[mlir][tosa] Add folding for TOSA ArgMax operator #88871
Conversation
TOSA ArgMax operator could be folded into a constant tensor filled with zeros when dimension of the selected axis equals one.
|
Thank you for submitting a Pull Request (PR) to the LLVM Project! This PR will be automatically labeled and the relevant teams will be If you wish to, you can add reviewers by using the "Reviewers" section on this page. If this is not working for you, it is probably because you do not have write If you have received no comments on your PR for a week, you can request a review If you have further questions, they may be answered by the LLVM GitHub User Guide. You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums. |
|
@llvm/pr-subscribers-mlir @llvm/pr-subscribers-mlir-tosa Author: None (d-agbv) ChangesTOSA ArgMax operator could be folded into a constant tensor filled with zeros when dimension of the selected axis equals one. Full diff: https://github.com/llvm/llvm-project/pull/88871.diff 3 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index 306e4a43952088..dde17e2dc8924d 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -49,6 +49,7 @@ def Tosa_ArgMaxOp : Tosa_InferShapedTypeOp<"argmax"> {
Tosa_Tensor: $output
);
+ let hasFolder = 1;
let hasVerifier = 1;
}
diff --git a/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp b/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
index d23c9fe824c94a..53ae7211f987e2 100644
--- a/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
+++ b/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
@@ -507,6 +507,20 @@ 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());
diff --git a/mlir/test/Dialect/Tosa/constant-op-fold.mlir b/mlir/test/Dialect/Tosa/constant-op-fold.mlir
index de752f31fcbaa1..1513d4f772330d 100644
--- a/mlir/test/Dialect/Tosa/constant-op-fold.mlir
+++ b/mlir/test/Dialect/Tosa/constant-op-fold.mlir
@@ -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: @argmax_fold_dim_1
+func.func @argmax_fold_dim_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_1
+func.func @argmax_dynamic_shape_no_fold_dim_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
@@ -1100,9 +1114,9 @@ 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>
@@ -1110,18 +1124,18 @@ 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>
}
|
Update after review: - Make code formatting consistent - Rename LIT test
Update after review: - Rename LIT test
|
@d-agbv Congratulations on having your first Pull Request (PR) merged into the LLVM Project! Your changes will be combined with recent changes from other authors, then tested Please check whether problems have been caused by your change specifically, as How to do this, and the rest of the post-merge process, is covered in detail here. If your change does cause a problem, it may be reverted, or you can revert it yourself. If you don't get any reports, no action is required from you. Your changes are working as expected, well done! |
TOSA ArgMax operator could be folded into a constant tensor filled with zeros when dimension of the selected axis equals one.