Skip to content

[mlir][Vector] Improve support for vector.extract(broadcast) #116234

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 9 commits into from
Feb 24, 2025
Merged

[mlir][Vector] Improve support for vector.extract(broadcast) #116234

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
e0475be
[mlir][Vector] Improve dynamic support for vector.extract(broadcast) …
Groverkss Nov 14, 2024
55c7536
s
Groverkss Jan 25, 2025
5785c7c
Allow dim-1 broadcasting + dynamic indices
Groverkss Feb 8, 2025
85c2d27
Test fixes
Groverkss Feb 9, 2025
0c7a84e
fix name
Groverkss Feb 9, 2025
e2a4e6d
be consistent in stye
Groverkss Feb 19, 2025
7900062
fix types and use dynamic indices
Groverkss Feb 19, 2025
110ccdc
fix more
Groverkss Feb 19, 2025
350f449
resolve comment
Groverkss Feb 20, 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
16 changes: 7 additions & 9 deletions mlir/lib/Dialect/Vector/IR/VectorOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1660,10 +1660,6 @@ static bool hasZeroDimVectors(Operation *op) {

/// Fold extractOp with scalar result coming from BroadcastOp or SplatOp.
static Value foldExtractFromBroadcast(ExtractOp extractOp) {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return Value();

Operation *defOp = extractOp.getVector().getDefiningOp();
if (!defOp || !isa<vector::BroadcastOp, SplatOp>(defOp))
return Value();
Expand Down Expand Up @@ -1700,20 +1696,22 @@ static Value foldExtractFromBroadcast(ExtractOp extractOp) {
// extract position to `0` when extracting from the source operand.
llvm::SetVector<int64_t> broadcastedUnitDims =
broadcastOp.computeBroadcastedUnitDims();
SmallVector<int64_t> extractPos(extractOp.getStaticPosition());
SmallVector<OpFoldResult> extractPos(extractOp.getMixedPosition());
OpBuilder b(extractOp.getContext());
int64_t broadcastRankDiff = broadcastDstRank - broadcastSrcRank;
for (int64_t i = broadcastRankDiff, e = extractPos.size(); i < e; ++i)
if (broadcastedUnitDims.contains(i))
extractPos[i] = 0;
extractPos[i] = b.getIndexAttr(0);
// `rankDiff` leading dimensions correspond to new broadcasted dims, drop the
// matching extract position when extracting from the source operand.
int64_t rankDiff = broadcastSrcRank - extractResultRank;
extractPos.erase(extractPos.begin(),
std::next(extractPos.begin(), extractPos.size() - rankDiff));
// OpBuilder is only used as a helper to build an I64ArrayAttr.
OpBuilder b(extractOp.getContext());
extractOp.setOperand(0, source);
extractOp.setStaticPosition(extractPos);
auto [staticPos, dynPos] = decomposeMixedValues(extractPos);
extractOp->setOperands(
llvm::to_vector(llvm::concat<Value>(ValueRange(source), dynPos)));
extractOp.setStaticPosition(staticPos);
return extractOp.getResult();
}

Expand Down
92 changes: 61 additions & 31 deletions mlir/test/Dialect/Vector/canonicalize.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -710,29 +710,61 @@ func.func @fold_extract_transpose(

// -----

// CHECK-LABEL: fold_extract_broadcast
// CHECK-LABEL: fold_extract_broadcast_same_input_output_scalar
// CHECK-SAME: %[[A:.*]]: f32
// CHECK: return %[[A]] : f32
func.func @fold_extract_broadcast(%a : f32) -> f32 {
func.func @fold_extract_broadcast_same_input_output_scalar(%a : f32,
%idx0 : index, idx1 : index) -> f32 {
%b = vector.broadcast %a : f32 to vector<1x2x4xf32>
%r = vector.extract %b[0, 1, 2] : f32 from vector<1x2x4xf32>
// The indices don't matter for this folder, so we use mixed indices.
%r = vector.extract %b[%idx0, %idx1, 2] : f32 from vector<1x2x4xf32>
return %r : f32
}

// -----

// CHECK-LABEL: fold_extract_broadcast_0dvec
// CHECK-LABEL: fold_extract_broadcast_same_input_output_vec
// CHECK-SAME: %[[A:.*]]: vector<4xf32>
// CHECK: return %[[A]] : vector<4xf32>
func.func @fold_extract_broadcast_same_input_output_vec(%a : vector<4xf32>,
%idx0 : index) -> vector<4xf32> {
%b = vector.broadcast %a : vector<4xf32> to vector<1x2x4xf32>
// The indices don't matter for this folder, so we use mixed indices.
%r = vector.extract %b[0, %idx0] : vector<4xf32> from vector<1x2x4xf32>
return %r : vector<4xf32>
}

// -----

// CHECK-LABEL: fold_extract_broadcast_0dvec_input_scalar_output
// CHECK-SAME: %[[A:.*]]: vector<f32>
// CHECK: %[[B:.+]] = vector.extractelement %[[A]][] : vector<f32>
// CHECK: return %[[B]] : f32
func.func @fold_extract_broadcast_0dvec(%a : vector<f32>) -> f32 {
func.func @fold_extract_broadcast_0dvec_input_scalar_output(%a : vector<f32>,
%idx0 : index, idx1 : index) -> f32 {
%b = vector.broadcast %a : vector<f32> to vector<1x2x4xf32>
%r = vector.extract %b[0, 1, 2] : f32 from vector<1x2x4xf32>
// The indices don't matter for this folder, so we use mixed indices.
%r = vector.extract %b[%idx0, %idx1, 2] : f32 from vector<1x2x4xf32>
return %r : f32
}

// -----

// CHECK-LABEL: fold_extract_broadcast_diff_input_output_vec
// CHECK-SAME: %[[A:.*]]: vector<2x4xf32>
// CHECK-SAME: %[[IDX0:.*]]: index, %[[IDX1:.*]]: index
// CHECK: %[[B:.+]] = vector.extract %[[A]][%[[IDX1]]] : vector<4xf32> from vector<2x4xf32>
// CHECK: return %[[B]] : vector<4xf32>
func.func @fold_extract_broadcast_diff_input_output_vec(%a : vector<2x4xf32>,
%idx0 : index, idx1 : index) -> vector<4xf32> {
%b = vector.broadcast %a : vector<2x4xf32> to vector<1x2x4xf32>
// The indices don't matter for this folder, so we use mixed indices.
%r = vector.extract %b[%idx0, %idx1] : vector<4xf32> from vector<1x2x4xf32>
return %r : vector<4xf32>
}

// -----

// CHECK-LABEL: fold_extract_broadcast_negative
// CHECK: vector.broadcast %{{.*}} : vector<1x1xf32> to vector<1x1x4xf32>
// CHECK: vector.extract %{{.*}}[0, 0] : vector<4xf32> from vector<1x1x4xf32>
Expand All @@ -747,57 +779,55 @@ func.func @fold_extract_broadcast_negative(%a : vector<1x1xf32>) -> vector<4xf32
// CHECK-LABEL: fold_extract_splat
// CHECK-SAME: %[[A:.*]]: f32
// CHECK: return %[[A]] : f32
func.func @fold_extract_splat(%a : f32) -> f32 {
func.func @fold_extract_splat(%a : f32, %idx0 : index, %idx1 : index) -> f32 {
%b = vector.splat %a : vector<1x2x4xf32>
%r = vector.extract %b[0, 1, 2] : f32 from vector<1x2x4xf32>
// The indices don't matter for this folder, so we use mixed indices.
%r = vector.extract %b[%idx0, %idx1, 2] : f32 from vector<1x2x4xf32>
return %r : f32
}

// -----

// CHECK-LABEL: fold_extract_broadcast_vector
// CHECK-SAME: %[[A:.*]]: vector<4xf32>
// CHECK: return %[[A]] : vector<4xf32>
func.func @fold_extract_broadcast_vector(%a : vector<4xf32>) -> vector<4xf32> {
%b = vector.broadcast %a : vector<4xf32> to vector<1x2x4xf32>
%r = vector.extract %b[0, 1] : vector<4xf32> from vector<1x2x4xf32>
return %r : vector<4xf32>
}

// -----

// CHECK-LABEL: fold_extract_broadcast
// CHECK-SAME: %[[A:.*]]: vector<4xf32>
// CHECK: %[[R:.*]] = vector.extract %[[A]][2] : f32 from vector<4xf32>
// CHECK-LABEL: fold_extract_broadcast_dim1_broadcasting
// CHECK-SAME: %[[A:.*]]: vector<2x1xf32>
// CHECK-SAME: %[[IDX:.*]]: index, %[[IDX1:.*]]: index, %[[IDX2:.*]]: index
// CHECK: %[[R:.*]] = vector.extract %[[A]][%[[IDX1]], 0] : f32 from vector<2x1xf32>
// CHECK: return %[[R]] : f32
func.func @fold_extract_broadcast(%a : vector<4xf32>) -> f32 {
%b = vector.broadcast %a : vector<4xf32> to vector<1x2x4xf32>
%r = vector.extract %b[0, 1, 2] : f32 from vector<1x2x4xf32>
func.func @fold_extract_broadcast_dim1_broadcasting(%a : vector<2x1xf32>,
%idx : index, idx1 : index, idx2 : index) -> f32 {
%b = vector.broadcast %a : vector<2x1xf32> to vector<1x2x4xf32>
// The indices don't matter for this folder, so we use mixed indices.
%r = vector.extract %b[%idx, %idx1, %idx2] : f32 from vector<1x2x4xf32>
return %r : f32
}

// -----

// CHECK-LABEL: fold_extract_broadcast
// CHECK-LABEL: fold_extract_broadcast_to_higher_rank
// CHECK: %[[B:.*]] = vector.broadcast %{{.*}} : f32 to vector<4xf32>
// CHECK: return %[[B]] : vector<4xf32>
func.func @fold_extract_broadcast(%a : f32) -> vector<4xf32> {
func.func @fold_extract_broadcast_to_higher_rank(%a : f32, idx0 : index)
-> vector<4xf32> {
%b = vector.broadcast %a : f32 to vector<1x2x4xf32>
%r = vector.extract %b[0, 1] : vector<4xf32> from vector<1x2x4xf32>
// The indices don't matter for this canonicalizer, so we use mixed indices.
%r = vector.extract %b[0, %idx0] : vector<4xf32> from vector<1x2x4xf32>
return %r : vector<4xf32>
}

// -----

// CHECK-LABEL: fold_extract_broadcast
// CHECK-LABEL: fold_extract_broadcast_to_equal_rank
// CHECK-SAME: %[[A:.*]]: vector<1xf32>
// CHECK: %[[R:.*]] = vector.broadcast %[[A]] : vector<1xf32> to vector<8xf32>
// CHECK: return %[[R]] : vector<8xf32>
func.func @fold_extract_broadcast(%a : vector<1xf32>) -> vector<8xf32> {
func.func @fold_extract_broadcast_to_equal_rank(%a : vector<1xf32>, idx0 : index)
-> vector<8xf32> {
%b = vector.broadcast %a : vector<1xf32> to vector<1x8xf32>
%r = vector.extract %b[0] : vector<8xf32> from vector<1x8xf32>
// The indices don't matter for this canonicalizer, so we use mixed indices.
%r = vector.extract %b[%idx0] : vector<8xf32> from vector<1x8xf32>
return %r : vector<8xf32>
}

// -----

// CHECK-LABEL: @fold_extract_shuffle
Expand Down
Loading