Skip to content

[mlir][vector] Add support for yielding loop bounds in scf.for distribution. #163443

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 4 commits into from
Oct 17, 2025
Merged

[mlir][vector] Add support for yielding loop bounds in scf.for distribution. #163443

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
d9ad36d
add fix
charithaintc Oct 14, 2025
c0fd573
Merge branch 'main' into loop_bounds_fix_in_scf_for
charithaintc Oct 15, 2025
7224394
fix
charithaintc Oct 15, 2025
e3c160f
Merge branch 'main' into loop_bounds_fix_in_scf_for
charithaintc Oct 17, 2025
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
26 changes: 19 additions & 7 deletions mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2032,11 +2032,19 @@ struct WarpOpScfForOp : public WarpDistributionPattern {
}

// Newly created `WarpOp` will yield values in following order:
// 1. All init args of the `ForOp`.
// 2. All escaping values.
// 3. All non-`ForOp` yielded values.
// 1. Loop bounds.
// 2. All init args of the `ForOp`.
// 3. All escaping values.
// 4. All non-`ForOp` yielded values.
SmallVector<Value> newWarpOpYieldValues;
SmallVector<Type> newWarpOpDistTypes;
newWarpOpYieldValues.insert(
newWarpOpYieldValues.end(),
{forOp.getLowerBound(), forOp.getUpperBound(), forOp.getStep()});
newWarpOpDistTypes.insert(newWarpOpDistTypes.end(),
{forOp.getLowerBound().getType(),
forOp.getUpperBound().getType(),
forOp.getStep().getType()});
for (auto [i, initArg] : llvm::enumerate(forOp.getInitArgs())) {
newWarpOpYieldValues.push_back(initArg);
// Compute the distributed type for this init arg.
Expand Down Expand Up @@ -2072,20 +2080,24 @@ struct WarpOpScfForOp : public WarpDistributionPattern {

// Next, we create a new `ForOp` with the init args yielded by the new
// `WarpOp`.
const unsigned initArgsStartIdx = 3; // After loop bounds.
const unsigned escapingValuesStartIdx =
initArgsStartIdx +
forOp.getInitArgs().size(); // `ForOp` init args are positioned before
// escaping values in the new `WarpOp`.
SmallVector<Value> newForOpOperands;
for (size_t i = 0; i < escapingValuesStartIdx; ++i)
for (size_t i = initArgsStartIdx; i < escapingValuesStartIdx; ++i)
newForOpOperands.push_back(newWarpOp.getResult(newIndices[i]));

// Create a new `ForOp` outside the new `WarpOp` region.
OpBuilder::InsertionGuard g(rewriter);
rewriter.setInsertionPointAfter(newWarpOp);
auto newForOp = scf::ForOp::create(
rewriter, forOp.getLoc(), forOp.getLowerBound(), forOp.getUpperBound(),
forOp.getStep(), newForOpOperands, /*bodyBuilder=*/nullptr,
forOp.getUnsignedCmp());
rewriter, forOp.getLoc(),
/**LowerBound=**/ newWarpOp.getResult(newIndices[0]),
/**UpperBound=**/ newWarpOp.getResult(newIndices[1]),
/**Step=**/ newWarpOp.getResult(newIndices[2]), newForOpOperands,
/*bodyBuilder=*/nullptr, forOp.getUnsignedCmp());
// Next, we insert a new `WarpOp` (called inner `WarpOp`) inside the
// newly created `ForOp`. This `WarpOp` will contain all ops that were
// contained within the original `ForOp` body.
Expand Down
35 changes: 35 additions & 0 deletions mlir/test/Dialect/Vector/vector-warp-distribute.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -473,6 +473,41 @@ func.func @warp_scf_for_use_from_above(%arg0: index) {
return
}

// -----
// CHECK-PROP-LABEL: func.func @warp_scf_for_local_loop_bounds
// CHECK-PROP: (%{{.*}}: index, %[[ARG1:[a-zA-Z0-9]+]]: index) {
// CHECK-PROP: %[[W:.*]] = gpu.warp_execute_on_lane_0(%{{.*}})[32] args(%[[ARG1]] : index) -> (vector<4xf32>) {
// CHECK-PROP: ^bb0(%{{.*}}: index):
// CHECK-PROP: %[[T2:.*]] = "some_def"() : () -> vector<128xf32>
// CHECK-PROP: gpu.yield %[[T2]] : vector<128xf32>
// CHECK-PROP: }
// CHECK-PROP: %[[FOR:.*]] = scf.for %{{.*}} to %[[ARG1]] step %{{.*}} iter_args(%{{.*}}) -> (vector<4xf32>) {
// CHECK-PROP: %[[W2:.*]] = gpu.warp_execute_on_lane_0(%{{.*}})[32]
// CHECK-PROP-SAME: args(%{{.*}} : vector<4xf32>) -> (vector<4xf32>) {
// CHECK-PROP: ^bb0(%{{.*}}: vector<128xf32>):
// CHECK-PROP: gpu.yield %{{.*}} : vector<128xf32>
// CHECK-PROP: }
// CHECK-PROP: scf.yield %[[W2]] : vector<4xf32>
// CHECK-PROP: }
// CHECK-PROP: "some_use"(%[[FOR]]) : (vector<4xf32>) -> ()
// CHECK-PROP: return
func.func @warp_scf_for_local_loop_bounds(%arg0: index, %bound: index) {
%c1 = arith.constant 1 : index
%c0 = arith.constant 0 : index
%0 = gpu.warp_execute_on_lane_0(%arg0)[32]
args(%bound : index) -> (vector<4xf32>) {
^bb0(%arg1: index):
%ini = "some_def"() : () -> (vector<128xf32>)
%3 = scf.for %arg3 = %c0 to %arg1 step %c1 iter_args(%arg4 = %ini) -> (vector<128xf32>) {
%acc = "some_def"(%arg4) : (vector<128xf32>) -> (vector<128xf32>)
scf.yield %acc : vector<128xf32>
}
gpu.yield %3 : vector<128xf32>
}
"some_use"(%0) : (vector<4xf32>) -> ()
return
}

// -----

// CHECK-PROP-LABEL: func @warp_scf_for_swap(
Expand Down