Fix incorrect pad indices in AveragePool count_include_pad computation

onnxruntime/core/providers/cpu/nn/pool_functors.h

@@ -462,7 +462,7 @@ struct AveragePool2DTask final {
     for (int64_t ph = 0; ph < pooled_height; ++ph) {
       int64_t hstart = ph * stride_h - pads[0];
       int64_t hend = hstart + kernel_shape[0] * dilation_h;
-      hend = std::min(hend, height + pads[1]);
+      hend = std::min(hend, height + pads[2]);
       for (int64_t pw = 0; pw < pooled_width; ++pw) {
         int64_t wstart = pw * stride_w - pads[1];
         int64_t wend = wstart + kernel_shape[1] * dilation_w;
@@ -535,11 +535,11 @@ struct AveragePool3DTask {
     for (int64_t ph = 0; ph < pooled_height; ++ph) {
       int64_t hstart = ph * stride_h - pads[0];
       int64_t hend = hstart + kernel_shape[0] * dilation_h;
-      hend = std::min(hend, height + pads[1]);
+      hend = std::min(hend, height + pads[3]);
       for (int64_t pw = 0; pw < pooled_width; ++pw) {
         int64_t wstart = pw * stride_w - pads[1];
         int64_t wend = wstart + kernel_shape[1] * dilation_w;
-        wend = std::min(wend, width + pads[3]);
+        wend = std::min(wend, width + pads[4]);
         for (int64_t pd = 0; pd < pooled_depth; ++pd) {
           int64_t dstart = pd * stride_d - pads[2];
           int64_t dend = dstart + kernel_shape[2] * dilation_d;

onnxruntime/test/providers/cpu/nn/pool_op_test.cc

@@ -966,6 +966,78 @@ TEST(PoolTest, AveragePool_IncludePadPixel) {
   test.Run(OpTester::ExpectResult::kExpectSuccess, "", {kTensorrtExecutionProvider});
 }
 
+// Regression test for https://github.com/microsoft/onnxruntime/issues/26708
+// AveragePool with count_include_pad=1 and asymmetric pads (only bottom/right)
+// was using incorrect pad index for hend, producing wrong results.
+TEST(PoolTest, AveragePool_CountIncludePad_AsymmetricPads) {
+  OpTester test("AveragePool", 19);
+
+  test.AddAttribute("auto_pad", "");
+  test.AddAttribute("strides", std::vector<int64_t>{1, 1});
+  test.AddAttribute("pads", vector<int64_t>{0, 0, 1, 1});  // no top/left, 1 bottom, 1 right
+  test.AddAttribute("kernel_shape", vector<int64_t>{2, 2});
+  test.AddAttribute("count_include_pad", (int64_t)1);
+
+  // Input: 2x2 all ones
+  std::vector<float> x_vals = {1.0f, 1.0f,
+                               1.0f, 1.0f};
+  std::vector<int64_t> x_dims = {1, 1, 2, 2};
+
+  // Output: 2x2
+  // Top-left:     (1+1+1+1)/4 = 1.0
+  // Top-right:    (1+0+1+0)/4 = 0.5
+  // Bottom-left:  (1+1+0+0)/4 = 0.5
+  // Bottom-right: (1+0+0+0)/4 = 0.25
+  std::vector<int64_t> expected_dims = {1, 1, 2, 2};
+  std::vector<float> expected_vals = {1.0f, 0.5f,
+                                      0.5f, 0.25f};
+
+  test.AddInput<float>("X", x_dims, x_vals);
+  test.AddOutput<float>("Y", expected_dims, expected_vals);
+  // This test targets the CPU fix only. Exclude EPs whose external libraries
+  // (cuDNN, CoreML, etc.) also produce wrong results for this case.
+  test.Run(OpTester::ExpectResult::kExpectSuccess, "",
+           {kCudaExecutionProvider, kCudaNHWCExecutionProvider,
+            kTensorrtExecutionProvider, kAclExecutionProvider, kOpenVINOExecutionProvider,
+            kDnnlExecutionProvider, kCoreMLExecutionProvider, kQnnExecutionProvider,
+            kDmlExecutionProvider});
+}
+
+// AveragePool3D with count_include_pad=1 and asymmetric pads (only back/bottom)
+// Regression test for 3D path of the pad-index bug
+TEST(PoolTest, AveragePool3D_CountIncludePad_AsymmetricPads) {
+  OpTester test("AveragePool", 19);
+  test.AddAttribute("auto_pad", "");
+  test.AddAttribute("strides", std::vector<int64_t>{1, 1, 1});
+  test.AddAttribute("pads", std::vector<int64_t>{0, 0, 0, 1, 1, 0});  // no front/top/left, 1 back, 1 bottom
+  test.AddAttribute("kernel_shape", std::vector<int64_t>{2, 2, 2});
+  test.AddAttribute("count_include_pad", (int64_t)1);
+  // Input: 2x2x2 all ones (N=1, C=1, D=2, H=2, W=2)
+  std::vector<float> x3d_vals = {
+      1.0f, 1.0f,
+      1.0f, 1.0f,
+      1.0f, 1.0f,
+      1.0f, 1.0f};
+  std::vector<int64_t> x3d_dims = {1, 1, 2, 2, 2};
+  // Output: 2x2x1 (D x H x W)
+  // D=0,H=0,W=0: (8 ones)/8 = 1.0
+  // D=1,H=0,W=0: (4 ones + 4 padded zeros)/8 = 0.5
+  // D=0,H=1,W=0: (4 ones + 4 padded zeros)/8 = 0.5
+  // D=1,H=1,W=0: (2 ones + 6 padded zeros)/8 = 0.25
+  std::vector<int64_t> expected3d_dims = {1, 1, 2, 2, 1};
+  std::vector<float> expected3d_vals = {1.0f, 0.5f,
+                                        0.5f, 0.25f};
+  test.AddInput<float>("X", x3d_dims, x3d_vals);
+  test.AddOutput<float>("Y", expected3d_dims, expected3d_vals);
+  // This test targets the CPU fix only. Exclude EPs whose external libraries
+  // (cuDNN, CoreML, etc.) also produce wrong results for this case.
+  test.Run(OpTester::ExpectResult::kExpectSuccess, "",
+           {kCudaExecutionProvider, kCudaNHWCExecutionProvider,
+            kTensorrtExecutionProvider, kAclExecutionProvider, kOpenVINOExecutionProvider,
+            kDnnlExecutionProvider, kCoreMLExecutionProvider, kQnnExecutionProvider,
+            kDmlExecutionProvider});
+}
+
 // test 'strides' attribute not specified
 TEST(PoolTest, AveragePool_DefaultStrides) {
   OpTester test("AveragePool");