support auto generate for p_norm (#51590)

* supoort auto generate p_norm

* fix bug in backward

Author: RedContritio

paddle/fluid/operators/p_norm_op.cc

@@ -1028,6 +1028,16 @@
     func : overlap_add_grad
     data_type : x
 
+- backward_op : p_norm_grad
+  forward : p_norm(Tensor x,  float porder=2,  int axis=-1,  float epsilon=1.0e-12f,  bool keepdim=false,  bool asvector=false) -> Tensor(out)
+  args : (Tensor x, Tensor out, Tensor out_grad,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : GeneralUnaryGradInferMeta
+    param: [x]
+  kernel :
+    func : p_norm_grad
+
 - backward_op : pixel_shuffle_grad
   forward : pixel_shuffle (Tensor x, int upscale_factor=1, str data_format="NCHW") -> Tensor(out)
   args : (Tensor out_grad, int upscale_factor, str data_format)

paddle/phi/api/yaml/backward.yaml

@@ -863,16 +863,6 @@
   kernel :
     func : norm_grad
 
-- backward_op : p_norm_grad
-  forward : p_norm(Tensor x,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector=false) -> Tensor(out)
-  args : (Tensor x, Tensor out, Tensor out_grad,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector)
-  output : Tensor(x_grad)
-  infer_meta :
-    func : UnchangedInferMeta
-    param: [x]
-  kernel :
-    func : p_norm_grad
-
 - backward_op : pad3d_double_grad
   forward : pad3d_grad(Tensor x, Tensor grad_out, IntArray paddings, str mode, float pad_value, str data_format) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray paddings, str mode, float pad_value, str data_format)

paddle/phi/api/yaml/legacy_backward.yaml

@@ -1235,15 +1235,6 @@
   output : Tensor(out)
   invoke : full_like(x, 1, dtype, place)
 
-- op : p_norm
-  args : (Tensor x,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector=false)
-  output : Tensor(out)
-  infer_meta :
-    func : PNormInferMeta
-  kernel :
-    func : p_norm
-  backward : p_norm_grad
-
 - op : pad
   args : (Tensor x, int[] paddings, Scalar pad_value)
   output : Tensor

paddle/phi/api/yaml/legacy_ops.yaml

@@ -1290,6 +1290,13 @@
   outputs :
     out : Out
 
+- op : p_norm
+  backward: p_norm_grad
+  inputs :
+    x : X
+  outputs :
+    out : Out
+
 - op : pad2d
   backward : pad2d_grad
   extra :

paddle/phi/api/yaml/op_compat.yaml

@@ -102,6 +102,14 @@
           comment : In order to force fill output variable to gpu memory.
           default : "false"
 
+- op : p_norm
+  version :
+    - checkpoint : Upgrade p_norm, add 1 attribute [asvector].
+      action :
+        - add_attr : asvector
+          comment : Compute as vector when axis is None and input is matrix.
+          default : "false"
+
 - op : pixel_shuffle
   version :
     - checkpoint : Compatible upgrade of pixel_shuffle, add a new attribute [data_format]

paddle/phi/api/yaml/op_version.yaml

@@ -1078,6 +1078,15 @@
     data_type : x
   backward: overlap_add_grad
 
+- op : p_norm
+  args : (Tensor x,  float porder=2,  int axis=-1,  float epsilon=1.0e-12f,  bool keepdim=false,  bool asvector=false)
+  output : Tensor(out)
+  infer_meta :
+    func : PNormInferMeta
+  kernel :
+    func : p_norm
+  backward : p_norm_grad
+
 - op : pixel_shuffle
   args : (Tensor x, int upscale_factor=1, str data_format="NCHW")
   output : Tensor