[Mlir-commits] [mlir] [mlir][tensor] Loosen restrictions on folding dynamic reshapes (PR #137963)

[llvmlistbot at llvm.org]

================
@@ -28,67 +32,324 @@ 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 isLastTargetDim = targetDimIdx == numTargetDims - 1;
----------------
MaheshRavishankar wrote:

Nit: Dont see a point of this being a separate variable. Could just be
```
shouldMatchGreedily = targetDimIdx == numTargetDims - 1
```

https://github.com/llvm/llvm-project/pull/137963