[Comp] Add view_shape_double_grad for eager mode (PaddlePaddle#76667)

* add view_shape_double_grad for eager mode

Author: HydrogenSulfate

paddle/fluid/eager/auto_code_generator/generator/eager_gen.py

@@ -94,6 +94,7 @@
     "put_along_axis_double_grad",
     "masked_fill_double_grad",
     "index_elementwise_put_with_tensor_double_grad",
+    "view_shape_double_grad",
 ]
 
 # white ops list whose kernel can automatically do type promotion.
@@ -3170,12 +3171,7 @@ def GenerateNodeDefinition(
             )
 
             grad_api_args[grad_api_position] = name
-            if (
-                not is_invoke_forward_api
-                or name in self.grad_api_contents['invoke']
-            ):
-                # NOTE: attr 'dims' is not necessary for 'invoke: view_shape(out_grad, input.shape())'
-                get_grad_in_args_list.append(get_attr_str)
+            get_grad_in_args_list.append(get_attr_str)
 
         get_grad_in_args_str = "\n".join(get_grad_in_args_list)
 

paddle/fluid/pir/dialect/op_generator/vjp_interface_black_list.py

@@ -43,4 +43,5 @@
     'masked_select_grad',
     'index_elementwise_get_grad',
     'index_elementwise_put_with_tensor_grad',
+    'view_shape_grad',
 ]

paddle/fluid/prim/api/composite_backward/composite_double_backward_api.h

@@ -1296,5 +1296,16 @@ void index_elementwise_put_with_tensor_double_grad(
   }
 }
 
+template <typename T>
+void view_shape_double_grad(const Tensor& grad_input_grad,
+                            const std::vector<int64_t> dims,
+                            Tensor* grad_out_grad) {
+  if (grad_out_grad) {
+    Tensor grad_out_grad_tmp;
+    grad_out_grad_tmp = reshape<T>(grad_input_grad, dims);
+    set_output<T>(grad_out_grad_tmp, grad_out_grad);
+  }
+}
+
 }  // namespace prim
 }  // namespace paddle

paddle/phi/ops/yaml/backward.yaml

@@ -4025,6 +4025,26 @@
     data_type : out_grad
   no_need_buffer: input
 
+- backward_op : view_shape_double_grad
+  forward : view_shape_grad (Tensor input, Tensor grad_out, int64_t[] dims) -> Tensor(grad_input)
+  args : (Tensor grad_input_grad, int64_t[] dims)
+  output : Tensor(grad_out_grad)
+  infer_meta :
+    func : StridedUnChangedInferMeta
+    param : [grad_input_grad]
+  composite: view_shape_double_grad(grad_input_grad, dims, grad_out_grad)
+
+- backward_op : view_shape_grad
+  forward : view_shape (Tensor input, int64_t[] dims = {}) -> Tensor(out)
+  args : (Tensor input, Tensor out_grad, int64_t[] dims = {})
+  output : Tensor(input_grad)
+  infer_meta :
+    func : StridedUnChangedInferMeta
+    param : [input]
+  kernel :
+    func : view_shape_grad
+  backward : view_shape_double_grad
+
 - backward_op : warpctc_grad
   forward : warpctc (Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank = 0, bool norm_by_times = false) -> Tensor(loss), Tensor(warpctcgrad)
   args : (Tensor logits, Tensor logits_length, Tensor warpctcgrad, Tensor loss_grad, int blank, bool norm_by_times)

paddle/phi/ops/yaml/inconsistent/dygraph_backward.yaml

@@ -361,13 +361,6 @@
   composite : tile_grad(x, out_grad, repeat_times, x_grad)
   backward : tile_double_grad
 
-- backward_op : view_shape_grad
-  forward : view_shape (Tensor input, int64_t[] dims = {}) -> Tensor(out)
-  args : (Tensor input, Tensor out_grad, int64_t[] dims = {})
-  output : Tensor(input_grad)
-  invoke: view_shape(out_grad, input.shape())
-  no_need_buffer: input
-
 - backward_op: fused_gemm_epilogue_grad
   forward : fused_gemm_epilogue(Tensor x, Tensor y, Tensor bias, bool trans_x, bool trans_y, str activation) -> Tensor(out), Tensor(reserve_space)
   args : (Tensor x, Tensor y, Tensor reserve_space, Tensor out_grad, bool trans_x, bool trans_y, str activation)

paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml

@@ -404,12 +404,3 @@
     data_type : x
   optional : indices, inverse, counts
   traits : paddle::dialect::ForwardOnlyTrait
-
-- op : view_shape
-  args : (Tensor input, int64_t[] dims = {})
-  output : Tensor(out)
-  infer_meta :
-    func : ViewShapeInferMeta
-  kernel :
-    func : view_shape
-  backward : view_shape_grad

paddle/phi/ops/yaml/op_compat.yaml

@@ -4044,6 +4044,9 @@
   get_expected_kernel_type :
     update_loss_scaling_ : GetUpdateLossScalingExpectedKernelType
 
+- op : view_shape
+  backward : view_shape_grad, view_shape_double_grad
+
 - op : viterbi_decode
   inputs :
     {potentials : Input, transition_params : Transition, lengths : Length}

paddle/phi/ops/yaml/ops.yaml

@@ -5817,6 +5817,15 @@
   no_need_buffer : input
   interfaces : paddle::dialect::InferSymbolicShapeInterface
 
+- op : view_shape
+  args : (Tensor input, int64_t[] dims = {})
+  output : Tensor(out)
+  infer_meta :
+    func : ViewShapeInferMeta
+  kernel :
+    func : view_shape
+  backward : view_shape_grad
+
 - op : view_slice
   args : (Tensor input, int64_t begin_idx, int64_t end_idx)
   output : Tensor

test/indexing/test_setitem.py

@@ -1024,7 +1024,7 @@ def test_boolean_mask_tensor_broadcast_v(self):
         np.testing.assert_allclose(res, tensor_np)
 
     def test_index_elementwise_put_with_tensor(self):
-        with dygraph_guard(), paddle.device("cpu"):
+        with dygraph_guard():
             x = paddle.randn(10, 4, requires_grad=True)
             xx = x + 0
 

test/legacy_test/test_stride.py

@@ -971,6 +971,7 @@ def call_stride(self):
         self.call_view5()
         self.call_view6()
         self.call_view7()
+        self.call_view8()
         self.call_view9()
         self.call_view10()
         self.call_view11()