[Mlir-commits] [mlir] [mlir][tosa] Stop folding pad into avg_pool2d (PR #164599)

[Vitalii Shutov]

https://github.com/Lallapallooza updated https://github.com/llvm/llvm-project/pull/164599

>From 44518644d3abd56d7d9f02e64df5228e7c969ac0 Mon Sep 17 00:00:00 2001
From: Vitalii Shutov <vitalii.shutov at arm.com>
Date: Tue, 21 Oct 2025 07:49:21 +0100
Subject: [PATCH] [mlir][tosa] Stop folding pad into avg_pool2d

Keep explicit padding ahead of tosa.avg_pool2d to preserve semantics.
Folding a pad into the op drops padded values from the average divisor.

Change-Id: I229bbdc0a8ef5d4ff4c6942788614c55593ce30f
---
 mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td  |  1 -
 .../Dialect/Tosa/IR/TosaCanonicalizations.cpp | 29 -------------------
 mlir/test/Dialect/Tosa/canonicalize.mlir      |  8 ++---
 3 files changed, 4 insertions(+), 34 deletions(-)

diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index 697a04e94441a..137554f49460d 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -108,7 +108,6 @@ def Tosa_AvgPool2dOp : Tosa_InferShapedTypeOp<"avg_pool2d"> {
     LogicalResult verifyOutputZeroPoint(int64_t zp);
   }];
 
-  let hasCanonicalizer = 1;
   let hasVerifier = 1;
 
   let assemblyFormat =
diff --git a/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp b/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
index caf80165fc640..6f32f601e5be0 100644
--- a/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
+++ b/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
@@ -75,28 +75,6 @@ namespace {
 template <typename OpTy>
 struct PoolPadFoldAdaptor;
 
-template <>
-struct PoolPadFoldAdaptor<tosa::AvgPool2dOp> {
-  using OpTy = tosa::AvgPool2dOp;
-  static bool checkKernelCompliance(OpTy op, const ArrayRef<int64_t> newPad) {
-    const llvm::ArrayRef<int64_t> kernel = op.getKernel();
-    if (newPad[2] >= kernel[1] || newPad[3] >= kernel[1] ||
-        newPad[0] >= kernel[0] || newPad[1] >= kernel[0])
-      return false;
-    return true;
-  }
-  static bool checkPadConstCompliance(OpTy op, Value padConst) {
-    return checkMatchingPadConstAndZp(padConst, op.getInputZp());
-  }
-  static void replaceOpWithNewPad(PatternRewriter &rewriter, OpTy op,
-                                  Value padInput, ArrayRef<int64_t> newPad) {
-    rewriter.replaceOpWithNewOp<tosa::AvgPool2dOp>(
-        op, op.getType(), padInput, op.getInputZp(), op.getOutputZp(),
-        op.getKernel(), op.getStride(), rewriter.getDenseI64ArrayAttr(newPad),
-        op.getAccType());
-  }
-};
-
 template <>
 struct PoolPadFoldAdaptor<tosa::MaxPool2dOp> {
   using OpTy = tosa::MaxPool2dOp;
@@ -245,13 +223,6 @@ struct FoldPadToTensorOp : public OpRewritePattern<OpTy> {
 };
 } // namespace
 
-void AvgPool2dOp::getCanonicalizationPatterns(RewritePatternSet &results,
-                                              MLIRContext *context) {
-  results.add<FoldPadToTensorOp<tosa::AvgPool2dOp,
-                                PoolPadFoldAdaptor<tosa::AvgPool2dOp>>>(
-      context);
-}
-
 void Conv2DOp::getCanonicalizationPatterns(RewritePatternSet &results,
                                            MLIRContext *context) {
   results.add<
diff --git a/mlir/test/Dialect/Tosa/canonicalize.mlir b/mlir/test/Dialect/Tosa/canonicalize.mlir
index e8525a5d2ed62..45d942bb92d6c 100644
--- a/mlir/test/Dialect/Tosa/canonicalize.mlir
+++ b/mlir/test/Dialect/Tosa/canonicalize.mlir
@@ -9,11 +9,11 @@ func.func @argmax_nofold(%arg0: tensor<?x1xf32>) -> tensor<1xi32> {
 
 // -----
 
-// CHECK-LABEL: @pad_wh_avg_pool2d_fold
-func.func @pad_wh_avg_pool2d_fold(%input: tensor<1x10x8x3xf32>) -> tensor<1x6x5x3xf32> {
-  // CHECK-NOT: tosa.pad
+// CHECK-LABEL: @pad_wh_avg_pool2d_nofold
+func.func @pad_wh_avg_pool2d_nofold(%input: tensor<1x10x8x3xf32>) -> tensor<1x6x5x3xf32> {
+  // CHECK: tosa.pad
   // CHECK: tosa.avg_pool2d
-  // CHECK-SAME: pad = array<i64: 1, 1, 1, 1>
+  // CHECK-SAME: pad = array<i64: 0, 1, 0, 1>
   %pad_shape = tosa.const_shape { values = dense<[0, 0, 1, 0, 1, 0, 0, 0]> : tensor<8xindex>} : () -> !tosa.shape<8>
   %pad_const = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : ()-> tensor<1xf32>
   %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : ()-> tensor<1xf32>