Skip to content

[mlir][scf] Considering defining operators of indices when fusing scf::ParallelOp #80145

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
Feb 1, 2024
Merged

[mlir][scf] Considering defining operators of indices when fusing scf::ParallelOp #80145

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
27 changes: 25 additions & 2 deletions mlir/lib/Dialect/SCF/Transforms/ParallelLoopFusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
#include "mlir/IR/Builders.h"
#include "mlir/IR/IRMapping.h"
#include "mlir/IR/OpDefinition.h"
#include "mlir/IR/OperationSupport.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"

namespace mlir {
Expand Down Expand Up @@ -102,8 +103,30 @@ static bool haveNoReadsAfterWriteExceptSameIndex(
return WalkResult::interrupt();
for (int i = 0, e = storeIndices.size(); i < e; ++i) {
if (firstToSecondPloopIndices.lookupOrDefault(storeIndices[i]) !=
loadIndices[i])
return WalkResult::interrupt();
loadIndices[i]) {
auto *storeIndexDefOp = storeIndices[i].getDefiningOp();
auto *loadIndexDefOp = loadIndices[i].getDefiningOp();
if (storeIndexDefOp && loadIndexDefOp) {
if (!isMemoryEffectFree(storeIndexDefOp))
return WalkResult::interrupt();
if (!isMemoryEffectFree(loadIndexDefOp))
return WalkResult::interrupt();
if (!OperationEquivalence::isEquivalentTo(
storeIndexDefOp, loadIndexDefOp,
[&](Value storeIndex, Value loadIndex) {
if (firstToSecondPloopIndices.lookupOrDefault(storeIndex) !=
firstToSecondPloopIndices.lookupOrDefault(loadIndex))
return failure();
else
return success();
},
/*markEquivalent=*/nullptr,
OperationEquivalence::Flags::IgnoreLocations)) {
return WalkResult::interrupt();
}
} else
return WalkResult::interrupt();
}
}
return WalkResult::advance();
});
Expand Down
95 changes: 95 additions & 0 deletions mlir/test/Dialect/SCF/parallel-loop-fusion.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -480,3 +480,98 @@ func.func @do_not_fuse_multiple_stores_on_diff_indices(
// CHECK: scf.reduce
// CHECK: }
// CHECK: memref.dealloc [[SUM]]

// -----

func.func @fuse_same_indices_by_affine_apply(
%A: memref<2x2xf32>, %B: memref<2x2xf32>) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%sum = memref.alloc() : memref<2x3xf32>
scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
%B_elem = memref.load %B[%i, %j] : memref<2x2xf32>
%1 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%i, %j)
memref.store %B_elem, %sum[%i, %1] : memref<2x3xf32>
scf.reduce
}
scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
%1 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%i, %j)
%sum_elem = memref.load %sum[%i, %1] : memref<2x3xf32>
%A_elem = memref.load %A[%i, %j] : memref<2x2xf32>
%product = arith.mulf %sum_elem, %A_elem : f32
memref.store %product, %B[%i, %j] : memref<2x2xf32>
scf.reduce
}
memref.dealloc %sum : memref<2x3xf32>
return
}
// CHECK: #[[$MAP:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
// CHECK-LABEL: fuse_same_indices_by_affine_apply
// CHECK-SAME: (%[[ARG0:.*]]: memref<2x2xf32>, %[[ARG1:.*]]: memref<2x2xf32>) {
// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
// CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index
// CHECK: %[[ALLOC:.*]] = memref.alloc() : memref<2x3xf32>
// CHECK-NEXT: scf.parallel (%[[ARG2:.*]], %[[ARG3:.*]]) = (%[[C0]], %[[C0]]) to (%[[C2]], %[[C2]]) step (%[[C1]], %[[C1]]) {
// CHECK-NEXT: %[[S0:.*]] = memref.load %[[ARG1]][%[[ARG2]], %[[ARG3]]] : memref<2x2xf32>
// CHECK-NEXT: %[[S1:.*]] = affine.apply #[[$MAP]](%[[ARG2]], %[[ARG3]])
// CHECK-NEXT: memref.store %[[S0]], %[[ALLOC]][%[[ARG2]], %[[S1]]] : memref<2x3xf32>
// CHECK-NEXT: %[[S2:.*]] = affine.apply #[[$MAP]](%[[ARG2]], %[[ARG3]])
// CHECK-NEXT: %[[S3:.*]] = memref.load %[[ALLOC]][%[[ARG2]], %[[S2]]] : memref<2x3xf32>
// CHECK-NEXT: %[[S4:.*]] = memref.load %[[ARG0]][%[[ARG2]], %[[ARG3]]] : memref<2x2xf32>
// CHECK-NEXT: %[[S5:.*]] = arith.mulf %[[S3]], %[[S4]] : f32
// CHECK-NEXT: memref.store %[[S5]], %[[ARG1]][%[[ARG2]], %[[ARG3]]] : memref<2x2xf32>
// CHECK-NEXT: scf.reduce
// CHECK-NEXT: }
// CHECK-NEXT: memref.dealloc %[[ALLOC]] : memref<2x3xf32>
// CHECK-NEXT: return

// -----

func.func @do_not_fuse_affine_apply_to_non_ind_var(
%A: memref<2x2xf32>, %B: memref<2x2xf32>, %OffsetA: index, %OffsetB: index) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%sum = memref.alloc() : memref<2x3xf32>
scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
%B_elem = memref.load %B[%i, %j] : memref<2x2xf32>
%1 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%i, %OffsetA)
memref.store %B_elem, %sum[%i, %1] : memref<2x3xf32>
scf.reduce
}
scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
%1 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%i, %OffsetB)
%sum_elem = memref.load %sum[%i, %1] : memref<2x3xf32>
%A_elem = memref.load %A[%i, %j] : memref<2x2xf32>
%product = arith.mulf %sum_elem, %A_elem : f32
memref.store %product, %B[%i, %j] : memref<2x2xf32>
scf.reduce
}
memref.dealloc %sum : memref<2x3xf32>
return
}
// CHECK: #[[$MAP:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
// CHECK-LABEL: do_not_fuse_affine_apply_to_non_ind_var
// CHECK-SAME: (%[[ARG0:.*]]: memref<2x2xf32>, %[[ARG1:.*]]: memref<2x2xf32>, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index) {
// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
// CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index
// CHECK: %[[ALLOC:.*]] = memref.alloc() : memref<2x3xf32>
// CHECK-NEXT: scf.parallel (%[[ARG4:.*]], %[[ARG5:.*]]) = (%[[C0]], %[[C0]]) to (%[[C2]], %[[C2]]) step (%[[C1]], %[[C1]]) {
// CHECK-NEXT: %[[S0:.*]] = memref.load %[[ARG1]][%[[ARG4]], %[[ARG5]]] : memref<2x2xf32>
// CHECK-NEXT: %[[S1:.*]] = affine.apply #[[$MAP]](%[[ARG4]], %[[ARG2]])
// CHECK-NEXT: memref.store %[[S0]], %[[ALLOC]][%[[ARG4]], %[[S1]]] : memref<2x3xf32>
// CHECK-NEXT: scf.reduce
// CHECK-NEXT: }
// CHECK-NEXT: scf.parallel (%[[ARG4:.*]], %[[ARG5:.*]]) = (%[[C0]], %[[C0]]) to (%[[C2]], %[[C2]]) step (%[[C1]], %[[C1]]) {
// CHECK-NEXT: %[[S0:.*]] = affine.apply #[[$MAP]](%[[ARG4]], %[[ARG3]])
// CHECK-NEXT: %[[S1:.*]] = memref.load %[[ALLOC]][%[[ARG4]], %[[S0]]] : memref<2x3xf32>
// CHECK-NEXT: %[[S2:.*]] = memref.load %[[ARG0]][%[[ARG4]], %[[ARG5]]] : memref<2x2xf32>
// CHECK-NEXT: %[[S3:.*]] = arith.mulf %[[S1]], %[[S2]] : f32
// CHECK-NEXT: memref.store %[[S3]], %[[ARG1]][%[[ARG4]], %[[ARG5]]] : memref<2x2xf32>
// CHECK-NEXT: scf.reduce
// CHECK-NEXT: }
// CHECK-NEXT: memref.dealloc %[[ALLOC]] : memref<2x3xf32>
// CHECK-NEXT: return