Google Git
Sign in
llvm / llvm-project / mlir / ab7113d26477b45ea4965bc7c9e5ff9bca4ffecc / . / lib / Dialect / Affine / Transforms / LoopFusion.cpp
blob: ff0157eb9e4f3a8220b2590c3ffc04ab9839de5a [file] [log] [blame]
//===- LoopFusion.cpp - Code to perform loop fusion -----------------------===//
//
// 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 file implements affine fusion.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Affine/Passes.h"
#include "mlir/Dialect/Affine/Analysis/AffineStructures.h"
#include "mlir/Dialect/Affine/Analysis/LoopAnalysis.h"
#include "mlir/Dialect/Affine/Analysis/Utils.h"
#include "mlir/Dialect/Affine/LoopFusionUtils.h"
#include "mlir/Dialect/Affine/LoopUtils.h"
#include "mlir/Dialect/Affine/Utils.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/AffineMap.h"
#include "mlir/IR/Builders.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/DebugLog.h"
#include "llvm/Support/raw_ostream.h"
#include <iomanip>
#include <optional>
#include <sstream>
namespace mlir {
namespace affine {
#define GEN_PASS_DEF_AFFINELOOPFUSION
#include "mlir/Dialect/Affine/Passes.h.inc"
} // namespace affine
} // namespace mlir
#define DEBUG_TYPE "affine-fusion"
using namespace mlir;
using namespace mlir::affine;
namespace {
/// Loop fusion pass. This pass currently supports a greedy fusion policy,
/// which fuses loop nests with single-writer/single-reader memref dependences
/// with the goal of improving locality.
// TODO: Support fusion of source loop nests which write to multiple
// memrefs, where each memref can have multiple users (if profitable).
struct LoopFusion : public affine::impl::AffineLoopFusionBase<LoopFusion> {
LoopFusion() = default;
LoopFusion(unsigned fastMemorySpace, uint64_t localBufSizeThresholdBytes,
bool maximalFusion, enum FusionMode affineFusionMode) {
this->fastMemorySpace = fastMemorySpace;
this->localBufSizeThreshold = localBufSizeThresholdBytes / 1024;
this->maximalFusion = maximalFusion;
this->affineFusionMode = affineFusionMode;
}
void runOnBlock(Block *block);
void runOnOperation() override;
};
} // namespace
/// Returns true if node 'srcId' can be removed after fusing it with node
/// 'dstId'. The node can be removed if any of the following conditions are met:
/// 1. 'srcId' has no output dependences after fusion and no escaping memrefs.
/// 2. 'srcId' has no output dependences after fusion, has escaping memrefs
/// and the fusion slice is maximal.
/// 3. 'srcId' has output dependences after fusion, the fusion slice is
/// maximal and the fusion insertion point dominates all the dependences.
static bool canRemoveSrcNodeAfterFusion(
unsigned srcId, unsigned dstId, const ComputationSliceState &fusionSlice,
Operation *fusedLoopInsPoint, const DenseSet<Value> &escapingMemRefs,
const MemRefDependenceGraph &mdg) {
Operation *dstNodeOp = mdg.getNode(dstId)->op;
bool hasOutDepsAfterFusion = false;
for (auto &outEdge : mdg.outEdges.lookup(srcId)) {
Operation *depNodeOp = mdg.getNode(outEdge.id)->op;
// Skip dependence with dstOp since it will be removed after fusion.
if (depNodeOp == dstNodeOp)
continue;
// Only fusion within the same block is supported. Use domination analysis
// when needed.
if (depNodeOp->getBlock() != dstNodeOp->getBlock())
return false;
// Check if the insertion point of the fused loop dominates the dependence.
// Otherwise, the src loop can't be removed.
if (fusedLoopInsPoint != depNodeOp &&
!fusedLoopInsPoint->isBeforeInBlock(depNodeOp)) {
LDBG() << "Src loop can't be removed: dst loop doesn't "
<< "dominate dependence";
return false;
}
hasOutDepsAfterFusion = true;
}
// If src loop has dependences after fusion or it writes to an live-out or
// escaping memref, we can only remove it if the fusion slice is maximal so
// that all the dependences are preserved.
if (hasOutDepsAfterFusion || !escapingMemRefs.empty()) {
std::optional<bool> isMaximal = fusionSlice.isMaximal();
if (!isMaximal) {
LDBG() << "Src loop can't be removed: can't determine "
<< "if fusion is maximal";
return false;
}
if (!*isMaximal) {
LDBG() << "Src loop can't be removed: fusion is not maximal";
return false;
}
}
return true;
}
/// Returns in 'srcIdCandidates' the producer fusion candidates for consumer
/// 'dstId'. Candidates are sorted by node id order. This order corresponds to
/// the program order when the 'mdg' is created. However, program order is not
/// guaranteed and must not be required by the client. Program order won't be
/// held if the 'mdg' is reused from a previous fusion step or if the node
/// creation order changes in the future to support more advance cases.
// TODO: Move this to a loop fusion utility once 'mdg' is also moved.
static void getProducerCandidates(unsigned dstId,
const MemRefDependenceGraph &mdg,
SmallVectorImpl<unsigned> &srcIdCandidates) {
// Skip if no input edges along which to fuse.
if (mdg.inEdges.count(dstId) == 0)
return;
// Gather memrefs from loads in 'dstId'.
auto *dstNode = mdg.getNode(dstId);
DenseSet<Value> consumedMemrefs;
for (Operation *load : dstNode->loads)
consumedMemrefs.insert(cast<AffineReadOpInterface>(load).getMemRef());
// Traverse 'dstId' incoming edges and gather the nodes that contain a store
// to one of the consumed memrefs.
for (const auto &srcEdge : mdg.inEdges.lookup(dstId)) {
const auto *srcNode = mdg.getNode(srcEdge.id);
// Skip if 'srcNode' is not a loop nest.
if (!isa<AffineForOp>(srcNode->op))
continue;
if (any_of(srcNode->stores, [&](Operation *op) {
auto storeOp = cast<AffineWriteOpInterface>(op);
return consumedMemrefs.count(storeOp.getMemRef()) > 0;
}))
srcIdCandidates.push_back(srcNode->id);
}
llvm::sort(srcIdCandidates);
srcIdCandidates.erase(llvm::unique(srcIdCandidates), srcIdCandidates.end());
}
/// Returns in 'producerConsumerMemrefs' the memrefs involved in a
/// producer-consumer dependence between 'srcId' and 'dstId'.
static void
gatherProducerConsumerMemrefs(unsigned srcId, unsigned dstId,
const MemRefDependenceGraph &mdg,
DenseSet<Value> &producerConsumerMemrefs) {
auto *dstNode = mdg.getNode(dstId);
auto *srcNode = mdg.getNode(srcId);
gatherProducerConsumerMemrefs(srcNode->stores, dstNode->loads,
producerConsumerMemrefs);
}
/// A memref escapes in the context of the fusion pass if either:
/// 1. it (or its alias) is a block argument, or
/// 2. created by an op not known to guarantee alias freedom,
/// 3. it (or its alias) are used by ops other than affine dereferencing ops
/// (e.g., by call op, memref load/store ops, alias creating ops, unknown ops,
/// terminator ops, etc.); such ops do not deference the memref in an affine
/// way.
static bool isEscapingMemref(Value memref, Block *block) {
Operation *defOp = memref.getDefiningOp();
// Check if 'memref' is a block argument.
if (!defOp)
return true;
// Check if this is defined to be an alias of another memref.
if (auto viewOp = dyn_cast<mlir::ViewLikeOpInterface>(defOp))
if (memref == viewOp.getViewDest() &&
isEscapingMemref(viewOp.getViewSource(), block))
return true;
// Any op besides allocating ops wouldn't guarantee alias freedom
if (!hasSingleEffect<mlir::MemoryEffects::Allocate>(defOp, memref))
return true;
// Check if 'memref' is used by a non-deferencing op (including unknown ones)
// (e.g., call ops, alias creating ops, etc.).
return llvm::any_of(memref.getUsers(), [&](Operation *user) {
// Ignore users outside of `block`.
Operation *ancestorOp = block->getParent()->findAncestorOpInRegion(*user);
if (!ancestorOp)
return true;
if (ancestorOp->getBlock() != block)
return false;
return !isa<AffineMapAccessInterface>(*user);
});
}
/// Returns in 'escapingMemRefs' the memrefs from affine store ops in node 'id'
/// that escape the block or are accessed in a non-affine way.
static void gatherEscapingMemrefs(unsigned id, const MemRefDependenceGraph &mdg,
DenseSet<Value> &escapingMemRefs) {
auto *node = mdg.getNode(id);
for (Operation *storeOp : node->stores) {
auto memref = cast<AffineWriteOpInterface>(storeOp).getMemRef();
if (escapingMemRefs.count(memref))
continue;
if (isEscapingMemref(memref, &mdg.block))
escapingMemRefs.insert(memref);
}
}
// Sinks all sequential loops to the innermost levels (while preserving
// relative order among them) and moves all parallel loops to the
// outermost (while again preserving relative order among them).
// This can increase the loop depth at which we can fuse a slice, since we are
// pushing loop carried dependence to a greater depth in the loop nest.
static void sinkSequentialLoops(MemRefDependenceGraph::Node *node) {
assert(isa<AffineForOp>(node->op));
AffineForOp newRootForOp = sinkSequentialLoops(cast<AffineForOp>(node->op));
node->op = newRootForOp;
}
/// Get the operation that should act as a dominance filter while replacing
/// memref uses with a private memref for which `producerStores` and
/// `sliceInsertionBlock` are provided. This effectively determines in what
/// part of the IR we should be performing the replacement.
static Operation *
getDominanceFilterForPrivateMemRefRepl(Block *sliceInsertionBlock,
ArrayRef<Operation *> producerStores) {
assert(!producerStores.empty() && "expected producer store");
// We first find the common block that contains the producer stores and
// the slice computation. The first ancestor among the ancestors of the
// producer stores in that common block is the dominance filter to use for
// replacement.
Block *commonBlock = nullptr;
// Find the common block of all relevant operations.
for (Operation *store : producerStores) {
Operation *otherOp =
!commonBlock ? &*sliceInsertionBlock->begin() : &*commonBlock->begin();
commonBlock = findInnermostCommonBlockInScope(store, otherOp);
}
assert(commonBlock &&
"common block of producer stores and slice should exist");
// Find the first ancestor among the ancestors of `producerStores` in
// `commonBlock`.
Operation *firstAncestor = nullptr;
for (Operation *store : producerStores) {
Operation *ancestor = commonBlock->findAncestorOpInBlock(*store);
assert(ancestor && "producer store should be contained in common block");
firstAncestor = !firstAncestor || ancestor->isBeforeInBlock(firstAncestor)
? ancestor
: firstAncestor;
}
return firstAncestor;
}
/// Returns the amount of additional (redundant) computation that will be done
/// as a fraction of the total computation if `srcForOp` is fused into
/// `dstForOp` at depth `depth`. The method returns the compute cost of the
/// slice and the fused nest's compute cost in the trailing output arguments.
static std::optional<double> getAdditionalComputeFraction(
AffineForOp srcForOp, AffineForOp dstForOp, unsigned depth,
ArrayRef<ComputationSliceState> depthSliceUnions, int64_t &sliceCost,
int64_t &fusedLoopNestComputeCost) {
LDBG() << "Determining additional compute fraction...";
// Compute cost of sliced and unsliced src loop nest.
// Walk src loop nest and collect stats.
LoopNestStats srcLoopNestStats;
if (!getLoopNestStats(srcForOp, &srcLoopNestStats)) {
LDBG() << "Failed to get source loop nest stats.";
return std::nullopt;
}
// Compute cost of dst loop nest.
LoopNestStats dstLoopNestStats;
if (!getLoopNestStats(dstForOp, &dstLoopNestStats)) {
LDBG() << "Failed to get destination loop nest stats.";
return std::nullopt;
}
// Compute op instance count for the src loop nest without iteration slicing.
uint64_t srcLoopNestCost = getComputeCost(srcForOp, srcLoopNestStats);
// Compute op cost for the dst loop nest.
uint64_t dstLoopNestCost = getComputeCost(dstForOp, dstLoopNestStats);
const ComputationSliceState &slice = depthSliceUnions[depth - 1];
// Skip slice union if it wasn't computed for this depth.
if (slice.isEmpty()) {
LDBG() << "Slice wasn't computed.";
return std::nullopt;
}
if (!getFusionComputeCost(srcForOp, srcLoopNestStats, dstForOp,
dstLoopNestStats, slice,
&fusedLoopNestComputeCost)) {
LDBG() << "Unable to compute fusion compute cost";
return std::nullopt;
}
double additionalComputeFraction =
fusedLoopNestComputeCost /
(static_cast<double>(srcLoopNestCost) + dstLoopNestCost) -
1;
return additionalComputeFraction;
}
// Creates and returns a private (single-user) memref for fused loop rooted at
// 'forOp', with (potentially reduced) memref size based on the memref region
// written to by `storeOps` at depth 'dstLoopDepth'. 'sliceInsertionBlock'
// specifies the block in which the slice was/will be inserted. The method
// expects that all stores ops to the memref have the same access function.
// Returns nullptr if the creation failed.
static Value createPrivateMemRef(AffineForOp forOp,
ArrayRef<Operation *> storeOps,
unsigned dstLoopDepth,
std::optional<unsigned> fastMemorySpace,
Block *sliceInsertionBlock,
uint64_t localBufSizeThreshold) {
assert(!storeOps.empty() && "no source stores supplied");
// Check if all stores have the same access function; we only support this
// case.
// TODO: Use union of memref write regions to compute private memref footprint
// for store ops with different access functions.
if (storeOps.size() > 1 &&
!std::equal(std::next(storeOps.begin()), storeOps.end(), storeOps.begin(),
[](Operation *a, Operation *b) {
MemRefAccess aM(cast<AffineWriteOpInterface>(a));
MemRefAccess bM(cast<AffineWriteOpInterface>(b));
return aM == bM;
})) {
LDBG() << "Private memref creation unsupported for multiple producer "
<< "stores with different access functions.";
return nullptr;
}
Operation *srcStoreOp = storeOps[0];
// Create builder to insert alloc op just before 'forOp'.
OpBuilder b(forOp);
// Builder to create constants at the top level.
OpBuilder top(forOp->getParentRegion());
// Create new memref type based on slice bounds.
auto oldMemRef = cast<AffineWriteOpInterface>(srcStoreOp).getMemRef();
auto oldMemRefType = cast<MemRefType>(oldMemRef.getType());
unsigned rank = oldMemRefType.getRank();
// Compute MemRefRegion for 'srcStoreOpInst' at depth 'dstLoopDepth'.
MemRefRegion region(srcStoreOp->getLoc());
bool validRegion = succeeded(
region.compute(srcStoreOp, dstLoopDepth, /*sliceState=*/nullptr,
/*addMemRefDimBounds=*/true, /*dropLocalVars=*/false));
(void)validRegion;
assert(validRegion && "unexpected memref region failure");
SmallVector<int64_t, 4> newShape;
SmallVector<AffineMap, 4> lbs;
lbs.reserve(rank);
// Query 'region' for 'newShape' and lower bounds of MemRefRegion accessed
// by 'srcStoreOpInst' at depth 'dstLoopDepth'.
std::optional<int64_t> numElements =
region.getConstantBoundingSizeAndShape(&newShape, &lbs);
assert(numElements && "non-constant number of elts in local buffer");
const FlatAffineValueConstraints *cst = region.getConstraints();
// 'outerIVs' holds the values that this memory region is symbolic/parametric
// on; this would correspond to loop IVs surrounding the level at which the
// slice is being materialized.
SmallVector<Value, 8> outerIVs;
cst->getValues(rank, cst->getNumDimAndSymbolVars(), &outerIVs);
// Build 'rank' AffineExprs from MemRefRegion 'lbs'
SmallVector<AffineExpr, 4> offsets;
offsets.reserve(rank);
// Outer IVs are considered symbols during memref region computation. Replace
// them uniformly with dims so that valid IR is guaranteed.
SmallVector<AffineExpr> replacements;
for (unsigned j = 0, e = lbs[0].getNumSymbols(); j < e; ++j)
replacements.push_back(mlir::getAffineDimExpr(j, forOp.getContext()));
for (unsigned d = 0; d < rank; ++d) {
assert(lbs[d].getNumResults() == 1 &&
"invalid private memref bound calculation");
offsets.push_back(lbs[d].getResult(0).replaceSymbols(replacements));
}
// Create 'newMemRefType' using 'newShape' from MemRefRegion accessed
// by 'srcStoreOpInst'.
auto eltSize = getMemRefIntOrFloatEltSizeInBytes(oldMemRefType);
assert(eltSize && "memrefs with size elt types expected");
uint64_t bufSize = *eltSize * *numElements;
Attribute newMemSpace;
if (bufSize <= localBufSizeThreshold && fastMemorySpace.has_value()) {
newMemSpace = b.getI64IntegerAttr(*fastMemorySpace);
} else {
newMemSpace = oldMemRefType.getMemorySpace();
}
auto newMemRefType = MemRefType::get(newShape, oldMemRefType.getElementType(),
/*map=*/AffineMap(), newMemSpace);
// Create new private memref for fused loop 'forOp'. 'newShape' is always
// a constant shape.
// TODO: Create/move alloc ops for private memrefs closer to their
// consumer loop nests to reduce their live range. Currently they are added
// at the beginning of the block, because loop nests can be reordered
// during the fusion pass.
Value newMemRef = memref::AllocOp::create(top, forOp.getLoc(), newMemRefType);
// Build an AffineMap to remap access functions based on lower bound offsets.
SmallVector<AffineExpr, 4> remapExprs;
remapExprs.reserve(rank);
for (unsigned i = 0; i < rank; i++) {
auto dimExpr = b.getAffineDimExpr(outerIVs.size() + i);
auto remapExpr =
simplifyAffineExpr(dimExpr - offsets[i], outerIVs.size() + rank, 0);
remapExprs.push_back(remapExpr);
}
auto indexRemap =
AffineMap::get(outerIVs.size() + rank, 0, remapExprs, forOp.getContext());
// Replace all users of 'oldMemRef' with 'newMemRef'.
Operation *domFilter =
getDominanceFilterForPrivateMemRefRepl(sliceInsertionBlock, storeOps);
auto userFilterFn = [&](Operation *user) {
auto domInfo = std::make_unique<DominanceInfo>(
domFilter->getParentOfType<FunctionOpInterface>());
return domInfo->dominates(domFilter, user);
};
LogicalResult res = replaceAllMemRefUsesWith(
oldMemRef, newMemRef, /*extraIndices=*/{}, indexRemap,
/*extraOperands=*/outerIVs,
/*symbolOperands=*/{}, userFilterFn);
assert(succeeded(res) &&
"replaceAllMemrefUsesWith should always succeed here");
(void)res;
LDBG() << "Created private memref of type: " << newMemRefType;
return newMemRef;
}
// Checks the profitability of fusing a backwards slice of the loop nest
// `srcForOp` into the loop nest surrounding 'dstLoadOpInsts'. The argument
// 'srcStoreOpInst' is used to calculate the storage reduction on the memref
// being produced and consumed, which is an input to the cost model. For
// producer-consumer fusion, 'srcStoreOpInst' will be the same as 'srcOpInst',
// as we are slicing w.r.t to that producer. For input-reuse fusion, 'srcOpInst'
// will be the src loop nest LoadOp which reads from the same memref as dst loop
// nest load ops, and 'srcStoreOpInst' will be the unique store op in the src
// node, which will be used to check that the write region is the same after
// input-reuse fusion. Computation slices are provided in 'depthSliceUnions' for
// each legal fusion depth. The maximal depth at which fusion is legal is
// provided in 'maxLegalFusionDepth'. Returns true if it is profitable to fuse
// the candidate loop nests. Returns false otherwise. `dstLoopDepth` is set to
// the most profitable depth at which to materialize the source loop nest slice.
// The profitability model executes the following steps:
// *) Computes the backward computation slice at 'srcOpInst'. This
// computation slice of the loop nest surrounding 'srcOpInst' is
// represented by modified src loop bounds in 'sliceState', which are
// functions of loop IVs in the loop nest surrounding 'srcOpInst'.
// *) Computes the cost of unfused src/dst loop nests (currently the cost of a
// loop nest is the total number of dynamic operation instances in the loop
// nest).
// *) Computes the cost of fusing a slice of the src loop nest into the dst
// loop nest at various values of dst loop depth, attempting to fuse
// the largest computation slice at the maximal dst loop depth (closest to
// the load) to minimize reuse distance and potentially enable subsequent
// load/store forwarding.
// NOTE: 'dstLoopDepth' refers to the loop depth within the destination loop
// nest, at which the src computation slice is inserted/fused.
// NOTE: We attempt to maximize the dst loop depth, but there are cases
// where a particular setting for 'dstLoopNest' might fuse an unsliced
// loop (within the src computation slice) at a depth which results in
// excessive recomputation (see unit tests for examples).
// *) Compares the total cost of the unfused loop nests to the min cost fused
// loop nest computed in the previous step, and returns true if the latter
// is lower.
// TODO: Extend profitability analysis to support scenarios with multiple
// stores.
static bool isFusionProfitable(AffineForOp srcForOp,
ArrayRef<Operation *> producerStores,
AffineForOp dstForOp,
ArrayRef<ComputationSliceState> depthSliceUnions,
unsigned maxLegalFusionDepth,
unsigned *dstLoopDepth,
double computeToleranceThreshold) {
LDBG() << "Checking whether fusion is profitable between source nest:";
LDBG() << ' ' << srcForOp << " and destination nest:";
LDBG() << dstForOp;
if (maxLegalFusionDepth == 0) {
LDBG() << "Can't fuse: maxLegalFusionDepth is 0";
return false;
}
// Compute cost of sliced and unsliced src loop nest.
// Walk src loop nest and collect stats.
LoopNestStats srcLoopNestStats;
if (!getLoopNestStats(srcForOp, &srcLoopNestStats))
return false;
// Compute cost of dst loop nest.
LoopNestStats dstLoopNestStats;
if (!getLoopNestStats(dstForOp, &dstLoopNestStats))
return false;
// We limit profitability analysis to only scenarios with
// a single producer store for now. Note that some multi-store
// producer scenarios will still go through profitability analysis
// if only one of the stores is involved in the producer-consumer
// relationship of the candidate loops.
// TODO: Suppport multiple producer stores in profitability
// analysis.
if (producerStores.size() > 1) {
LDBG() << "Limited profitability analysis. Not "
<< "supported for multiple producer store case.";
int64_t sliceCost;
int64_t fusedLoopNestComputeCost;
// We will still fuse if fusion obeys the specified compute
// tolerance at the max legal depth.
auto fraction = getAdditionalComputeFraction(
srcForOp, dstForOp, maxLegalFusionDepth, depthSliceUnions, sliceCost,
fusedLoopNestComputeCost);
if (!fraction || fraction > computeToleranceThreshold) {
LDBG() << "Additional computation exceeds "
<< "compute tolerance. Not fusing.";
return false;
}
LDBG() << "Considering fusion profitable at max legal depth.";
return true;
}
Operation *srcStoreOp = producerStores.front();
// Search for min cost value for 'dstLoopDepth'. At each value of
// 'dstLoopDepth' from 'maxLegalLoopDepth' to '1', compute computation slice
// bounds between 'srcOpInst' and each op in 'dstOpinsts' (taking the union
// of these bounds). Next the union slice bounds are used to calculate
// the cost of the slice and the cost of the slice inserted into the dst
// loop nest at 'dstLoopDepth'.
uint64_t minFusedLoopNestComputeCost = std::numeric_limits<uint64_t>::max();
double maxStorageReduction = 0.0;
std::optional<uint64_t> sliceMemEstimate;
// The best loop depth at which to materialize the slice.
std::optional<unsigned> bestDstLoopDepth;
// Compute src loop nest write region size.
MemRefRegion srcWriteRegion(srcStoreOp->getLoc());
if (failed(srcWriteRegion.compute(srcStoreOp, /*loopDepth=*/0))) {
LDBG() << "Unable to compute MemRefRegion for source operation";
return false;
}
std::optional<int64_t> maybeSrcWriteRegionSizeBytes =
srcWriteRegion.getRegionSize();
if (!maybeSrcWriteRegionSizeBytes.has_value())
return false;
int64_t srcWriteRegionSizeBytes = *maybeSrcWriteRegionSizeBytes;
// Compute op instance count for the src loop nest without iteration slicing.
uint64_t srcLoopNestCost = getComputeCost(srcForOp, srcLoopNestStats);
// Compute op instance count for the destination loop nest.
uint64_t dstLoopNestCost = getComputeCost(dstForOp, dstLoopNestStats);
// Evaluate all depth choices for materializing the slice in the destination
// loop nest.
for (unsigned i = maxLegalFusionDepth; i >= 1; --i) {
const ComputationSliceState &slice = depthSliceUnions[i - 1];
// Skip slice union if it wasn't computed for this depth.
if (slice.isEmpty())
continue;
// Compute cost of the slice separately, i.e, the compute cost of the slice
// if all outer trip counts are one.
int64_t sliceCost;
int64_t fusedLoopNestComputeCost;
auto mayAdditionalComputeFraction =
getAdditionalComputeFraction(srcForOp, dstForOp, i, depthSliceUnions,
sliceCost, fusedLoopNestComputeCost);
if (!mayAdditionalComputeFraction) {
LDBG() << "Can't determine additional compute fraction.";
continue;
}
double additionalComputeFraction = *mayAdditionalComputeFraction;
// Determine what the slice write MemRefRegion would be, if the src loop
// nest slice 'slice' were to be inserted into the dst loop nest at loop
// depth 'i'.
MemRefRegion sliceWriteRegion(srcStoreOp->getLoc());
if (failed(sliceWriteRegion.compute(srcStoreOp, /*loopDepth=*/0, &slice))) {
LDBG() << "Failed to compute slice write region at loopDepth: " << i;
continue;
}
std::optional<int64_t> maybeSliceWriteRegionSizeBytes =
sliceWriteRegion.getRegionSize();
if (!maybeSliceWriteRegionSizeBytes.has_value() ||
*maybeSliceWriteRegionSizeBytes == 0) {
LDBG() << "Failed to get slice write region size at loopDepth: " << i;
continue;
}
int64_t sliceWriteRegionSizeBytes = *maybeSliceWriteRegionSizeBytes;
double storageReduction = static_cast<double>(srcWriteRegionSizeBytes) /
static_cast<double>(sliceWriteRegionSizeBytes);
LLVM_DEBUG({
std::stringstream msg;
msg << " evaluating fusion profitability at depth : " << i << "\n"
<< std::fixed << std::setprecision(2)
<< " additional compute fraction: "
<< 100.0 * additionalComputeFraction << "%\n"
<< " storage reduction factor: " << storageReduction << "x\n"
<< " fused nest cost: " << fusedLoopNestComputeCost << "\n"
<< " src write region size: " << srcWriteRegionSizeBytes << "\n"
<< " slice write region size: " << sliceWriteRegionSizeBytes;
LDBG() << msg.str();
});
// TODO: This is a placeholder cost model.
// Among all choices that add an acceptable amount of redundant computation
// (as per computeToleranceThreshold), we will simply pick the one that
// reduces the intermediary size the most.
if ((storageReduction > maxStorageReduction) &&
(additionalComputeFraction <= computeToleranceThreshold)) {
maxStorageReduction = storageReduction;
bestDstLoopDepth = i;
minFusedLoopNestComputeCost = fusedLoopNestComputeCost;
sliceMemEstimate = sliceWriteRegionSizeBytes;
}
}
// A simple cost model: fuse if it reduces the memory footprint.
if (!bestDstLoopDepth) {
LDBG() << "All fusion choices involve more than the threshold amount of "
<< "redundant computation; NOT fusing.";
return false;
}
if (!bestDstLoopDepth) {
LDBG() << "no fusion depth could be evaluated.";
return false;
}
// Set dstLoopDepth based on best values from search.
*dstLoopDepth = *bestDstLoopDepth;
LDBG() << " LoopFusion fusion stats:";
LDBG() << " best loop depth: " << bestDstLoopDepth;
LDBG() << " src loop nest compute cost: " << srcLoopNestCost;
LDBG() << " dst loop nest compute cost: " << dstLoopNestCost;
LDBG() << " fused loop nest compute cost: " << minFusedLoopNestComputeCost;
auto dstMemSize = getMemoryFootprintBytes(dstForOp);
auto srcMemSize = getMemoryFootprintBytes(srcForOp);
std::optional<double> storageReduction;
if (!dstMemSize || !srcMemSize) {
LDBG() << " fusion memory benefit cannot be evaluated; NOT fusing.";
return false;
}
auto srcMemSizeVal = *srcMemSize;
auto dstMemSizeVal = *dstMemSize;
assert(sliceMemEstimate && "expected value");
auto fusedMem = dstMemSizeVal + *sliceMemEstimate;
LDBG() << " src mem: " << srcMemSizeVal;
LDBG() << " dst mem: " << dstMemSizeVal;
LDBG() << " fused mem: " << fusedMem;
LDBG() << " slice mem: " << sliceMemEstimate;
if (static_cast<long>(fusedMem) > srcMemSizeVal + dstMemSizeVal) {
LDBG() << "Fusion is not profitable; NOT fusing.";
return false;
}
storageReduction =
100.0 *
(1.0 - fusedMem / (static_cast<double>(srcMemSizeVal) + dstMemSizeVal));
double additionalComputeFraction =
100.0 * (minFusedLoopNestComputeCost /
(static_cast<double>(srcLoopNestCost) + dstLoopNestCost) -
1);
(void)additionalComputeFraction;
LLVM_DEBUG({
std::stringstream msg;
msg << " fusion is most profitable at depth " << *dstLoopDepth << " with "
<< std::setprecision(2) << additionalComputeFraction
<< "% redundant computation and a ";
msg << (storageReduction ? std::to_string(*storageReduction) : "<unknown>");
msg << "% storage reduction.";
LDBG() << msg.str();
});
return true;
}
namespace {
// GreedyFusion greedily fuses loop nests which have a producer/consumer or
// input-reuse relationship on a memref, with the goal of improving locality.
//
// The steps of the producer-consumer fusion algorithm are as follows:
//
// *) A worklist is initialized with node ids from the dependence graph.
// *) For each node id in the worklist:
// *) Pop an AffineForOp of the worklist. This 'dstAffineForOp' will be a
// candidate destination AffineForOp into which fusion will be attempted.
// *) Add each LoadOp currently in 'dstAffineForOp' into list 'dstLoadOps'.
// *) For each LoadOp in 'dstLoadOps' do:
// *) Look up dependent loop nests which have a single store op to the same
// memref.
// *) Check if dependences would be violated by the fusion.
// *) Get a computation slice of 'srcLoopNest', which adjusts its loop
// bounds to be functions of 'dstLoopNest' IVs and symbols.
// *) Fuse the 'srcLoopNest' computation slice into the 'dstLoopNest',
// at a loop depth determined by the cost model in 'isFusionProfitable'.
// *) Add the newly fused load/store operations to the state,
// and also add newly fused load ops to 'dstLoopOps' to be considered
// as fusion dst load ops in another iteration.
// *) Remove old src loop nest and its associated state.
//
// The steps of the input-reuse fusion algorithm are as follows:
//
// *) Initialize 'worklist' with node ids from the dependence graph.
// *) For each 'dstNode' in the worklist:
// *) Find a candidate sibling node 'sibNode' to fuse with 'dstNode' which
// loads from the same memref, but which has no dependence paths to/from.
// *) Get a computation slice of 'sibLoopNest', which adjusts its loop
// bounds to be functions of 'dstLoopNest' IVs and symbols.
// *) Fuse the 'sibLoopNest' computation slice into the 'dstLoopNest',
// at a loop depth determined by the cost model in 'isFusionProfitable'.
// This function also checks that the memref write region of 'sibLoopNest',
// is preserved in the fused loop nest.
// *) Update graph state to reflect the fusion of 'sibNode' into 'dstNode'.
//
// Given a graph where top-level operations are vertices in the set 'V' and
// edges in the set 'E' are dependences between vertices, this algorithm
// takes O(V) time for initialization, and has runtime O(V + E).
//
// This greedy algorithm is not 'maximal' due to the current restriction of
// fusing along single producer consumer edges, but there is a TODO: to fix
// this.
//
// TODO: Experiment with other fusion policies.
struct GreedyFusion {
public:
// The data dependence graph to traverse during fusion.
MemRefDependenceGraph *mdg;
// Worklist of graph nodes visited during the fusion pass.
SmallVector<unsigned, 8> worklist;
// Parameter for local buffer size threshold.
unsigned localBufSizeThreshold;
// Parameter for fast memory space.
std::optional<unsigned> fastMemorySpace;
// If true, ignore any additional (redundant) computation tolerance threshold
// that would have prevented fusion.
bool maximalFusion;
// The amount of additional computation that is tolerated while fusing
// pair-wise as a fraction of the total computation.
double computeToleranceThreshold;
using Node = MemRefDependenceGraph::Node;
GreedyFusion(MemRefDependenceGraph *mdg, unsigned localBufSizeThreshold,
std::optional<unsigned> fastMemorySpace, bool maximalFusion,
double computeToleranceThreshold)
: mdg(mdg), localBufSizeThreshold(localBufSizeThreshold),
fastMemorySpace(fastMemorySpace), maximalFusion(maximalFusion),
computeToleranceThreshold(computeToleranceThreshold) {}
/// Initializes 'worklist' with nodes from 'mdg'.
void init() {
// TODO: Add a priority queue for prioritizing nodes by different
// metrics (e.g. arithmetic intensity/flops-to-bytes ratio).
worklist.clear();
for (auto &idAndNode : mdg->nodes) {
const Node &node = idAndNode.second;
worklist.push_back(node.id);
}
}
/// Run only sibling fusion on the `mdg`.
void runSiblingFusionOnly() {
fuseSiblingNodes();
eraseUnusedMemRefAllocations();
}
/// Run only producer/consumer fusion on the `mdg`.
void runProducerConsumerFusionOnly() {
fuseProducerConsumerNodes(
/*maxSrcUserCount=*/std::numeric_limits<unsigned>::max());
eraseUnusedMemRefAllocations();
}
// Run the GreedyFusion pass.
// *) First pass through the nodes fuses single-use producer nodes into their
// unique consumer.
// *) Second pass fuses sibling nodes which share no dependence edges.
// *) Third pass fuses any remaining producer nodes into their users.
void runGreedyFusion() {
// TODO: Run this repeatedly until a fixed-point is reached.
fuseProducerConsumerNodes(/*maxSrcUserCount=*/1);
fuseSiblingNodes();
fuseProducerConsumerNodes(
/*maxSrcUserCount=*/std::numeric_limits<unsigned>::max());
eraseUnusedMemRefAllocations();
}
/// Returns true if a private memref can be created for `memref` given
/// the fusion scenario reflected by the other arguments.
bool canCreatePrivateMemRef(Value memref,
const DenseSet<Value> &srcEscapingMemRefs,
unsigned producerId, unsigned consumerId,
bool removeSrcNode) {
// We can't generate private memrefs if their size can't be computed.
if (!getMemRefIntOrFloatEltSizeInBytes(cast<MemRefType>(memref.getType())))
return false;
const Node *consumerNode = mdg->getNode(consumerId);
// If `memref` is an escaping one, do not create a private memref
// for the below scenarios, since doing so will leave the escaping
// memref unmodified as all the writes originally meant for the
// escaping memref would be performed on the private memref:
// 1. The source is to be removed after fusion,
// OR
// 2. The destination writes to `memref`.
if (srcEscapingMemRefs.count(memref) > 0 &&
(removeSrcNode || consumerNode->getStoreOpCount(memref) > 0))
return false;
// Don't create a private memref if 'srcNode' has in edges on
// 'memref' or 'dstNode' has out edges on 'memref'.
if (mdg->getIncomingMemRefAccesses(producerId, memref) > 0 ||
mdg->getOutEdgeCount(consumerId, memref) > 0)
return false;
// If 'srcNode' will be removed but it has out edges on 'memref' to
// nodes other than 'dstNode', we have to preserve dependences and
// cannot create a private memref.
if (removeSrcNode &&
any_of(mdg->outEdges[producerId], [&](const auto &edge) {
return edge.value == memref && edge.id != consumerId;
}))
return false;
return true;
}
/// Perform fusions with node `dstId` as the destination of fusion, with
/// No fusion is performed when producers with a user count greater than
/// `maxSrcUserCount` for any of the memrefs involved.
void performFusionsIntoDest(unsigned dstId, unsigned maxSrcUserCount) {
LDBG() << "Evaluating dst loop " << dstId;
// Skip if this node was removed (fused into another node).
if (mdg->nodes.count(dstId) == 0)
return;
// Get 'dstNode' into which to attempt fusion.
auto *dstNode = mdg->getNode(dstId);
// Skip if 'dstNode' is not a loop nest.
if (!isa<AffineForOp>(dstNode->op))
return;
// Skip if 'dstNode' is a loop nest returning values.
// TODO: support loop nests that return values.
if (dstNode->op->getNumResults() > 0)
return;
LDBG() << "Evaluating dst loop " << dstId;
// Sink sequential loops in 'dstNode' (and thus raise parallel loops)
// while preserving relative order. This can increase the maximum loop
// depth at which we can fuse a slice of a producer loop nest into a
// consumer loop nest.
sinkSequentialLoops(dstNode);
auto dstAffineForOp = cast<AffineForOp>(dstNode->op);
// Try to fuse 'dstNode' with candidate producer loops until a fixed point
// is reached. Fusing two loops may expose new fusion opportunities.
bool dstNodeChanged;
do {
// Gather src loop candidates for 'dstNode' and visit them in "quasi"
// reverse program order to minimize the number of iterations needed to
// reach the fixed point. Note that this is a best effort approach since
// 'getProducerCandidates' does not always guarantee that program order
// in 'srcIdCandidates'.
dstNodeChanged = false;
SmallVector<unsigned, 16> srcIdCandidates;
getProducerCandidates(dstId, *mdg, srcIdCandidates);
for (unsigned srcId : llvm::reverse(srcIdCandidates)) {
// Get 'srcNode' from which to attempt fusion into 'dstNode'.
auto *srcNode = mdg->getNode(srcId);
auto srcAffineForOp = cast<AffineForOp>(srcNode->op);
LDBG() << "Trying to fuse producer loop nest " << srcId
<< " with consumer loop nest " << dstId;
LDBG() << "Compute tolerance threshold: " << computeToleranceThreshold;
LDBG() << "Producer loop nest:";
LDBG() << *srcNode->op << " and consumer loop nest:";
LDBG() << *dstNode->op;
LDBG() << "Evaluating src loop " << srcId << " for dst loop " << dstId;
// Skip if 'srcNode' is a loop nest returning values.
// TODO: support loop nests that return values.
if (isa<AffineForOp>(srcNode->op) && srcNode->op->getNumResults() > 0)
continue;
DenseSet<Value> producerConsumerMemrefs;
gatherProducerConsumerMemrefs(srcId, dstId, *mdg,
producerConsumerMemrefs);
// Skip if 'srcNode' out edge count on any memref is greater than
// 'maxSrcUserCount'.
if (any_of(producerConsumerMemrefs, [&](Value memref) {
return mdg->getOutEdgeCount(srcNode->id, memref) >
maxSrcUserCount;
}))
continue;
// Gather memrefs in 'srcNode' that are written and escape out of the
// block (e.g., memref block arguments, returned memrefs,
// memrefs passed to function calls, etc.).
DenseSet<Value> srcEscapingMemRefs;
gatherEscapingMemrefs(srcNode->id, *mdg, srcEscapingMemRefs);
// Compute an operation list insertion point for the fused loop
// nest which preserves dependences.
Operation *fusedLoopInsPoint =
mdg->getFusedLoopNestInsertionPoint(srcNode->id, dstNode->id);
if (fusedLoopInsPoint == nullptr)
continue;
// It's possible this fusion is at an inner depth (i.e., there are
// common surrounding affine loops for the source and destination for
// ops). We need to get this number because the call to canFuseLoops
// needs to be passed the absolute depth. The max legal depth and the
// depths we try below are however *relative* and as such don't include
// the common depth.
SmallVector<AffineForOp, 4> surroundingLoops;
getAffineForIVs(*dstAffineForOp, &surroundingLoops);
unsigned numSurroundingLoops = surroundingLoops.size();
// Compute the innermost common loop depth for dstNode
// producer-consumer loads/stores.
SmallVector<Operation *, 2> dstMemrefOps;
for (Operation *op : dstNode->loads)
if (producerConsumerMemrefs.count(
cast<AffineReadOpInterface>(op).getMemRef()) > 0)
dstMemrefOps.push_back(op);
for (Operation *op : dstNode->stores)
if (producerConsumerMemrefs.count(
cast<AffineWriteOpInterface>(op).getMemRef()))
dstMemrefOps.push_back(op);
if (dstMemrefOps.empty())
continue;
unsigned dstLoopDepthTest =
getInnermostCommonLoopDepth(dstMemrefOps) - numSurroundingLoops;
// Check the feasibility of fusing src loop nest into dst loop nest
// at loop depths in range [1, dstLoopDepthTest].
unsigned maxLegalFusionDepth = 0;
SmallVector<ComputationSliceState, 8> depthSliceUnions;
depthSliceUnions.resize(dstLoopDepthTest);
FusionStrategy strategy(FusionStrategy::ProducerConsumer);
for (unsigned i = 1; i <= dstLoopDepthTest; ++i) {
FusionResult result =
affine::canFuseLoops(srcAffineForOp, dstAffineForOp,
/*dstLoopDepth=*/i + numSurroundingLoops,
&depthSliceUnions[i - 1], strategy);
if (result.value == FusionResult::Success) {
maxLegalFusionDepth = i;
LDBG() << "Found valid slice for depth: " << i;
}
}
if (maxLegalFusionDepth == 0) {
LDBG() << "Can't fuse: fusion is not legal at any depth";
continue;
}
LDBG() << "Max legal depth for fusion: " << maxLegalFusionDepth;
double computeToleranceThresholdToUse = computeToleranceThreshold;
// Cyclic dependences in the source nest may be violated when performing
// slicing-based fusion. They aren't actually violated in cases where no
// redundant execution of the source happens (1:1 pointwise dep on the
// producer-consumer memref access for example). Check this and allow
// fusion accordingly.
if (hasCyclicDependence(srcAffineForOp)) {
LDBG() << "Source nest has a cyclic dependence.";
// Maximal fusion does not check for compute tolerance threshold; so
// perform the maximal fusion only when the redundanation computation
// is zero.
if (maximalFusion) {
auto srcForOp = cast<AffineForOp>(srcNode->op);
auto dstForOp = cast<AffineForOp>(dstNode->op);
int64_t sliceCost;
int64_t fusedLoopNestComputeCost;
auto fraction = getAdditionalComputeFraction(
srcForOp, dstForOp, maxLegalFusionDepth, depthSliceUnions,
sliceCost, fusedLoopNestComputeCost);
if (!fraction || fraction > 0) {
LDBG() << "Can't perform maximal fusion with a cyclic dependence "
<< "and non-zero additional compute.";
return;
}
} else {
// Set redundant computation tolerance to zero regardless of what
// the user specified. Without this, fusion would be invalid.
LDBG() << "Setting compute tolerance to zero since "
<< "source has a cylic dependence.";
computeToleranceThresholdToUse = 0;
}
}
// Check if fusion would be profitable. We skip profitability analysis
// for maximal fusion since we already know the maximal legal depth to
// fuse.
unsigned bestDstLoopDepth = maxLegalFusionDepth;
if (!maximalFusion) {
// Retrieve producer stores from the src loop.
SmallVector<Operation *, 2> producerStores;
for (Operation *op : srcNode->stores)
if (producerConsumerMemrefs.count(
cast<AffineWriteOpInterface>(op).getMemRef()))
producerStores.push_back(op);
assert(!producerStores.empty() && "Expected producer store");
if (!isFusionProfitable(srcAffineForOp, producerStores,
dstAffineForOp, depthSliceUnions,
maxLegalFusionDepth, &bestDstLoopDepth,
computeToleranceThresholdToUse)) {
continue;
}
}
assert(bestDstLoopDepth > 0 && "Unexpected loop fusion depth");
ComputationSliceState &bestSlice =
depthSliceUnions[bestDstLoopDepth - 1];
assert(!bestSlice.isEmpty() && "Missing slice union for depth");
// Determine if 'srcId' can be removed after fusion, taking into
// account remaining dependences, escaping memrefs and the fusion
// insertion point.
bool removeSrcNode = canRemoveSrcNodeAfterFusion(
srcId, dstId, bestSlice, fusedLoopInsPoint, srcEscapingMemRefs,
*mdg);
DenseSet<Value> privateMemrefs;
for (Value memref : producerConsumerMemrefs) {
if (canCreatePrivateMemRef(memref, srcEscapingMemRefs, srcId, dstId,
removeSrcNode)) {
// Create a private version of this memref.
LDBG() << "Creating private memref for " << memref;
// Create a private version of this memref.
privateMemrefs.insert(memref);
}
}
// Fuse computation slice of 'srcLoopNest' into 'dstLoopNest'.
fuseLoops(srcAffineForOp, dstAffineForOp, bestSlice);
dstNodeChanged = true;
LDBG() << "Fused src loop " << srcId << " into dst loop " << dstId
<< " at depth " << bestDstLoopDepth << ":";
LDBG() << dstAffineForOp;
// Move 'dstAffineForOp' before 'insertPointInst' if needed.
if (fusedLoopInsPoint != dstAffineForOp)
dstAffineForOp->moveBefore(fusedLoopInsPoint);
// Update edges between 'srcNode' and 'dstNode'.
mdg->updateEdges(srcNode->id, dstNode->id, privateMemrefs,
removeSrcNode);
// Create private memrefs.
if (!privateMemrefs.empty()) {
// Note the block into which fusion was performed. This can be used to
// place `alloc`s that create private memrefs.
Block *sliceInsertionBlock = bestSlice.insertPoint->getBlock();
// Gather stores for all the private-to-be memrefs.
DenseMap<Value, SmallVector<Operation *, 4>> privateMemRefToStores;
dstAffineForOp.walk([&](AffineWriteOpInterface storeOp) {
Value storeMemRef = storeOp.getMemRef();
if (privateMemrefs.count(storeMemRef) > 0)
privateMemRefToStores[storeMemRef].push_back(storeOp);
});
// Replace original memrefs with private memrefs. Note that all the
// loads and stores on these memrefs will be replaced with a new
// loads and stores. Any reference to the original ones becomes
// invalid after this point.
for (auto &memrefToStoresPair : privateMemRefToStores) {
ArrayRef<Operation *> storesForMemref = memrefToStoresPair.second;
Value newMemRef = createPrivateMemRef(
dstAffineForOp, storesForMemref, bestDstLoopDepth,
fastMemorySpace, sliceInsertionBlock, localBufSizeThreshold);
if (!newMemRef)
continue;
// Create new node in dependence graph for 'newMemRef' alloc op.
unsigned newMemRefNodeId = mdg->addNode(newMemRef.getDefiningOp());
// Add edge from 'newMemRef' node to dstNode.
mdg->addEdge(newMemRefNodeId, dstId, newMemRef);
}
// One or more entries for 'newMemRef' alloc op are inserted into
// the DenseMap mdg->nodes. Since an insertion may cause DenseMap to
// reallocate, update dstNode.
dstNode = mdg->getNode(dstId);
}
// Collect dst loop stats after memref privatization transformation.
LoopNestStateCollector dstLoopCollector;
dstLoopCollector.collect(dstAffineForOp);
// Clear and add back loads and stores.
mdg->clearNodeLoadAndStores(dstNode->id);
mdg->addToNode(
dstId, dstLoopCollector.loadOpInsts, dstLoopCollector.storeOpInsts,
dstLoopCollector.memrefLoads, dstLoopCollector.memrefStores,
dstLoopCollector.memrefFrees);
if (removeSrcNode) {
LDBG() << "Removing src loop " << srcId << " after fusion";
// srcNode is no longer valid after it is removed from mdg.
srcAffineForOp.erase();
mdg->removeNode(srcId);
srcNode = nullptr;
}
}
} while (dstNodeChanged);
}
/// Visit each node in the graph, and for each node, attempt to fuse it with
/// producer-consumer candidates. No fusion is performed when producers with a
/// user count greater than `maxSrcUserCount` for any of the memrefs involved
/// are encountered.
void fuseProducerConsumerNodes(unsigned maxSrcUserCount) {
LDBG() << "--- Producer/Consumer Fusion ---";
init();
while (!worklist.empty()) {
unsigned dstId = worklist.back();
worklist.pop_back();
performFusionsIntoDest(dstId, maxSrcUserCount);
}
}
// Visits each node in the graph, and for each node, attempts to fuse it with
// its sibling nodes (nodes which share a parent, but no dependence edges).
void fuseSiblingNodes() {
LDBG() << "--- Sibling Fusion ---";
init();
while (!worklist.empty()) {
unsigned dstId = worklist.back();
worklist.pop_back();
// Skip if this node was removed (fused into another node).
if (mdg->nodes.count(dstId) == 0)
continue;
// Get 'dstNode' into which to attempt fusion.
auto *dstNode = mdg->getNode(dstId);
// Skip if 'dstNode' is not a loop nest.
if (!isa<AffineForOp>(dstNode->op))
continue;
// Attempt to fuse 'dstNode' with its sibling nodes in the graph.
fuseWithSiblingNodes(dstNode);
}
}
// Attempt to fuse 'dstNode' with sibling nodes in the graph.
void fuseWithSiblingNodes(Node *dstNode) {
DenseSet<unsigned> visitedSibNodeIds;
std::pair<unsigned, Value> idAndMemref;
auto dstAffineForOp = cast<AffineForOp>(dstNode->op);
while (findSiblingNodeToFuse(dstNode, &visitedSibNodeIds, &idAndMemref)) {
unsigned sibId = idAndMemref.first;
Value memref = idAndMemref.second;
// TODO: Check that 'sibStoreOpInst' post-dominates all other
// stores to the same memref in 'sibNode' loop nest.
auto *sibNode = mdg->getNode(sibId);
// Compute an operation list insertion point for the fused loop
// nest which preserves dependences.
assert(sibNode->op->getBlock() == dstNode->op->getBlock());
Operation *insertPointInst =
sibNode->op->isBeforeInBlock(dstNode->op)
? mdg->getFusedLoopNestInsertionPoint(sibNode->id, dstNode->id)
: mdg->getFusedLoopNestInsertionPoint(dstNode->id, sibNode->id);
if (insertPointInst == nullptr)
continue;
// Check if fusion would be profitable and at what depth.
// Get unique 'sibNode' load op to 'memref'.
SmallVector<Operation *, 2> sibLoadOpInsts;
sibNode->getLoadOpsForMemref(memref, &sibLoadOpInsts);
// Currently findSiblingNodeToFuse searches for siblings with one load.
Operation *sibLoadOpInst = llvm::getSingleElement(sibLoadOpInsts);
// Gather 'dstNode' load ops to 'memref'.
SmallVector<Operation *, 2> dstLoadOpInsts;
dstNode->getLoadOpsForMemref(memref, &dstLoadOpInsts);
// It's possible this fusion is at an inner depth (i.e., there are common
// surrounding affine loops for the source and destination for ops). We
// need to get this number because the call to canFuseLoops needs to be
// passed the absolute depth. The max legal depth and the depths we try
// below are however *relative* and as such don't include the common
// depth.
SmallVector<AffineForOp, 4> surroundingLoops;
getAffineForIVs(*dstAffineForOp, &surroundingLoops);
unsigned numSurroundingLoops = surroundingLoops.size();
SmallVector<AffineForOp, 4> dstLoopIVs;
getAffineForIVs(*dstLoadOpInsts[0], &dstLoopIVs);
unsigned dstLoopDepthTest = dstLoopIVs.size() - numSurroundingLoops;
auto sibAffineForOp = cast<AffineForOp>(sibNode->op);
// Compute loop depth and slice union for fusion.
SmallVector<ComputationSliceState, 8> depthSliceUnions;
depthSliceUnions.resize(dstLoopDepthTest);
unsigned maxLegalFusionDepth = 0;
FusionStrategy strategy(memref);
for (unsigned i = 1; i <= dstLoopDepthTest; ++i) {
FusionResult result =
affine::canFuseLoops(sibAffineForOp, dstAffineForOp,
/*dstLoopDepth=*/i + numSurroundingLoops,
&depthSliceUnions[i - 1], strategy);
if (result.value == FusionResult::Success)
maxLegalFusionDepth = i;
}
LDBG() << "Max legal depth for fusion: " << maxLegalFusionDepth;
// Skip if fusion is not feasible at any loop depths.
if (maxLegalFusionDepth == 0)
continue;
double computeToleranceThresholdToUse = computeToleranceThreshold;
// Cyclic dependences in the source nest may be violated when performing
// slicing-based fusion. They aren't actually violated in cases where no
// redundant execution of the source happens (1:1 pointwise dep on the
// producer-consumer memref access for example). Check this and allow
// fusion accordingly.
if (hasCyclicDependence(sibAffineForOp)) {
LDBG() << "Source nest has a cyclic dependence.";
// Maximal fusion does not check for compute tolerance threshold; so
// perform the maximal fusion only when the redundanation computation is
// zero.
if (maximalFusion) {
auto dstForOp = cast<AffineForOp>(dstNode->op);
int64_t sliceCost;
int64_t fusedLoopNestComputeCost;
auto fraction = getAdditionalComputeFraction(
sibAffineForOp, dstForOp, maxLegalFusionDepth, depthSliceUnions,
sliceCost, fusedLoopNestComputeCost);
if (!fraction || fraction > 0) {
LDBG() << "Can't perform maximal fusion with a cyclic dependence "
<< "and non-zero additional compute.";
return;
}
} else {
// Set redundant computation tolerance to zero regardless of what the
// user specified. Without this, fusion would be invalid.
LDBG() << "Setting compute tolerance to zero since "
<< "source has a cyclic dependence.";
computeToleranceThresholdToUse = 0.0;
}
}
unsigned bestDstLoopDepth = maxLegalFusionDepth;
if (!maximalFusion) {
// Check if fusion would be profitable. For sibling fusion, the sibling
// load op is treated as the src "store" op for fusion profitability
// purposes. The footprint of the load in the slice relative to the
// unfused source's determines reuse.
if (!isFusionProfitable(sibAffineForOp, sibLoadOpInst, dstAffineForOp,
depthSliceUnions, maxLegalFusionDepth,
&bestDstLoopDepth,
computeToleranceThresholdToUse))
continue;
}
assert(bestDstLoopDepth > 0 && "Unexpected loop fusion depth");
const ComputationSliceState &bestSlice =
depthSliceUnions[bestDstLoopDepth - 1];
assert(!bestSlice.isEmpty() &&
"Fusion depth has no computed slice union");
// Do not perform sibling fusion if it isn't maximal. We always remove the
// sibling node and as such fusion shouldn't be performed if a part of the
// slice is used in the destination.
auto isMaximal = bestSlice.isMaximal();
if (!isMaximal.value_or(false)) {
LDBG() << "Slice isn't maximal; not performing sibling fusion.";
continue;
}
// Check if source loop is being inserted in the innermost
// destination loop. Based on this, the fused loop may be optimized
// further inside `fuseLoops`.
bool isInnermostInsertion = (bestDstLoopDepth == dstLoopDepthTest);
// Fuse computation slice of 'sibLoopNest' into 'dstLoopNest'.
affine::fuseLoops(sibAffineForOp, dstAffineForOp, bestSlice,
isInnermostInsertion);
auto dstForInst = cast<AffineForOp>(dstNode->op);
// Update operation position of fused loop nest (if needed).
if (insertPointInst != dstForInst)
dstForInst->moveBefore(insertPointInst);
LDBG() << "Fused sibling nest " << sibId << " into destination nest "
<< dstNode->id << " at depth " << bestDstLoopDepth << ":";
LDBG() << dstAffineForOp;
// Update data dependence graph state post fusion.
updateStateAfterSiblingFusion(sibNode, dstNode);
// Remove old sibling loop nest.
// Get op before we invalidate the MDG node.
Operation *op = sibNode->op;
mdg->removeNode(sibNode->id);
op->erase();
}
}
// Searches block argument uses and the graph from 'dstNode' looking for a
// fusion candidate sibling node which shares no dependences with 'dstNode'
// but which loads from the same memref. Returns true and sets
// 'idAndMemrefToFuse' on success. Returns false otherwise.
bool findSiblingNodeToFuse(Node *dstNode,
DenseSet<unsigned> *visitedSibNodeIds,
std::pair<unsigned, Value> *idAndMemrefToFuse) {
// Returns true if 'sibNode' can be fused with 'dstNode' for input reuse
// on 'memref'.
auto canFuseWithSibNode = [&](Node *sibNode, Value memref) {
// Skip if 'outEdge' is not a read-after-write dependence.
// TODO: Remove restrict to single load op restriction.
if (sibNode->getLoadOpCount(memref) != 1)
return false;
// Skip if there exists a path of dependent edges between
// 'sibNode' and 'dstNode'.
if (mdg->hasDependencePath(sibNode->id, dstNode->id) ||
mdg->hasDependencePath(dstNode->id, sibNode->id))
return false;
// Skip sib node if it loads to (and stores from) the same memref on
// which it also has an input dependence edge.
DenseSet<Value> loadAndStoreMemrefSet;
sibNode->getLoadAndStoreMemrefSet(&loadAndStoreMemrefSet);
if (llvm::any_of(loadAndStoreMemrefSet, [=](Value memref) {
return mdg->getIncomingMemRefAccesses(sibNode->id, memref) > 0;
}))
return false;
// Check that all stores are to the same memref if any.
DenseSet<Value> storeMemrefs;
for (auto *storeOpInst : sibNode->stores) {
storeMemrefs.insert(
cast<AffineWriteOpInterface>(storeOpInst).getMemRef());
}
return storeMemrefs.size() <= 1;
};
// Search for siblings which load the same memref block argument.
Block *block = dstNode->op->getBlock();
for (unsigned i = 0, e = block->getNumArguments(); i != e; ++i) {
for (Operation *user : block->getArgument(i).getUsers()) {
auto loadOp = dyn_cast<AffineReadOpInterface>(user);
if (!loadOp)
continue;
// Gather loops surrounding 'use'.
SmallVector<AffineForOp, 4> loops;
getAffineForIVs(*user, &loops);
// Skip 'use' if it is not within a loop nest.
// Find the surrounding affine.for nested immediately within the
// block.
auto *it = llvm::find_if(loops, [&](AffineForOp loop) {
return loop->getBlock() == &mdg->block;
});
// Skip 'use' if it is not within a loop nest in `block`.
if (it == loops.end())
continue;
Node *sibNode = mdg->getForOpNode(*it);
assert(sibNode != nullptr);
// Skip 'use' if it not a sibling to 'dstNode'.
if (sibNode->id == dstNode->id)
continue;
// Skip 'use' if it has been visited.
if (visitedSibNodeIds->count(sibNode->id) > 0)
continue;
// Skip 'use' if it does not load from the same memref as 'dstNode'.
auto memref = loadOp.getMemRef();
if (dstNode->getLoadOpCount(memref) == 0)
continue;
// Check if 'sibNode/dstNode' can be input-reuse fused on 'memref'.
if (canFuseWithSibNode(sibNode, memref)) {
visitedSibNodeIds->insert(sibNode->id);
idAndMemrefToFuse->first = sibNode->id;
idAndMemrefToFuse->second = memref;
return true;
}
}
}
// Search for siblings by following edges through an intermediate src node.
// Collect candidate 'dstNode' input edges in 'inEdges'.
SmallVector<MemRefDependenceGraph::Edge, 2> inEdges;
mdg->forEachMemRefInputEdge(
dstNode->id, [&](MemRefDependenceGraph::Edge inEdge) {
// Add 'inEdge' if it is a read-after-write dependence or an edge
// from a memref defining op (e.g. view-like op or alloc op).
if (dstNode->getLoadOpCount(inEdge.value) > 0 &&
(mdg->getNode(inEdge.id)->getStoreOpCount(inEdge.value) > 0 ||
inEdge.value.getDefiningOp() == mdg->getNode(inEdge.id)->op))
inEdges.push_back(inEdge);
});
// Search for sibling nodes to fuse by visiting output edges from each input
// edge in 'inEdges'.
for (auto &inEdge : inEdges) {
// Collect candidate output edges from each node 'inEdge.id' in 'inEdges'.
SmallVector<MemRefDependenceGraph::Edge, 2> outEdges;
mdg->forEachMemRefOutputEdge(
inEdge.id, [&](MemRefDependenceGraph::Edge outEdge) {
unsigned sibNodeId = outEdge.id;
if (visitedSibNodeIds->count(sibNodeId) > 0)
return;
// Skip output edge if not a sibling using the same memref.
if (outEdge.id == dstNode->id || outEdge.value != inEdge.value)
return;
auto *sibNode = mdg->getNode(sibNodeId);
if (!isa<AffineForOp>(sibNode->op))
return;
// Check if 'sibNode/dstNode' can be input-reuse fused on 'memref'.
if (canFuseWithSibNode(sibNode, outEdge.value)) {
// Add candidate 'outEdge' to sibling node.
outEdges.push_back(outEdge);
}
});
// Add first candidate if any were returned.
if (!outEdges.empty()) {
visitedSibNodeIds->insert(outEdges[0].id);
idAndMemrefToFuse->first = outEdges[0].id;
idAndMemrefToFuse->second = outEdges[0].value;
return true;
}
}
return false;
}
/// Update data dependence graph state to reflect sibling fusion of 'sibNode'
/// into 'dstNode'.
void updateStateAfterSiblingFusion(Node *sibNode, Node *dstNode) {
// Update 'sibNode' and 'dstNode' input/output edges to reflect fusion.
mdg->updateEdges(sibNode->id, dstNode->id);
// Collect dst loop stats after memref privatization transformation.
auto dstForInst = cast<AffineForOp>(dstNode->op);
LoopNestStateCollector dstLoopCollector;
dstLoopCollector.collect(dstForInst);
// Clear and add back loads and stores
mdg->clearNodeLoadAndStores(dstNode->id);
mdg->addToNode(dstNode->id, dstLoopCollector.loadOpInsts,
dstLoopCollector.storeOpInsts, dstLoopCollector.memrefLoads,
dstLoopCollector.memrefStores, dstLoopCollector.memrefFrees);
}
// Clean up any allocs with no users.
void eraseUnusedMemRefAllocations() {
for (auto &pair : mdg->memrefEdgeCount) {
if (pair.second > 0)
continue;
auto memref = pair.first;
// Skip if there exist other uses (return operation or function calls).
if (!memref.use_empty())
continue;
// Use list expected to match the dep graph info.
auto *op = memref.getDefiningOp();
if (isa_and_nonnull<memref::AllocOp>(op))
op->erase();
}
}
};
} // namespace
/// Run fusion on `block`.
void LoopFusion::runOnBlock(Block *block) {
MemRefDependenceGraph g(*block);
if (!g.init()) {
LDBG() << "MDG init failed";
return;
}
std::optional<unsigned> fastMemorySpaceOpt;
if (fastMemorySpace.hasValue())
fastMemorySpaceOpt = fastMemorySpace;
unsigned localBufSizeThresholdBytes = localBufSizeThreshold * 1024;
GreedyFusion fusion(&g, localBufSizeThresholdBytes, fastMemorySpaceOpt,
maximalFusion, computeToleranceThreshold);
if (affineFusionMode == FusionMode::ProducerConsumer)
fusion.runProducerConsumerFusionOnly();
else if (affineFusionMode == FusionMode::Sibling)
fusion.runSiblingFusionOnly();
else
fusion.runGreedyFusion();
}
void LoopFusion::runOnOperation() {
// Call fusion on every op that has at least two affine.for nests (in post
// order).
getOperation()->walk([&](Operation *op) {
for (Region &region : op->getRegions()) {
for (Block &block : region.getBlocks()) {
auto affineFors = block.getOps<AffineForOp>();
if (!affineFors.empty() && !llvm::hasSingleElement(affineFors))
runOnBlock(&block);
}
}
});
}
std::unique_ptr<Pass> mlir::affine::createLoopFusionPass(
unsigned fastMemorySpace, uint64_t localBufSizeThreshold,
bool maximalFusion, enum FusionMode affineFusionMode) {
return std::make_unique<LoopFusion>(fastMemorySpace, localBufSizeThreshold,
maximalFusion, affineFusionMode);
}