[mlir][mesh] Use tensor shape notation for the shape of a cluster (#73826)

Examle:

substitute
mesh.cluster @Mesh0(rank = 2, dim_sizes = [0, 4])

with
mesh.cluster @Mesh0(rank = 2, dim_sizes = ?x4)

Same as tensor/memref shapes. The only difference is for 0-rank shapes.
With tensors you would have something like `tensor<f32>`. Here to avoid
matching an empty string a 0-rank shape is denoted by `[]`.

Author: sogartar

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

@@ -40,26 +40,27 @@ def Mesh_ClusterOp : Mesh_Op<"cluster", [Symbol]> {
     determine the layout and the addressing space of the computation distributed
     across the mesh.
 
-    3. `dim_sizes`: This attribute represents the device assignment along the
-    axes of the cluster. Each integer in the array corresponds to the number of
-    devices along a specific axis. If an integer value is 0, it implies that the
-    number of devices along that axis is unknown. This flexibility allows for
-    dynamic device assignment or configurations where the exact number of
-    devices might not be determined during compile time.
+    3. `dim_sizes`: This attribute represents the shape of the device cluster.
+    It uses the same notation as a tensor shape. Also allowing for dynamic
+    dimensions.
+    This flexibility allows for dynamic device assignment or configurations
+    where the exact number of devices might not be determined during compile
+    time.
+    For example `2x?x4`.
 
     Example:
     ```
     // A device mesh cluster with 3 axes, the total device number is 4 * 8 * 12
     // The dimension sizes are 4, 8, 12 
-    mesh.cluster @mesh0(rank = 3, dim_sizes = [4, 8, 12])
+    mesh.cluster @mesh0(rank = 3, dim_sizes = 4x8x12)
 
     // A device mesh cluster with 2 axes, the total device number is unknown
     // The first dimension size is 4 and the second is unknown
-    mesh.cluster @mesh1(rank = 2, dim_sizes = [4])
+    mesh.cluster @mesh1(rank = 2, dim_sizes = 4)
 
     // A device mesh cluster with 2 axes, the total device number is unknown
     // The first dimension size is unknown and the second is 4
-    mesh.cluster @mesh2(rank = 2, dim_sizes = [0, 4])
+    mesh.cluster @mesh2(rank = 2, dim_sizes = ?x4)
 
     // A device mesh cluster with 2 axes, the number of devices along both axes
     // is unknown
@@ -76,7 +77,7 @@ def Mesh_ClusterOp : Mesh_Op<"cluster", [Symbol]> {
     DefaultValuedAttr<DenseI64ArrayAttr, "{}">:$dim_sizes
   );
   let assemblyFormat = [{
-    $sym_name `(` `rank` `=` $rank (`,` `dim_sizes` `=` $dim_sizes^)? `)`
+    $sym_name `(` `rank` `=` $rank (`,` `dim_sizes` `=` custom<DimensionList>($dim_sizes)^)? `)`
       attr-dict
   }];
   let extraClassDeclaration = [{
@@ -88,7 +89,7 @@ def Mesh_ClusterOp : Mesh_Op<"cluster", [Symbol]> {
     template <typename OutIt>
     void canonicalDimSizes(OutIt outIt) {
       std::copy(getDimSizes().begin(), getDimSizes().end(), outIt);
-      std::fill_n(outIt, getRank() - getDimSizes().size(), 0);
+      std::fill_n(outIt, getRank() - getDimSizes().size(), ::mlir::ShapedType::kDynamic);
     }
   }];
   let hasVerifier = 1;
@@ -210,7 +211,7 @@ def Mesh_AllGatherOp : Mesh_CollectiveCommunicationOpBase<"all_gather", [
 
     Example:
     ```mlir
-    mesh.cluster @mesh0(rank = 2, dim_sizes = [2, 2])
+    mesh.cluster @mesh0(rank = 2, dim_sizes = 2x2)
     ...
     %1 = mesh.all_gather %0 on @mesh0 mesh_axes = [1] gather_axis = 1
       : tensor<2x2xi8> -> tensor<2x4xi8>
@@ -295,7 +296,7 @@ def Mesh_AllToAllOp : Mesh_CollectiveCommunicationOpBase<"all_to_all", [
 
     Example:
     ```
-    mesh.cluster @mesh0(rank = 1, dim_sizes = [3])
+    mesh.cluster @mesh0(rank = 1, dim_sizes = 3)
     ...
     %1 = mesh.all_to_all %0 on @mesh0 mesh_axes = [0]
       split_axis = 0 concat_axis = 0
@@ -527,7 +528,7 @@ def Mesh_ReduceScatterOp : Mesh_CollectiveCommunicationOpBase<"reduce_scatter",
     across the device group.
     Example:
     ```
-    mesh.cluster @mesh0(rank = 1, dim_sizes = [2, 2])
+    mesh.cluster @mesh0(rank = 1, dim_sizes = 2x2)
     ...
     %1 = mesh.reduce_scatter %0 on @mesh0 mesh_axes = [1]
       reduction = <max> scatter_axis = 0

mlir/include/mlir/IR/OpImplementation.h

@@ -226,6 +226,8 @@ class AsmPrinter {
     printArrowTypeList(results);
   }
 
+  void printDimensionList(ArrayRef<int64_t> shape);
+
   /// Class used to automatically end a cyclic region on destruction.
   class CyclicPrintReset {
   public:
@@ -1762,6 +1764,17 @@ class OpAsmDialectInterface
                  const SetVector<AsmDialectResourceHandle> &referencedResources,
                  AsmResourceBuilder &builder) const {}
 };
+
+//===--------------------------------------------------------------------===//
+// Custom printers and parsers.
+//===--------------------------------------------------------------------===//
+
+// Handles custom<DimensionList>(...) in TableGen.
+void printDimensionList(OpAsmPrinter &printer, Operation *op,
+                        ArrayRef<int64_t> dimensions);
+ParseResult parseDimensionList(OpAsmParser &parser,
+                               DenseI64ArrayAttr &dimensions);
+
 } // namespace mlir
 
 //===--------------------------------------------------------------------===//

mlir/lib/Dialect/Mesh/IR/MeshOps.cpp

@@ -58,11 +58,6 @@ static SmallVector<T> &canonicalizeSetAsVector(SmallVector<T> &vec) {
   return vec;
 }
 
-template <typename DimSize>
-static bool isMeshDimensionDynamic(DimSize size) {
-  return size <= DimSize(0);
-}
-
 using MeshAxis = int16_t;
 
 namespace {
@@ -161,9 +156,9 @@ LogicalResult ClusterOp::verify() {
         "rank of dim_sizes is not expected to be larger than rank of cluster");
 
   for (int64_t dimSize : dimSizes) {
-    if (dimSize < 0)
-      return emitOpError(
-          "dimension size of a mesh cluster is expected to be non-negative");
+    if (dimSize < 0 && !ShapedType::isDynamic(dimSize))
+      return emitOpError("dimension size of a mesh cluster is expected to be "
+                         "non-negative or dynamic");
   }
 
   return success();
@@ -316,7 +311,7 @@ static int64_t collectiveDeviceGroupSize(ArrayRef<MeshAxis> meshAxes,
   int64_t res = 1;
 
   for (MeshAxis axis : meshAxes) {
-    if (isMeshDimensionDynamic(meshShape[axis])) {
+    if (ShapedType::isDynamic(meshShape[axis])) {
       return ShapedType::kDynamic;
     }
     assert(size_t(axis) < meshShape.size());

mlir/lib/IR/AsmPrinter.cpp

@@ -16,7 +16,9 @@
 #include "mlir/IR/AsmState.h"
 #include "mlir/IR/Attributes.h"
 #include "mlir/IR/Builders.h"
+#include "mlir/IR/BuiltinAttributes.h"
 #include "mlir/IR/BuiltinDialect.h"
+#include "mlir/IR/BuiltinTypeInterfaces.h"
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Dialect.h"
 #include "mlir/IR/DialectImplementation.h"
@@ -27,6 +29,7 @@
 #include "mlir/IR/Operation.h"
 #include "mlir/IR/Verifier.h"
 #include "llvm/ADT/APFloat.h"
+#include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/MapVector.h"
 #include "llvm/ADT/STLExtras.h"
@@ -44,6 +47,7 @@
 #include "llvm/Support/SaveAndRestore.h"
 #include "llvm/Support/Threading.h"
 #include "llvm/Support/raw_ostream.h"
+#include <type_traits>
 
 #include <optional>
 #include <tuple>
@@ -425,6 +429,8 @@ class AsmPrinter::Impl {
 
   void popCyclicPrinting();
 
+  void printDimensionList(ArrayRef<int64_t> shape);
+
 protected:
   void printOptionalAttrDict(ArrayRef<NamedAttribute> attrs,
                              ArrayRef<StringRef> elidedAttrs = {},
@@ -1860,6 +1866,20 @@ class AsmStateImpl {
   // Allow direct access to the impl fields.
   friend AsmState;
 };
+
+template <typename Range>
+void printDimensionList(raw_ostream &stream, Range &&shape) {
+  llvm::interleave(
+      shape, stream,
+      [&stream](const auto &dimSize) {
+        if (ShapedType::isDynamic(dimSize))
+          stream << "?";
+        else
+          stream << dimSize;
+      },
+      "x");
+}
+
 } // namespace detail
 } // namespace mlir
 
@@ -2576,13 +2596,9 @@ void AsmPrinter::Impl::printTypeImpl(Type type) {
       })
       .Case<RankedTensorType>([&](RankedTensorType tensorTy) {
         os << "tensor<";
-        for (int64_t dim : tensorTy.getShape()) {
-          if (ShapedType::isDynamic(dim))
-            os << '?';
-          else
-            os << dim;
+        printDimensionList(tensorTy.getShape());
+        if (!tensorTy.getShape().empty())
           os << 'x';
-        }
         printType(tensorTy.getElementType());
         // Only print the encoding attribute value if set.
         if (tensorTy.getEncoding()) {
@@ -2598,13 +2614,9 @@ void AsmPrinter::Impl::printTypeImpl(Type type) {
       })
       .Case<MemRefType>([&](MemRefType memrefTy) {
         os << "memref<";
-        for (int64_t dim : memrefTy.getShape()) {
-          if (ShapedType::isDynamic(dim))
-            os << '?';
-          else
-            os << dim;
+        printDimensionList(memrefTy.getShape());
+        if (!memrefTy.getShape().empty())
           os << 'x';
-        }
         printType(memrefTy.getElementType());
         MemRefLayoutAttrInterface layout = memrefTy.getLayout();
         if (!llvm::isa<AffineMapAttr>(layout) || !layout.isIdentity()) {
@@ -2735,6 +2747,10 @@ LogicalResult AsmPrinter::Impl::pushCyclicPrinting(const void *opaquePointer) {
 
 void AsmPrinter::Impl::popCyclicPrinting() { state.popCyclicPrinting(); }
 
+void AsmPrinter::Impl::printDimensionList(ArrayRef<int64_t> shape) {
+  detail::printDimensionList(os, shape);
+}
+
 //===--------------------------------------------------------------------===//
 // AsmPrinter
 //===--------------------------------------------------------------------===//
@@ -2800,6 +2816,10 @@ void AsmPrinter::printResourceHandle(const AsmDialectResourceHandle &resource) {
   impl->printResourceHandle(resource);
 }
 
+void AsmPrinter::printDimensionList(ArrayRef<int64_t> shape) {
+  detail::printDimensionList(getStream(), shape);
+}
+
 LogicalResult AsmPrinter::pushCyclicPrinting(const void *opaquePointer) {
   return impl->pushCyclicPrinting(opaquePointer);
 }
@@ -3911,3 +3931,47 @@ void Block::printAsOperand(raw_ostream &os, AsmState &state) {
   OperationPrinter printer(os, state.getImpl());
   printer.printBlockName(this);
 }
+
+//===--------------------------------------------------------------------===//
+// Custom printers
+//===--------------------------------------------------------------------===//
+namespace mlir {
+
+void printDimensionList(OpAsmPrinter &printer, Operation *op,
+                        ArrayRef<int64_t> dimensions) {
+  if (dimensions.empty())
+    printer << "[";
+  printer.printDimensionList(dimensions);
+  if (dimensions.empty())
+    printer << "]";
+}
+
+ParseResult parseDimensionList(OpAsmParser &parser,
+                               DenseI64ArrayAttr &dimensions) {
+  // Empty list case denoted by "[]".
+  if (succeeded(parser.parseOptionalLSquare())) {
+    if (failed(parser.parseRSquare())) {
+      return parser.emitError(parser.getCurrentLocation())
+             << "Failed parsing dimension list.";
+    }
+    dimensions =
+        DenseI64ArrayAttr::get(parser.getContext(), ArrayRef<int64_t>());
+    return success();
+  }
+
+  // Non-empty list case.
+  SmallVector<int64_t> shapeArr;
+  if (failed(parser.parseDimensionList(shapeArr, true, false))) {
+    return parser.emitError(parser.getCurrentLocation())
+           << "Failed parsing dimension list.";
+  }
+  if (shapeArr.empty()) {
+    return parser.emitError(parser.getCurrentLocation())
+           << "Failed parsing dimension list. Did you mean an empty list? It "
+              "must be denoted by \"[]\".";
+  }
+  dimensions = DenseI64ArrayAttr::get(parser.getContext(), shapeArr);
+  return success();
+}
+
+} // namespace mlir

mlir/test/Dialect/Mesh/canonicalization.mlir

@@ -1,6 +1,6 @@
 // RUN: mlir-opt --canonicalize %s | FileCheck %s
 
-mesh.cluster @mesh0(rank = 2, dim_sizes = [2, 4])
+mesh.cluster @mesh0(rank = 2, dim_sizes = 2x4)
 
 // CHECK-LABEL: func @all_reduce_empty_mesh_axes
 func.func @all_reduce_empty_mesh_axes(