Skip to content

[mlir][Vector] Support vector.extract(xfer_read) folding with dynamic indices #143269

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 2 commits into from
Jun 16, 2025
Merged

[mlir][Vector] Support vector.extract(xfer_read) folding with dynamic indices #143269

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6bc6e3a
[mlir][Vector] Support `xfer_read(vector.extract))` folding with dyna…
dcaballe Jun 7, 2025
f98e348
Feedback
dcaballe Jun 16, 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
28 changes: 21 additions & 7 deletions mlir/lib/Dialect/Vector/Transforms/VectorTransferOpTransforms.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -886,17 +886,31 @@ class RewriteScalarExtractOfTransferRead
SmallVector<Value> newIndices(xferOp.getIndices().begin(),
xferOp.getIndices().end());
for (auto [i, pos] : llvm::enumerate(extractOp.getMixedPosition())) {
assert(isa<Attribute>(pos) && "Unexpected non-constant index");
int64_t offset = cast<IntegerAttr>(cast<Attribute>(pos)).getInt();
int64_t idx = newIndices.size() - extractOp.getNumIndices() + i;
OpFoldResult ofr = affine::makeComposedFoldedAffineApply(
rewriter, extractOp.getLoc(),
rewriter.getAffineSymbolExpr(0) + offset, {newIndices[idx]});
if (auto value = dyn_cast<Value>(ofr)) {

// Compute affine expression `newIndices[idx] + pos` where `pos` can be
// either a constant or a value.
OpFoldResult composedIdx;
if (auto attr = dyn_cast<Attribute>(pos)) {
int64_t offset = cast<IntegerAttr>(attr).getInt();
composedIdx = affine::makeComposedFoldedAffineApply(
rewriter, extractOp.getLoc(),
rewriter.getAffineSymbolExpr(0) + offset, {newIndices[idx]});
} else {
Value dynamicOffset = cast<Value>(pos);
AffineExpr sym0, sym1;
bindSymbols(rewriter.getContext(), sym0, sym1);
composedIdx = affine::makeComposedFoldedAffineApply(
rewriter, extractOp.getLoc(), sym0 + sym1,
{newIndices[idx], dynamicOffset});
}

// Update the corresponding index with the folded result.
if (auto value = dyn_cast<Value>(composedIdx)) {
newIndices[idx] = value;
} else {
newIndices[idx] = rewriter.create<arith::ConstantIndexOp>(
extractOp.getLoc(), *getConstantIntValue(ofr));
extractOp.getLoc(), *getConstantIntValue(composedIdx));
}
}
if (isa<MemRefType>(xferOp.getBase().getType())) {
Expand Down
30 changes: 30 additions & 0 deletions mlir/test/Dialect/Vector/scalar-vector-transfer-to-memref.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,3 +148,33 @@ func.func @subvector_extract(%m: memref<?x?xf32>, %idx: index) -> vector<16xf32>
return %1 : vector<16xf32>
}

// -----

// CHECK: #[[$MAP:.*]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK-LABEL: func @transfer_read_1d_extract_dynamic(
// CHECK-SAME: %[[MEMREF:.*]]: memref<?xf32>, %[[M_IDX:.*]]: index, %[[E_IDX:.*]]: index
// CHECK: %[[APPLY:.*]] = affine.apply #[[$MAP]]()[%[[M_IDX]], %[[E_IDX]]]
// CHECK: %[[RES:.*]] = memref.load %[[MEMREF]][%[[APPLY]]]
func.func @transfer_read_1d_extract_dynamic(%m: memref<?xf32>, %idx: index,
%offset: index) -> f32 {
%cst = arith.constant 0.0 : f32
%vec = vector.transfer_read %m[%idx], %cst {in_bounds = [true]} : memref<?xf32>, vector<5xf32>
%elem = vector.extract %vec[%offset] : f32 from vector<5xf32>
return %elem : f32
}

// -----

// CHECK: #[[$MAP:.*]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK-LABEL: func @transfer_read_2d_extract_dynamic(
// CHECK-SAME: %[[MEMREF:.*]]: memref<?x?xf32>, %[[ROW_IDX:.*]]: index, %[[COL_IDX:.*]]: index, %[[ROW_OFFSET:.*]]: index, %[[COL_OFFSET:.*]]: index
// CHECK: %[[ROW_APPLY:.*]] = affine.apply #[[$MAP]]()[%[[ROW_IDX]], %[[ROW_OFFSET]]]
// CHECK: %[[COL_APPLY:.*]] = affine.apply #[[$MAP]]()[%[[COL_IDX]], %[[COL_OFFSET]]]
// CHECK: %[[RES:.*]] = memref.load %[[MEMREF]][%[[ROW_APPLY]], %[[COL_APPLY]]]
func.func @transfer_read_2d_extract_dynamic(%m: memref<?x?xf32>, %row_idx: index, %col_idx: index,
%row_offset: index, %col_offset: index) -> f32 {
%cst = arith.constant 0.0 : f32
%vec = vector.transfer_read %m[%row_idx, %col_idx], %cst {in_bounds = [true, true]} : memref<?x?xf32>, vector<10x5xf32>
%elem = vector.extract %vec[%row_offset, %col_offset] : f32 from vector<10x5xf32>
return %elem : f32
}
Loading