SCFToGPU.cpp

Go to the documentation of this file.

//===- SCFToGPU.cpp - Convert an affine loop nest to a GPU kernel -------===//
//
// Part of the LLVM Project, 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
//
//===----------------------------------------------------------------------===//
//
// This implements a straightforward conversion of an loop nest into a GPU
// kernel.  The caller is expected to guarantee that the conversion is correct
// or to further transform the kernel to ensure correctness.
//
//===----------------------------------------------------------------------===//
 
#include "mlir/Conversion/SCFToGPU/SCFToGPU.h"
 
#include "mlir/Conversion/AffineToStandard/AffineToStandard.h"
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
#include "mlir/Dialect/GPU/Transforms/ParallelLoopMapper.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/IRMapping.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/DialectConversion.h"
#include "mlir/Transforms/Passes.h"
#include "mlir/Transforms/RegionUtils.h"
#include "llvm/ADT/Sequence.h"
#include "llvm/Support/Debug.h"
#include <optional>
 
#define DEBUG_TYPE "loops-to-gpu"
 
using namespace mlir;
using namespace mlir::affine;
using namespace mlir::scf;
 
// Name of internal attribute to mark visited operations during conversion.
//
// NOTE: The conversion originally used the following legality criteria:
//   `!parallelOp->hasAttr(gpu::getMappingAttrName())`
// But the provided pattern might reject some cases based on more detailed
// analysis of the `mapping` attribute.
// To avoid dialect conversion failure due to non-converted illegal operation
// we use this extra Unit attribute as a marker, that the operation was checked
// by the pattern and is should be considered as legal in the following legality
// checks. The `finalizeParallelLoopToGPUConversion` function performs clean up
// of this extra attributes ans is supposed to be called after the dialect
// conversion.
//
// TODO: Implement a cleaner solution, factoring out the "matching" logic
// from the pattern and its callees into a separate function that can be called
// from both the pattern and the op legality check.
static constexpr StringLiteral kVisitedAttrName = "SCFToGPU_visited";
 
// Extract an indexed value from KernelDim3.
static Value getDim3Value(const gpu::KernelDim3 &dim3, unsigned pos) {
  switch (pos) {
  case 0:
    return dim3.x;
  case 1:
    return dim3.y;
  case 2:
    return dim3.z;
  default:
    llvm_unreachable("dim3 position out of bounds");
  }
  return nullptr;
}
 
// Get the lower bound-related operands of a loop operation.
static Operation::operand_range getLowerBoundOperands(AffineForOp forOp) {
  return forOp.getLowerBoundOperands();
}
 
// Get the upper bound-related operands of a loop operation.
static Operation::operand_range getUpperBoundOperands(AffineForOp forOp) {
  return forOp.getUpperBoundOperands();
}
 
// Get a Value that corresponds to the loop step.  If the step is an attribute,
// materialize a corresponding constant using builder.
static Value getOrCreateStep(AffineForOp forOp, OpBuilder &builder) {
  return builder.create<arith::ConstantIndexOp>(forOp.getLoc(),
                                                forOp.getStepAsInt());
}
 
// Get a Value for the loop lower bound.  If the value requires computation,
// materialize the instructions using builder.
static Value getOrEmitLowerBound(AffineForOp forOp, OpBuilder &builder) {
  return lowerAffineLowerBound(forOp, builder);
}
 
// Get a Value for the loop upper bound.  If the value requires computation,
// materialize the instructions using builder.
static Value getOrEmitUpperBound(AffineForOp forOp, OpBuilder &builder) {
  return lowerAffineUpperBound(forOp, builder);
}
 
// Check the structure of the loop nest:
//   - there are enough loops to map to numDims;
//   - the loops are perfectly nested;
//   - the loop bounds can be computed above the outermost loop.
// This roughly corresponds to the "matcher" part of the pattern-based
// rewriting infrastructure.
static LogicalResult checkAffineLoopNestMappableImpl(AffineForOp forOp,
                                                     unsigned numDims) {
  Region &limit = forOp.getRegion();
  for (unsigned i = 0, e = numDims; i < e; ++i) {
    Operation *nested = &forOp.getBody()->front();
    if (!areValuesDefinedAbove(getLowerBoundOperands(forOp), limit) ||
        !areValuesDefinedAbove(getUpperBoundOperands(forOp), limit))
      return forOp.emitError(
          "loops with bounds depending on other mapped loops "
          "are not supported");
 
    // The innermost loop can have an arbitrary body, skip the perfect nesting
    // check for it.
    if (i == e - 1)
      break;
 
    auto begin = forOp.getBody()->begin(), end = forOp.getBody()->end();
    if (forOp.getBody()->empty() || std::next(begin, 2) != end)
      return forOp.emitError("expected perfectly nested loops in the body");
 
    if (!(forOp = dyn_cast<AffineForOp>(nested)))
      return nested->emitError("expected a nested loop");
  }
  return success();
}
 
static LogicalResult checkAffineLoopNestMappable(AffineForOp forOp,
                                                 unsigned numBlockDims,
                                                 unsigned numThreadDims) {
  if (numBlockDims < 1 || numThreadDims < 1) {
    LLVM_DEBUG(llvm::dbgs() << "nothing to map");
    return success();
  }
 
  if (numBlockDims > 3) {
    return forOp.emitError("cannot map to more than 3 block dimensions");
  }
  if (numThreadDims > 3) {
    return forOp.emitError("cannot map to more than 3 thread dimensions");
  }
  return checkAffineLoopNestMappableImpl(forOp, numBlockDims + numThreadDims);
}
 
namespace {
// Helper structure that holds common state of the loop to GPU kernel
// conversion.
struct AffineLoopToGpuConverter {
  std::optional<AffineForOp> collectBounds(AffineForOp forOp,
                                           unsigned numLoops);
 
  void createLaunch(AffineForOp rootForOp, AffineForOp innermostForOp,
                    unsigned numBlockDims, unsigned numThreadDims);
 
  // Ranges of the loops mapped to blocks or threads.
  SmallVector<Value, 6> dims;
  // Lower bounds of the loops mapped to blocks or threads.
  SmallVector<Value, 6> lbs;
  // Induction variables of the loops mapped to blocks or threads.
  SmallVector<Value, 6> ivs;
  // Steps of the loops mapped to blocks or threads.
  SmallVector<Value, 6> steps;
};
} // namespace
 
// Collect ranges, bounds, steps and induction variables in preparation for
// mapping a loop nest of depth "numLoops" rooted at "forOp" to a GPU kernel.
// This may fail if the IR for computing loop bounds cannot be constructed, for
// example if an affine loop uses semi-affine maps. Return the last loop to be
// mapped on success, std::nullopt on failure.
std::optional<AffineForOp>
AffineLoopToGpuConverter::collectBounds(AffineForOp forOp, unsigned numLoops) {
  OpBuilder builder(forOp.getOperation());
  dims.reserve(numLoops);
  lbs.reserve(numLoops);
  ivs.reserve(numLoops);
  steps.reserve(numLoops);
  AffineForOp currentLoop = forOp;
  for (unsigned i = 0; i < numLoops; ++i) {
    Value lowerBound = getOrEmitLowerBound(currentLoop, builder);
    Value upperBound = getOrEmitUpperBound(currentLoop, builder);
    if (!lowerBound || !upperBound) {
      return std::nullopt;
    }
 
    Value range = builder.create<arith::SubIOp>(currentLoop.getLoc(),
                                                upperBound, lowerBound);
    Value step = getOrCreateStep(currentLoop, builder);
    if (getConstantIntValue(step) != static_cast<int64_t>(1))
      range =
          builder.create<arith::CeilDivSIOp>(currentLoop.getLoc(), range, step);
    dims.push_back(range);
 
    lbs.push_back(lowerBound);
    ivs.push_back(currentLoop.getInductionVar());
    steps.push_back(step);
 
    if (i != numLoops - 1)
      currentLoop = cast<AffineForOp>(&currentLoop.getBody()->front());
  }
  return currentLoop;
}
 
// Replace the rooted at "rootForOp" with a GPU launch operation.  This expects
// "innermostForOp" to point to the last loop to be transformed to the kernel,
// and to have (numBlockDims + numThreadDims) perfectly nested loops between
// "rootForOp" and "innermostForOp".
void AffineLoopToGpuConverter::createLaunch(AffineForOp rootForOp,
                                            AffineForOp innermostForOp,
                                            unsigned numBlockDims,
                                            unsigned numThreadDims) {
  OpBuilder builder(rootForOp.getOperation());
  // Prepare the grid and block sizes for the launch operation.  If there is
  // no loop mapped to a specific dimension, use constant "1" as its size.
  Value constOne =
      (numBlockDims < 3 || numThreadDims < 3)
          ? builder.create<arith::ConstantIndexOp>(rootForOp.getLoc(), 1)
          : nullptr;
  Value gridSizeX = numBlockDims > 0 ? dims[0] : constOne;
  Value gridSizeY = numBlockDims > 1 ? dims[1] : constOne;
  Value gridSizeZ = numBlockDims > 2 ? dims[2] : constOne;
  Value blockSizeX = numThreadDims > 0 ? dims[numBlockDims] : constOne;
  Value blockSizeY = numThreadDims > 1 ? dims[numBlockDims + 1] : constOne;
  Value blockSizeZ = numThreadDims > 2 ? dims[numBlockDims + 2] : constOne;
 
  // Create a launch op and move the body region of the innermost loop to the
  // launch op.
  auto launchOp = builder.create<gpu::LaunchOp>(
      rootForOp.getLoc(), gridSizeX, gridSizeY, gridSizeZ, blockSizeX,
      blockSizeY, blockSizeZ);
 
  // Replace the loop terminator (loops contain only a single block) with the
  // gpu terminator and move the operations from the loop body block to the gpu
  // launch body block.  Do not move the entire block because of the difference
  // in block arguments.
  Operation &terminator = innermostForOp.getBody()->back();
  Location terminatorLoc = terminator.getLoc();
  terminator.erase();
  builder.setInsertionPointToEnd(innermostForOp.getBody());
  builder.create<gpu::TerminatorOp>(terminatorLoc, std::nullopt);
  launchOp.getBody().front().getOperations().splice(
      launchOp.getBody().front().begin(),
      innermostForOp.getBody()->getOperations());
 
  // Remap the loop iterators to use block/thread identifiers instead.  Loops
  // may iterate from LB with step S whereas GPU thread/block ids always iterate
  // from 0 to N with step 1.  Therefore, loop induction variables are replaced
  // with (gpu-thread/block-id * S) + LB.
  builder.setInsertionPointToStart(&launchOp.getBody().front());
  auto *lbArgumentIt = lbs.begin();
  auto *stepArgumentIt = steps.begin();
  for (const auto &en : llvm::enumerate(ivs)) {
    Value id =
        en.index() < numBlockDims
            ? getDim3Value(launchOp.getBlockIds(), en.index())
            : getDim3Value(launchOp.getThreadIds(), en.index() - numBlockDims);
    Value step = steps[en.index()];
    if (getConstantIntValue(step) != static_cast<int64_t>(1))
      id = builder.create<arith::MulIOp>(rootForOp.getLoc(), step, id);
 
    Value ivReplacement =
        builder.create<arith::AddIOp>(rootForOp.getLoc(), *lbArgumentIt, id);
    en.value().replaceAllUsesWith(ivReplacement);
    std::advance(lbArgumentIt, 1);
    std::advance(stepArgumentIt, 1);
  }
 
  // We are done and can erase the original outermost loop.
  rootForOp.erase();
}
 
// Generic loop to GPU kernel conversion function.
static LogicalResult convertAffineLoopNestToGPULaunch(AffineForOp forOp,
                                                      unsigned numBlockDims,
                                                      unsigned numThreadDims) {
  if (failed(checkAffineLoopNestMappable(forOp, numBlockDims, numThreadDims)))
    return failure();
 
  AffineLoopToGpuConverter converter;
  auto maybeInnerLoop =
      converter.collectBounds(forOp, numBlockDims + numThreadDims);
  if (!maybeInnerLoop)
    return failure();
  converter.createLaunch(forOp, *maybeInnerLoop, numBlockDims, numThreadDims);
 
  return success();
}
 
LogicalResult mlir::convertAffineLoopNestToGPULaunch(AffineForOp forOp,
                                                     unsigned numBlockDims,
                                                     unsigned numThreadDims) {
  return ::convertAffineLoopNestToGPULaunch(forOp, numBlockDims, numThreadDims);
}
 
namespace {
struct ParallelToGpuLaunchLowering : public OpRewritePattern<ParallelOp> {
  using OpRewritePattern<ParallelOp>::OpRewritePattern;
 
  LogicalResult matchAndRewrite(ParallelOp parallelOp,
                                PatternRewriter &rewriter) const override;
};
} // namespace
 
/// Tries to derive a static upper bound from the defining operation of
/// `upperBound`.
static Value deriveStaticUpperBound(Value upperBound,
                                    PatternRewriter &rewriter) {
  if (auto op = upperBound.getDefiningOp<arith::ConstantIndexOp>()) {
    return op;
  }
 
  if (auto minOp = upperBound.getDefiningOp<AffineMinOp>()) {
    for (const AffineExpr &result : minOp.getMap().getResults()) {
      if (auto constExpr = dyn_cast<AffineConstantExpr>(result)) {
        return rewriter.create<arith::ConstantIndexOp>(minOp.getLoc(),
                                                       constExpr.getValue());
      }
    }
  }
 
  if (auto minOp = upperBound.getDefiningOp<arith::MinSIOp>()) {
    for (Value operand : {minOp.getLhs(), minOp.getRhs()}) {
      if (auto staticBound = deriveStaticUpperBound(operand, rewriter))
        return staticBound;
    }
  }
 
  if (auto multiplyOp = upperBound.getDefiningOp<arith::MulIOp>()) {
    if (auto lhs = dyn_cast_or_null<arith::ConstantIndexOp>(
            deriveStaticUpperBound(multiplyOp.getOperand(0), rewriter)
                .getDefiningOp()))
      if (auto rhs = dyn_cast_or_null<arith::ConstantIndexOp>(
              deriveStaticUpperBound(multiplyOp.getOperand(1), rewriter)
                  .getDefiningOp())) {
        // Assumptions about the upper bound of minimum computations no longer
        // work if multiplied by mixed signs, so abort in this case.
        if ((lhs.value() < 0) != (rhs.value() < 0))
          return {};
 
        return rewriter.create<arith::ConstantIndexOp>(
            multiplyOp.getLoc(), lhs.value() * rhs.value());
      }
  }
 
  return {};
}
 
static bool isMappedToProcessor(gpu::Processor processor) {
  return processor != gpu::Processor::Sequential;
}
 
static unsigned getLaunchOpArgumentNum(gpu::Processor processor) {
  switch (processor) {
  case gpu::Processor::BlockX:
    return 0;
  case gpu::Processor::BlockY:
    return 1;
  case gpu::Processor::BlockZ:
    return 2;
  case gpu::Processor::ThreadX:
    return 3;
  case gpu::Processor::ThreadY:
    return 4;
  case gpu::Processor::ThreadZ:
    return 5;
  default:;
  }
  llvm_unreachable(
      "invalid processor type while retrieving launch op argument number");
}
 
/// Modifies the current transformation state to capture the effect of the given
/// `scf.parallel` operation on index substitutions and the operations to be
/// inserted.
/// Specifically, if a dimension of a parallel loop is mapped to a hardware id,
/// this function will
/// - compute the loop index based on the hardware id and affine map from the
///   mapping and update `cloningMap` to substitute all uses.
/// - derive a new upper bound for the hardware id and augment the provided
///   `gpu.launch operation` accordingly.
/// - if the upper bound is imprecise, insert a conditional in the `gpu.launch`
///   and update the rewriter to insert into the conditional's body.
/// If the dimension is mapped to sequential,
/// - insert a for loop into the body and update the rewriter to insert into
///   the for loop's body.
/// - update the `cloningMap` to replace uses of the index with the index of
///   the new for loop.
/// In either case,
/// - append the instructions from the loops body to worklist, in reverse order.
/// To note the end of the current scope in case a loop or conditional was
/// inserted, a sentinel (the `gpu.launch` operation) is inserted into the
/// worklist. This signals the processor of the worklist to pop the rewriter
/// one scope-level up.
static LogicalResult processParallelLoop(
    ParallelOp parallelOp, gpu::LaunchOp launchOp, IRMapping &cloningMap,
    SmallVectorImpl<Operation *> &worklist,
    DenseMap<gpu::Processor, Value> &bounds, PatternRewriter &rewriter) {
  // TODO: Verify that this is a valid GPU mapping.
  // processor ids: 0-2 block [x/y/z], 3-5 -> thread [x/y/z], 6-> sequential
  ArrayAttr mapping =
      parallelOp->getAttrOfType<ArrayAttr>(gpu::getMappingAttrName());
 
  // TODO: Support multiple reductions.
  if (!mapping || parallelOp.getNumResults() > 1)
    return failure();
 
  Location loc = parallelOp.getLoc();
 
  auto launchIndependent = [&launchOp](Value val) {
    return val.getParentRegion()->isAncestor(launchOp->getParentRegion());
  };
 
  auto ensureLaunchIndependent = [&rewriter,
                                  launchIndependent](Value val) -> Value {
    if (launchIndependent(val))
      return val;
    if (auto constOp = val.getDefiningOp<arith::ConstantOp>())
      return rewriter.create<arith::ConstantOp>(constOp.getLoc(),
                                                constOp.getValue());
    return {};
  };
 
  for (auto config : llvm::zip(
           mapping, parallelOp.getInductionVars(), parallelOp.getLowerBound(),
           parallelOp.getUpperBound(), parallelOp.getStep())) {
    Attribute mappingAttribute;
    Value iv, lowerBound, upperBound, step;
    std::tie(mappingAttribute, iv, lowerBound, upperBound, step) = config;
    auto annotation =
        dyn_cast<gpu::ParallelLoopDimMappingAttr>(mappingAttribute);
    if (!annotation)
      return parallelOp.emitOpError()
             << "expected mapping attribute for lowering to GPU";
    Value newIndex;
    gpu::Processor processor = annotation.getProcessor();
 
    if (isMappedToProcessor(processor)) {
      // Use the corresponding thread/grid index as replacement for the loop iv.
      Value operand =
          launchOp.getBody().getArgument(getLaunchOpArgumentNum(processor));
      // Take the indexmap and add the lower bound and step computations in.
      // This computes operand * step + lowerBound.
      // Use an affine map here so that it composes nicely with the provided
      // annotation.
      AffineMap lowerAndStep = AffineMap::get(
          1, 2,
          rewriter.getAffineDimExpr(0) * rewriter.getAffineSymbolExpr(0) +
              rewriter.getAffineSymbolExpr(1));
      newIndex = rewriter.create<AffineApplyOp>(
          loc, annotation.getMap().compose(lowerAndStep),
          ValueRange{operand, ensureLaunchIndependent(step),
                     ensureLaunchIndependent(lowerBound)});
      // If there was also a bound, insert that, too.
      // TODO: Check that we do not assign bounds twice.
      if (annotation.getBound()) {
        // We pass as the single operand to the bound-map the number of
        // iterations, which is (upperBound - lowerBound) ceilDiv step. To
        // support inner loops with dynamic upper bounds (as generated by e.g.
        // tiling), try to derive a max for the bounds. If the used bound for
        // the hardware id is imprecise, wrap the contained code into a
        // conditional. If the lower-bound is constant or defined before the
        // launch, we can use it in the launch bounds. Otherwise fail.
        if (!launchIndependent(lowerBound) &&
            !isa_and_nonnull<arith::ConstantOp>(lowerBound.getDefiningOp()))
          return failure();
        // The step must also be constant or defined outside of the loop nest.
        if (!launchIndependent(step) &&
            !isa_and_nonnull<arith::ConstantOp>(step.getDefiningOp()))
          return failure();
        // If the upper-bound is constant or defined before the launch, we can
        // use it in the launch bounds directly. Otherwise try derive a bound.
        bool boundIsPrecise =
            launchIndependent(upperBound) ||
            isa_and_nonnull<arith::ConstantOp>(upperBound.getDefiningOp());
        {
          PatternRewriter::InsertionGuard guard(rewriter);
          rewriter.setInsertionPoint(launchOp);
          if (!boundIsPrecise) {
            upperBound = deriveStaticUpperBound(upperBound, rewriter);
            if (!upperBound) {
              return rewriter.notifyMatchFailure(
                  parallelOp,
                  "cannot derive loop-invariant upper bound for number of"
                  "iterations");
            }
          }
          // Compute the number of iterations needed. We compute this as an
          // affine expression ceilDiv (upperBound - lowerBound) step. We use
          // affine.apply here so that it composes nicely with the provided map.
          AffineMap stepMap = AffineMap::get(
              1, 2,
              ((rewriter.getAffineDimExpr(0) - rewriter.getAffineSymbolExpr(0))
                   .ceilDiv(rewriter.getAffineSymbolExpr(1))));
          Value launchBound = rewriter.create<AffineApplyOp>(
              loc, annotation.getBound().compose(stepMap),
              ValueRange{
                  ensureLaunchIndependent(
                      cloningMap.lookupOrDefault(upperBound)),
                  ensureLaunchIndependent(
                      cloningMap.lookupOrDefault(lowerBound)),
                  ensureLaunchIndependent(cloningMap.lookupOrDefault(step))});
          // todo(herhut,ravishankarm): Update the behavior of setMappingAttr
          // when this condition is relaxed.
          if (!bounds.try_emplace(processor, launchBound).second) {
            return rewriter.notifyMatchFailure(
                parallelOp, "cannot redefine the bound for processor " +
                                Twine(static_cast<int64_t>(processor)));
          }
        }
        if (!boundIsPrecise) {
          // We are using an approximation, create a surrounding conditional.
          Value originalBound = std::get<3>(config);
          arith::CmpIOp pred = rewriter.create<arith::CmpIOp>(
              loc, arith::CmpIPredicate::slt, newIndex,
              cloningMap.lookupOrDefault(originalBound));
          scf::IfOp ifOp = rewriter.create<scf::IfOp>(loc, pred, false);
          rewriter.setInsertionPointToStart(&ifOp.getThenRegion().front());
          // Put a sentinel into the worklist so we know when to pop out of the
          // if body again. We use the launchOp here, as that cannot be part of
          // the bodies instruction.
          worklist.push_back(launchOp.getOperation());
        }
      }
    } else {
      // Create a sequential for loop.
      auto loopOp = rewriter.create<scf::ForOp>(
          loc, cloningMap.lookupOrDefault(lowerBound),
          cloningMap.lookupOrDefault(upperBound),
          cloningMap.lookupOrDefault(step));
      newIndex = loopOp.getInductionVar();
      rewriter.setInsertionPointToStart(loopOp.getBody());
      // Put a sentinel into the worklist so we know when to pop out of the loop
      // body again. We use the launchOp here, as that cannot be part of the
      // bodies instruction.
      worklist.push_back(launchOp.getOperation());
    }
    cloningMap.map(iv, newIndex);
  }
 
  // Propagate custom user defined optional attributes, that can be used at
  // later stage, such as extension data for GPU kernel dispatch
  for (const auto &namedAttr : parallelOp->getAttrs()) {
    if (namedAttr.getName() == gpu::getMappingAttrName() ||
        namedAttr.getName() == ParallelOp::getOperandSegmentSizeAttr())
      continue;
    launchOp->setAttr(namedAttr.getName(), namedAttr.getValue());
  }
 
  Block *body = parallelOp.getBody();
  worklist.reserve(worklist.size() + body->getOperations().size());
  // Include scf.reduce terminator if exists and has an operand.
  if (auto terminator = body->getTerminator();
      isa<scf::ReduceOp>(terminator) && terminator->getOperands().size() == 1) {
    worklist.push_back(terminator);
  }
  for (Operation &op : llvm::reverse(body->without_terminator()))
    worklist.push_back(&op);
  return success();
}
 
/// Lower a `scf.parallel` operation into a corresponding `gpu.launch`
/// operation.
///
/// This essentially transforms a loop nest into a corresponding SIMT function.
/// The conversion is driven by mapping annotations on the `scf.parallel`
/// operations. The mapping is provided via a `DictionaryAttribute` named
/// `mapping`, which has three entries:
///  - processor: the hardware id to map to. 0-2 are block dimensions, 3-5 are
///               thread dimensions and 6 is sequential.
///  - map : An affine map that is used to pre-process hardware ids before
///          substitution.
///  - bound : An affine map that is used to compute the bound of the hardware
///            id based on an upper bound of the number of iterations.
/// If the `scf.parallel` contains nested `scf.parallel` operations, those
/// need to be annotated, as well. Structurally, the transformation works by
/// splicing all operations from nested `scf.parallel` operations into a single
/// sequence. Indices mapped to hardware ids are substituted with those ids,
/// wheras sequential mappings result in a sequential for-loop. To have more
/// flexibility when mapping code to hardware ids, the transform supports two
/// affine maps. The first `map` is used to compute the actual index for
/// substitution from the hardware id. The second `bound` is used to compute the
/// launch dimension for the hardware id from the number of iterations the
/// mapped loop is performing. Note that the number of iterations might be
/// imprecise if the corresponding loop-bounds are loop-dependent. In such case,
/// the hardware id might iterate over additional indices. The transformation
/// caters for this by predicating the created sequence of instructions on
/// the actual loop bound. This only works if an static upper bound for the
/// dynamic loop bound can be derived, currently via analyzing `affine.min`
/// operations.
LogicalResult
ParallelToGpuLaunchLowering::matchAndRewrite(ParallelOp parallelOp,
                                             PatternRewriter &rewriter) const {
  // Mark the operation as visited for recursive legality check.
  parallelOp->setAttr(kVisitedAttrName, rewriter.getUnitAttr());
 
  // We can only transform starting at the outer-most loop. Launches inside of
  // parallel loops are not supported.
  if (auto parentLoop = parallelOp->getParentOfType<ParallelOp>())
    return failure();
  // Create a launch operation. We start with bound one for all grid/block
  // sizes. Those will be refined later as we discover them from mappings.
  Location loc = parallelOp.getLoc();
  Value constantOne =
      rewriter.create<arith::ConstantIndexOp>(parallelOp.getLoc(), 1);
  gpu::LaunchOp launchOp = rewriter.create<gpu::LaunchOp>(
      parallelOp.getLoc(), constantOne, constantOne, constantOne, constantOne,
      constantOne, constantOne);
  rewriter.setInsertionPointToEnd(&launchOp.getBody().front());
  rewriter.create<gpu::TerminatorOp>(loc);
  rewriter.setInsertionPointToStart(&launchOp.getBody().front());
 
  IRMapping cloningMap;
  llvm::DenseMap<gpu::Processor, Value> launchBounds;
  SmallVector<Operation *, 16> worklist;
  if (failed(processParallelLoop(parallelOp, launchOp, cloningMap, worklist,
                                 launchBounds, rewriter)))
    return failure();
 
  // Whether we have seen any side-effects. Reset when leaving an inner scope.
  bool seenSideeffects = false;
  // Whether we have left a nesting scope (and hence are no longer innermost).
  bool leftNestingScope = false;
  while (!worklist.empty()) {
    Operation *op = worklist.pop_back_val();
    // Now walk over the body and clone it.
    // TODO: This is only correct if there either is no further scf.parallel
    //       nested or this code is side-effect free. Otherwise we might need
    //       predication. We are overly conservative for now and only allow
    //       side-effects in the innermost scope.
    if (auto nestedParallel = dyn_cast<ParallelOp>(op)) {
      // Before entering a nested scope, make sure there have been no
      // sideeffects until now.
      if (seenSideeffects)
        return failure();
      // A nested scf.parallel needs insertion of code to compute indices.
      // Insert that now. This will also update the worklist with the loops
      // body.
      if (failed(processParallelLoop(nestedParallel, launchOp, cloningMap,
                                     worklist, launchBounds, rewriter)))
        return failure();
    } else if (op == launchOp.getOperation()) {
      // Found our sentinel value. We have finished the operations from one
      // nesting level, pop one level back up.
      auto *parent = rewriter.getInsertionPoint()->getParentOp();
      rewriter.setInsertionPointAfter(parent);
      leftNestingScope = true;
      seenSideeffects = false;
    } else if (auto reduceOp = dyn_cast<scf::ReduceOp>(op)) {
      // Convert scf.reduction op
      auto parentLoop = op->getParentOfType<ParallelOp>();
      if (!parentLoop || op->getOperands().size() != 1)
        return failure();
      auto operand = op->getOperands().front();
      auto newValue = cloningMap.lookupOrNull(operand);
      if (!newValue || !operand.getType().isSignlessIntOrFloat())
        return failure();
      // Ensure reduction region is isolated from above.
      llvm::SetVector<Value> externalValues;
      getUsedValuesDefinedAbove(reduceOp.getRegion(0), externalValues);
      if (externalValues.size())
        return failure();
      // Replace by gpu.all_reduce.
      auto gpuRedOp = rewriter.create<gpu::AllReduceOp>(loc, newValue);
      cloningMap.map(parentLoop->getResult(0), gpuRedOp.getResult());
      // Copy region.
      rewriter.inlineRegionBefore(reduceOp.getRegion(0), gpuRedOp.getRegion(),
                                  gpuRedOp.getRegion().begin());
      // Replace src.reduce.return with gpu.yield.
      auto scfReturn = gpuRedOp.getRegion().front().getTerminator();
      auto ip = rewriter.saveInsertionPoint();
      rewriter.setInsertionPointToEnd(&gpuRedOp.getRegion().front());
      rewriter.replaceOpWithNewOp<gpu::YieldOp>(
          scfReturn, scfReturn->getOperands().front());
      rewriter.restoreInsertionPoint(ip);
    } else {
      // Otherwise we copy it over.
      Operation *clone = rewriter.clone(*op, cloningMap);
      cloningMap.map(op->getResults(), clone->getResults());
      // Check for side effects.
      // TODO: Handle region side effects properly.
      seenSideeffects |=
          !isMemoryEffectFree(clone) || clone->getNumRegions() != 0;
      // If we are no longer in the innermost scope, sideeffects are disallowed.
      if (seenSideeffects && leftNestingScope)
        return failure();
    }
  }
 
  // Now that we succeeded creating the launch operation, also update the
  // bounds.
  for (auto bound : launchBounds)
    launchOp.setOperand(getLaunchOpArgumentNum(std::get<0>(bound)),
                        std::get<1>(bound));
 
  rewriter.eraseOp(parallelOp);
  return success();
}
 
void mlir::populateParallelLoopToGPUPatterns(RewritePatternSet &patterns) {
  patterns.add<ParallelToGpuLaunchLowering>(patterns.getContext());
}
 
void mlir::configureParallelLoopToGPULegality(ConversionTarget &target) {
  target.addLegalDialect<memref::MemRefDialect>();
  target.addDynamicallyLegalOp<scf::ParallelOp>([](scf::ParallelOp parallelOp) {
    return !parallelOp->hasAttr(gpu::getMappingAttrName()) ||
           parallelOp->hasAttr(kVisitedAttrName);
  });
}
 
void mlir::finalizeParallelLoopToGPUConversion(Operation *op) {
  op->walk([](scf::ParallelOp parallelOp) {
    parallelOp->removeAttr(kVisitedAttrName);
  });
}