Skip to content

[mlir][tosa] Add folding for TOSA ArgMax operator #88871

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
Apr 18, 2024
Merged

[mlir][tosa] Add folding for TOSA ArgMax operator #88871

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
3580389
[mlir][tosa] Add folding for TOSA ArgMax operator
d-agbv Apr 16, 2024
a55e1fd
[mlir][tosa] Add folding for TOSA ArgMax operator
d-agbv Apr 17, 2024
4451365
[mlir][tosa] Add folding for TOSA ArgMax operator
d-agbv Apr 17, 2024
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
1 change: 1 addition & 0 deletions mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,7 @@ def Tosa_ArgMaxOp : Tosa_InferShapedTypeOp<"argmax"> {
Tosa_Tensor: $output
);

let hasFolder = 1;
let hasVerifier = 1;
}

Expand Down
13 changes: 13 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 @@ -507,6 +507,19 @@ OpFoldResult AddOp::fold(FoldAdaptor adaptor) {
resultTy);
}

OpFoldResult ArgMaxOp::fold(FoldAdaptor adaptor) {
auto inputTy = llvm::dyn_cast<RankedTensorType>(getInput().getType());
auto outputTy = llvm::dyn_cast<RankedTensorType>(getType());
if (!inputTy || !outputTy || !inputTy.hasStaticShape() ||
!outputTy.hasStaticShape())
return {};

if (inputTy.getDimSize(getAxis()) == 1)
return DenseElementsAttr::get(outputTy, 0);

return {};
}

OpFoldResult DivOp::fold(FoldAdaptor adaptor) {
auto lhsTy = llvm::dyn_cast<RankedTensorType>(getInput1().getType());
auto rhsTy = llvm::dyn_cast<RankedTensorType>(getInput2().getType());
Expand Down
26 changes: 20 additions & 6 deletions mlir/test/Dialect/Tosa/constant-op-fold.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,20 @@

// RUN: mlir-opt --split-input-file --tosa-layerwise-constant-fold="aggressive-reduce-constant=true" %s | FileCheck %s --check-prefix=AGGRESIVE

// CHECK-LABEL: @armax_fold_dim_size_1
func.func @armax_fold_dim_size_1(%arg0: tensor<2x1x3xf32>) -> tensor<2x3xi32> {
// CHECK: "tosa.const"() <{value = dense<0> : tensor<2x3xi32>}> : () -> tensor<2x3xi32>
%0 = tosa.argmax %arg0 {axis = 1 : i32}: (tensor<2x1x3xf32>) -> tensor<2x3xi32>
return %0 : tensor<2x3xi32>
}

// CHECK-LABEL: @argmax_dynamic_shape_no_fold_dim_size_1
func.func @argmax_dynamic_shape_no_fold_dim_size_1(%arg0: tensor<?x1x3xf32>) -> tensor<?x3xi32> {
// CHECK: tosa.argmax
%0 = tosa.argmax %arg0 {axis = 1 : i32}: (tensor<?x1x3xf32>) -> tensor<?x3xi32>
return %0 : tensor<?x3xi32>
}

// CHECK-LABEL: @transpose_fold
func.func @transpose_fold(%arg0: tensor<3x4xf32>) -> tensor<3x4xf32> {
// CHECK: return %arg0
Expand Down Expand Up @@ -1100,28 +1114,28 @@ func.func @reduce_sum_constant_aggressive() -> tensor<2x3xi32> {
// AGGRESIVE-DAG: %[[VAL_0:.*]] = "tosa.const"() <{value = dense<2> : tensor<1x2x3xi32>}> : () -> tensor<1x2x3xi32>
// AGGRESIVE-DAG: %[[VAL_1:.*]] = "tosa.const"() <{value = dense<1> : tensor<2x2x3xi32>}> : () -> tensor<2x2x3xi32>
// AGGRESIVE-DAG: %[[VAL_2:.*]] = "tosa.const"() <{value = dense<2> : tensor<2x3xi32>}> : () -> tensor<2x3xi32>
// AGGRESIVE: %[[VAL_3:.*]] = tosa.argmax %[[VAL_0]] {axis = 0 : i32} : (tensor<1x2x3xi32>) -> tensor<2x3xi32>
// AGGRESIVE: %[[VAL_3:.*]] = tosa.argmax %[[VAL_0]] {axis = 1 : i32} : (tensor<1x2x3xi32>) -> tensor<1x3xi32>
// AGGRESIVE: %[[VAL_4:.*]] = tosa.argmax %[[VAL_1]] {axis = 0 : i32} : (tensor<2x2x3xi32>) -> tensor<2x3xi32>
// AGGRESIVE: %[[VAL_5:.*]] = tosa.add %[[VAL_3]], %[[VAL_2]] : (tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
// AGGRESIVE: %[[VAL_5:.*]] = tosa.add %[[VAL_3]], %[[VAL_2]] : (tensor<1x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
// AGGRESIVE: %[[VAL_6:.*]] = tosa.add %[[VAL_5]], %[[VAL_4]] : (tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
// AGGRESIVE: return %[[VAL_6]] : tensor<2x3xi32>

// CHECK-LABEL: func.func @reduce_sum_constant_aggressive() -> tensor<2x3xi32> {
// CHECK: %[[VAL_0:.*]] = "tosa.const"() <{value = dense<1> : tensor<2x2x3xi32>}> : () -> tensor<2x2x3xi32>
// CHECK: %[[VAL_1:.*]] = "tosa.const"() <{value = dense<2> : tensor<2x3xi32>}> : () -> tensor<2x3xi32>
// CHECK: %[[VAL_2:.*]] = tosa.reduce_sum %[[VAL_0]] {axis = 0 : i32} : (tensor<2x2x3xi32>) -> tensor<1x2x3xi32>
// CHECK: %[[VAL_3:.*]] = tosa.argmax %[[VAL_2]] {axis = 0 : i32} : (tensor<1x2x3xi32>) -> tensor<2x3xi32>
// CHECK: %[[VAL_3:.*]] = tosa.argmax %[[VAL_2]] {axis = 1 : i32} : (tensor<1x2x3xi32>) -> tensor<1x3xi32>
// CHECK: %[[VAL_4:.*]] = tosa.argmax %[[VAL_0]] {axis = 0 : i32} : (tensor<2x2x3xi32>) -> tensor<2x3xi32>
// CHECK: %[[VAL_5:.*]] = tosa.add %[[VAL_3]], %[[VAL_1]] : (tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
// CHECK: %[[VAL_5:.*]] = tosa.add %[[VAL_3]], %[[VAL_1]] : (tensor<1x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
// CHECK: %[[VAL_6:.*]] = tosa.add %[[VAL_5]], %[[VAL_4]] : (tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
// CHECK: return %[[VAL_6]] : tensor<2x3xi32>

%const0 = "tosa.const"() {value = dense<1> : tensor<2x2x3xi32>} : () -> tensor<2x2x3xi32>
%const1 = "tosa.const"() {value = dense<2> : tensor<2x3xi32>} : () -> tensor<2x3xi32>
%reduce0 = tosa.reduce_sum %const0 {axis = 0 : i32} : (tensor<2x2x3xi32>) -> tensor<1x2x3xi32>
%argmax0 = tosa.argmax %reduce0 {axis = 0 : i32} : (tensor<1x2x3xi32>) -> tensor<2x3xi32>
%argmax0 = tosa.argmax %reduce0 {axis = 1 : i32} : (tensor<1x2x3xi32>) -> tensor<1x3xi32>
%argmax1 = tosa.argmax %const0 {axis = 0 : i32} : (tensor<2x2x3xi32>) -> tensor<2x3xi32>
%res0 = tosa.add %argmax0, %const1 : (tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
%res0 = tosa.add %argmax0, %const1 : (tensor<1x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
%res1 = tosa.add %res0, %argmax1 : (tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
return %res1 : tensor<2x3xi32>
}