refine conv add for xpu (#48432)

python/paddle/nn/functional/conv.py

@@ -145,20 +145,9 @@ def _conv_nd(
             data_format,
         )
         if bias is not None:
-            channel_dim = (
-                channel_dim + len(x.shape) if channel_dim < 0 else channel_dim
-            )
-            if isinstance(x, tuple):
-                x = x[0]
-            if isinstance(bias, tuple):
-                bias = bias[0]
-            if len(bias.shape) < len(x.shape):
-                bias = _C_ops.reshape(
-                    bias,
-                    [1 for i in range(channel_dim)]
-                    + bias.shape
-                    + [1 for i in range(len(x.shape) - channel_dim - 1)],
-                )
+            new_shape = [1] * len(x.shape)
+            new_shape[channel_dim] = -1
+            bias = bias.reshape(new_shape)
             # TODO(qili93): temporary for ascned npu performance to be removed along with npu_identity op
             if 'npu' in get_all_custom_device_type():
                 with no_grad():
@@ -182,16 +171,10 @@ def _conv_nd(
             data_format,
         )
         if bias is not None:
-            channel_dim = (
-                channel_dim + len(x.shape) if channel_dim < 0 else channel_dim
-            )
-            tmp_bias = _C_ops.reshape(
-                bias,
-                [1 for i in range(channel_dim)]
-                + bias.shape
-                + [1 for i in range(len(x.shape) - channel_dim - 1)],
-            )
-            return _C_ops.add(pre_bias, tmp_bias)
+            new_shape = [1] * len(x.shape)
+            new_shape[channel_dim] = -1
+            bias = bias.reshape(new_shape)
+            return _C_ops.add(pre_bias, bias)
         else:
             return pre_bias
 
@@ -207,14 +190,10 @@ def _conv_nd(
             data_format,
         )
         if bias is not None:
-            channel_dim = (
-                channel_dim + len(x.shape) if channel_dim < 0 else channel_dim
-            )
-            tmp_bias = _C_ops.reshape(
-                bias,
-                bias.shape + [1 for i in range(len(x.shape) - channel_dim - 1)],
-            )
-            return _C_ops.add(pre_bias, tmp_bias)
+            new_shape = [1] * len(x.shape)
+            new_shape[channel_dim] = -1
+            bias = bias.reshape(new_shape)
+            return _C_ops.add(pre_bias, bias)
         else:
             return pre_bias