[mlir][transform] Add support for expressing scalable tile sizes

This patch enables specifying scalable tile sizes when using the
Transform dialect to drive tiling, e.g.:
```
%1, %loop = transform.structured.tile %0 [[4]]
```
This is implemented by extending the TileOp with a dedicated attribute
for "scalability" and by updating various parsing hooks. At the moment,
only the trailing tile size can be scalable. The following is not yet
supported:
```
%1, %loop = transform.structured.tile %0 [[4], [4]]
```

This change is a part of larger effort to enable scalable vectorisation
in Linalg. See this RFC for more context:
  * https://discourse.llvm.org/t/rfc-scalable-vectorisation-in-linalg/

Differential Revision: https://reviews.llvm.org/D150944

mlir/include/mlir/Dialect/Linalg/TransformOps/LinalgTransformOps.td

@@ -1528,7 +1528,8 @@ def TileOp : Op<Transform_Dialect, "structured.tile",
   let arguments = (ins TransformHandleTypeInterface:$target,
                    Variadic<TransformParamTypeOrAnyHandle>:$dynamic_sizes,
                    DefaultValuedOptionalAttr<DenseI64ArrayAttr, "{}">:$static_sizes,
-                   DefaultValuedOptionalAttr<DenseI64ArrayAttr, "{}">:$interchange);
+                   DefaultValuedOptionalAttr<DenseI64ArrayAttr, "{}">:$interchange,
+                   DefaultValuedOptionalAttr<BoolAttr, "false">:$last_tile_size_scalable);
   let results = (outs TransformHandleTypeInterface:$tiled_linalg_op,
                       Variadic<TransformHandleTypeInterface>:$loops);
   let builders = [

mlir/include/mlir/Interfaces/ViewLikeInterface.h

@@ -72,17 +72,27 @@ void printDynamicIndexList(
 ///   1. `result` is filled with the i64 ArrayAttr "[`kDynamic`, 7, 42,
 ///   `kDynamic`]"
 ///   2. `ssa` is filled with "[%arg0, %arg1]".
+///
+/// Trailing indices can be scalable. For example, "42" in "[7, [42]]" is
+/// scalable. This notation is similar to how scalable dims are marked when
+/// defining Vectors. If /p isTrailingIdxScalable is null, scalable indices are
+/// not allowed/expected. When it's not null, this hook will set the
+/// corresponding value to:
+///   * true if the trailing idx is scalable,
+///   * false otherwise.
 ParseResult parseDynamicIndexList(
     OpAsmParser &parser,
     SmallVectorImpl<OpAsmParser::UnresolvedOperand> &values,
-    DenseI64ArrayAttr &integers, SmallVectorImpl<Type> *valueTypes = nullptr,
+    DenseI64ArrayAttr &integers, bool *isTrailingIdxScalable = nullptr,
+    SmallVectorImpl<Type> *valueTypes = nullptr,
     AsmParser::Delimiter delimiter = AsmParser::Delimiter::Square);
 inline ParseResult parseDynamicIndexList(
     OpAsmParser &parser,
     SmallVectorImpl<OpAsmParser::UnresolvedOperand> &values,
     DenseI64ArrayAttr &integers, SmallVectorImpl<Type> &valueTypes,
     AsmParser::Delimiter delimiter = AsmParser::Delimiter::Square) {
-  return parseDynamicIndexList(parser, values, integers, &valueTypes,
+  return parseDynamicIndexList(parser, values, integers,
+                               /*isTrailingIdxScalable=*/nullptr, &valueTypes,
                                delimiter);
 }
 

mlir/lib/Dialect/Linalg/TransformOps/LinalgTransformOps.cpp

@@ -2391,6 +2391,7 @@ transform::TileOp::apply(TransformResults &transformResults,
   SmallVector<Operation *> tiled;
   SmallVector<SmallVector<Operation *, 4>, 4> loops;
   loops.resize(getLoops().size());
+  bool scalable = getLastTileSizeScalable();
   for (auto [i, op] : llvm::enumerate(targets)) {
     auto tilingInterface = dyn_cast<TilingInterface>(op);
     auto dpsInterface = dyn_cast<DestinationStyleOpInterface>(op);
@@ -2409,10 +2410,21 @@ transform::TileOp::apply(TransformResults &transformResults,
         SmallVector<Value, 4> sizes;
         sizes.reserve(tileSizes.size());
         unsigned dynamicIdx = 0;
-        for (OpFoldResult ofr : getMixedSizes()) {
+        unsigned trailingIdx = getMixedSizes().size() - 1;
+
+        for (auto [ofrIdx, ofr] : llvm::enumerate(getMixedSizes())) {
           if (auto attr = llvm::dyn_cast_if_present<Attribute>(ofr)) {
-            sizes.push_back(b.create<arith::ConstantIndexOp>(
+            // Only the trailing tile size is allowed to be scalable atm.
-                getLoc(), cast<IntegerAttr>(attr).getInt()));
+            if (scalable && (ofrIdx == trailingIdx)) {
+              auto val = b.create<arith::ConstantIndexOp>(
+                  getLoc(), attr.cast<IntegerAttr>().getInt());
+              Value vscale =
+                  b.create<vector::VectorScaleOp>(getLoc(), b.getIndexType());
+              sizes.push_back(b.create<arith::MulIOp>(getLoc(), val, vscale));
+            } else {
+              sizes.push_back(b.create<arith::ConstantIndexOp>(
+                  getLoc(), cast<IntegerAttr>(attr).getInt()));
+            }
             continue;
           }
           ArrayRef<Operation *> dynamicSizes = dynamicSizeProducers[dynamicIdx];
@@ -2507,8 +2519,9 @@ ParseResult transform::TileOp::parse(OpAsmParser &parser,
   DenseI64ArrayAttr staticSizes;
   FunctionType functionalType;
   llvm::SMLoc operandLoc;
+  bool scalable = false;
   if (parser.parseOperand(target) || parser.getCurrentLocation(&operandLoc) ||
-      parseDynamicIndexList(parser, dynamicSizes, staticSizes) ||
+      parseDynamicIndexList(parser, dynamicSizes, staticSizes, &scalable) ||
       parseOptionalInterchange(parser, result) ||
       parser.parseColonType(functionalType))
     return ParseResult::failure();
@@ -2531,6 +2544,10 @@ ParseResult transform::TileOp::parse(OpAsmParser &parser,
     return failure();
   }
 
+  auto scalableAttr = parser.getBuilder().getBoolAttr(scalable);
+  result.addAttribute(getLastTileSizeScalableAttrName(result.name),
+                      scalableAttr);
+
   result.addAttribute(getStaticSizesAttrName(result.name), staticSizes);
   result.addTypes(functionalType.getResults());
   return success();

mlir/lib/Dialect/SCF/IR/SCF.cpp

@@ -1261,9 +1261,9 @@ ParseResult ForallOp::parse(OpAsmParser &parser, OperationState &result) {
       dynamicSteps;
   if (succeeded(parser.parseOptionalKeyword("in"))) {
     // Parse upper bounds.
-    if (parseDynamicIndexList(parser, dynamicUbs, staticUbs,
+    if (parseDynamicIndexList(
-                              /*valueTypes=*/nullptr,
+            parser, dynamicUbs, staticUbs, /*scalable=*/nullptr,
-                              OpAsmParser::Delimiter::Paren) ||
+            /*valueTypes=*/nullptr, OpAsmParser::Delimiter::Paren) ||
         parser.resolveOperands(dynamicUbs, indexType, result.operands))
       return failure();
 
@@ -1273,26 +1273,26 @@ ParseResult ForallOp::parse(OpAsmParser &parser, OperationState &result) {
   } else {
     // Parse lower bounds.
     if (parser.parseEqual() ||
-        parseDynamicIndexList(parser, dynamicLbs, staticLbs,
+        parseDynamicIndexList(
-                              /*valueTypes=*/nullptr,
+            parser, dynamicLbs, staticLbs, /*scalable=*/nullptr,
-                              OpAsmParser::Delimiter::Paren) ||
+            /*valueTypes=*/nullptr, OpAsmParser::Delimiter::Paren) ||
 
         parser.resolveOperands(dynamicLbs, indexType, result.operands))
       return failure();
 
     // Parse upper bounds.
     if (parser.parseKeyword("to") ||
-        parseDynamicIndexList(parser, dynamicUbs, staticUbs,
+        parseDynamicIndexList(
-                              /*valueTypes=*/nullptr,
+            parser, dynamicUbs, staticUbs, /*scalable=*/nullptr,
-                              OpAsmParser::Delimiter::Paren) ||
+            /*valueTypes=*/nullptr, OpAsmParser::Delimiter::Paren) ||
         parser.resolveOperands(dynamicUbs, indexType, result.operands))
       return failure();
 
     // Parse step values.
     if (parser.parseKeyword("step") ||
-        parseDynamicIndexList(parser, dynamicSteps, staticSteps,
+        parseDynamicIndexList(
-                              /*valueTypes=*/nullptr,
+            parser, dynamicSteps, staticSteps, /*scalable=*/nullptr,
-                              OpAsmParser::Delimiter::Paren) ||
+            /*valueTypes=*/nullptr, OpAsmParser::Delimiter::Paren) ||
         parser.resolveOperands(dynamicSteps, indexType, result.operands))
       return failure();
   }

mlir/lib/Dialect/Transform/Utils/Utils.cpp

@@ -42,5 +42,6 @@ ParseResult mlir::transform::parsePackedOrDynamicIndexList(
     return success();
   }
 
-  return parseDynamicIndexList(parser, values, integers, &valueTypes);
+  return parseDynamicIndexList(parser, values, integers, /*scalable=*/nullptr,
+                               &valueTypes);
 }

mlir/lib/Interfaces/ViewLikeInterface.cpp

@@ -128,13 +128,26 @@ void mlir::printDynamicIndexList(OpAsmPrinter &printer, Operation *op,
 ParseResult mlir::parseDynamicIndexList(
     OpAsmParser &parser,
     SmallVectorImpl<OpAsmParser::UnresolvedOperand> &values,
-    DenseI64ArrayAttr &integers, SmallVectorImpl<Type> *valueTypes,
+    DenseI64ArrayAttr &integers, bool *isTrailingIdxScalable,
-    AsmParser::Delimiter delimiter) {
+    SmallVectorImpl<Type> *valueTypes, AsmParser::Delimiter delimiter) {
 
   SmallVector<int64_t, 4> integerVals;
+  bool foundScalable = false;
   auto parseIntegerOrValue = [&]() {
     OpAsmParser::UnresolvedOperand operand;
     auto res = parser.parseOptionalOperand(operand);
+
+    // If `foundScalable` has already been set to `true` then a non-trailing
+    // tile size was identified as scalable.
+    if (foundScalable) {
+      parser.emitError(parser.getNameLoc())
+          << "non-trailing tile size cannot be scalable";
+      return failure();
+    }
+
+    if (isTrailingIdxScalable && parser.parseOptionalLSquare().succeeded())
+      foundScalable = true;
+
     if (res.has_value() && succeeded(res.value())) {
       values.push_back(operand);
       integerVals.push_back(ShapedType::kDynamic);
@@ -146,6 +159,8 @@ ParseResult mlir::parseDynamicIndexList(
         return failure();
       integerVals.push_back(integer);
     }
+    if (foundScalable && parser.parseOptionalRSquare().failed())
+      return failure();
     return success();
   };
   if (parser.parseCommaSeparatedList(delimiter, parseIntegerOrValue,
@@ -153,6 +168,8 @@ ParseResult mlir::parseDynamicIndexList(
     return parser.emitError(parser.getNameLoc())
            << "expected SSA value or integer";
   integers = parser.getBuilder().getDenseI64ArrayAttr(integerVals);
+  if (isTrailingIdxScalable)
+    *isTrailingIdxScalable = foundScalable;
   return success();
 }
 

mlir/test/Dialect/Linalg/transform-op-tile.mlir

@@ -1,4 +1,4 @@
-// RUN: mlir-opt --test-transform-dialect-interpreter --split-input-file --verify-diagnostics %s | FileCheck %s
+// RUN: mlir-opt --test-transform-dialect-interpreter --mlir-print-local-scope --split-input-file --verify-diagnostics %s | FileCheck %s
 
 transform.sequence failures(propagate) {
 ^bb0(%arg1: !transform.any_op):
@@ -149,3 +149,96 @@ transform.sequence failures(propagate) {
   transform.structured.tile_to_forall_op %0 tile_sizes[1, 1]
          : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
 }
+
+// -----
+
+#map = affine_map<(d0) -> (d0)>
+
+module {
+  func.func @scalable_tile(%arg0: tensor<?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<?xf32>, %arg3: f32) -> tensor<?xf32> {
+    %0 = linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = ["parallel"]} ins(%arg0, %arg1 : tensor<?xf32>, tensor<?xf32>) outs(%arg2 : tensor<?xf32>) {
+    ^bb0(%in_1: f32, %in_2: f32, %out: f32):
+      %1 = arith.addf %in_1, %in_2 : f32
+      %2 = arith.mulf %arg3, %1 : f32
+      linalg.yield %2 : f32
+    } -> tensor<?xf32>
+    return %0 : tensor<?xf32>
+  }
+}
+
+// CHECK-LABEL:   func.func @scalable_tile(
+// CHECK-SAME:      %[[ARG_0:.*]]: tensor<?xf32>, %[[ARG_1:.*]]: tensor<?xf32>, %[[ARG_2:.*]]: tensor<?xf32>,
+// CHECK:           %[[C4:.*]] = arith.constant 0 : index
+// CHECK:           %[[DIM:.*]] = tensor.dim %[[ARG_0]], %[[C4]] : tensor<?xf32>
+// CHECK:           %[[VEC_SIZE:.*]] = arith.constant 4 : index
+// CHECK:           %[[VS:.*]] = vector.vscale
+// CHECK:           %[[STEP:.*]] = arith.muli %[[VEC_SIZE]], %[[VS]] : index
+// CHECK:           %[[C0:.*]] = arith.constant 0 : index
+// CHECK:           scf.for %[[IV:.*]] = %[[C0]] to %[[DIM]] step %[[STEP]] iter_args(%[[VAL:.*]] = %[[ARG_2]]) -> (tensor<?xf32>) {
+// CHECK:             %[[SIZE:.*]] = affine.min affine_map<(d0)[s0, s1] -> (s0, -d0 + s1)>(%[[IV]])[%[[STEP]], %[[DIM]]]
+// CHECK:             %[[SLICE_ARG0:.*]] = tensor.extract_slice %[[ARG_0]][%[[IV]]] [%[[SIZE]]] [1] : tensor<?xf32> to tensor<?xf32>
+// CHECK:             %[[SLICE_ARG1:.*]] = tensor.extract_slice %[[ARG_1]][%[[IV]]] [%[[SIZE]]] [1] : tensor<?xf32> to tensor<?xf32>
+// CHECK:             %[[SLICE_ARG2:.*]] = tensor.extract_slice %[[VAL]][%[[IV]]] [%[[SIZE]]] [1] : tensor<?xf32> to tensor<?xf32>
+// CHECK:             linalg.generic {indexing_maps = {{.*}}, iterator_types = ["parallel"]} ins(%[[SLICE_ARG0]], %[[SLICE_ARG1]] : tensor<?xf32>, tensor<?xf32>) outs(%[[SLICE_ARG2]] : tensor<?xf32>) {
+
+transform.sequence failures(propagate) {
+  ^bb0(%arg1: !transform.any_op):
+    %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+    %1, %loop = transform.structured.tile %0 [[4]] : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+}
+
+// -----
+
+// CHECK-LABEL:   func.func @scalable_and_fixed_length_tile
+// CHECK:           %[[STEP_0:.*]] = arith.constant 4 : index
+// CHECK:           %[[STEP_1:.*]] = arith.constant 4 : index
+// CHECK:           %[[C4:.*]] = arith.constant 4 : index
+// CHECK:           %[[VS:.*]] = vector.vscale
+// CHECK:           %[[STEP_2:.*]] = arith.muli %[[C4]], %[[VS]] : index
+// CHECK:           %[[C0:.*]] = arith.constant 0 : index
+// CHECK:           %[[C128:.*]] = arith.constant 128 : index
+// CHECK:           scf.for %[[VAL_11:.*]] = %[[C0]] to %[[C128]] step %[[STEP_0]]
+// CHECK:             %[[C0_1:.*]] = arith.constant 0 : index
+// CHECK:             %[[C128_1:.*]] = arith.constant 128 : index
+// CHECK:             scf.for %[[VAL_16:.*]] = %[[C0_1]] to %[[C128_1]] step %[[STEP_1]]
+// CHECK:               %[[C0_2:.*]] = arith.constant 0 : index
+// CHECK:               %[[C128_2:.*]] = arith.constant 128 : index
+// CHECK:               scf.for %{{.*}} = %[[C0_2]] to %[[C128_2]] step %[[STEP_2]]
+
+func.func @scalable_and_fixed_length_tile(
+  %arg0: tensor<128x128xf32>, %arg1: tensor<128x128xf32>, %arg2: tensor<128x128xf32>)
+    -> tensor<128x128xf32> {
+  %0 = linalg.matmul  ins(%arg0, %arg1: tensor<128x128xf32>, tensor<128x128xf32>)
+                     outs(%arg2: tensor<128x128xf32>)
+    -> tensor<128x128xf32>
+
+  return %0 : tensor<128x128xf32>
+}
+
+transform.sequence failures(propagate) {
+^bb0(%arg1: !transform.any_op):
+  %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+  %1, %loops:3 = transform.structured.tile %0 [4, 4, [4]] : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op)
+}
+
+// -----
+
+// TODO: Add support for for specyfying more than one scalable tile size
+
+func.func @scalable_and_fixed_length_tile(
+  %arg0: tensor<128x128xf32>, %arg1: tensor<128x128xf32>, %arg2: tensor<128x128xf32>)
+    -> tensor<128x128xf32> {
+  %0 = linalg.matmul  ins(%arg0, %arg1: tensor<128x128xf32>, tensor<128x128xf32>)
+                     outs(%arg2: tensor<128x128xf32>)
+    -> tensor<128x128xf32>
+
+  return %0 : tensor<128x128xf32>
+}
+
+transform.sequence failures(propagate) {
+^bb0(%arg1: !transform.any_op):
+  %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+  // expected-error @below {{non-trailing tile size cannot be scalable}}
+  // expected-error @below {{expected SSA value or integer}}
+  %1, %loops:3 = transform.structured.tile %0 [4, [4], [4]] : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op)
+}