[Tensor API] Support multiple Tensor C++ api (#50731)

* change phi tensor_gen->tensor_operants_gen

* [Tensor API] Support multiple Tensor C++ api

paddle/fluid/prim/api/auto_code_generated/CMakeLists.txt

@@ -78,30 +78,22 @@ execute_process(
           ${static_prim_api_cc_path})
 message("copy tmp_xxx_prim_api to xxx_prim_api")
 
-set(tmp_eager_tensor_operants_cc_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/eager/eager_tensor_operants.cc.tmp"
-)
-set(tmp_eager_tensor_operants_h_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/eager/eager_tensor_operants.h.tmp"
-)
-set(tmp_static_tensor_operants_cc_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/static/static_tensor_operants.cc.tmp"
-)
-set(tmp_static_tensor_operants_h_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/static/static_tensor_operants.h.tmp"
-)
 set(eager_tensor_operants_cc_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/eager/eager_tensor_operants.cc"
-)
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/eager/eager_tensor_operants.cc)
 set(eager_tensor_operants_h_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/eager/eager_tensor_operants.h"
-)
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/eager/eager_tensor_operants.h)
 set(static_tensor_operants_cc_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/static/static_tensor_operants.cc"
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/static/static_tensor_operants.cc
 )
 set(static_tensor_operants_h_path
-    "${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/static/static_tensor_operants.h"
-)
+    ${PADDLE_SOURCE_DIR}/paddle/fluid/prim/utils/static/static_tensor_operants.h
+)
+set(tmp_eager_tensor_operants_cc_path ${eager_tensor_operants_cc_path}.tmp)
+set(tmp_eager_tensor_operants_h_path ${eager_tensor_operants_h_path}.tmp)
+set(tmp_static_tensor_operants_cc_path ${static_tensor_operants_cc_path}.tmp)
+set(tmp_static_tensor_operants_h_path ${static_tensor_operants_h_path}.tmp)
+set(tensor_api_yaml_path
+    ${PADDLE_SOURCE_DIR}/paddle/phi/api/yaml/tensor_api.yaml)
 
 message("Prim tensor operants code generator")
 execute_process(
@@ -114,7 +106,7 @@ execute_process(
     ${tmp_eager_tensor_operants_cc_path} --static_tensor_operants_header_path
     ${tmp_static_tensor_operants_h_path} --static_tensor_operants_source_path
     ${tmp_static_tensor_operants_cc_path} --api_prim_yaml_path
-    ${api_prim_yaml_path}
+    ${tensor_api_yaml_path}
   RESULT_VARIABLE _result)
 if(${_result})
   message(FATAL_ERROR "Prim tensor operants generate failed, exiting.")

paddle/fluid/prim/api/auto_code_generated/tensor_operants_gen.py

@@ -25,6 +25,8 @@ eager_header_include = """// Generated by paddle/fluid/prim/api/auto_code_genera
 
 #include "paddle/phi/api/include/operants_base.h"
 #include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/common/scalar.h"
+#include "paddle/phi/common/int_array.h"
 #include "paddle/phi/core/macros.h"
 
 """
@@ -35,6 +37,8 @@ namespace paddle {
 namespace prim {
 
 using Tensor = paddle::experimental::Tensor;
+using Scalar = paddle::experimental::Scalar;
+using IntArray = paddle::experimental::IntArray;
 using TensorOperantsBase = paddle::operants::TensorOperantsBase;
 
 class EagerTensorOperants : public TensorOperantsBase {
@@ -85,6 +89,8 @@ static_header_include = """// Generated by paddle/fluid/prim/api/auto_code_gener
 
 #include "paddle/phi/api/include/operants_base.h"
 #include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/common/scalar.h"
+#include "paddle/phi/common/int_array.h"
 #include "paddle/phi/core/macros.h"
 
 """
@@ -95,6 +101,8 @@ namespace paddle {
 namespace prim {
 
 using Tensor = paddle::experimental::Tensor;
+using Scalar = paddle::experimental::Scalar;
+using IntArray = paddle::experimental::IntArray;
 using TensorOperantsBase = paddle::operants::TensorOperantsBase;
 
 class StaticTensorOperants : public TensorOperantsBase {
@@ -269,8 +277,6 @@ def generate_tensor_operants_api(
 
     with open(api_prim_path, 'rt') as f:
         api_prims = yaml.safe_load(f)
-        # white list temporarily
-        api_prims = ('add', 'subtract', 'multiply', 'divide')
 
     for api in apis:
         eager_api = PrimTensorAPI(api, api_prims)

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

@@ -61,7 +61,7 @@ void gather_grad(const Tensor& x,
 template <typename T>
 void tanh_grad(const Tensor& out, const Tensor& grad_out, Tensor* grad_x) {
   if (!grad_x) return;
-  auto tmp = pow<T>(out, 2.0);
+  auto tmp = out.pow(2.0);
   tmp = scale<T>(tmp, -1.0, 1.0, true);
   auto grad_x_tmp = grad_out * tmp;
   set_output<T>(grad_x_tmp, grad_x);
@@ -82,8 +82,8 @@ void subtract_grad(const Tensor& x,
       if (!reduce_dim.size()) {
         by_pass<T>(scale_out_grad, dy);
       } else {
-        auto dy_reduce_res = sum<T>(
-            scale_out_grad, phi::vectorize(reduce_dim), y.dtype(), false);
+        auto dy_reduce_res =
+            scale_out_grad.sum(phi::vectorize(reduce_dim), y.dtype(), false);
         auto dy_tmp = reshape<T>(dy_reduce_res, phi::vectorize(y.dims()));
         set_output<T>(dy_tmp, dy);
       }
@@ -99,7 +99,7 @@ void subtract_grad(const Tensor& x,
         by_pass<T>(out_grad, dx);
       } else {
         auto dx_reduce_res =
-            sum<T>(out_grad, phi::vectorize(reduce_dim), x.dtype(), false);
+            out_grad.sum(phi::vectorize(reduce_dim), x.dtype(), false);
         auto dx_tmp = reshape<T>(dx_reduce_res, phi::vectorize(x.dims()));
         set_output<T>(dx_tmp, dx);
       }
@@ -124,7 +124,7 @@ void add_grad(const Tensor& x,
         by_pass<T>(out_grad, dy);
       } else {
         auto dy_reduce_res =
-            sum<T>(out_grad, phi::vectorize(reduce_dim), y.dtype(), false);
+            out_grad.sum(phi::vectorize(reduce_dim), y.dtype(), false);
         auto dy_tmp = reshape<T>(dy_reduce_res, phi::vectorize(y.dims()));
         set_output<T>(dy_tmp, dy);
       }
@@ -141,7 +141,7 @@ void add_grad(const Tensor& x,
         by_pass<T>(out_grad, dx);
       } else {
         auto dx_reduce_res =
-            sum<T>(out_grad, phi::vectorize(reduce_dim), x.dtype(), false);
+            out_grad.sum(phi::vectorize(reduce_dim), x.dtype(), false);
         auto dx_tmp = reshape<T>(dx_reduce_res, phi::vectorize(x.dims()));
         set_output<T>(dx_tmp, dx);
       }
@@ -172,7 +172,7 @@ void sum_grad(const Tensor& x,
   }
   auto x_grad_tmp = Tensor();
   if (x_dim_size == 1) {
-    x_grad_tmp = expand<T>(out_grad, IntArray(x_dim));
+    x_grad_tmp = out_grad.expand(IntArray(x_dim));
   } else {
     if (!keepdim) {
       auto axis_ = std::vector<int64_t>();
@@ -184,9 +184,9 @@ void sum_grad(const Tensor& x,
         axis_ = axis.GetData();
       }
       auto out_grad_ = unsqueeze<T>(out_grad, axis_);
-      x_grad_tmp = expand<T>(out_grad_, IntArray(x_dim));
+      x_grad_tmp = out_grad_.expand(IntArray(x_dim));
     } else {
-      x_grad_tmp = expand<T>(out_grad, IntArray(x_dim));
+      x_grad_tmp = out_grad.expand(IntArray(x_dim));
     }
   }
 
@@ -203,7 +203,7 @@ void divide_grad(const Tensor& x,
                  Tensor* dy) {
   if (dy) {
     // dy = -(x/y^2) * dout
-    auto tmp0 = pow<T>(y, 2.0);
+    auto tmp0 = y.pow(2.0);
     auto tmp1 = x / tmp0;
     auto tmp2 = scale<T>(tmp1, -1.0, 0.0, true);
     auto dy_res = tmp2 * out_grad;
@@ -214,7 +214,7 @@ void divide_grad(const Tensor& x,
         set_output<T>(dy_res, dy);
       } else {
         auto dy_reduce_res =
-            sum<T>(dy_res, phi::vectorize(reduce_dim), y.dtype(), false);
+            dy_res.sum(phi::vectorize(reduce_dim), y.dtype(), false);
         auto dy_tmp = reshape<T>(dy_reduce_res, phi::vectorize(y.dims()));
         set_output<T>(dy_tmp, dy);
       }
@@ -233,7 +233,7 @@ void divide_grad(const Tensor& x,
         set_output<T>(dx_res, dx);
       } else {
         auto dx_reduce_res =
-            sum<T>(dx_res, phi::vectorize(reduce_dim), x.dtype(), false);
+            dx_res.sum(phi::vectorize(reduce_dim), x.dtype(), false);
         auto dx_tmp = reshape<T>(dx_reduce_res, phi::vectorize(x.dims()));
         set_output<T>(dx_tmp, dx);
       }
@@ -267,10 +267,8 @@ void multiply_grad(const Tensor& x,
       if (!axes.size()) {
         set_output<T>(x_grad_unreduce, x_grad);
       } else {
-        auto x_grad_reduced = sum<T>(x_grad_unreduce,
-                                     phi::vectorize(axes),
-                                     x_grad_unreduce.dtype(),
-                                     false);
+        auto x_grad_reduced = x_grad_unreduce.sum(
+            phi::vectorize(axes), x_grad_unreduce.dtype(), false);
         if (x_grad_reduced.dims().size() != x.dims().size()) {
           x_grad_reduced = reshape<T>(x_grad_reduced, x.shape());
         }
@@ -287,10 +285,8 @@ void multiply_grad(const Tensor& x,
       if (!axes.size()) {
         set_output<T>(y_grad_unreduce, y_grad);
       } else {
-        auto y_grad_reduced = sum<T>(y_grad_unreduce,
-                                     phi::vectorize(axes),
-                                     y_grad_unreduce.dtype(),
-                                     false);
+        auto y_grad_reduced = y_grad_unreduce.sum(
+            phi::vectorize(axes), y_grad_unreduce.dtype(), false);
         if (y_grad_reduced.dims().size() != y.dims().size()) {
           y_grad_reduced = reshape<T>(y_grad_reduced, y.shape());
         }
@@ -314,7 +310,7 @@ void expand_grad(const Tensor& x,
       if (!axes.size()) {
         by_pass<T>(out_grad, x_grad);
       } else {
-        auto reduced = sum<T>(out_grad, phi::vectorize(axes), x.dtype(), false);
+        auto reduced = out_grad.sum(phi::vectorize(axes), x.dtype(), false);
         if (reduced.dims().size() != x.dims().size()) {
           reduced = reshape<T>(reduced, x.shape());
         }

paddle/phi/api/include/tensor.h

@@ -47,6 +47,16 @@ namespace paddle {
 
 namespace experimental {
 
+class Tensor;
+
+template <typename T>
+class ScalarBase;
+using Scalar = paddle::experimental::ScalarBase<Tensor>;
+
+template <typename T>
+class IntArrayBase;
+using IntArray = paddle::experimental::IntArrayBase<Tensor>;
+
 class AbstractAutogradMeta {
  public:
   // No AbstractAutogradMeta should be created
@@ -646,12 +656,35 @@ class PADDLE_API Tensor final {
    */
   std::string name_{""};
 
+ public:
   // Tensor C++ APIs
   // Example: Tensor add(const Tensor& other) const;
   Tensor add(const Tensor& y) const;
   Tensor divide(const Tensor& y) const;
   Tensor multiply(const Tensor& y) const;
   Tensor subtract(const Tensor& y) const;
+  Tensor exp() const;
+  Tensor floor() const;
+  Tensor gather_nd(const Tensor& index) const;
+  Tensor log() const;
+  Tensor pow(const Scalar& y) const;
+  Tensor roll(const IntArray& shifts, const std::vector<int64_t>& axis) const;
+  Tensor scatter(const Tensor& index,
+                 const Tensor& updates,
+                 bool overwrite) const;
+  Tensor scatter_nd_add(const Tensor& index, const Tensor& updates) const;
+  Tensor abs() const;
+  Tensor assign() const;
+  Tensor elementwise_pow(const Tensor& y) const;
+  Tensor expand(const IntArray& shape) const;
+  Tensor matmul(const Tensor& y, bool transpose_x, bool transpose_y) const;
+  Tensor max(const IntArray& axis, bool keepdim) const;
+  Tensor maximum(const Tensor& y) const;
+  Tensor minimum(const Tensor& y) const;
+  Tensor prod(const IntArray& dims, bool keep_dim, bool reduce_all) const;
+  Tensor scale(const Scalar& scale, float bias, bool bias_after_scale) const;
+  Tensor sum(const IntArray& axis, DataType dtype, bool keepdim) const;
+  Tensor tile(const IntArray& repeat_times) const;
 };
 
 }  // namespace experimental

paddle/phi/api/lib/CMakeLists.txt

@@ -98,8 +98,10 @@ set(wrapped_infermeta_source_file
     ${CMAKE_SOURCE_DIR}/paddle/phi/infermeta/generated.cc)
 
 # tensor and tensor operants file
+set(tensor_api_yaml_path
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/yaml/tensor_api.yaml)
 set(tensor_gen_file
-    ${CMAKE_SOURCE_DIR}/paddle/phi/api/yaml/generator/tensor_gen.py)
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/yaml/generator/tensor_operants_gen.py)
 set(operants_base_file
     ${CMAKE_SOURCE_DIR}/paddle/phi/api/include/operants_base.h)
 set(tensor_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/tensor_api.cc)
@@ -249,6 +251,7 @@ execute_process(
     --phi_tensor_operants_source_path ${phi_tensor_operants_source_file_tmp}
     --operants_manager_header_path ${operants_manager_header_file_tmp}
     --operants_manager_source_path ${operants_manager_source_file_tmp}
+    --tensor_api_yaml_path ${tensor_api_yaml_path}
   RESULT_VARIABLE _result)
 if(${_result})
   message(FATAL_ERROR "tensor codegen failed, exiting.")

paddle/phi/api/yaml/generator/tensor_gen.py → paddle/phi/api/yaml/generator/tensor_operants_gen.py

@@ -30,11 +30,13 @@ inplace_optional_out_type_map = {
 indent = "  "
 
 
-operants_base_include = """// Generated by paddle/phi/api/yaml/generator/tensor_gen.py
+operants_base_include = """// Generated by paddle/phi/api/yaml/generator/tensor_operants_gen.py
 
 #pragma once
 
 #include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/common/scalar.h"
+#include "paddle/phi/common/int_array.h"
 
 """
 
@@ -44,6 +46,8 @@ namespace paddle {
 namespace operants {
 
 using Tensor = paddle::experimental::Tensor;
+using Scalar = paddle::experimental::Scalar;
+using IntArray = paddle::experimental::IntArray;
 
 class TensorOperantsBase {
  public:
@@ -58,7 +62,7 @@ operants_base_end = """};
 
 """
 
-tensor_api_source_include = """// Generated by paddle/phi/api/yaml/generator/tensor_gen.py
+tensor_api_source_include = """// Generated by paddle/phi/api/yaml/generator/tensor_operants_gen.py
 
 #include "paddle/phi/api/include/tensor.h"
 
@@ -96,12 +100,14 @@ tensor_api_source_end = """
 """
 
 
-operants_header_include = """// Generated by paddle/phi/api/yaml/generator/tensor_gen.py
+operants_header_include = """// Generated by paddle/phi/api/yaml/generator/tensor_operants_gen.py
 
 #pragma once
 
 #include "paddle/phi/api/include/operants_base.h"
 #include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/common/scalar.h"
+#include "paddle/phi/common/int_array.h"
 #include "paddle/phi/core/macros.h"
 
 """
@@ -111,6 +117,9 @@ namespace paddle {
 
 namespace operants {
 
+using Scalar = paddle::experimental::Scalar;
+using IntArray = paddle::experimental::IntArray;
+
 class PhiTensorOperants : public TensorOperantsBase {
  private:
   DISABLE_COPY_AND_ASSIGN(PhiTensorOperants);
@@ -128,7 +137,7 @@ operants_header_end = """};
 """
 
 
-operants_source_include = """// Generated by paddle/phi/api/yaml/generator/tensor_gen.py
+operants_source_include = """// Generated by paddle/phi/api/yaml/generator/tensor_operants_gen.py
 
 #include "paddle/phi/api/include/tensor_operants.h"
 
@@ -151,12 +160,14 @@ operants_source_end = """
 """
 
 
-operants_manager_header_include = """// Generated by paddle/phi/api/yaml/generator/tensor_gen.py
+operants_manager_header_include = """// Generated by paddle/phi/api/yaml/generator/tensor_operants_gen.py
 
 #pragma once
 
 #include "paddle/phi/api/include/operants_base.h"
 #include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/common/scalar.h"
+#include "paddle/phi/common/int_array.h"
 #include "paddle/phi/core/macros.h"
 
 """
@@ -165,6 +176,8 @@ operants_manager_header_start = """
 namespace paddle {
 
 using Tensor = paddle::experimental::Tensor;
+using Scalar = paddle::experimental::Scalar;
+using IntArray = paddle::experimental::IntArray;
 using TensorOperantsBase = paddle::operants::TensorOperantsBase;
 
 /**
@@ -222,7 +235,7 @@ operants_manager_header_end = """};
 """
 
 
-operants_manager_source_include = """// Generated by paddle/phi/api/yaml/generator/tensor_gen.py
+operants_manager_source_include = """// Generated by paddle/phi/api/yaml/generator/tensor_operants_gen.py
 
 #include "paddle/phi/api/include/operants_manager.h"
 
@@ -271,11 +284,17 @@ class OperantsAPI(ForwardAPI):
 """
 
     def get_define_args_without_first_tensor(self, inplace_flag=False):
-        # NOTE(HongyuJia): consider vector<Tensor> becomes first input argument.
+        func_name = self.get_api_func_name()
         define_args = self.get_input_tensor_args(inplace_flag)
-        assert (
-            len(define_args) > 1
-        ), "Can't use tensor api without Tensor inputs"
+        assert len(define_args) >= 1, (
+            "Error! Api %s has no Tensor inputs" % func_name
+        )
+        first_input_type = " ".join(define_args[0].split(" ")[:-1])
+        # NOTE(HongyuJia): Do not consider "const paddle::optional<Tensor>&"
+        assert first_input_type == "const Tensor&", (
+            "Error! The first argument of Tensor Api %s must be Tensor, but received %s"
+            % (func_name, first_input_type)
+        )
         for name in self.attrs['names']:
             define_args.append(self.attrs['attr_info'][name][0] + ' ' + name)
         # remove first Tensor argument
@@ -283,23 +302,26 @@ class OperantsAPI(ForwardAPI):
 
     def gene_tensor_api_implementation(self):
         func_name = self.get_api_func_name()
-        assert (
-            len(self.inputs['names']) > 1
-        ), "Can't use tensor api without Tensor inputs"
+        assert len(self.inputs['names']) >= 1, (
+            "Error! Api %s has no Tensor inputs" % func_name
+        )
         # remove first Tensor argument
         func_args = self.inputs['names'][1:] + self.attrs['names']
-        func_args_code = ", ".join(func_args)
+        if len(func_args) > 0:
+            func_args_code = ", ".join([""] + func_args)
+        else:
+            func_args_code = ""
         # func decalaration
         if func_name[-1] != '_':
             return f"""
 {self.get_return_type()} Tensor::{func_name}({self.get_define_args_without_first_tensor()}) const {{
-{indent}return paddle::OperantsManager::Instance().{func_name}(static_cast<const Tensor &>(*this), {func_args_code});
+{indent}return paddle::OperantsManager::Instance().{func_name}(static_cast<const Tensor &>(*this){func_args_code});
 }}
 """
         else:
             return f"""
 {self.get_return_type(inplace_flag=True)} Tensor::{func_name}({self.get_define_args_without_first_tensor(inplace_flag=True)}) const {{
-{indent}return paddle::OperantsManager::Instance().{func_name}(static_cast<const Tensor &>(*this), {func_args_code});
+{indent}return paddle::OperantsManager::Instance().{func_name}(static_cast<const Tensor &>(*this){func_args_code});
 }}
 
 """
@@ -394,6 +416,7 @@ def generate_tensor_operants_api(
     operants_source_path,
     operants_manager_header_path,
     operants_manager_source_path,
+    tensor_api_yaml_path,
 ):
     apis = []
 
@@ -423,8 +446,8 @@ def generate_tensor_operants_api(
     operants_manager_source_file.write(operants_manager_source_include)
     operants_manager_source_file.write(operants_manager_source_start)
 
-    # white list temporarily
-    api_prims = ('add', 'subtract', 'multiply', 'divide')
+    with open(tensor_api_yaml_path, 'rt') as f:
+        api_prims = yaml.safe_load(f)
 
     for api in apis:
         operants_api = OperantsAPI(api, api_prims)
@@ -506,6 +529,12 @@ def main():
         default='paddle/phi/api/lib/operants_manager.cc',
     )
 
+    parser.add_argument(
+        '--tensor_api_yaml_path',
+        help='path to tensor_api yaml file',
+        default='paddle/phi/api/yaml/tensor_api.yaml',
+    )
+
     options = parser.parse_args()
 
     api_yaml_path = options.api_yaml_path
@@ -515,6 +544,7 @@ def main():
     operants_source_path = options.phi_tensor_operants_source_path
     operants_manager_header_path = options.operants_manager_header_path
     operants_manager_source_path = options.operants_manager_source_path
+    tensor_api_yaml_path = options.tensor_api_yaml_path
 
     generate_tensor_operants_api(
         api_yaml_path,
@@ -524,6 +554,7 @@ def main():
         operants_source_path,
         operants_manager_header_path,
         operants_manager_source_path,
+        tensor_api_yaml_path,
     )
 
 

paddle/phi/api/yaml/tensor_api.yaml

+- unsqueeze
+- pow
+- exp
+- scale
+- multiply
+- matmul
+- expand
+- divide
+- sum
+- add
+- abs
+- assign
+- elementwise_pow
+- floor
+- gather_nd
+- log
+- max
+- maximum
+- minimum
+- prod
+- roll
+- scatter
+- scatter_nd_add
+- tile
+- subtract