mlir-hlo/lib/Dialect/mhlo/transforms/rank_specialization.cc

/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

==============================================================================*/

#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "mlir-hlo/Dialect/mhlo/IR/chlo_ops.h"
#include "mlir-hlo/Dialect/mhlo/IR/hlo_ops.h"
#include "mlir-hlo/Dialect/mhlo/transforms/passes.h"
#include "mlir-hlo/Dialect/mhlo/transforms/rewriters.h"
#include "mlir/Dialect/StandardOps/IR/Ops.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/IR/BlockAndValueMapping.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/Operation.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/Interfaces/InferTypeOpInterface.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"

namespace mlir {

// Needed to build `llvm::SmallSet`s of `mlir::Value`s.
static bool operator<(const Value &lhs, const Value &rhs) {
  return lhs.getAsOpaquePointer() < rhs.getAsOpaquePointer();
}

namespace mhlo {
namespace {

/// Identify clusters of operations that can be rank-specialized together. The
/// required traits for clustered operations are:
///   - Element-wise: All operations in the group must be element-wise. This
///     allows to reshape operands before applying the operations as well as
///     reshaping the result to the desired shape afterwards. This way, we can,
///     e.g., apply unary ops to a completely flattened operand and restore the
///     original shape afterwards.
///   - Broadcasting semantics: All operations must implement broadcasting
///     semantics. Most importantly, this allows extending operand shapes such
///     that they match in rank.
///   - Shape reification: All operations must implement
///     `InferShapedTypeOpInterface`. This is later needed to compute and to
///     restore the desired result shape.

bool IsClusterable(Operation *op) {
  if (!llvm::isa<InferShapedTypeOpInterface>(op)) return false;
  unsigned int num_operands = op->getNumOperands();
  if (num_operands == 0) return false;
  if (num_operands == 1) return op->hasTrait<OpTrait::Elementwise>();
  return op->hasTrait<chlo::OpTrait::BroadcastingElementwise>() &&
         op->hasTrait<chlo::OpTrait::Broadcasting>();
}

struct RankSpecializationClusterPattern : public RewritePattern {
  explicit RankSpecializationClusterPattern(MLIRContext *ctx)
      : RewritePattern(MatchAnyOpTypeTag(), /*benefit=*/1, ctx) {}

  LogicalResult matchAndRewrite(Operation *root_op,
                                PatternRewriter &rewriter) const override {
    // Only apply to operations that have not been clustered yet.
    if (root_op->getParentOfType<chlo::RankSpecializationClusterOp>()) {
      return failure();
    }

    // Only cluster when rank specialization is needed.
    if (!IsClusterable(root_op) ||
        !llvm::any_of(root_op->getOperandTypes(),
                      [](Type ty) { return ty.isa<UnrankedTensorType>(); })) {
      return failure();
    }

    // Collect all collectively rank specializable ops.
    SmallVector<Operation *, 16> cluster;
    llvm::SmallSet<Value, 16> operand_set;
    llvm::SmallSet<Value, 16> result_set;
    Operation *new_op = root_op;
    while (new_op != nullptr && IsClusterable(new_op)) {
      // Find results that escape the cluster.
      for (OpOperand &use : new_op->getUses()) {
        if (!llvm::is_contained(cluster, use.getOwner()))
          result_set.insert(use.get());
      }

      // Update cluster operands.
      for (OpResult v : new_op->getResults()) operand_set.erase(Value(v));
      for (OpOperand &v : new_op->getOpOperands()) operand_set.insert(v.get());

      cluster.push_back(new_op);
      new_op = new_op->getPrevNode();
    }

    // Create `RankSpecializationClusterOp`.
    auto operands = llvm::to_vector<16>(operand_set);
    auto results = llvm::to_vector<16>(result_set);
    auto result_types = llvm::to_vector<16>(
        llvm::map_range(result_set, [](Value v) { return v.getType(); }));
    Location loc = root_op->getLoc();
    auto cluster_op = rewriter.create<chlo::RankSpecializationClusterOp>(
        loc, result_types, operands);

    // Create body block.
    auto operand_types = llvm::to_vector<16>(
        llvm::map_range(operand_set, [](Value v) { return v.getType(); }));
    Block *block = rewriter.createBlock(&cluster_op.body(), {}, operand_types);

    // Copy operations into the body.
    BlockAndValueMapping bvm;
    for (auto it : llvm::zip(operands, block->getArguments()))
      bvm.map(std::get<0>(it), std::get<1>(it));
    rewriter.setInsertionPointToStart(block);
    for (Operation *it : llvm::reverse(cluster)) rewriter.clone(*it, bvm);

    // Create `RankSpecializationClusterYieldOp`.
    auto mapped_results = llvm::to_vector<16>(
        llvm::map_range(results, [&](Value v) { return bvm.lookup(v); }));
    rewriter.create<chlo::RankSpecializationClusterYieldOp>(loc,
                                                            mapped_results);

    // Replace original ops with the new results.
    for (auto it : llvm::zip(results, cluster_op.results()))
      bvm.map(std::get<0>(it), std::get<1>(it));
    for (Operation *it : cluster) {
      if (it->getUses().empty()) {
        rewriter.eraseOp(it);
        continue;
      }
      auto replacements = llvm::to_vector<16>(llvm::map_range(
          it->getResults(), [&](Value v) { return bvm.lookup(v); }));
      rewriter.replaceOp(root_op, replacements);
    }

    return success();
  }
};

struct RankSpecializationClusterPass
    : public PassWrapper<RankSpecializationClusterPass, FunctionPass> {
  void getDependentDialects(DialectRegistry &registry) const override {
    registry.insert<mhlo::MhloDialect, chlo::HloClientDialect>();
  }

  void runOnFunction() override {
    MLIRContext *ctx = &getContext();
    RewritePatternSet patterns(ctx);
    mhlo::PopulateRankSpecializationClusterPatterns(ctx, &patterns);
    if (failed(
            applyPatternsAndFoldGreedily(getFunction(), std::move(patterns)))) {
      return signalPassFailure();
    }
  }
};

}  // namespace

void PopulateRankSpecializationClusterPatterns(
    MLIRContext *context, OwningRewritePatternList *patterns) {
  patterns->insert<RankSpecializationClusterPattern>(context);
}

std::unique_ptr<FunctionPass> createRankSpecializationClusterPass() {
  return std::make_unique<RankSpecializationClusterPass>();
}

}  // namespace mhlo
}  // namespace mlir