Skip to content

[MLIR][Affine] Fix fusion private memref creation for multiple producer stores #130365

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Mar 8, 2025
Merged

[MLIR][Affine] Fix fusion private memref creation for multiple producer stores #130365

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 26 additions & 5 deletions mlir/lib/Dialect/Affine/Transforms/LoopFusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -328,14 +328,34 @@ static std::optional<double> getAdditionalComputeFraction(
// 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.
// 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;
})) {
LLVM_DEBUG(llvm::dbgs()
<< "Private memref creation unsupported for multiple producer "
"stores with different access functions.\n");
return nullptr;
}

Operation *srcStoreOp = storeOps[0];

// Create builder to insert alloc op just before 'forOp'.
Expand Down Expand Up @@ -432,6 +452,8 @@ static Value createPrivateMemRef(AffineForOp forOp,
assert(succeeded(res) &&
"replaceAllMemrefUsesWith should always succeed here");
(void)res;
LLVM_DEBUG(llvm::dbgs() << "Created private memref of type: " << newMemRefType
<< '\n');
return newMemRef;
}

Expand Down Expand Up @@ -1123,13 +1145,12 @@ struct GreedyFusion {
// loads and stores. Any reference to the original ones becomes
// invalid after this point.
for (auto &memrefToStoresPair : privateMemRefToStores) {
// TODO: Use union of memref write regions to compute
// private memref footprint.
SmallVector<Operation *, 4> &storesForMemref =
memrefToStoresPair.second;
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.
Expand Down
16 changes: 10 additions & 6 deletions mlir/test/Dialect/Affine/loop-fusion-2.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,16 +39,20 @@ func.func @should_fuse_at_depth_above_loop_carried_dependence(%arg0: memref<64x4

// We can fuse source loop nest '%i0' into dst loop nest '%i2', but the
// depth at which we can insert the src loop nest slice into the dst loop
// lest must be decreased because of a loop carried dependence on loop '%i3'.
// nest must be decreased because of a loop carried dependence on loop '%i3'.
// As a result, the source loop nest is inserted at dst loop nest depth 1,
// just above the loop with the carried dependence. In addition, the source
// loop nest iteration bounds on its loop '%i1' are reduced to 1, so the
// memref size can be reduced to 128x1xf32.
// memref size can be reduced to 64x1xf32.

// CHECK: memref.alloc() : memref<64x1xf32>
// In this case, since we have multiple producer stores (for %out) with
// different access functions and we don't yet support private memref
// computation in such cases, a 64x1 private memref isn't created.

// CHECK: memref.alloc() : memref<64x4xf32>
// CHECK: affine.for %{{.*}} = 0 to 4 {
// CHECK-NEXT: affine.for %{{.*}} = 0 to 64 {
// CHECK-NEXT: affine.store %{{.*}}, %{{.*}}[%{{.*}}, 0] : memref<64x1xf32>
// CHECK-NEXT: affine.store %{{.*}}, %{{.*}}[%{{.*}}, %{{.*}}] : memref<64x4xf32>
// CHECK-NEXT: }
// CHECK-NEXT: affine.for %{{.*}} = 0 to 4 {
// CHECK-NEXT: affine.for %{{.*}} = 0 to 16 {
Expand All @@ -62,9 +66,9 @@ func.func @should_fuse_at_depth_above_loop_carried_dependence(%arg0: memref<64x4
// CHECK-NEXT: }
// CHECK-NEXT: affine.for %{{.*}} = 0 to 16 {
// CHECK-NEXT: %{{.*}} = "op2"() : () -> f32
// CHECK: affine.load %{{.*}}[%{{.*}} * 16 + %{{.*}}, 0] : memref<64x1xf32>
// CHECK: affine.load %{{.*}}[%{{.*}} * 16 + %{{.*}}, %{{.*}}] : memref<64x4xf32>
// CHECK-NEXT: arith.addf %{{.*}}, %{{.*}} : f32
// CHECK: affine.store %{{.*}}, %{{.*}}[%{{.*}} * 16 + %{{.*}}, 0] : memref<64x1xf32>
// CHECK: affine.store %{{.*}}, %{{.*}}[%{{.*}} * 16 + %{{.*}}, %{{.*}}] : memref<64x4xf32>
// CHECK-NEXT: }
// CHECK-NEXT: }
// CHECK-NEXT: }
Expand Down
44 changes: 44 additions & 0 deletions mlir/test/Dialect/Affine/loop-fusion-4.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -622,3 +622,47 @@ func.func @zero_tolerance(%arg0: memref<65536xcomplex<f64>>, %arg1: memref<30x13
}
func.func private @__external_levelwise_forward_ntt(memref<30x131072xi64>)
func.func private @__external_reduce_barrett(i64, i64, i64, i64, i128) -> i64

// An unrolled loop nest. Fusion here should correctly fuse while preserving
// dependences between store-load pairs of the same memref. A private memref
// of size 1x1x1 can't be created.

// PRODUCER-CONSUMER-MAXIMAL-LABEL: func @unrolled
func.func @unrolled(%arg0: memref<2x4xf32>, %arg1: memref<1x2x4xf32>) {
%alloc = memref.alloc() : memref<1x2x4xf32>
affine.for %i = 0 to 1 {
%0 = affine.load %arg0[0, 0] : memref<2x4xf32>
%1 = affine.load %arg0[0, 1] : memref<2x4xf32>
%2 = affine.load %arg0[0, 2] : memref<2x4xf32>
%3 = affine.load %arg0[0, 3] : memref<2x4xf32>
%4 = affine.load %arg0[1, 0] : memref<2x4xf32>
%5 = affine.load %arg0[1, 1] : memref<2x4xf32>
%6 = affine.load %arg0[1, 2] : memref<2x4xf32>
%7 = affine.load %arg0[1, 3] : memref<2x4xf32>

affine.store %0, %alloc[0, 0, 0] : memref<1x2x4xf32>
affine.store %1, %alloc[0, 0, 1] : memref<1x2x4xf32>
affine.store %2, %alloc[0, 0, 2] : memref<1x2x4xf32>
affine.store %3, %alloc[0, 0, 3] : memref<1x2x4xf32>
affine.store %4, %alloc[0, 1, 0] : memref<1x2x4xf32>
affine.store %5, %alloc[0, 1, 1] : memref<1x2x4xf32>
affine.store %6, %alloc[0, 1, 2] : memref<1x2x4xf32>
affine.store %7, %alloc[0, 1, 3] : memref<1x2x4xf32>
}

affine.for %i = 0 to 2 {
affine.for %j = 0 to 4 {
%8 = affine.load %alloc[0, %i, %j] : memref<1x2x4xf32>
%9 = arith.negf %8 : f32
affine.store %9, %arg1[0, %i, %j] : memref<1x2x4xf32>
}
}
// PRODUCER-CONSUMER-MAXIMAL: affine.for %{{.*}} = 0 to 2 {
// PRODUCER-CONSUMER-MAXIMAL-NEXT: affine.for %{{.*}} = 0 to 4 {
// PRODUCER-CONSUMER-MAXIMAL-NEXT: affine.load %{{.*}}[0, 0]
// PRODUCER-CONSUMER-MAXIMAL: affine.load %{{.*}}[1, 3]
// PRODUCER-CONSUMER-MAXIMAL: affine.store %{{.*}}[0, 0, 0]
// PRODUCER-CONSUMER-MAXIMAL: affine.store %{{.*}}[0, 1, 3]
// PRODUCER-CONSUMER-MAXIMAL: affine.load %{{.*}}[0, %{{.*}}, %{{.*}}]
return
}