  DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmResultNames"]>
+]> {
+  let summary =
+      "For given mesh index get the linear indices of the direct neighbor processes along the given split.";
+  let description = [{
+    Example:
+    ```
+    mesh.mesh @mesh0(shape = 10x20x30)
+    %c1 = arith.constant 1 : index
+    %c2 = arith.constant 2 : index
+    %c3 = arith.constant 3 : index
+    %idx = mesh.neighbors_linear_indices on @mesh[%c1, %c2, %c3] split_axes = [1] : index
+    ```
+    The above returns two indices, `633` and `693`, which correspond to the
+    index of the previous process `(1, 1, 3)`, and the next process 
+    `(1, 3, 3) along the split axis `1`.
+
+    A negative value is returned if there is no neighbor in the respective
+    direction along the given `split_axes`.
+  }];
+  let arguments = (ins FlatSymbolRefAttr:$mesh,
+                       Variadic<Index>:$device,
+                       Mesh_MeshAxesAttr:$split_axes);
+  let results = (outs Index:$neighbor_down, Index:$neighbor_up);
+  let assemblyFormat =  [{
+      `on` $mesh `[` $device `]`
+      `split_axes` `=` $split_axes
+      attr-dict `:` type(results)
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // Sharding operations.
 //===----------------------------------------------------------------------===//
@@ -1058,12 +1092,12 @@ def Mesh_ShiftOp : Mesh_CollectiveCommunicationOpBase<"shift", [
 }
 
 def Mesh_UpdateHaloOp : Mesh_Op<"update_halo", [
+  Pure,
   DestinationStyleOpInterface,
   TypesMatchWith<
     "result has same type as destination",
     "result", "destination", "$_self">,
-  DeclareOpInterfaceMethods<SymbolUserOpInterface>,
-  AttrSizedOperandSegments
+  DeclareOpInterfaceMethods<SymbolUserOpInterface>
 ]> {
   let summary = "Update halo data.";
   let description = [{
@@ -1072,7 +1106,7 @@ def Mesh_UpdateHaloOp : Mesh_Op<"update_halo", [
     on the remote devices. Changes might be caused by mutating operations
     and/or if the new halo regions are larger than the existing ones.
 
-    Source and destination might have different halo sizes.
+    Destination is supposed to be initialized with the local data (not halos).
 
     Assumes all devices hold tensors with same-sized halo data as specified
     by `source_halo_sizes/static_source_halo_sizes` and
@@ -1084,25 +1118,21 @@ def Mesh_UpdateHaloOp : Mesh_Op<"update_halo", [
 
   }];
   let arguments = (ins
-    AnyTypeOf<[AnyNon0RankedMemRef, AnyNon0RankedTensor]>:$source,
     AnyTypeOf<[AnyNon0RankedMemRef, AnyNon0RankedTensor]>:$destination,
     FlatSymbolRefAttr:$mesh,
     Mesh_MeshAxesArrayAttr:$split_axes,
-    Variadic<I64>:$source_halo_sizes,
-    DefaultValuedAttr<DenseI64ArrayAttr, "{}">:$static_source_halo_sizes,
-    Variadic<I64>:$destination_halo_sizes,
-    DefaultValuedAttr<DenseI64ArrayAttr, "{}">:$static_destination_halo_sizes
+    Variadic<I64>:$halo_sizes,
+    DefaultValuedAttr<DenseI64ArrayAttr, "{}">:$static_halo_sizes
   );
   let results = (outs
     AnyTypeOf<[AnyNon0RankedMemRef, AnyNon0RankedTensor]>:$result
   );
   let assemblyFormat = [{
-    $source `into` $destination
+    $destination
     `on` $mesh
     `split_axes` `=` $split_axes
-    (`source_halo_sizes` `=` custom<DynamicIndexList>($source_halo_sizes, $static_source_halo_sizes)^)?
-    (`destination_halo_sizes` `=` custom<DynamicIndexList>($destination_halo_sizes, $static_destination_halo_sizes)^)?
-    attr-dict `:` type($source) `->` type($result)
+    (`halo_sizes` `=` custom<DynamicIndexList>($halo_sizes, $static_halo_sizes)^)?
+    attr-dict `:` type($result)
   }];
   let extraClassDeclaration = [{
     MutableOperandRange getDpsInitsMutable() { return getDestinationMutable(); }

mlir/lib/Conversion/CMakeLists.txt

@@ -41,6 +41,7 @@ add_subdirectory(MathToSPIRV)
 add_subdirectory(MemRefToEmitC)
 add_subdirectory(MemRefToLLVM)
 add_subdirectory(MemRefToSPIRV)
+add_subdirectory(MeshToMPI)
 add_subdirectory(NVGPUToNVVM)
 add_subdirectory(NVVMToLLVM)
 add_subdirectory(OpenACCToSCF)

mlir/lib/Conversion/MeshToMPI/CMakeLists.txt

@@ -0,0 +1,22 @@
+add_mlir_conversion_library(MLIRMeshToMPI
+  MeshToMPI.cpp
+
+  ADDITIONAL_HEADER_DIRS
+  ${MLIR_MAIN_INCLUDE_DIR}/mlir/Conversion/MeshToMPI
+
+  DEPENDS
+  MLIRConversionPassIncGen
+
+  LINK_COMPONENTS
+  Core
+
+  LINK_LIBS PUBLIC
+  MLIRFuncDialect
+  MLIRIR
+  MLIRLinalgTransforms
+  MLIRMemRefDialect
+  MLIRPass
+  MLIRMeshDialect
+  MLIRMPIDialect
+  MLIRTransforms
+  )