Skip to content

feat: Do not inline operand concat inputs on folder if it is being used by another operation #611

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
roberteg16 merged 1 commit into from
Aug 8, 2025
Merged

feat: Do not inline operand concat inputs on folder if it is being used by another operation #611

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
8 changes: 8 additions & 0 deletions mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1693,6 +1693,14 @@ OpFoldResult ConcatOp::fold(FoldAdaptor adaptor) {
if (getAxis() != producer.getAxis())
continue;

// If there are multiple uses of this operand concat and they are different
// operations, this means that operand concat will have to happen, so do not
// add its operands to us to avoid repeating data concatenation
const bool allConcatUsersAreThisConcat = llvm::all_of(
producer->getUsers(), [&](Operation *user) { return *this == user; });
if (!allConcatUsersAreThisConcat)
continue;

// Replace the original operand with all incoming operands
foundFoldableConcat = true;
concatOperands.pop_back();
Expand Down
39 changes: 38 additions & 1 deletion mlir/test/Dialect/Tosa/fold_concats.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,43 @@ func.func @nested_fold(%arg0: tensor<1x1x7x7xf32>) -> tensor<1x8x7x7xf32> {

// -----

func.func @concat_multiple_users(%arg0: tensor<1x1x7x7xf32>, %arg1: tensor<1x1x7x7xf32>) -> (tensor<1x3x7x7xf32>, tensor<1x2x7x7xf32>) {
%tmp = tensor.empty() : tensor<1x1x7x7xf32>
%0 = tosa.concat %arg0, %arg1 {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x2x7x7xf32>
%1 = tosa.concat %tmp, %0 {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x2x7x7xf32>) -> tensor<1x3x7x7xf32>
%2 = tosa.add %0, %0 : (tensor<1x2x7x7xf32>, tensor<1x2x7x7xf32>) -> tensor<1x2x7x7xf32>
return %1, %2 : tensor<1x3x7x7xf32>, tensor<1x2x7x7xf32>
}

// CHECK-LABEL: func.func @concat_multiple_users
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<1x1x7x7xf32>, [[PARAM_1_:%.+]]: tensor<1x1x7x7xf32>) -> (tensor<1x3x7x7xf32>, tensor<1x2x7x7xf32>) {
// CHECK-DAG: [[VAR_0_:%.+]] = tensor.empty() : tensor<1x1x7x7xf32>
// CHECK-DAG: [[VAR_1_:%.+]] = tosa.concat [[PARAM_0_]], [[PARAM_1_]] {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x2x7x7xf32>
// CHECK-NOT: separator of consecutive DAGs
// CHECK-DAG: [[VAR_2_:%.+]] = tosa.concat [[VAR_0_]], [[VAR_1_]] {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x2x7x7xf32>) -> tensor<1x3x7x7xf32>
// CHECK-DAG: [[VAR_3_:%.+]] = tosa.add [[VAR_1_]], [[VAR_1_]] : (tensor<1x2x7x7xf32>, tensor<1x2x7x7xf32>) -> tensor<1x2x7x7xf32>
// CHECK: return [[VAR_2_]], [[VAR_3_]] : tensor<1x3x7x7xf32>, tensor<1x2x7x7xf32>
// CHECK: }

// -----

func.func @concat_diamond_shape(%arg0: tensor<1x1x7x7xf32>, %arg1: tensor<1x1x7x7xf32>, %arg2: tensor<1x1x7x7xf32>, %arg3: tensor<1x1x7x7xf32>) -> tensor<1x6x7x7xf32> {
%tmp = tensor.empty() : tensor<1x1x7x7xf32>
%0 = tosa.concat %arg0, %arg1 {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x2x7x7xf32>
%1 = tosa.concat %0, %arg2 {axis = 1 : i32} : (tensor<1x2x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x3x7x7xf32>
%2 = tosa.concat %0, %arg3 {axis = 1 : i32} : (tensor<1x2x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x3x7x7xf32>
%3 = tosa.concat %1, %2 {axis = 1 : i32} : (tensor<1x3x7x7xf32>, tensor<1x3x7x7xf32>) -> tensor<1x6x7x7xf32>
return %3 : tensor<1x6x7x7xf32>
}

// CHECK-LABEL: func.func @concat_diamond_shape
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<1x1x7x7xf32>, [[PARAM_1_:%.+]]: tensor<1x1x7x7xf32>, [[PARAM_2_:%.+]]: tensor<1x1x7x7xf32>, [[PARAM_3_:%.+]]: tensor<1x1x7x7xf32>) -> tensor<1x6x7x7xf32> {
// CHECK: [[VAR_0_:%.+]] = tosa.concat [[PARAM_0_]], [[PARAM_1_]], [[PARAM_2_]], [[PARAM_0_]], [[PARAM_1_]], [[PARAM_3_]] {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x6x7x7xf32>
// CHECK: return [[VAR_0_]] : tensor<1x6x7x7xf32>
// CHECK: }

// -----

func.func @wide_fold(%arg0: tensor<1x1x7x7xf32>, %arg1: tensor<1x1x7x7xf32>) -> tensor<1x4x7x7xf32> {
%0 = tosa.concat %arg0, %arg0 {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x2x7x7xf32>
%1 = tosa.concat %arg1, %arg1 {axis = 1 : i32} : (tensor<1x1x7x7xf32>, tensor<1x1x7x7xf32>) -> tensor<1x2x7x7xf32>
Expand Down Expand Up @@ -91,4 +128,4 @@ func.func @partially_foldable(%arg0: tensor<1x1x8x8xf32>, %arg1: tensor<1x2x4x8x
// CHECK: [[VAR_1_:%.+]] = tosa.tile [[PARAM_1_]], [[VAR_0_]] : (tensor<1x2x4x8xf32>, !tosa.shape<4>) -> tensor<1x2x8x8xf32>
// CHECK: [[VAR_2_:%.+]] = tosa.concat [[PARAM_0_]], [[PARAM_0_]], [[VAR_1_]] {axis = 1 : i32} : (tensor<1x1x8x8xf32>, tensor<1x1x8x8xf32>, tensor<1x2x8x8xf32>) -> tensor<1x4x8x8xf32>
// CHECK: return [[VAR_2_]] : tensor<1x4x8x8xf32>
// CHECK: }
// CHECK: }