Skip to content

AffineCFG: check lockstep dependences on the parallel check's relations #3405

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
wsmoses merged 1 commit into from
Oct 7, 2026
Merged

AffineCFG: check lockstep dependences on the parallel check's relations #3405

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
22 changes: 18 additions & 4 deletions src/enzyme_ad/jax/Passes/AffineCFG.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8843,17 +8843,31 @@ static bool isLoopMemoryLockStepExecutable(AffineForOp forOp) {
// Dep check depth would be number of enclosing loops + 1.
unsigned depth = ::getNestingDepth(forOp) + 1;

// Check dependences between all pairs of ops in 'loadAndStoreOps'.
// Check dependences between all pairs of ops in 'loadAndStoreOps', on the
// access relations isLoopMemoryParallel builds (loop-invariant terms
// abstracted, with what the nest's in-bounds accesses and the checks on the
// way to it establish). Accesses to different memrefs, or that only read,
// have none.
InvariantTerms terms(forOp);
for (auto *srcOp : loadAndStoreOps) {
MemRefAccess srcAccess(srcOp);
for (auto *dstOp : loadAndStoreOps) {
LLVM_DEBUG(llvm::dbgs() << "Checking dep\n"
<< "src: " << *srcOp << "\n"
<< "dst: " << *dstOp << "\n");
MemRefAccess dstAccess(dstOp);
SmallVector<DependenceComponent, 2> dcs;
DependenceResult result = checkMemrefAccessDependence(
srcAccess, dstAccess, depth, nullptr, &dcs);
DependenceResult result(DependenceResult::NoDependence);
if (srcAccess.memref == dstAccess.memref &&
(isa<AffineWriteOpInterface>(srcOp) ||
isa<AffineWriteOpInterface>(dstOp))) {
presburger::IntegerRelation srcRel(
presburger::PresburgerSpace::getRelationSpace()),
dstRel(presburger::PresburgerSpace::getRelationSpace());
if (failed(terms.accessRelation(srcOp, srcRel)) ||
failed(terms.accessRelation(dstOp, dstRel)))
return false;
result = checkAccessDependence(srcRel, dstRel, depth);
}

if (result.value == DependenceResult::Failure) {
LLVM_DEBUG(llvm::dbgs() << "Failed\n");
Expand Down
69 changes: 28 additions & 41 deletions test/lit_tests/raising/affine_to_stablehlo_if_masking.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,47 +78,34 @@ func.func @test_affine_if_masking(%arg0: memref<10xf32>) {
// CHECK-NEXT: return %[[a21]] : tensor<10xf32>
// CHECK-NEXT: }
// CHECK-NEXT: func.func private @test_if_else_masking_raised(%[[a1]]: tensor<10xf32>, %[[a22:.+]]: tensor<10xi1>) -> (tensor<10xf32>, tensor<10xi1>) {
// CHECK-NEXT: %[[a3]] = stablehlo.constant dense<0> : tensor<i64>
// CHECK-NEXT: %[[a5]] = stablehlo.constant dense<10> : tensor<i64>
// CHECK-NEXT: %[[a7]] = stablehlo.constant dense<1> : tensor<i64>
// CHECK-NEXT: %[[a2]]:3 = stablehlo.while(%[[a23:.+]] = %[[a3]], %[[a24:.+]] = %[[a1]], %[[a25:.+]] = %[[a22]]) : tensor<i64>, tensor<10xf32>, tensor<10xi1>
// CHECK-NEXT: cond {
// CHECK-NEXT: %[[a4]] = stablehlo.compare LT, %[[a23]], %[[a5]] : (tensor<i64>, tensor<i64>) -> tensor<i1>
// CHECK-NEXT: stablehlo.return %[[a4]] : tensor<i1>
// CHECK-NEXT: } do {
// CHECK-NEXT: %[[a4]] = stablehlo.dynamic_slice %[[a25]], %[[a23]], sizes = [1] : (tensor<10xi1>, tensor<i64>) -> tensor<1xi1>
// CHECK-NEXT: %[[a6]] = stablehlo.reshape %[[a4]] : (tensor<1xi1>) -> tensor<i1>
// CHECK-NEXT: %[[a8]] = stablehlo.dynamic_slice %[[a24]], %[[a23]], sizes = [1] : (tensor<10xf32>, tensor<i64>) -> tensor<1xf32>
// CHECK-NEXT: %[[a9]] = stablehlo.reshape %[[a8]] : (tensor<1xf32>) -> tensor<f32>
// CHECK-NEXT: %[[a11]] = arith.addf %[[a9]], %[[a9]] : tensor<f32>
// CHECK-NEXT: %[[a14]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a15]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a16]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a17]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a18]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a13]] = stablehlo.broadcast_in_dim %[[a11]], dims = [] : (tensor<f32>) -> tensor<1xf32>
// CHECK-NEXT: %[[a20]] = stablehlo.dynamic_slice %[[a24]], %[[a23]], sizes = [1] : (tensor<10xf32>, tensor<i64>) -> tensor<1xf32>
// CHECK-NEXT: %[[a21]] = stablehlo.broadcast_in_dim %[[a6]], dims = [] : (tensor<i1>) -> tensor<1xi1>
// CHECK-NEXT: %[[a26:.+]] = stablehlo.select %[[a21]], %[[a13]], %[[a20]] : tensor<1xi1>, tensor<1xf32>
// CHECK-NEXT: %[[a27:.+]] = stablehlo.dynamic_update_slice %[[a24]], %[[a26]], %[[a23]] : (tensor<10xf32>, tensor<1xf32>, tensor<i64>) -> tensor<10xf32>
// CHECK-NEXT: %[[a28:.+]] = stablehlo.not %[[a6]] : tensor<i1>
// CHECK-NEXT: %[[a29:.+]] = stablehlo.dynamic_slice %[[a27]], %[[a23]], sizes = [1] : (tensor<10xf32>, tensor<i64>) -> tensor<1xf32>
// CHECK-NEXT: %[[a30:.+]] = stablehlo.reshape %[[a29]] : (tensor<1xf32>) -> tensor<f32>
// CHECK-NEXT: %[[a31:.+]] = arith.mulf %[[a30]], %[[a30]] : tensor<f32>
// CHECK-NEXT: %[[a19]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a32:.+]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a33:.+]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a34:.+]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a35:.+]] = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %[[a36:.+]] = stablehlo.broadcast_in_dim %[[a31]], dims = [] : (tensor<f32>) -> tensor<1xf32>
// CHECK-NEXT: %[[a37:.+]] = stablehlo.dynamic_slice %[[a27]], %[[a23]], sizes = [1] : (tensor<10xf32>, tensor<i64>) -> tensor<1xf32>
// CHECK-NEXT: %[[a38:.+]] = stablehlo.broadcast_in_dim %[[a28]], dims = [] : (tensor<i1>) -> tensor<1xi1>
// CHECK-NEXT: %[[a39:.+]] = stablehlo.select %[[a38]], %[[a36]], %[[a37]] : tensor<1xi1>, tensor<1xf32>
// CHECK-NEXT: %[[a40:.+]] = stablehlo.dynamic_update_slice %[[a27]], %[[a39]], %[[a23]] : (tensor<10xf32>, tensor<1xf32>, tensor<i64>) -> tensor<10xf32>
// CHECK-NEXT: %[[a41:.+]] = stablehlo.add %[[a23]], %[[a7]] : tensor<i64>
// CHECK-NEXT: stablehlo.return %[[a41]], %[[a40]], %[[a25]] : tensor<i64>, tensor<10xf32>, tensor<10xi1>
// CHECK-NEXT: }
// CHECK-NEXT: return %[[a2]]#1, %[[a2]]#2 : tensor<10xf32>, tensor<10xi1>
// CHECK-NEXT: %0 = stablehlo.iota dim = 0 : tensor<10xi64>
// CHECK-NEXT: %c = stablehlo.constant dense<0> : tensor<10xi64>
// CHECK-NEXT: %1 = stablehlo.add %0, %c : tensor<10xi64>
// CHECK-NEXT: %c_0 = stablehlo.constant dense<1> : tensor<10xi64>
// CHECK-NEXT: %2 = stablehlo.multiply %1, %c_0 : tensor<10xi64>
// CHECK-NEXT: %c_1 = stablehlo.constant dense<0> : tensor<i64>
// CHECK-NEXT: %c_2 = stablehlo.constant dense<0> : tensor<i64>
// CHECK-NEXT: %3 = arith.addf %arg0, %arg0 : tensor<10xf32>
// CHECK-NEXT: %c_3 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_4 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_5 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_6 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_7 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_8 = stablehlo.constant dense<0> : tensor<i64>
// CHECK-NEXT: %4 = stablehlo.select %arg1, %3, %arg0 : tensor<10xi1>, tensor<10xf32>
// CHECK-NEXT: %5 = stablehlo.dynamic_update_slice %arg0, %4, %c_8 : (tensor<10xf32>, tensor<10xf32>, tensor<i64>) -> tensor<10xf32>
// CHECK-NEXT: %6 = stablehlo.not %arg1 : tensor<10xi1>
// CHECK-NEXT: %c_9 = stablehlo.constant dense<0> : tensor<i64>
// CHECK-NEXT: %7 = arith.mulf %5, %5 : tensor<10xf32>
// CHECK-NEXT: %c_10 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_11 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_12 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_13 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_14 = stablehlo.constant dense<0> : tensor<1xi64>
// CHECK-NEXT: %c_15 = stablehlo.constant dense<0> : tensor<i64>
// CHECK-NEXT: %8 = stablehlo.select %6, %7, %5 : tensor<10xi1>, tensor<10xf32>
// CHECK-NEXT: %9 = stablehlo.dynamic_update_slice %5, %8, %c_15 : (tensor<10xf32>, tensor<10xf32>, tensor<i64>) -> tensor<10xf32>
// CHECK-NEXT: return %9, %arg1 : tensor<10xf32>, tensor<10xi1>
// CHECK-NEXT: }
// CHECK-NEXT: func.func private @test_nested_if_masking_raised(%[[a1]]: tensor<10xf32>, %[[a22]]: tensor<10xi1>, %[[a42:.+]]: tensor<10xi1>) -> (tensor<10xf32>, tensor<10xi1>, tensor<10xi1>) {
// CHECK-NEXT: %[[a2]] = stablehlo.iota dim = 0 : tensor<10xi64>
Expand Down
64 changes: 64 additions & 0 deletions test/lit_tests/raising/affine_to_stablehlo_lockstep_invariant_terms.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,64 @@
// RUN: enzymexlamlir-opt %s "--pass-pipeline=builtin.module(raise-affine-to-stablehlo{enable_lockstep_for=true})" | FileCheck %s

// A loop of constant count whose accesses are offset by a product of
// symbols: the lockstep check now builds its dependence problems as the
// parallel check does, with that product abstracted, and the loop runs in
// lock step (one gather and one scatter) rather than as a while.
module {
func.func private @copy(%a: memref<64xf64, 1>, %b: memref<64xf64, 1>, %p: memref<2xi64, 1>) {
%s64 = affine.load %p[0] : memref<2xi64, 1>
%t64 = affine.load %p[1] : memref<2xi64, 1>
%s = arith.index_cast %s64 : i64 to index
%t = arith.index_cast %t64 : i64 to index
affine.parallel (%i) = (0) to (4) {
affine.for %k = 0 to 3 {
%v = affine.load %a[%k + %i * 3 + symbol(%s) * symbol(%t)] : memref<64xf64, 1>
affine.store %v, %b[%k + %i * 3 + symbol(%s) * symbol(%t)] : memref<64xf64, 1>
}
}
return
}
}

// CHECK: func.func private @copy_raised(%arg0: tensor<64xf64>, %arg1: tensor<64xf64>, %arg2: tensor<2xi64>) -> (tensor<64xf64>, tensor<64xf64>, tensor<2xi64>) {
// CHECK-NEXT: %0 = stablehlo.slice %arg2 [0:1] : (tensor<2xi64>) -> tensor<1xi64>
// CHECK-NEXT: %1 = stablehlo.reshape %0 : (tensor<1xi64>) -> tensor<i64>
// CHECK-NEXT: %2 = stablehlo.slice %arg2 [1:2] : (tensor<2xi64>) -> tensor<1xi64>
// CHECK-NEXT: %3 = stablehlo.reshape %2 : (tensor<1xi64>) -> tensor<i64>
// CHECK-NEXT: %4 = stablehlo.iota dim = 0 : tensor<4xi64>
// CHECK-NEXT: %c = stablehlo.constant dense<0> : tensor<4xi64>
// CHECK-NEXT: %5 = stablehlo.add %4, %c : tensor<4xi64>
// CHECK-NEXT: %c_0 = stablehlo.constant dense<1> : tensor<4xi64>
// CHECK-NEXT: %6 = stablehlo.multiply %5, %c_0 : tensor<4xi64>
// CHECK-NEXT: %7 = stablehlo.iota dim = 0 : tensor<3xi64>
// CHECK-NEXT: %c_1 = stablehlo.constant dense<0> : tensor<3xi64>
// CHECK-NEXT: %8 = stablehlo.add %7, %c_1 : tensor<3xi64>
// CHECK-NEXT: %c_2 = stablehlo.constant dense<1> : tensor<3xi64>
// CHECK-NEXT: %9 = stablehlo.multiply %8, %c_2 : tensor<3xi64>
// CHECK-NEXT: %c_3 = stablehlo.constant dense<3> : tensor<i64>
// CHECK-NEXT: %10 = stablehlo.broadcast_in_dim %c_3, dims = [] : (tensor<i64>) -> tensor<4xi64>
// CHECK-NEXT: %11 = stablehlo.multiply %6, %10 : tensor<4xi64>
// CHECK-NEXT: %12 = stablehlo.broadcast_in_dim %9, dims = [0] : (tensor<3xi64>) -> tensor<3x4xi64>
// CHECK-NEXT: %13 = stablehlo.broadcast_in_dim %11, dims = [1] : (tensor<4xi64>) -> tensor<3x4xi64>
// CHECK-NEXT: %14 = stablehlo.add %12, %13 : tensor<3x4xi64>
// CHECK-NEXT: %15 = stablehlo.multiply %1, %3 : tensor<i64>
// CHECK-NEXT: %16 = stablehlo.broadcast_in_dim %15, dims = [] : (tensor<i64>) -> tensor<3x4xi64>
// CHECK-NEXT: %17 = stablehlo.add %14, %16 : tensor<3x4xi64>
// CHECK-NEXT: %18 = stablehlo.reshape %17 : (tensor<3x4xi64>) -> tensor<3x4x1xi64>
// CHECK-NEXT: %19 = "stablehlo.gather"(%arg0, %18) <{dimension_numbers = #stablehlo.gather<collapsed_slice_dims = [0], start_index_map = [0], index_vector_dim = 2>, indices_are_sorted = false, slice_sizes = array<i64: 1>}> : (tensor<64xf64>, tensor<3x4x1xi64>) -> tensor<3x4xf64>
// CHECK-NEXT: %c_4 = stablehlo.constant dense<3> : tensor<i64>
// CHECK-NEXT: %20 = stablehlo.broadcast_in_dim %c_4, dims = [] : (tensor<i64>) -> tensor<4xi64>
// CHECK-NEXT: %21 = stablehlo.multiply %6, %20 : tensor<4xi64>
// CHECK-NEXT: %22 = stablehlo.broadcast_in_dim %9, dims = [0] : (tensor<3xi64>) -> tensor<3x4xi64>
// CHECK-NEXT: %23 = stablehlo.broadcast_in_dim %21, dims = [1] : (tensor<4xi64>) -> tensor<3x4xi64>
// CHECK-NEXT: %24 = stablehlo.add %22, %23 : tensor<3x4xi64>
// CHECK-NEXT: %25 = stablehlo.multiply %1, %3 : tensor<i64>
// CHECK-NEXT: %26 = stablehlo.broadcast_in_dim %25, dims = [] : (tensor<i64>) -> tensor<3x4xi64>
// CHECK-NEXT: %27 = stablehlo.add %24, %26 : tensor<3x4xi64>
// CHECK-NEXT: %28 = stablehlo.reshape %27 : (tensor<3x4xi64>) -> tensor<3x4x1xi64>
// CHECK-NEXT: %29 = "stablehlo.scatter"(%arg1, %28, %19) <{indices_are_sorted = false, scatter_dimension_numbers = #stablehlo.scatter<inserted_window_dims = [0], scatter_dims_to_operand_dims = [0], index_vector_dim = 2>, unique_indices = true}> ({
// CHECK-NEXT: ^bb0(%arg3: tensor<f64>, %arg4: tensor<f64>):
// CHECK-NEXT: stablehlo.return %arg4 : tensor<f64>
// CHECK-NEXT: }) : (tensor<64xf64>, tensor<3x4x1xi64>, tensor<3x4xf64>) -> tensor<64xf64>
// CHECK-NEXT: return %arg0, %29, %arg2 : tensor<64xf64>, tensor<64xf64>, tensor<2xi64>
// CHECK-NEXT: }
Loading