pnnx: fix Conv2d padding tuple normalization

Normalize 4-element aten pad values from (left, right, top, bottom) to nn.Conv2d padding order (height, width) when fusing Conv2d modules.

Also add regression coverage for asymmetric ZeroPad2d + depthwise Conv2d to keep the output shape and ncnn pad params correct.

Author: Yeuvoir

tools/pnnx/src/pass_level1/nn_Conv2d.cpp

@@ -6,6 +6,27 @@
 
 namespace pnnx {
 
+static void normalize_conv2d_padding(Parameter& padding_param)
+{
+    std::vector<int>& padding = padding_param.ai;
+    if (padding.size() != 4)
+        return;
+
+    // aten::pad uses (left, right, top, bottom), while nn.Conv2d
+    // stores padding as (top/bottom, left/right).
+    if (padding[0] == padding[1] && padding[2] == padding[3])
+    {
+        padding = {padding[2], padding[0]};
+        return;
+    }
+
+    if (padding[0] == padding[2] && padding[1] == padding[3] && padding[0] != padding[1])
+    {
+        padding.resize(0);
+        padding_param.s = "same";
+    }
+}
+
 class Conv2d : public FuseModulePass
 {
 public:
@@ -53,58 +74,19 @@ class Conv2d : public FuseModulePass
         {
             op->params["padding_mode"] = pad->namedInput("mode");
             op->params["padding"] = pad->namedInput("pad");
-            std::vector<int>& padding = op->params["padding"].ai;
-            if (padding.size() == 4)
-            {
-                // Conv2d only accepts tuple of two integers
-                if (padding[0] == padding[1] && padding[1] == padding[2] && padding[2] == padding[3])
-                {
-                    padding.resize(2);
-                }
-                else if (padding[0] == padding[2] && padding[1] == padding[3] && padding[0] != padding[1])
-                {
-                    padding.resize(0);
-                    op->params["padding"].s = "same";
-                }
-            }
+            normalize_conv2d_padding(op->params["padding"]);
         }
         else if (reflection_pad2d)
         {
             op->params["padding_mode"] = "reflect";
             op->params["padding"] = reflection_pad2d->namedInput("padding");
-            std::vector<int>& padding = op->params["padding"].ai;
-            if (padding.size() == 4)
-            {
-                // Conv2d only accepts tuple of two integers
-                if (padding[0] == padding[1] && padding[1] == padding[2] && padding[2] == padding[3])
-                {
-                    padding.resize(2);
-                }
-                else if (padding[0] == padding[2] && padding[1] == padding[3] && padding[0] != padding[1])
-                {
-                    padding.resize(0);
-                    op->params["padding"].s = "same";
-                }
-            }
+            normalize_conv2d_padding(op->params["padding"]);
         }
         else if (replication_pad2d)
         {
             op->params["padding_mode"] = "replicate";
             op->params["padding"] = replication_pad2d->namedInput("padding");
-            std::vector<int>& padding = op->params["padding"].ai;
-            if (padding.size() == 4)
-            {
-                // Conv2d only accepts tuple of two integers
-                if (padding[0] == padding[1] && padding[1] == padding[2] && padding[2] == padding[3])
-                {
-                    padding.resize(2);
-                }
-                else if (padding[0] == padding[2] && padding[1] == padding[3] && padding[0] != padding[1])
-                {
-                    padding.resize(0);
-                    op->params["padding"].s = "same";
-                }
-            }
+            normalize_conv2d_padding(op->params["padding"]);
         }
         else
         {

tools/pnnx/tests/CMakeLists.txt

@@ -360,6 +360,7 @@ pnnx_add_test(pnnx_fuse_slice_to_tensor_split)
 pnnx_add_test(pnnx_fuse_adjacent_permute)
 pnnx_add_test(pnnx_fuse_adjacent_reshape)
 pnnx_add_test(pnnx_fuse_pad_conv1d)
+pnnx_add_test(pnnx_fuse_asymmetric_pad_conv2d)
 pnnx_add_test(pnnx_fuse_pad_conv2d)
 pnnx_add_test(pnnx_fuse_pixel_shuffle)
 pnnx_add_test(pnnx_fuse_pixel_unshuffle)

tools/pnnx/tests/test_nn_Conv2d.py

@@ -21,6 +21,7 @@ def __init__(self):
             self.conv_4 = nn.Conv2d(in_channels=28, out_channels=32, kernel_size=3, stride=1, padding='same', dilation=(1,2), groups=2, bias=False, padding_mode='zeros')
         self.conv_5 = nn.Conv2d(in_channels=32, out_channels=32, kernel_size=2, stride=2, padding=3, dilation=1, groups=32, bias=True, padding_mode='reflect')
         self.conv_6 = nn.Conv2d(in_channels=32, out_channels=28, kernel_size=2, stride=1, padding=2, dilation=1, groups=1, bias=False, padding_mode='replicate')
+        self.conv_7 = nn.Conv2d(in_channels=28, out_channels=28, kernel_size=3, stride=1, padding=(1,2), dilation=1, groups=1, bias=False, padding_mode='reflect')
         #self.conv_7 = nn.Conv2d(in_channels=28, out_channels=24, kernel_size=3, stride=2, padding=(5,6), dilation=2, groups=1, bias=True, padding_mode='circular')
 
         self.conv_8 = nn.Conv2d(in_channels=28, out_channels=24, kernel_size=3, stride=2, padding=(5,6), dilation=2, groups=1, bias=True)
@@ -37,6 +38,7 @@ def forward(self, x):
         x = self.conv_4(x)
         x = self.conv_5(x)
         x = self.conv_6(x)
+        x = self.conv_7(x)
         #x = self.conv_7(x)
         x = self.conv_8(x)
 

tools/pnnx/tests/test_pnnx_fuse_asymmetric_pad_conv2d.py

@@ -0,0 +1,63 @@
+# Copyright 2026 Tencent
+# SPDX-License-Identifier: BSD-3-Clause
+
+import os
+
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+        self.pad = nn.ZeroPad2d((1, 2, 1, 2))
+        self.conv = nn.Conv2d(4, 4, kernel_size=5, stride=2, groups=4, bias=False)
+
+    def forward(self, x):
+        x = self.pad(x)
+        x = self.conv(x)
+        return x
+
+
+def test():
+    net = Model()
+    net.eval()
+
+    torch.manual_seed(0)
+    x = torch.rand(1, 4, 56, 56)
+
+    a = net(x)
+    if a.shape != (1, 4, 28, 28):
+        return False
+
+    mod = torch.jit.trace(net, x)
+    mod.save("test_pnnx_fuse_asymmetric_pad_conv2d.pt")
+
+    os.system("../src/pnnx test_pnnx_fuse_asymmetric_pad_conv2d.pt inputshape=[1,4,56,56]")
+
+    import test_pnnx_fuse_asymmetric_pad_conv2d_pnnx
+    b = test_pnnx_fuse_asymmetric_pad_conv2d_pnnx.test_inference()
+
+    if not torch.equal(a, b):
+        return False
+
+    with open("test_pnnx_fuse_asymmetric_pad_conv2d.pnnx.param", "r") as f:
+        pnnx_param = f.read()
+    if "#2=(1,4,28,28)f32" not in pnnx_param:
+        return False
+
+    with open("test_pnnx_fuse_asymmetric_pad_conv2d.ncnn.param", "r") as f:
+        ncnn_param = f.read()
+
+    if "ConvolutionDepthWise" in ncnn_param:
+        return " 4=1 " in ncnn_param and " 14=1 " in ncnn_param and " 15=2 " in ncnn_param and " 16=2 " in ncnn_param
+
+    return "Padding" in ncnn_param and " 0=1 " in ncnn_param and " 1=2 " in ncnn_param and " 2=1 " in ncnn_param and " 3=2 " in ncnn_param
+
+
+if __name__ == "__main__":
+    if test():
+        exit(0)
+    else:
+        exit(1)