llvm-external-projects/iree-dialects/lib/Dialect/LinalgExt/Passes/SplitReduction.cpp

// Copyright 2022 The IREE Authors
//
// Licensed under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception

#include "iree-dialects/Dialect/LinalgExt/IR/LinalgExtDialect.h"
#include "iree-dialects/Dialect/LinalgExt/IR/LinalgExtOps.h"
#include "iree-dialects/Dialect/LinalgExt/Passes/PassDetail.h"
#include "iree-dialects/Dialect/LinalgExt/Passes/Passes.h"
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Dialect/Math/IR/Math.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/Operation.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
#include "llvm/ADT/STLExtras.h"

using namespace mlir;
using namespace mlir::iree_compiler::IREE::LinalgExt;

namespace {

SmallVector<int64_t> getExpandedShape(ArrayRef<int64_t> shape,
                                      int64_t splitReductionRatio,
                                      int64_t splitDimParallel) {
  SmallVector<int64_t> ans;
  ans.reserve(shape.size() + 1);
  ans.assign(shape.begin(), shape.end());
  ans[splitDimParallel] = splitReductionRatio;
  ans.insert(std::next(ans.begin(), splitDimParallel + 1),
             shape[splitDimParallel] / splitReductionRatio);

  return ans;
}

SmallVector<int64_t> getCollapsedShape(ArrayRef<int64_t> shape,
                                       int64_t splitReductionRatio, int64_t k,
                                       int64_t targetDim) {
  SmallVector<int64_t> ans(shape.begin(), shape.end());
  ans[targetDim] = k * splitReductionRatio;
  return ans;
}

SmallVector<ReassociationIndices>
getReassociationIndices(int64_t rank, int64_t splitDimParallel) {
  SmallVector<ReassociationIndices> reassociationIndices;
  for (int i = 0; i < rank; ++i) {
    if (i < splitDimParallel) {
      reassociationIndices.push_back({i});
    } else if (i == splitDimParallel) {
      reassociationIndices.push_back({i, i + 1});
    } else if (i > splitDimParallel) {
      reassociationIndices.push_back({i + 1});
    }
  }
  return reassociationIndices;
}

LogicalResult shouldParallelTopk(iree_compiler::IREE::LinalgExt::TopkOp topkOp,
                                 PatternRewriter &rewriter, int64_t kDimOrig,
                                 int64_t splitReductionRatio,
                                 int64_t splitReductionDepth) {
  // Determine if we should split the reduction. Requires aligned static shapes
  // and no input indicies.
  auto valuesOrigType = topkOp.getInputType();
  if (valuesOrigType.isDynamicDim(kDimOrig)) {
    return rewriter.notifyMatchFailure(topkOp,
                                       "cannot split dynamic dimension");
  }
  if (topkOp.indices() && splitReductionDepth == 0) {
    return rewriter.notifyMatchFailure(
        topkOp, "input indices aren't supported for first split");
  }
  if (splitReductionRatio <= 1) {
    return rewriter.notifyMatchFailure(topkOp, "reduction ratio <= 1");
  }
  if (valuesOrigType.getDimSize(kDimOrig) % splitReductionRatio != 0) {
    return rewriter.notifyMatchFailure(
        topkOp,
        "reduction dimension must be perfectly aligned to (divisible by) the "
        "split ratio");
  }
  return success();
}

// Creates the first phase of the topk split reduction by reshaping the input
// into parallel computations then feeding them into a topk op.
iree_compiler::IREE::LinalgExt::TopkOp
computeParallelTopk(Location loc, PatternRewriter &rewriter,
                    iree_compiler::IREE::LinalgExt::TopkOp topkOp,
                    ArrayRef<ReassociationIndices> reassociationIndices,
                    int64_t splitReductionRatio, int64_t splitDimParallel,
                    int64_t kDimParallel, int64_t kSize) {
  Value valuesOrig = topkOp.values();
  auto valuesOrigType = valuesOrig.getType().cast<ShapedType>();
  Type valueElementType = valuesOrigType.getElementType();
  Type indicesElementType =
      topkOp.getResultTypes()[1].cast<ShapedType>().getElementType();

  SmallVector<int64_t> expandedShape = getExpandedShape(
      valuesOrigType.getShape(), splitReductionRatio, splitDimParallel);
  auto valuesExpandedType =
      RankedTensorType::get(expandedShape, valueElementType);

  // Expand input values shape for parallel processing
  Value valuesExpanded = rewriter.create<tensor::ExpandShapeOp>(
      loc, valuesExpandedType, valuesOrig, reassociationIndices);

  // Expand input indices shape for parallel processing if they exist
  Optional<Value> indicesExpanded;
  if (Optional<Value> inputIndices = topkOp.indices()) {
    // Type inputElementType = inputIndices->getType().cast<ShapedType>();
    Type indicesExpandedType =
        RankedTensorType::get(expandedShape, indicesElementType);
    indicesExpanded = rewriter.create<tensor::ExpandShapeOp>(
        loc, indicesExpandedType, inputIndices.value(), reassociationIndices);
  }

  // Define the expanded output types
  SmallVector<int64_t> expandedResultShape = expandedShape;
  expandedResultShape[kDimParallel] = kSize;
  auto outputValuesExpandedType =
      RankedTensorType::get(expandedResultShape, valueElementType);
  auto outputIndicesExpandedType =
      RankedTensorType::get(expandedResultShape, indicesElementType);

  // Initialize the expanded output values
  SmallVector<Value> dynSizes;
  for (auto i : llvm::seq<int64_t>(0, valuesExpandedType.getRank())) {
    if (valuesExpandedType.isDynamicDim(i)) {
      dynSizes.push_back(
          rewriter.create<tensor::DimOp>(loc, valuesExpanded, i));
    }
  }
  Value emptyTensorOutputValues = rewriter.create<mlir::tensor::EmptyOp>(
      loc, outputValuesExpandedType.getShape(), valueElementType, dynSizes);
  Value emptyTensorOutputIndices = rewriter.create<mlir::tensor::EmptyOp>(
      loc, outputIndicesExpandedType.getShape(), indicesElementType, dynSizes);

  // Initialize indices to positive infinity and values to negative infinity
  // for a top (maxk) comparison.
  Attribute negInfAttr;
  if (auto intType = valueElementType.dyn_cast<IntegerType>()) {
    negInfAttr = rewriter.getIntegerAttr(
        intType, APInt::getSignedMinValue(intType.getWidth()));
  } else {
    auto negApFloat =
        APFloat::getInf(valueElementType.cast<FloatType>().getFloatSemantics(),
                        /*Negative=*/true);
    negInfAttr = rewriter.getFloatAttr(valueElementType, negApFloat);
  }
  Value negInf = rewriter.create<arith::ConstantOp>(loc, negInfAttr);
  Attribute posInfAttr =
      rewriter.getIntegerAttr(indicesElementType, APInt::getSignedMaxValue(32));
  Value posInf = rewriter.create<arith::ConstantOp>(loc, posInfAttr);
  Value negInfTensor =
      rewriter.create<linalg::FillOp>(loc, negInf, emptyTensorOutputValues)
          .result();
  Value posInfTensor =
      rewriter.create<linalg::FillOp>(loc, posInf, emptyTensorOutputIndices)
          .result();

  SmallVector<Type> parallelTopkResultTypes = {outputValuesExpandedType,
                                               outputIndicesExpandedType};
  SmallVector<Value> parallelTopkIns = {valuesExpanded};
  if (indicesExpanded) {
    parallelTopkIns.push_back(indicesExpanded.value());
  }
  SmallVector<Value> parallelTopkOuts = {negInfTensor, posInfTensor};

  // Parallel topk
  auto parallelTopkOp = rewriter.create<iree_compiler::IREE::LinalgExt::TopkOp>(
      loc,
      /*resultTypes=*/
      parallelTopkResultTypes,
      /*ins=*/parallelTopkIns,
      /*outs=*/parallelTopkOuts, kDimParallel);
  rewriter.cloneRegionBefore(topkOp.getRegion(), parallelTopkOp.getRegion(),
                             parallelTopkOp.getRegion().end());

  return parallelTopkOp;
}

// Update the output indices from the parallel TopK with the correct offsets.
// Each parallel computation uses implicit indices (starting from 0) during
// selection, but the values are part of the large input space split into M =
// splitReductionFn() ways. The following linalg.generic adds the appropriate
// offset to reflect to values original position. "Updated pos" = "initial
// pos" + "splitDimParallel size * "splitDimParallel index"
Value offsetParallelIndices(Location loc, PatternRewriter &rewriter,
                            Value parallelIndices, int64_t kDimParallelSize,
                            int64_t splitDimParallel) {
  auto parallelIndicesType = parallelIndices.getType().cast<ShapedType>();
  size_t parallelIndicesRank = parallelIndicesType.getRank();
  AffineMap mapIdentity = rewriter.getMultiDimIdentityMap(parallelIndicesRank);
  SmallVector<AffineMap> indexingMaps = {mapIdentity};
  SmallVector<utils::IteratorType> iterators(parallelIndicesRank,
                                             utils::IteratorType::parallel);
  Value mSplitVal = rewriter.create<arith::ConstantIntOp>(
      loc, kDimParallelSize, parallelIndicesType.getElementType());
  return rewriter
      .create<linalg::GenericOp>(
          loc,
          /*resultType=*/parallelIndicesType,
          /*inputs=*/ValueRange{},
          /*outputs=*/ValueRange{parallelIndices}, indexingMaps, iterators,
          [&](OpBuilder &b, Location loc, ValueRange args) {
            Value splitIndex = b.create<linalg::IndexOp>(loc, splitDimParallel);
            Value splitIndexInt = b.create<arith::IndexCastOp>(
                loc, parallelIndicesType.getElementType(), splitIndex);
            Value mOffset =
                b.create<arith::MulIOp>(loc, mSplitVal, splitIndexInt);
            Value updatedParallelIndex =
                b.create<arith::AddIOp>(loc, mOffset, args[0]);
            b.create<linalg::YieldOp>(loc, updatedParallelIndex);
          })
      .getResult(0);
}

// Creates the second phase of the topk split reduction by collapsing output
// from parallel topk and computing the final combined result.
TopkOp computeReductionTopk(Location loc, PatternRewriter &rewriter,
                            TopkOp topkOp, TopkOp parallelTopkOp,
                            Value updatedParallelIndices,
                            ArrayRef<ReassociationIndices> reassociationIndices,
                            int64_t splitReductionRatio, int64_t kDimOrig,
                            int64_t kSize) {
  Value valuesOrig = topkOp.values();
  auto valuesOrigType = valuesOrig.getType().cast<ShapedType>();
  Type valueElementType = valuesOrigType.getElementType();
  Type indicesElementType =
      topkOp.getResultTypes()[1].cast<ShapedType>().getElementType();

  // Define the collapsed input shapes
  SmallVector<int64_t> collapsedShape = getCollapsedShape(
      valuesOrigType.getShape(), splitReductionRatio, kSize, kDimOrig);
  auto valuesCollapsedType =
      RankedTensorType::get(collapsedShape, valueElementType);
  auto indicesCollapsedType =
      RankedTensorType::get(collapsedShape, indicesElementType);

  // Collapse collapse parallel output for the input of final reduction
  Value valuesCollapsed = rewriter.create<tensor::CollapseShapeOp>(
      loc, valuesCollapsedType, parallelTopkOp.getResults()[0],
      reassociationIndices);
  Value indicesCollapsed = rewriter.create<tensor::CollapseShapeOp>(
      loc, indicesCollapsedType, updatedParallelIndices, reassociationIndices);

  // Combined final topk
  auto reductionTopkOp =
      rewriter.create<iree_compiler::IREE::LinalgExt::TopkOp>(
          loc,
          /*resultTypes=*/topkOp->getResultTypes(),
          /*ins=*/ValueRange{valuesCollapsed, indicesCollapsed},
          /*outs=*/topkOp.getOutputs(), kDimOrig);
  rewriter.cloneRegionBefore(topkOp.getRegion(), reductionTopkOp.getRegion(),
                             reductionTopkOp.getRegion().end());
  return reductionTopkOp;
}

int64_t getSplitReductionDepth(TopkOp topkOp) {
  auto attr =
      topkOp->template getAttrOfType<IntegerAttr>(kSplitReductionDepthMarker);
  if (attr) {
    return attr.getInt();
  } else {
    return 0;
  }
}

void setSplitReductionDepth(TopkOp topkOp, PatternRewriter &rewriter,
                            int64_t depth) {
  topkOp->setAttr(kSplitReductionDepthMarker,
                  rewriter.getI64IntegerAttr(depth));
}

struct TopkOpSplitReduction : public OpRewritePattern<TopkOp> {
  using OpRewritePattern::OpRewritePattern;

  TopkOpSplitReduction(MLIRContext *context, TopkSplitReductionControlFn fn,
                       LinalgTransformationFilter filt)
      : OpRewritePattern<TopkOp>(context), splitReductionFn(std::move(fn)),
        filter(std::move(filt)) {}

  // Transforms an applicable standard single reduction TopkOp into a parallel
  // reduction TopkOp with a reduce step following.
  //
  // Handles parallel reductions in 2 phases: A "map" parallel phase and the a
  // single "reduce" reduction phase. The first phase expands the input tensor
  // shape by breaking the reduction dimension into multiple parallel reductions
  // (upping the rank of the input). Topk is run on these dimensions in parallel
  // The second phase collapses the parallel results into a single final reduce.
  // Topk is run again on the combined output to produce a final output.
  //
  // Currently only topk operations without input indices are supported.
  LogicalResult matchAndRewrite(TopkOp topkOp,
                                PatternRewriter &rewriter) const override {
    if (failed(filter.checkAndNotify(rewriter, topkOp))) {
      return rewriter.notifyMatchFailure(topkOp, "preconditions not met");
    }
    Location loc = topkOp.getLoc();
    // Original reduction dimension used for the final combined reduction
    int64_t kDimOrig = topkOp.getDimension();
    // For parallel topk: the dimension that we compute parallel reductions
    int64_t splitDimParallel = kDimOrig;
    // For parallel topk: the dimension that we reduce
    int64_t kDimParallel = kDimOrig + 1;
    int64_t kSize =
        topkOp.getResult(0).getType().cast<ShapedType>().getDimSize(kDimOrig);
    int64_t splitReductionDepth = getSplitReductionDepth(topkOp);
    int64_t splitReductionRatio = splitReductionFn(splitReductionDepth);
    SmallVector<ReassociationIndices> reassociationIndices =
        getReassociationIndices(topkOp.getInputRank(), splitDimParallel);

    // Determine if should compute parallel topk
    LogicalResult shouldParallelTopkResult = shouldParallelTopk(
        topkOp, rewriter, kDimOrig, splitReductionRatio, splitReductionDepth);
    if (shouldParallelTopkResult.failed()) {
      return shouldParallelTopkResult;
    }

    // Topk parallel reduction
    TopkOp parallelTopkOp = computeParallelTopk(
        loc, rewriter, topkOp, reassociationIndices, splitReductionRatio,
        splitDimParallel, kDimParallel, kSize);

    // Update parallel indices to correct offsets if input indices weren't
    // provided. If input indices were provided, no offsetting is needed as
    // original original indices are already known.
    Value updatedParallelIndices = parallelTopkOp.getResult(1);
    if (!topkOp.indices()) {
      Value parallelIndices = parallelTopkOp.getResult(1);
      SmallVector<int64_t> expandedShape = getExpandedShape(
          topkOp.values().getType().cast<ShapedType>().getShape(),
          splitReductionRatio, splitDimParallel);
      int64_t kDimParallelSize = expandedShape[kDimParallel];
      updatedParallelIndices = offsetParallelIndices(
          loc, rewriter, parallelIndices, kDimParallelSize, splitDimParallel);
    }

    // Topk final reduction
    TopkOp reductionTopkOp = computeReductionTopk(
        loc, rewriter, topkOp, parallelTopkOp, updatedParallelIndices,
        reassociationIndices, splitReductionRatio, kDimOrig, kSize);

    // Replace and update result
    rewriter.replaceOp(topkOp, reductionTopkOp.getResults());
    filter.replaceLinalgTransformationFilter(rewriter, parallelTopkOp);
    setSplitReductionDepth(reductionTopkOp, rewriter, splitReductionDepth + 1);
    return success();
  }

private:
  TopkSplitReductionControlFn splitReductionFn;
  LinalgTransformationFilter filter;
};

} // namespace

//===----------------------------------------------------------------------===//
// Pass
//===----------------------------------------------------------------------===//

namespace {
struct TopkSplitReductionPass
    : public TopkSplitReductionBase<TopkSplitReductionPass> {
  void getDependentDialects(DialectRegistry &registry) const override {
    registry.insert<linalg::LinalgDialect, func::FuncDialect,
                    mlir::arith::ArithDialect, math::MathDialect,
                    memref::MemRefDialect, scf::SCFDialect>();
  }

  void runOnOperation() override {
    if (splitRatios.empty()) {
      return;
    }
    RewritePatternSet patterns(&getContext());
    TopkSplitReductionControlFn splitReductionFn =
        [&](int64_t splitReductionDepth) -> int64_t {
      SmallVector<int64_t, 4> reductionRatios(splitRatios.begin(),
                                              splitRatios.end());
      if (splitReductionDepth >= reductionRatios.size()) {
        return -1;
      } else {
        return reductionRatios[splitReductionDepth];
      }
    };

    patterns.add<TopkOpSplitReduction>(
        patterns.getContext(), splitReductionFn,
        LinalgTransformationFilter(
            ArrayRef<StringAttr>{},
            StringAttr::get(patterns.getContext(), "SPLIT_REDUCTION")));
    if (failed(applyPatternsAndFoldGreedily(getOperation(),
                                            std::move(patterns)))) {
      return signalPassFailure();
    }

    // Remove all the markers at the end.
    auto funcOp = getOperation();
    funcOp->walk([&](TopkOp op) {
      op->removeAttr(LinalgTransforms::kLinalgTransformMarker);
      op->removeAttr(kSplitReductionDepthMarker);
    });
  }
};
} // namespace

void mlir::iree_compiler::IREE::LinalgExt::populateTopkSplitReductionPattern(
    RewritePatternSet &patterns,
    const TopkSplitReductionControlFn &splitReductionFn,
    const LinalgTransformationFilter &f) {
  patterns.add<TopkOpSplitReduction>(patterns.getContext(), splitReductionFn,
                                     f);
}

std::unique_ptr<OperationPass<func::FuncOp>>
mlir::iree_compiler::IREE::LinalgExt::createTopkSplitReductionPass() {
  return std::make_unique<TopkSplitReductionPass>();
}