[IR] Generate pd_op.parsed.yaml from pd_op.yaml (#56674)

* Generate pd_op.parsed.yaml from pd_op.yaml

* Generate pd_op.parsed.yaml from pd_op.yaml

* fix bug

* bug fix

* bug fix

* bug fix

* 向pd_ops.yaml中新增算子 & 修改pd_ops.parsed.yaml存放路径

* 修复路径依赖bug & 添加 .gitignore文件

* fix bug - compat input args in save_combine op

* fix compat file

* fix set_value_with_tensor yaml

* split backward op in original yaml file

* add send_v2 & recv_v2

paddle/fluid/ir/dialect/paddle_dialect/ir/.gitignore

+generated/*
+generated/**

paddle/fluid/ir/dialect/paddle_dialect/ir/CMakeLists.txt

@@ -2,6 +2,9 @@ set(PD_DIALECT_BINARY_DIR
     "${PADDLE_BINARY_DIR}/paddle/fluid/ir/dialect/paddle_dialect/ir")
 
 # Generate pd_dialect files defining op using op_gen_file
+set(op_gen_parsed_yaml_file
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/operators/generator/parse_op.py)
+
 set(op_gen_file
     ${PADDLE_SOURCE_DIR}/paddle/fluid/ir/dialect/op_generator/op_gen.py)
 set(op_compat_yaml_file ${PADDLE_SOURCE_DIR}/paddle/phi/api/yaml/op_compat.yaml)
@@ -23,11 +26,22 @@ set(fused_op_forward_yaml_file
 set(fused_op_backward_yaml_file
     ${PADDLE_SOURCE_DIR}/paddle/fluid/operators/generator/parsed_ops/fused_backward.parsed.yaml
 )
-set(op_yaml_file3
-    ${PADDLE_SOURCE_DIR}/paddle/fluid/ir/dialect/paddle_dialect/ir/pd_op.yaml)
+
+set(pd_op_forward_yaml_file
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/ir/dialect/paddle_dialect/ir/pd_ops.yaml)
+
+set(pd_op_backward_yaml_file
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/ir/dialect/paddle_dialect/ir/pd_ops_backward.yaml
+)
+
+set(parsed_op_dir
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/ir/dialect/paddle_dialect/ir/generated)
+
+set(op_yaml_file3 ${parsed_op_dir}/pd_ops.parsed.yaml)
+set(op_yaml_file4 ${parsed_op_dir}/pd_ops_backward.parsed.yaml)
 
 set(op_yaml_files
-    ${op_forward_yaml_file1},${op_forward_yaml_file2},${op_backward_yaml_file1},${op_backward_yaml_file2},${fused_op_forward_yaml_file},${fused_op_backward_yaml_file},${op_yaml_file3}
+    ${op_forward_yaml_file1},${op_forward_yaml_file2},${op_backward_yaml_file1},${op_backward_yaml_file2},${fused_op_forward_yaml_file},${fused_op_backward_yaml_file},${op_yaml_file3},${op_yaml_file4}
 )
 set(op_namespace paddle,dialect)
 set(dialect_name pd)
@@ -39,6 +53,15 @@ set(op_source_file_tmp ${op_source_file}.tmp)
 set(op_vjp_source_file ${PD_DIALECT_BINARY_DIR}/pd_op_vjp.cc)
 set(op_vjp_source_file_tmp ${op_vjp_source_file}.tmp)
 
+add_custom_command(
+  OUTPUT ${op_yaml_file3} ${op_yaml_file4}
+  COMMAND ${CMAKE_COMMAND} -E make_directory ${parsed_op_dir}
+  COMMAND ${PYTHON_EXECUTABLE} ${op_gen_parsed_yaml_file} --op_yaml_path
+          ${pd_op_forward_yaml_file} --output_path ${op_yaml_file3}
+  COMMAND ${PYTHON_EXECUTABLE} ${op_gen_parsed_yaml_file} --op_yaml_path
+          ${pd_op_backward_yaml_file} --output_path ${op_yaml_file4} --backward
+  VERBATIM)
+
 add_custom_command(
   OUTPUT ${op_header_file} ${op_source_file} ${op_vjp_source_file}
   COMMAND
@@ -55,9 +78,14 @@ add_custom_command(
           ${op_vjp_source_file}
   COMMENT
     "copy_if_different ${op_header_file} ${op_source_file} ${op_vjp_source_file}"
-  DEPENDS ${op_gen_file} ${op_forward_yaml_file1} ${op_forward_yaml_file2}
-          ${op_backward_yaml_file1} ${op_backward_yaml_file2}
+  DEPENDS ${op_gen_file}
+          ${op_forward_yaml_file1}
+          ${op_forward_yaml_file2}
+          ${op_backward_yaml_file1}
+          ${op_backward_yaml_file2}
           ${op_compat_yaml_file}
+          ${op_yaml_file3}
+          ${op_yaml_file4}
   VERBATIM)
 
 set(api_gen_file
@@ -79,9 +107,14 @@ add_custom_command(
   COMMAND ${CMAKE_COMMAND} -E copy_if_different ${api_source_file_tmp}
           ${api_source_file}
   COMMENT "copy_if_different ${api_header_file} ${api_source_file}"
-  DEPENDS ${api_gen_file} ${op_forward_yaml_file1} ${op_forward_yaml_file2}
-          ${op_backward_yaml_file1} ${op_backward_yaml_file2}
+  DEPENDS ${api_gen_file}
+          ${op_forward_yaml_file1}
+          ${op_forward_yaml_file2}
+          ${op_backward_yaml_file1}
+          ${op_backward_yaml_file2}
           ${op_compat_yaml_file}
+          ${op_yaml_file3}
+          ${op_yaml_file4}
   VERBATIM)
 
 set(python_c_gen_file
@@ -105,9 +138,14 @@ add_custom_command(
   COMMAND ${CMAKE_COMMAND} -E copy_if_different ${python_c_source_file_tmp}
           ${python_c_source_file}
   COMMENT "copy_if_different ${python_c_header_file} ${python_c_source_file}"
-  DEPENDS ${python_c_gen_file} ${op_forward_yaml_file1} ${op_forward_yaml_file2}
-          ${op_backward_yaml_file1} ${op_backward_yaml_file2}
+  DEPENDS ${python_c_gen_file}
+          ${op_forward_yaml_file1}
+          ${op_forward_yaml_file2}
+          ${op_backward_yaml_file1}
+          ${op_backward_yaml_file2}
           ${op_compat_yaml_file}
+          ${op_yaml_file3}
+          ${op_yaml_file4}
   VERBATIM)
 
 add_custom_target(static_op_function_gen ALL DEPENDS ${python_c_header_file}

paddle/fluid/ir/dialect/paddle_dialect/ir/pd_ops.yaml

+- op : add_n_
+  args : (Tensor[] inputs)
+  output : Tensor(out)
+  infer_meta:
+    func: AddNInferMeta
+    param: [inputs]
+  kernel:
+    func: add_n
+    param: [inputs]
+  backward: add_n_grad
+
+- op : add_n_with_kernel
+  args : (Tensor[] inputs)
+  output : Tensor(out)
+  infer_meta:
+    func: AddNInferMeta
+    param: [inputs]
+  kernel:
+    func: add_n
+    param: [inputs]
+  backward: add_n_grad
+
+- op : assert
+  args : (Tensor cond, Tensor[] data, int64_t summarize = -1)
+  output :
+  kernel :
+    func : assert
+    param : [cond, data, summarize]
+    data_type : cond
+
+- op : assign_value
+  args : (int[] shape, DataType dtype, Scalar[] values, Place place = {})
+  output : Tensor(out)
+  infer_meta :
+    func : AssignValueInferMeta
+    param: [shape, dtype]
+  kernel :
+    func : assign_value
+    param : [shape, dtype, values]
+    backend: place>
+    data_type : dtype
+
+- op : embedding_grad_sparse
+  args : (Tensor x, Tensor weight, Tensor out_grad, int64_t padding_idx = -1, bool sparse = false)
+  output : SelectedRows(weight_grad)
+  infer_meta:
+    func: EmbeddingGradSparseInferMeta
+    param: [weight]
+  kernel:
+    func: embedding_sparse_grad
+    param: [x, weight, out_grad, padding_idx, sparse]
+    data_type : weight
+
+- op : feed
+  args : (str name, int col)
+  output : Tensor(out)
+
+- op : fetch
+  args : (Tensor x, str name, int col)
+  output : Tensor(out)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : fetch
+    param : [x]
+
+- op : load_combine
+  args : (str file_path, bool load_as_fp16, bool model_from_memory)
+  output : Tensor[](Out)
+  kernel:
+    func: load_combine
+    param: [file_path, load_as_fp16, model_from_memory]
+  optional : Out
+
+- op : lod_array_length
+  args : (Tensor[] x)
+  output : Tensor(out)
+
+- op : print
+  args : (Tensor in, int first_n, str message, int summarize, bool print_tensor_name = true, bool print_tensor_type = true, bool print_tensor_shape = true, bool print_tensor_layout = true, bool print_tensor_lod = true, str print_phase = "BOTH", bool is_forward = true)
+  output : Tensor(out)
+  infer_meta:
+    func: UnchangedInferMeta
+    param: [in]
+  kernel :
+    func : print_kernel
+    param: [in, first_n, message, summarize, print_tensor_name, print_tensor_type, print_tensor_shape, print_tensor_layout, print_tensor_lod, print_phase, is_forward]
+
+- op : recv_v2
+  args : (int[] out_shape = {}, DataType dtype = DataType::FLOAT32, int peer = 0, int ring_id = 0, bool use_calc_stream = false, bool dynamic_shape = false)
+  output : Tensor(out)
+  infer_meta:
+    func: RecvV2InferMeta
+    param: [peer, dtype, out_shape]
+  kernel :
+    func : recv_v2
+    param : [ring_id, dynamic_shape, peer, out_shape, dtype, use_calc_stream]
+    data_type : dtype
+
+- op : save_combine
+  args : (Tensor[] x, str file_path, bool overwrite, bool save_as_fp16, bool save_to_memory)
+  output : Tensor(out)
+  kernel:
+    func: save_combine_tensor
+    param: [x, file_path, overwrite, save_as_fp16, save_to_memory]
+  optional : out
+
+- op : send_v2
+  args : (Tensor x, int ring_id = 0, int peer = 0, bool use_calc_stream = false, bool dynamic_shape = false)
+  output :
+  infer_meta:
+    func: SendV2InferMeta
+    param: [peer, ring_id]
+  kernel :
+    func : send_v2
+    param : [x, ring_id, dynamic_shape, peer, use_calc_stream]
+
+- op : set_value
+  args : (Tensor x, int64_t[] starts, int64_t[] ends, int64_t[] steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes, int64_t[] shape, Scalar[] values)
+  output : Tensor(out)
+  infer_meta:
+    func: SetValueInferMeta
+    param: [x]
+  kernel:
+    func: set_value
+    param: [x, starts, ends, steps, axes, decrease_axes, none_axes, shape, values]
+  inplace: (x -> out)
+  backward: set_value_grad
+
+- op : set_value_with_tensor
+  args : (Tensor x, Tensor values, int64_t[] starts, int64_t[] ends, int64_t[] steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
+  output : Tensor(out)
+  infer_meta:
+    func: SetValueInferMeta
+    param: [x]
+  kernel:
+    func: set_value_with_tensor
+    param: [x, values, starts, ends, steps, axes, decrease_axes, none_axes]
+  inplace: (x -> out)
+  backward: set_value_grad
+
+- op : shadow_feed
+  args : (Tensor x)
+  output : Tensor(out)
+  infer_meta:
+    func: UnchangedInferMeta
+    param: [x]
+  kernel:
+    func: shadow_feed
+    param: [x]
+
+- op : share_buffer_
+  args : (Tensor[] x, bool[] share_dims_and_dtype = {})
+  output : Tensor[](out){x.size()}, Tensor[](xout){x.size()}
+
+- op : write_to_array
+  args : (Tensor i, Tensor x)
+  output : Tensor[](out)
+  backward: write_to_array_grad

paddle/fluid/ir/dialect/paddle_dialect/ir/pd_ops_backward.yaml

+- backward_op : set_value_grad
+  args : (Tensor out_grad, Tensor values, int64_t[] starts, int64_t[] ends, int64_t[] steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
+  output : Tensor(x_grad), Tensor(values_grad)
+  infer_meta:
+    func: SetValueGradInferMeta
+    param: [out_grad, values]
+  kernel:
+    func: set_value_grad
+    param: [out_grad, starts, ends, steps, axes, decrease_axes, none_axes]

paddle/fluid/operators/generator/CMakeLists.txt

@@ -122,7 +122,6 @@ set(generated_static_argument_mapping_path
     ${CMAKE_SOURCE_DIR}/paddle/phi/ops/compat/generated_static_sig.cc)
 set(generated_sparse_argument_mapping_path
     ${CMAKE_SOURCE_DIR}/paddle/phi/ops/compat/generated_sparse_sig.cc)
-
 execute_process(
   WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/generator
   COMMAND ${CMAKE_COMMAND} -E make_directory ${parsed_op_dir} RESULTS_VARIABLE
@@ -139,6 +138,7 @@ message(
 - ${legacy_bw_op_yaml_file}
 - ${fused_op_yaml_file}
 - ${static_op_yaml_file}")
+
 execute_process(
   WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/generator
   COMMAND ${PYTHON_EXECUTABLE} parse_op.py --op_yaml_path ${op_yaml_file}

paddle/fluid/operators/generator/parse_utils.py

@@ -150,6 +150,8 @@ def parse_output(op_name: str, s: str) -> Dict[str, str]:
 
 
 def parse_outputs(op_name: str, outputs: str) -> List[Dict]:
+    if outputs is None:
+        return []
     outputs = parse_plain_list(outputs, sep=",")
     output_items = []
     for output in outputs:
@@ -169,6 +171,7 @@ def parse_candidates(s: str) -> Dict[str, Any]:
     delimiter = ">" if ">" in s else ","
     ordered = delimiter == ">"
     candidates = parse_plain_list(s, delimiter)
+    candidates = list(filter(None, candidates))
     return {"ordered": ordered, "candidates": candidates}
 
 
@@ -523,14 +526,20 @@ def parse_op_entry(op_entry: Dict[str, Any], name_field="op"):
 
     if is_base_op:
         # kernel
+        if "kernel" in op_entry:
             kernel = parse_kernel(op_name, op_entry["kernel"])
             if kernel["param"] is None:
                 kernel["param"] = input_names + attr_names
+            op.update({"kernel": kernel})
 
         # infer meta
+        if "infer_meta" in op_entry:
             infer_meta = parse_infer_meta(op_entry["infer_meta"])
             if infer_meta["param"] is None:
                 infer_meta["param"] = copy(kernel["param"])
+            op.update({"infer_meta": infer_meta})
+        # else:
+        #     assert(outputs == []), f"No infer_meta is given in {op_name}."
 
         # inplace
         if "inplace" in op_entry:
@@ -544,8 +553,6 @@ def parse_op_entry(op_entry: Dict[str, Any], name_field="op"):
             view_pairs = None
         op.update(
             {
-                "infer_meta": infer_meta,
-                "kernel": kernel,
                 "inplace": inplace_pairs,
                 "view": view_pairs,
             }

paddle/fluid/operators/generator/type_mapping.py

@@ -79,7 +79,11 @@ opmaker_attr_types_map = {
     'str[]': 'std::vector<std::string>',
 }
 
-output_type_map = {'Tensor': 'Tensor', 'Tensor[]': 'std::vector<Tensor>'}
+output_type_map = {
+    'Tensor': 'Tensor',
+    'Tensor[]': 'std::vector<Tensor>',
+    'SelectedRows': 'SelectedRows',
+}
 
 # ------------------------------ phi attr ------------------------------
 phi_attr_types_map = attr_types_map.copy()

paddle/phi/api/yaml/op_compat.yaml

@@ -2375,6 +2375,10 @@
   extra :
     attrs : [bool use_mkldnn = false, bool use_cudnn = false]
 
+- op : save_combine
+  inputs :
+    {x : X}
+
 - op : scale
   backward : scale_grad
   inputs :