[MLIR] Add support for multiway split in SplitOp

Add functionality that enables SplitOp to do a multiway split of
a traget op along a given dimension. With multiway attribute,
SplitOp takes a list of chunk sizes and applies it to a single
target along the given dimension to generate multiple
structured ops extracted from the target.

Author: muneebkhan85

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

@@ -1396,29 +1396,43 @@ def SplitOp : Op<Transform_Dialect, "structured.split",
      DeclareOpInterfaceMethods<TransformOpInterface>,
      ReportTrackingListenerFailuresOpTrait]> {
   let description = [{
-    Indicates that the given `target` op should be split into two complementary
+    Splits the given `target` op into two or more complementary
     parts, which combined cover the entire iteration domain of the original op.
     The split is performed along the iteration space dimension provided as
-    attribute. In case of dimension overflow, the transformation fails. The
-    split is performed at the dimension iterator value specified as either the
-    static split point attribute when it is known at transform IR construction
-    time or as the handle to an operation producing a single index-typed value
-    when it is computed by payload IR. In the latter case, the static split
+    chunk size attribute specifying the size of the lower part; the remaining
+    range in the iteration space is assigned as the upper part. In case of
+    dimension overflow, the transformation fails. The split is performed at the
+    dimension iterator value specified as either the static chunk size
+    attribute when it is known at transform IR construction time or
+    as the handle to an operation producing a single index-typed value
+    when it is computed by payload IR. In the latter case, the chunk size
     point must be set to `ShapedType::kDynamic` and the dynamic size handle
     must point to as many value-producing operations as there are structured
     operations pointed to by the target handle.
 
-    The operation consumes the target handle, but preserves the split point
-    handle if provided. It produces two new handles pointing to the two parts
-    of the structured op after splitting, in the same order as the target
-    operand, with the first handle corresponding to the part with lower
-    iteration space indices.
+    The operation consumes the target handle, but preserves the chunk size
+    handle if provided. Without the `multiway` attribute, it produces two
+    new handles pointing to the two parts of the structured op after splitting,
+    in the same order as the target operand, with the first handle
+    corresponding to the part with lower iteration space indices.
+
+    Multiway split mode is enabled by specifying the `multiway` attribute.
+    In this mode a single `target` op is split into multiple parts covering
+    the iteration space of the specified dimension. `static_chunk_sizes` and
+    `dynamic_chunk_sizes` in this case is a list of chunk sizes that the given
+    dimension should be split into. With `multiway` it produces two handles;
+    the first handle is a list of the multiple parts of the structured op
+    after splitting, where the target dimensions for each linalg op in the
+    list corresponds to the chunk sizes specfied in the input split list.
+    If the chunk sizes do not cover the entire iteration space, the leftover
+    chunk is the last payload in the first handle. The second handle is empty.
   }];
 
   let arguments = (ins TransformHandleTypeInterface:$target,
                        I64Attr:$dimension,
-                       Optional<TransformAnyParamTypeOrAnyHandle>:$dynamic_split_point,
-                       I64Attr:$static_split_point);
+                       Optional<TransformAnyParamTypeOrAnyHandle>:$dynamic_chunk_sizes,
+                       I64Attr:$static_chunk_sizes,
+                       UnitAttr:$multiway);
   let results = (outs TransformHandleTypeInterface:$first,
                       TransformHandleTypeInterface:$second);
   let hasCustomAssemblyFormat = 1;

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

@@ -2269,13 +2269,26 @@ SplitOp::apply(transform::TransformRewriter &rewriter,
   // Collect the dynamic split points if provided.
   SmallVector<Operation *> payload =
       llvm::to_vector(state.getPayloadOps(getTarget()));
-  SmallVector<OpFoldResult> splitPoints;
-  splitPoints.reserve(payload.size());
-  if (getDynamicSplitPoint()) {
+
+  bool isMultiwaySplit = getMultiway();
+
+  if (isMultiwaySplit && !llvm::hasSingleElement(payload)) {
+    return mlir::emitSilenceableFailure(getLoc())
+           << "requires exactly one target when "
+              "multiway split is enabled (got "
+           << llvm::range_size(payload) << ")";
+  }
+
+  SmallVector<OpFoldResult> chunkSizes;
+
+  if (!isMultiwaySplit)
+    chunkSizes.reserve(payload.size());
+
+  if (getDynamicChunkSizes()) {
     auto diag = DiagnosedSilenceableFailure::success();
-    if (isa<TransformHandleTypeInterface>(getDynamicSplitPoint().getType())) {
-      splitPoints = llvm::to_vector(llvm::map_range(
-          state.getPayloadOps(getDynamicSplitPoint()), [&](Operation *op) {
+    if (isa<TransformHandleTypeInterface>(getDynamicChunkSizes().getType())) {
+      chunkSizes = llvm::to_vector(llvm::map_range(
+          state.getPayloadOps(getDynamicChunkSizes()), [&](Operation *op) {
             if (op->getNumResults() != 1 ||
                 !op->getResult(0).getType().isIndex()) {
               diag = emitSilenceableError()
@@ -2286,103 +2299,171 @@ SplitOp::apply(transform::TransformRewriter &rewriter,
             return OpFoldResult(op->getResult(0));
           }));
     } else {
-      splitPoints = llvm::to_vector(
-          llvm::map_range(state.getParams(getDynamicSplitPoint()),
+      chunkSizes = llvm::to_vector(
+          llvm::map_range(state.getParams(getDynamicChunkSizes()),
                           [](Attribute attr) { return OpFoldResult(attr); }));
     }
     if (diag.isSilenceableFailure())
       return diag;
 
-    if (splitPoints.size() != payload.size()) {
+    // For multiway split, a single payload is expected to have multiple
+    // split points.
+    if (!isMultiwaySplit && chunkSizes.size() != payload.size()) {
       return emitDefiniteFailure()
              << "expected the dynamic split point handle to point to as "
                 "many operations ("
-             << splitPoints.size() << ") as the target handle ("
+             << chunkSizes.size() << ") as the target handle ("
              << payload.size() << ")";
     }
   } else {
-    splitPoints.resize(payload.size(),
-                       rewriter.getIndexAttr(getStaticSplitPoint()));
+    chunkSizes.resize(payload.size(),
+                       rewriter.getIndexAttr(getStaticChunkSizes()));
   }
 
-  // Split each target operation.
-  SmallVector<Operation *> first, second;
-  Operation *noSecondPart = nullptr;
-  for (const auto &pair : llvm::zip(payload, splitPoints)) {
-    Operation *target = std::get<0>(pair);
-    auto linalgOp = dyn_cast<LinalgOp>(target);
+  auto checkStructuredOpAndDimensions =
+      [&](LinalgOp linalgOp, Location loc) -> DiagnosedSilenceableFailure {
     if (!linalgOp) {
       auto diag = emitSilenceableError() << "only applies to structured ops";
-      diag.attachNote(target->getLoc()) << "target op";
+      diag.attachNote(loc) << "target op";
       return diag;
     }
 
     if (getDimension() >= linalgOp.getNumLoops()) {
       auto diag = emitSilenceableError() << "dimension " << getDimension()
-                                         << " does not exist in target op";
-      diag.attachNote(target->getLoc()) << "target op";
+                                          << " does not exist in target op";
+      diag.attachNote(loc) << "target op";
       return diag;
     }
+    return DiagnosedSilenceableFailure::success();
+  };
 
-    rewriter.setInsertionPoint(linalgOp);
-    std::tie(first.emplace_back(), second.emplace_back()) = linalg::splitOp(
-        rewriter, cast<TilingInterface>(linalgOp.getOperation()),
-        getDimension(), std::get<1>(pair));
-
-    // Propagate errors.
-    if (!first.back() && !second.back()) {
+  auto checkFailureInSplitting =
+      [&](bool hasFailed, Location loc) -> DiagnosedSilenceableFailure {
+    if (hasFailed) {
       auto diag = emitDefiniteFailure() << "internal failure in splitting";
-      diag.attachNote(target->getLoc()) << "target op";
+      diag.attachNote(loc) << "target op";
       return diag;
     }
+    return DiagnosedSilenceableFailure::success();
+  };
+
+  if (isMultiwaySplit) {
+
+    // Split a single target operation at multiple points.
+    SmallVector<Operation *> opList;
+    TilingInterface head, tail;
+    Operation *target = payload.front();
+
+    LinalgOp linalgOp = dyn_cast<LinalgOp>(target);
+    DiagnosedSilenceableFailure diag =
+        checkStructuredOpAndDimensions(linalgOp, target->getLoc());
+
+    if (diag.isSilenceableFailure())
+      return diag;
+
+    for (auto &&[idx, chunkSize] : llvm::enumerate(chunkSizes)) {
+
+      if (idx > 0)
+        target = tail.getOperation();
+
+      if (!target)
+        break;
 
-    // Do not add null second parts.
-    if (!second.back()) {
-      noSecondPart = target;
-      second.pop_back();
+      linalgOp = cast<LinalgOp>(target);
+
+      rewriter.setInsertionPoint(linalgOp);
+      std::tie(head, tail) = linalg::splitOp(
+          rewriter, cast<TilingInterface>(linalgOp.getOperation()),
+          getDimension(), chunkSize);
+
+      // Propagate errors.
+      DiagnosedSilenceableFailure diag =
+          checkFailureInSplitting(!head && !tail, target->getLoc());
+      if (diag.isDefiniteFailure())
+        return diag;
+
+      opList.push_back(head.getOperation());
     }
-  }
 
-  if (second.size() != first.size() && !second.empty()) {
-    auto diag = emitSilenceableError()
-                << "splitting does not produce the second part for a subset "
-                   "of targets";
-    diag.attachNote() << "expected splitting to produce the second part of all "
-                         "or none of the targets";
-    diag.attachNote(noSecondPart->getLoc())
-        << "first target with no second part";
-    return diag;
-  }
+    // Append any leftover parts to the end of the result list.
+    if (tail)
+      opList.push_back(tail);
+    results.set(cast<OpResult>(getFirst()), opList);
+    results.set(cast<OpResult>(getSecond()), {});
+
+  } else {
+    // Split each target operation.
+    SmallVector<Operation *> first, second;
+    Operation *noSecondPart = nullptr;
+    for (const auto &pair : llvm::zip(payload, chunkSizes)) {
+      Operation *target = std::get<0>(pair);
+      LinalgOp linalgOp = dyn_cast<LinalgOp>(target);
+      DiagnosedSilenceableFailure diag =
+          checkStructuredOpAndDimensions(linalgOp, target->getLoc());
+
+      if (diag.isSilenceableFailure())
+        return diag;
+
+      rewriter.setInsertionPoint(linalgOp);
+      std::tie(first.emplace_back(), second.emplace_back()) = linalg::splitOp(
+          rewriter, cast<TilingInterface>(linalgOp.getOperation()),
+          getDimension(), std::get<1>(pair));
+
+      // Propagate errors.
+      DiagnosedSilenceableFailure diagSplit = checkFailureInSplitting(
+          !first.back() && !second.back(), target->getLoc());
+      if (diagSplit.isDefiniteFailure())
+        return diag;
+
+      // Do not add null second parts.
+      if (!second.back()) {
+        noSecondPart = target;
+        second.pop_back();
+      }
+    }
+
+    if (second.size() != first.size() && !second.empty()) {
+      auto diag = emitSilenceableError()
+                  << "splitting does not produce the second part for a subset "
+                     "of targets";
+      diag.attachNote()
+          << "expected splitting to produce the second part of all "
+             "or none of the targets";
+      diag.attachNote(noSecondPart->getLoc())
+          << "first target with no second part";
+      return diag;
+    }
 
-  results.set(cast<OpResult>(getFirst()), first);
-  results.set(cast<OpResult>(getSecond()), second);
+    results.set(cast<OpResult>(getFirst()), first);
+    results.set(cast<OpResult>(getSecond()), second);
+  }
   return DiagnosedSilenceableFailure::success();
 }
 
 void SplitOp::getEffects(
     SmallVectorImpl<MemoryEffects::EffectInstance> &effects) {
   consumesHandle(getTarget(), effects);
-  if (getDynamicSplitPoint())
-    onlyReadsHandle(getDynamicSplitPoint(), effects);
+  if (getDynamicChunkSizes())
+    onlyReadsHandle(getDynamicChunkSizes(), effects);
   producesHandle(getResults(), effects);
   modifiesPayload(effects);
 }
 
 ParseResult SplitOp::parse(OpAsmParser &parser, OperationState &result) {
-  OpAsmParser::UnresolvedOperand target, dynamicSplitPoint;
-  IntegerAttr staticSplitPoint;
+  OpAsmParser::UnresolvedOperand target, dynamicChunkSizes;
+  IntegerAttr staticChunkSizes;
   if (parser.parseOperand(target) || parser.parseKeyword("after"))
     return failure();
 
   OptionalParseResult dynamicPointParseResult =
-      parser.parseOptionalOperand(dynamicSplitPoint);
+      parser.parseOptionalOperand(dynamicChunkSizes);
   if (!dynamicPointParseResult.has_value()) {
-    int64_t staticSplitPointValue;
-    if (failed(parser.parseInteger(staticSplitPointValue)))
+    int64_t staticChunkSizesValue;
+    if (failed(parser.parseInteger(staticChunkSizesValue)))
       return failure();
 
-    staticSplitPoint =
-        parser.getBuilder().getI64IntegerAttr(staticSplitPointValue);
+    staticChunkSizes =
+        parser.getBuilder().getI64IntegerAttr(staticChunkSizesValue);
   }
 
   Type targetType;
@@ -2392,43 +2473,43 @@ ParseResult SplitOp::parse(OpAsmParser &parser, OperationState &result) {
     return failure();
   }
   if (dynamicPointParseResult.has_value()) {
-    Type splitPointType;
+    Type ChunkSizesType;
     if (failed(*dynamicPointParseResult) || parser.parseComma() ||
-        parser.parseType(splitPointType) ||
-        parser.resolveOperand(dynamicSplitPoint, splitPointType,
+        parser.parseType(ChunkSizesType) ||
+        parser.resolveOperand(dynamicChunkSizes, ChunkSizesType,
                               result.operands)) {
       return failure();
     }
 
-    staticSplitPoint =
+    staticChunkSizes =
         parser.getBuilder().getI64IntegerAttr(ShapedType::kDynamic);
   }
 
   result.addAttribute(
-      SplitOp::getStaticSplitPointAttrName(result.name).getValue(),
-      staticSplitPoint);
+      SplitOp::getStaticChunkSizesAttrName(result.name).getValue(),
+      staticChunkSizes);
   result.addTypes({targetType, targetType});
   return success();
 }
 
 void SplitOp::print(OpAsmPrinter &printer) {
   printer << " " << getTarget() << " after ";
-  int64_t staticSplitSize = static_cast<int64_t>(getStaticSplitPoint());
-  if (staticSplitSize != ShapedType::kDynamic)
-    printer << staticSplitSize;
+  int64_t staticChunkSize = static_cast<int64_t>(getStaticChunkSizes());
+  if (staticChunkSize != ShapedType::kDynamic)
+    printer << staticChunkSize;
   else
-    printer << getDynamicSplitPoint();
+    printer << getDynamicChunkSizes();
   printer << " ";
   printer.printOptionalAttrDict(getOperation()->getAttrs(),
-                                {getStaticSplitPointAttrName()});
+                                {getStaticChunkSizesAttrName()});
   printer << " : " << getTarget().getType();
-  if (staticSplitSize == ShapedType::kDynamic)
-    printer << ", " << getDynamicSplitPoint().getType();
+  if (staticChunkSize == ShapedType::kDynamic)
+    printer << ", " << getDynamicChunkSizes().getType();
 }
 
 LogicalResult SplitOp::verify() {
-  if ((static_cast<int64_t>(getStaticSplitPoint()) != ShapedType::kDynamic) ^
-      (getDynamicSplitPoint() == nullptr)) {
+  if ((static_cast<int64_t>(getStaticChunkSizes()) != ShapedType::kDynamic) ^
+      (getDynamicChunkSizes() == nullptr)) {
     return emitOpError() << "expects either a dynamic or a static split "
                             "point to be provided";
   }