[mlir] Reapply "Loosen restrictions on folding dynamic reshapes" (llvm#142827)

The original PR llvm#137963 had a
nvidia bot failure. This appears to be a flaky test because rerunning
the build was successful.

This change needs commit 6f2ba47 to fix incorrect usage of
`getReassociationIndicesForCollapse`.

Reverts llvm#142639

Co-authored-by: Artem Gindinson <[email protected]>

Author: 2 people

mlir/lib/Dialect/Utils/ReshapeOpsUtils.cpp

@@ -10,6 +10,10 @@
 
 #include "mlir/IR/AffineMap.h"
 #include "mlir/IR/Builders.h"
+#include "mlir/IR/BuiltinTypeInterfaces.h"
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/LogicalResult.h"
 
 #include <numeric>
 #include <optional>
@@ -28,67 +32,329 @@ mlir::getReassociationIndicesForReshape(ShapedType sourceType,
   return std::nullopt;
 }
 
-std::optional<SmallVector<ReassociationIndices>>
-mlir::getReassociationIndicesForCollapse(ArrayRef<int64_t> sourceShape,
-                                         ArrayRef<int64_t> targetShape) {
-  if (sourceShape.size() <= targetShape.size())
-    return std::nullopt;
-  unsigned sourceDim = 0;
-  SmallVector<ReassociationIndices> reassociationMap;
-  reassociationMap.reserve(targetShape.size());
+namespace {
+/// A simple struct to represent ReassociationIndices as an inclusive interval.
+/// It's designed to be feasibly minimal, so the call sites should manage the
+/// validity of the range manually.
+struct ReassociationIndexRange {
+  /// FIXME: Signed type is used for consistency with ReassociationIndices.
+  /// We should consider refactoring all reassociation utilities to use unsigned
+  /// types.
+  int64_t leftIdx = 0, rightIdx = 0;
+
+  /// Util for manual checks of the range's validity
+  LogicalResult verify() const {
+    return leftIdx >= 0 && (leftIdx <= rightIdx) ? success() : failure();
+  }
+
+  /// Checks range's containment within another range. Treats the edges
+  /// non-exclusively.
+  bool isInRange(const ReassociationIndexRange &outerRange) const {
+    return leftIdx >= outerRange.leftIdx && rightIdx <= outerRange.rightIdx;
+  }
+
+  unsigned size() const {
+    assert(succeeded(verify()));
+    return rightIdx - leftIdx + 1;
+  }
+  bool containsSingleIndex() const { return size() == 1; }
+
+  /// Collects indices that do not overlap between this and another range.
+  ReassociationIndices
+  getNonOverlappingIndicesWith(ReassociationIndexRange &rhs) const {
+    if (rightIdx < rhs.leftIdx) {
+      // The intervals do not overlap - concatenate the indices from both.
+      auto jointFullIndices = getFullIndices();
+      jointFullIndices.append(rhs.getFullIndices());
+      return jointFullIndices;
+    }
+    ReassociationIndices result;
+    // Handle the chunk left of the overlapping range.
+    int64_t leftStart = std::min(leftIdx, rhs.leftIdx);
+    int64_t leftEnd = std::max(leftIdx, rhs.leftIdx);
+    llvm::append_range(result, llvm::seq(leftStart, leftEnd));
+    // Handle the chunk right of the overlapping range. Symmetrically, we should
+    // skip the edge of the overlap AND include the rightmost index.
+    int64_t rightStart = std::min(rightIdx, rhs.rightIdx) + 1;
+    int64_t rightEnd = std::max(rightIdx, rhs.rightIdx);
+    if (rightStart < rightEnd)
+      llvm::append_range(result, llvm::seq_inclusive(rightStart, rightEnd));
+    return result;
+  }
+
+  /// Converts the range into ReassociationIndices.
+  ReassociationIndices getFullIndices() const {
+    ReassociationIndices result;
+    for (int64_t idx = leftIdx; idx <= rightIdx; ++idx) {
+      result.push_back(idx);
+    }
+    return result;
+  }
+};
+} // namespace
+
+/// Starting from `sourceStartIdx`, searches `sourceShape` for the first
+/// sequence that can be collapsed into a dynamic dimension (at least one must
+/// be present in the source).
+/// By default, lazily returns once the first dynamic dimension has been found.
+/// Setting `matchGreedily` as `true` will also mark all subsequent
+/// source dimensions for collapsing into the target.
+static FailureOr<ReassociationIndexRange>
+findReassociationRangeForDynamicDim(ArrayRef<int64_t> sourceShape,
+                                    int64_t sourceStartIdx,
+                                    bool matchGreedily = false) {
+  const unsigned numSourceDims = sourceShape.size();
+  ReassociationIndexRange sourceShapeAsRange{0, numSourceDims - 1};
+  std::optional<ReassociationIndexRange> resultRange = std::nullopt;
+
+  ReassociationIndexRange iterationRange{sourceStartIdx, sourceStartIdx};
+  for (; iterationRange.isInRange(sourceShapeAsRange);
+       iterationRange.rightIdx++) {
+    int64_t sourceSize = sourceShape[iterationRange.rightIdx];
+    if (sourceSize == ShapedType::kDynamic) {
+      resultRange = iterationRange;
+      break;
+    }
+  }
+  if (!resultRange)
+    return failure();
+  if (matchGreedily)
+    resultRange->rightIdx = sourceShapeAsRange.rightIdx;
+  return *resultRange;
+}
 
-  ReassociationIndices currIndices;
+/// Starting from `sourceStartIdx`, searches `sourceShape` for the first
+/// sequence of static dimensions such that their product matches `targetSize`.
+/// By default, lazily returns once the product matches the target size. Setting
+/// `matchGreedily` as `true` will append all neighboring unit dimensions
+/// (dimensions of 1) to the match.
+static FailureOr<ReassociationIndexRange>
+findReassociationRangeForSize(ArrayRef<int64_t> sourceShape,
+                              int64_t sourceStartIdx, int64_t targetSize,
+                              bool matchGreedily = false) {
+  const unsigned numSourceDims = sourceShape.size();
+  ReassociationIndexRange sourceShapeAsRange{0, numSourceDims - 1};
+  std::optional<ReassociationIndexRange> resultRange = std::nullopt;
+
+  ReassociationIndexRange iterationRange{sourceStartIdx, sourceStartIdx};
   int64_t prodOfCollapsedDims = 1;
-  while (sourceDim < sourceShape.size()) {
-    unsigned targetDim = reassociationMap.size();
-    // If we have mapped all the target dimensions stop and handle the remaining
-    // tail of size-1 dimensions explicitly.
-    if (targetDim == targetShape.size())
+  while (iterationRange.isInRange(sourceShapeAsRange)) {
+    int64_t sourceSize = sourceShape[iterationRange.rightIdx];
+    if (sourceSize == ShapedType::kDynamic) {
+      // Reassociation for a static dim cannot include a dynamic dim. Reset
+      // induction variables to essentially restart the loop from the next
+      // source dimension.
+      prodOfCollapsedDims = 1;
+      iterationRange = {iterationRange.rightIdx + 1,
+                        iterationRange.rightIdx + 1};
+      continue;
+    }
+    prodOfCollapsedDims *= sourceSize;
+    // If the target size has been exceeded without matching, we need to shift
+    // the range start right. From the start of the range, roll back the
+    // multiplication until the target size exceeds the product again.
+    while (prodOfCollapsedDims > targetSize &&
+           !iterationRange.containsSingleIndex()) {
+      int64_t frontSourceSize = sourceShape[iterationRange.leftIdx];
+      prodOfCollapsedDims /= frontSourceSize;
+      // Shrink the range rightwards
+      iterationRange.leftIdx++;
+    }
+    // We could've reached the target size with the current dimension,
+    // also as a result of the above shift to right.
+    if (prodOfCollapsedDims == targetSize) {
+      resultRange = iterationRange;
       break;
+    }
+    // Increment the iteration range
+    iterationRange.rightIdx++;
+  }
+  if (!resultRange)
+    return failure();
+  if (matchGreedily) {
+    // We now want to collect all unit dimensions directly after the target
+    // product match. Advance the iterator to avoid OOB when the product match
+    // happens at the last element.
+    iterationRange.rightIdx++;
+    while (iterationRange.isInRange(sourceShapeAsRange) &&
+           sourceShape[iterationRange.rightIdx] == 1) {
+      resultRange = iterationRange;
+      iterationRange.rightIdx++;
+    }
+  }
+  return *resultRange;
+}
 
-    int64_t currTargetShape = targetShape[targetDim];
-    while (sourceDim < (sourceShape.size() - 1) &&
-           sourceShape[sourceDim] != ShapedType::kDynamic &&
-           prodOfCollapsedDims * sourceShape[sourceDim] < currTargetShape) {
-      prodOfCollapsedDims *= sourceShape[sourceDim];
-      currIndices.push_back(sourceDim++);
+/// Attempts to find a valid collapsing reassociation of `sourceShape` into
+/// `targetShape` through a simple traversal. If successful, an array of source
+/// index ranges is returned, correspondingly to each dimension in the target
+/// shape. The resulting indices shall fully cover the `sourceShape` without
+/// overlaps.
+///
+/// The algorithm is essentially a lazy one, searching for non-greedy matches -
+/// it will only yield a greedy match for the last target dimension.
+/// FIXME: The algorithm can only backtrack when it needs to append an offset
+/// for a static target dimension to the preceding dynamic one (this retains the
+/// linear complexity). As feasible, consider adding further backtracking
+/// routines to enable more reassociations, e.g.:
+/// - ?x2x?x2 into ?x2
+static FailureOr<SmallVector<ReassociationIndexRange>>
+findReassociationRangesForCollapse(ArrayRef<int64_t> sourceShape,
+                                   ArrayRef<int64_t> targetShape) {
+  unsigned numSourceDims = sourceShape.size(),
+           numTargetDims = targetShape.size();
+  assert(numSourceDims > numTargetDims);
+  ReassociationIndexRange sourceShapeAsRange{0, numSourceDims - 1};
+
+  SmallVector<ReassociationIndexRange> reassocRanges;
+  reassocRanges.reserve(numTargetDims);
+  // We'll iterate in strides of 2 to enable pseudo-backtracking for simple
+  // cases, e.g.:
+  // - ?x2x3x5 into ?x15
+  std::optional<int64_t> prevTargetSize = std::nullopt;
+  for (unsigned targetDimIdx = 0, sourceDimIdx = 0;
+       targetDimIdx < numTargetDims; ++targetDimIdx) {
+    int64_t targetSize = targetShape[targetDimIdx];
+    // Simply check if there are any subsequent target dimensions left - if not,
+    // the match must be made greedily.
+    bool shouldMatchGreedily = targetDimIdx == numTargetDims - 1;
+    FailureOr<ReassociationIndexRange> sourceRange;
+    if (targetSize == ShapedType::kDynamic) {
+      sourceRange = findReassociationRangeForDynamicDim(
+          sourceShape, sourceDimIdx, shouldMatchGreedily);
+    } else {
+      sourceRange = findReassociationRangeForSize(
+          sourceShape, sourceDimIdx, targetSize, shouldMatchGreedily);
     }
 
-    // If the current expanded dimension is dynamic, then the collapsed
-    // dimensions should also be dynamic and product of all previous unprocessed
-    // dimensions of the expanded shape should be 1.
-    if (sourceShape[sourceDim] == ShapedType::kDynamic &&
-        (currTargetShape != ShapedType::kDynamic || prodOfCollapsedDims != 1))
-      return std::nullopt;
-
-    // If the collapsed dim is dynamic, the current expanded dim should also
-    // be dynamic.
-    if (currTargetShape == ShapedType::kDynamic &&
-        sourceShape[sourceDim] != ShapedType::kDynamic)
-      return std::nullopt;
-
-    // For static shapes, if the product of dimensions of the expanded shape
-    // should match the collapsed dimension shape.
-    if (prodOfCollapsedDims * sourceShape[sourceDim] != currTargetShape)
-      return std::nullopt;
-
-    currIndices.push_back(sourceDim++);
-    reassociationMap.emplace_back(ReassociationIndices{});
-    std::swap(reassociationMap.back(), currIndices);
-    prodOfCollapsedDims = 1;
+    // Run sanity checks on the returned index range.
+    if (failed(sourceRange) || failed(sourceRange->verify()) ||
+        !sourceRange->isInRange(sourceShapeAsRange))
+      return failure();
+    if (sourceRange->leftIdx > sourceDimIdx) {
+      // If some source dimensions had to be skipped in order to find a match,
+      // they must be collapsed into the directly preceding dynamic dimension.
+      if (!prevTargetSize || prevTargetSize != ShapedType::kDynamic)
+        return failure();
+      reassocRanges.back().rightIdx = sourceRange->leftIdx - 1;
+    }
+
+    // Store the gathered information as required for the next iteration.
+    prevTargetSize = targetSize;
+    sourceDimIdx = sourceRange->rightIdx + 1;
+    reassocRanges.push_back(*sourceRange);
   }
-  // All the dimensions in the target must have been processed.
-  if (reassociationMap.size() != targetShape.size())
+  // Fail if the source shape wasn't a full match for the target shape. We only
+  // need to check the last recorded index - any other gaps should have been
+  // mended by the main loop.
+  if (reassocRanges.back().rightIdx < sourceShapeAsRange.rightIdx)
+    return failure();
+  return reassocRanges;
+}
+
+/// A variant of `findReassociationRangesForCollapse(...)` that can also scan
+/// the shapes right-to-left.
+static FailureOr<SmallVector<ReassociationIndexRange>>
+findReassociationRangesForCollapse(ArrayRef<int64_t> sourceShape,
+                                   ArrayRef<int64_t> targetShape,
+                                   bool iterateRightToLeft) {
+  if (!iterateRightToLeft)
+    return findReassociationRangesForCollapse(sourceShape, targetShape);
+  // NB: To iterate right-to-left, we currently reverse the shapes and then
+  // reverse the result back. The reversed shapes must not be temporary, as
+  // we're passing through an ArrayRef.
+  // FIXME: It would be preferable to avoid the expensive copies. At the moment,
+  // this approach is chosen for readability of the main implementation.
+  std::vector<int64_t> sourceToReverse = sourceShape.vec(),
+                       targetToReverse = targetShape.vec();
+  std::reverse(sourceToReverse.begin(), sourceToReverse.end());
+  std::reverse(targetToReverse.begin(), targetToReverse.end());
+  auto invertedRanges =
+      findReassociationRangesForCollapse(sourceToReverse, targetToReverse);
+  if (failed(invertedRanges))
+    return failure();
+  SmallVector<ReassociationIndexRange> &rangesToInvert = *invertedRanges;
+  unsigned numSourceDims = sourceShape.size();
+  // We have received the ranges for inverted shapes. Now we have to invert
+  // the ranges back to correspond with the original source shape.
+  for (auto &range : rangesToInvert) {
+    int64_t invLeftIdx = range.leftIdx, invRightIdx = range.rightIdx;
+    range.leftIdx = numSourceDims - 1 - invRightIdx;
+    range.rightIdx = numSourceDims - 1 - invLeftIdx;
+  }
+  // Also invert the ordering of the ranges to correspond with the original
+  // target shape.
+  std::reverse(rangesToInvert.begin(), rangesToInvert.end());
+  return rangesToInvert;
+}
+
+std::optional<SmallVector<ReassociationIndices>>
+mlir::getReassociationIndicesForCollapse(ArrayRef<int64_t> sourceShape,
+                                         ArrayRef<int64_t> targetShape) {
+  unsigned numSourceDims = sourceShape.size(),
+           numTargetDims = targetShape.size();
+  // We're supposed to search for a collapsing reassociation. If the sizes
+  // match, there's no actual collapsing taking place - it's either a no-op or a
+  // `tensor.reshape`-style reassociation (that would be beyond the scope of
+  // this utility).
+  if (numSourceDims <= numTargetDims)
+    return std::nullopt;
+  // Early handling for scalar target types.
+  if (numTargetDims == 0) {
+    ReassociationIndices allSourceIndices;
+    allSourceIndices.reserve(numSourceDims);
+    for (unsigned sourceDimIdx = 0; sourceDimIdx < numSourceDims;
+         ++sourceDimIdx) {
+      int64_t sourceSize = sourceShape[sourceDimIdx];
+      // All source dimensions must be unit or dynamic.
+      if (sourceSize != 1 && sourceSize != ShapedType::kDynamic)
+        return std::nullopt;
+      allSourceIndices.push_back(sourceDimIdx);
+    }
+    return SmallVector<ReassociationIndices>{allSourceIndices};
+  }
+
+  // Collect source ranges by iterating over the target shape left-to-right.
+  FailureOr<SmallVector<ReassociationIndexRange>> maybeForwardRanges =
+      findReassociationRangesForCollapse(sourceShape, targetShape);
+  if (failed(maybeForwardRanges))
+    return std::nullopt;
+  auto &ranges = *maybeForwardRanges;
+  // Now do the same in reverse. We need to get another valid reassociation
+  // through some other strategy, and then compare the results in order to
+  // disambiguate mixed subshapes, such as:
+  // ?x?x? into ?x?, ?x2x? into ?x?, ?x2x3x6x? into ?x6x?
+  // This leads us to lose some of the reassociation opportunities that can only
+  // be found by iterating in a certain direction, e.g. 2x2x? into 2x? - without
+  // backtracking, the algorithm will fail right-to-left. However, this is the
+  // best way to preserve correctness.
+  FailureOr<SmallVector<ReassociationIndexRange>> maybeReverseRanges =
+      findReassociationRangesForCollapse(sourceShape, targetShape,
+                                         /*iterateRightToLeft=*/true);
+  if (failed(maybeReverseRanges))
+    return std::nullopt;
+  auto &reverseRanges = *maybeReverseRanges;
+
+  if (ranges.size() != numTargetDims || reverseRanges.size() != numTargetDims)
     return std::nullopt;
-  // Process any remaining entries in the source shape. They all need to be
-  // 1 or dynamic.
-  for (; sourceDim < sourceShape.size(); sourceDim++) {
-    if (sourceShape[sourceDim] != ShapedType::kDynamic &&
-        sourceShape[sourceDim] != 1)
-      return std::nullopt;
-    // The map is empty when the target type is a scalar.
-    if (!reassociationMap.empty())
-      reassociationMap.back().push_back(sourceDim);
+  // Now we can check for ambiguity of each target dimension's reassociation. If
+  // successful, we put the full indices into our result map for the target
+  // shape.
+  SmallVector<ReassociationIndices> reassociationMap(numTargetDims);
+  for (unsigned targetDimIdx = 0; targetDimIdx < numTargetDims;
+       ++targetDimIdx) {
+    ReassociationIndexRange &range = ranges[targetDimIdx];
+    ReassociationIndexRange &reverseRange = reverseRanges[targetDimIdx];
+    // Get non-overlapping indices between the ranges
+    ReassociationIndices nonMatchingIndices =
+        range.getNonOverlappingIndicesWith(reverseRange);
+    // Unit dimensions can be collapsed wherever - this is the only ambiguity
+    // that we allow.
+    for (int64_t sourceDimIdx : nonMatchingIndices) {
+      if (sourceShape[sourceDimIdx] != 1)
+        return std::nullopt;
+    }
+    reassociationMap[targetDimIdx] = range.getFullIndices();
   }
   return reassociationMap;
 }