Skip to content

[TorchToLinalg] Fix multi batch matmul conversion to Linalg #4319

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 3 commits into from
Oct 10, 2025
+25 −5
Merged

[TorchToLinalg] Fix multi batch matmul conversion to Linalg #4319

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
cf49ea0
[TorchToLinalg] Fix multi batch matmul conversion to Linalg
mingzheTerapines Sep 22, 2025
d35aadf
use upstream utils functions and add lit tests
mingzheTerapines Sep 26, 2025
ec5ca99
simplify test
mingzheTerapines Oct 10, 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
14 changes: 9 additions & 5 deletions lib/Conversion/TorchToLinalg/Linear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -505,9 +505,11 @@ class ConvertAtenMatmulOp : public OpConversionPattern<AtenMatmulOp> {

// Broadcast the batch dimensions of both the matrices.
Value broadcastedLhs, broadcastedRhs;
// TODO: Improve usage of static shape information.
SmallVector<int64_t> lhsTargetShape(lhsBroadcastToShape.size(),
ShapedType::kDynamic);
SmallVector<int64_t> lhsTargetShape =
llvm::to_vector(llvm::map_range(lhsBroadcastToShape, [](Value v) {
return getConstantIntValue(v).value_or(ShapedType::kDynamic);
}));

auto lhsBroadcastType = RankedTensorType::get(
lhsTargetShape, lhsType.getElementType(), lhsType.getEncoding());
if (failed(torch_to_linalg::broadcastToGivenShape(
Expand All @@ -516,8 +518,10 @@ class ConvertAtenMatmulOp : public OpConversionPattern<AtenMatmulOp> {
return rewriter.notifyMatchFailure(
op, "unable to perform broadcast operation");
}
SmallVector<int64_t> rhsTargetShape(rhsBroadcastToShape.size(),
ShapedType::kDynamic);
SmallVector<int64_t> rhsTargetShape =
llvm::to_vector(llvm::map_range(rhsBroadcastToShape, [](Value v) {
return getConstantIntValue(v).value_or(ShapedType::kDynamic);
}));
auto rhsBroadcastType = RankedTensorType::get(
rhsTargetShape, rhsType.getElementType(), rhsType.getEncoding());
if (failed(torch_to_linalg::broadcastToGivenShape(
Expand Down
16 changes: 16 additions & 0 deletions test/Conversion/TorchToLinalg/basic.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,22 @@ func.func @torch.aten.matmul.2d(%arg0: !torch.vtensor<[8,16],f32>, %arg1: !torch

// -----

// CHECK-LABEL: func.func @torch.aten.matmul.4d
// CHECK-DAG: %[[LHS:.+]] = torch_c.to_builtin_tensor %arg0 : !torch.vtensor<[1,2,32,400],f32> -> tensor<1x2x32x400xf32>
// CHECK-DAG: %[[RHS:.+]] = torch_c.to_builtin_tensor %arg1 : !torch.vtensor<[1,2,400,32],f32> -> tensor<1x2x400x32xf32>
// CHECK-DAG: %[[LHS_CAST:.*]] = tensor.cast %[[LHS]] : tensor<1x2x32x400xf32> to tensor<1x2x32x400xf32>
// CHECK-DAG: %[[RHS_CAST:.*]] = tensor.cast %[[RHS]] : tensor<1x2x400x32xf32> to tensor<1x2x400x32xf32>
// CHECK-DAG: %[[COLLAPSED_LHS:.+]] = tensor.collapse_shape %[[LHS_CAST]] {{\[\[}}0, 1], [2], [3]] : tensor<1x2x32x400xf32> into tensor<2x32x400xf32>
// CHECK-DAG: %[[COLLAPSED_RHS:.+]] = tensor.collapse_shape %[[RHS_CAST]] {{\[\[}}0, 1], [2], [3]] : tensor<1x2x400x32xf32> into tensor<2x400x32xf32>
// CHECK: %[[MATMUL:.+]] = linalg.batch_matmul ins(%[[COLLAPSED_RHS]], %[[COLLAPSED_LHS]] : tensor<2x400x32xf32>, tensor<2x32x400xf32>) outs(%{{.*}} : tensor<2x400x400xf32>) -> tensor<2x400x400xf32>
// CHECK: %[[EXPANDED:.+]] = tensor.expand_shape %[[MATMUL]] {{\[\[}}0, 1], [2], [3]] output_shape [1, 2, 400, 400] : tensor<2x400x400xf32> into tensor<1x2x400x400xf32>
func.func @torch.aten.matmul.4d(%arg0: !torch.vtensor<[1,2,32,400],f32>, %arg1: !torch.vtensor<[1,2,400,32],f32>) -> !torch.vtensor<[1,2,400,400],f32> {
%0 = torch.aten.matmul %arg1, %arg0 : !torch.vtensor<[1,2,400,32],f32>, !torch.vtensor<[1,2,32,400],f32> -> !torch.vtensor<[1,2,400,400],f32>
return %0 : !torch.vtensor<[1,2,400,400],f32>
}

// -----

// CHECK-LABEL: func.func @torch.aten.mm$basic_strict(
// CHECK-NOT: assert
func.func @torch.aten.mm$basic_strict(%arg0: !torch.vtensor<[?,?],f32>, %arg1: !torch.vtensor<[?,?],f32>) -> !torch.vtensor<[?,2],f32>
Expand Down
Loading