diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index b87cdf6f6df19314342a24c98032d3856f0d3779..a279c76430f1b046a4c3ca05485824d5e3b62de2 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -1,5 +1,9 @@
 include(operators)
 
+# solve "math constants not defined" problems caused by the order of inclusion 
+# of <cmath> and the definition of macro _USE_MATH_DEFINES
+add_definitions(-D_USE_MATH_DEFINES)
+
 # clean cache and pybind_file content first when rebuild
 unset(GLOB_OP_LIB CACHE)
 unset(OP_LIBRARY CACHE)
diff --git a/paddle/fluid/operators/angle_op.h b/paddle/fluid/operators/angle_op.h
index 093a04f03df95681d7837d5c44717c678589e679..1e0dc803d76123573332b040bc29ece263c11d80 100644
--- a/paddle/fluid/operators/angle_op.h
+++ b/paddle/fluid/operators/angle_op.h
@@ -17,7 +17,7 @@
 #define _USE_MATH_DEFINES
 #endif
 #include <cmath>
-#include "paddle/fluid/operators/math/complex_functors.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
@@ -26,81 +26,6 @@
 namespace paddle {
 namespace operators {
 
-namespace math {
-template <typename T, typename Enable = void>
-struct AngleFunctor;
-
-// angel function for complex
-template <typename T>
-struct AngleFunctor<T, Complex<T, Real<T>>> {
-  AngleFunctor(const T* input, Real<T>* output, int64_t numel)
-      : input_(input), output_(output), numel_(numel) {}
-
-  HOSTDEVICE void operator()(int64_t idx) const {
-    output_[idx] = arg(input_[idx]);
-  }
-
-  const T* input_;
-  Real<T>* output_;
-  int64_t numel_;
-};
-
-// angel function for real
-template <typename T>
-struct AngleFunctor<T, NoComplex<T, Real<T>>> {
-  AngleFunctor(const T* input, T* output, int64_t numel)
-      : input_(input), output_(output), numel_(numel) {}
-
-  HOSTDEVICE void operator()(int64_t idx) const {
-    output_[idx] = input_[idx] < static_cast<T>(0) ? M_PI : 0;
-  }
-
-  const T* input_;
-  T* output_;
-  int64_t numel_;
-};
-
-template <typename T, typename Enable = void>
-struct AngleGradFunctor;
-
-// angle grad for complex
-template <typename T>
-struct AngleGradFunctor<T, Complex<T, Real<T>>> {
-  AngleGradFunctor(const math::Real<T>* dout, const T* x, T* dx, int64_t numel)
-      : dout_(dout), x_(x), dx_(dx), numel_(numel) {}
-
-  HOSTDEVICE void operator()(int64_t idx) const {
-    if (x_[idx] == T(0)) {
-      dx_[idx] = T(0);
-    } else {
-      const math::Real<T> r_square =
-          x_[idx].real * x_[idx].real + x_[idx].imag * x_[idx].imag;
-      dx_[idx] = T(-dout_[idx] * x_[idx].imag / r_square,
-                   dout_[idx] * x_[idx].real / r_square);
-    }
-  }
-
-  const math::Real<T>* dout_;
-  const T* x_;
-  T* dx_;
-  int64_t numel_;
-};
-
-// angle grad for real
-template <typename T>
-struct AngleGradFunctor<T, NoComplex<T, Real<T>>> {
-  AngleGradFunctor(const math::Real<T>* dout, const T* x, T* dx, int64_t numel)
-      : dout_(dout), x_(x), dx_(dx), numel_(numel) {}
-
-  HOSTDEVICE void operator()(int64_t idx) const { dx_[idx] = 0; }
-
-  const math::Real<T>* dout_;
-  const T* x_;
-  T* dx_;
-  int64_t numel_;
-};
-}  // namespace math
-
 using Tensor = framework::Tensor;
 template <typename DeviceContext, typename T>
 class AngleKernel : public framework::OpKernel<T> {
@@ -111,12 +36,12 @@ class AngleKernel : public framework::OpKernel<T> {
 
     auto numel = x->numel();
     auto* x_data = x->data<T>();
-    auto* out_data = out->mutable_data<math::Real<T>>(
-        context.GetPlace(), size_t(x->numel() * sizeof(math::Real<T>)));
+    auto* out_data = out->mutable_data<pten::funcs::Real<T>>(
+        context.GetPlace(), size_t(x->numel() * sizeof(pten::funcs::Real<T>)));
 
     auto& dev_ctx = context.template device_context<DeviceContext>();
     platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::AngleFunctor<T> functor(x_data, out_data, numel);
+    pten::funcs::AngleFunctor<T> functor(x_data, out_data, numel);
     for_range(functor);
   }
 };
@@ -132,14 +57,14 @@ class AngleGradKernel : public framework::OpKernel<T> {
         ctx.Output<framework::Tensor>(framework::GradVarName("X"));
 
     auto numel = d_out->numel();
-    auto* dout_data = d_out->data<math::Real<T>>();
+    auto* dout_data = d_out->data<pten::funcs::Real<T>>();
     auto* x_data = x->data<T>();
     auto* dx_data = d_x->mutable_data<T>(
         ctx.GetPlace(), static_cast<size_t>(numel * sizeof(T)));
 
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::AngleGradFunctor<T> functor(dout_data, x_data, dx_data, numel);
+    pten::funcs::AngleGradFunctor<T> functor(dout_data, x_data, dx_data, numel);
     for_range(functor);
   }
 };
diff --git a/paddle/fluid/operators/cholesky_solve_op.h b/paddle/fluid/operators/cholesky_solve_op.h
index 5004aad7c59bc4ad194bf961aaad6326ad03fd38..2c92969225f3bcbb8008c24c21e4a6f80dd03fd4 100644
--- a/paddle/fluid/operators/cholesky_solve_op.h
+++ b/paddle/fluid/operators/cholesky_solve_op.h
@@ -64,7 +64,7 @@ void cholesky_solve_fn(const paddle::framework::ExecutionContext &ctx,
   // calculate u's conjugate for complex
   framework::Tensor u_conj(u_bst.type());
   platform::ForRange<DeviceContext> u_for_range(dev_ctx, u_bst.numel());
-  math::ConjFunctor<T> u_functor(
+  pten::funcs::ConjFunctor<T> u_functor(
       u_bst.data<T>(), u_bst.numel(),
       u_conj.mutable_data<T>(u_bst.dims(), dev_ctx.GetPlace()));
   u_for_range(u_functor);
@@ -73,7 +73,7 @@ void cholesky_solve_fn(const paddle::framework::ExecutionContext &ctx,
   // calculate b's conjugate for complex
   framework::Tensor b_conj(b_bst.type());
   platform::ForRange<DeviceContext> b_for_range(dev_ctx, b_bst.numel());
-  math::ConjFunctor<T> b_functor(
+  pten::funcs::ConjFunctor<T> b_functor(
       b_bst.data<T>(), b_bst.numel(),
       b_conj.mutable_data<T>(b_bst.dims(), dev_ctx.GetPlace()));
   b_for_range(b_functor);
@@ -113,7 +113,7 @@ void cholesky_solve_fn(const paddle::framework::ExecutionContext &ctx,
 
   // calculate out's conjugate for complex
   platform::ForRange<DeviceContext> out_for_range(dev_ctx, out->numel());
-  math::ConjFunctor<T> out_functor(
+  pten::funcs::ConjFunctor<T> out_functor(
       out->data<T>(), out->numel(),
       out->mutable_data<T>(out->dims(), dev_ctx.GetPlace()));
   out_for_range(out_functor);
@@ -173,7 +173,7 @@ class CholeskySolveGradKernel : public framework::OpKernel<T> {
       // calculate out's conjugate for complex
       framework::Tensor out_conj(out->type());
       platform::ForRange<DeviceContext> out_for_range(dev_ctx, out->numel());
-      math::ConjFunctor<T> out_functor(
+      pten::funcs::ConjFunctor<T> out_functor(
           out->data<T>(), out->numel(),
           out_conj.mutable_data<T>(out->dims(), dev_ctx.GetPlace()));
       out_for_range(out_functor);
@@ -195,7 +195,7 @@ class CholeskySolveGradKernel : public framework::OpKernel<T> {
       framework::Tensor commonterm_conj(commonterm.type());
       platform::ForRange<DeviceContext> commonterm_for_range(
           dev_ctx, commonterm.numel());
-      math::ConjFunctor<T> commonterm_functor(
+      pten::funcs::ConjFunctor<T> commonterm_functor(
           commonterm.data<T>(), commonterm.numel(),
           commonterm_conj.mutable_data<T>(commonterm.dims(),
                                           dev_ctx.GetPlace()));
diff --git a/paddle/fluid/operators/complex_op.h b/paddle/fluid/operators/complex_op.h
index 3dd5ea9f7e83dbfaa353378cfee10231c445c222..fb324277fb004b93718793346957b9adbb10143b 100644
--- a/paddle/fluid/operators/complex_op.h
+++ b/paddle/fluid/operators/complex_op.h
@@ -16,8 +16,8 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/complex.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/fluid/operators/complex_view_op.h b/paddle/fluid/operators/complex_view_op.h
index 9a8d89db4020828cc9fdba90dd99ab7e5395864b..98ba732e2400421073fdbefc76ff4207fe5a9a8d 100644
--- a/paddle/fluid/operators/complex_view_op.h
+++ b/paddle/fluid/operators/complex_view_op.h
@@ -17,9 +17,9 @@
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/complex.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/fluid/operators/cumprod_op.cu b/paddle/fluid/operators/cumprod_op.cu
index 2b69db7d24a128de77e8508970705ac3b9a1fb66..3a63bd99ad57d8c91235ddf50219e6b015321972 100644
--- a/paddle/fluid/operators/cumprod_op.cu
+++ b/paddle/fluid/operators/cumprod_op.cu
@@ -14,9 +14,9 @@
 
 #include <thrust/transform.h>
 #include "paddle/fluid/operators/cumprod_op.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/math/inclusive_scan.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -243,12 +243,12 @@ class CumprodGradOpCUDAKernel : public framework::OpKernel<T> {
 
       platform::ForRange<platform::CUDADeviceContext> for_range_x(dev_ctx,
                                                                   numel);
-      math::ConjFunctor<T> functor_x(x_data, numel, x_data_conj);
+      pten::funcs::ConjFunctor<T> functor_x(x_data, numel, x_data_conj);
       for_range_x(functor_x);
 
       platform::ForRange<platform::CUDADeviceContext> for_range_y(dev_ctx,
                                                                   numel);
-      math::ConjFunctor<T> functor_y(y_data, numel, y_data_conj);
+      pten::funcs::ConjFunctor<T> functor_y(y_data, numel, y_data_conj);
       for_range_y(functor_y);
       x_data_deal = x_data_conj;
       y_data_deal = y_data_conj;
diff --git a/paddle/fluid/operators/cumprod_op.h b/paddle/fluid/operators/cumprod_op.h
index d8c3c1febdcf3ef336bc7c68ad9636bd9989c22e..15c3d514331b671817c97b0036a00b1279263dbb 100644
--- a/paddle/fluid/operators/cumprod_op.h
+++ b/paddle/fluid/operators/cumprod_op.h
@@ -18,8 +18,8 @@
 #include <type_traits>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -124,12 +124,12 @@ class CumprodGradOpCPUKernel : public framework::OpKernel<T> {
 
       platform::ForRange<platform::CPUDeviceContext> for_range_x(dev_ctx,
                                                                  numel);
-      math::ConjFunctor<T> functor_x(x_data, numel, x_data_conj);
+      pten::funcs::ConjFunctor<T> functor_x(x_data, numel, x_data_conj);
       for_range_x(functor_x);
 
       platform::ForRange<platform::CPUDeviceContext> for_range_out(dev_ctx,
                                                                    numel);
-      math::ConjFunctor<T> functor_out(out_data, numel, out_data_conj);
+      pten::funcs::ConjFunctor<T> functor_out(out_data, numel, out_data_conj);
       for_range_out(functor_out);
 
       x_data_deal = x_data_conj;
diff --git a/paddle/fluid/operators/determinant_op.h b/paddle/fluid/operators/determinant_op.h
index 90443e0928ba2535498122ea00df479b83acb56f..1da680fbd953a7f86cef4e9db13d1e63336f9a29 100644
--- a/paddle/fluid/operators/determinant_op.h
+++ b/paddle/fluid/operators/determinant_op.h
@@ -19,11 +19,11 @@
 #include <cmath>
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/math/matrix_inverse.h"
 #include "paddle/fluid/operators/svd_helper.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -395,7 +395,7 @@ class SlogDeterminantGradKernel : public framework::OpKernel<T> {
                                                      size_t(numel * sizeof(T)));
 
     platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::ConjFunctor<T> functor(inverse_A.data<T>(), numel, conj_data);
+    pten::funcs::ConjFunctor<T> functor(inverse_A.data<T>(), numel, conj_data);
     for_range(functor);
 
     VLOG(3) << "inverse(A).conj() dims: " << conj_inverse_A.dims();
diff --git a/paddle/fluid/operators/dot_op.h b/paddle/fluid/operators/dot_op.h
index c5d43ef01264b097ffb9c17bc716bd3dcedf8ce0..52fc26342a1b441dff032c13e0b423504a77265e 100644
--- a/paddle/fluid/operators/dot_op.h
+++ b/paddle/fluid/operators/dot_op.h
@@ -16,8 +16,8 @@
 
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 // only can include the headers in paddle/pten/api dirs
 #include "paddle/pten/api/lib/utils/tensor_utils.h"
diff --git a/paddle/fluid/operators/eig_op.h b/paddle/fluid/operators/eig_op.h
index b4b6e2ce2fc5664e8016e7815e037933d461e80d..f822802d305e9b03ac9e00604121033c7a1faa6d 100644
--- a/paddle/fluid/operators/eig_op.h
+++ b/paddle/fluid/operators/eig_op.h
@@ -17,12 +17,12 @@
 #include <math.h>
 #include <algorithm>
 #include <complex>
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/math/lapack_function.h"
 #include "paddle/fluid/operators/math/matrix_solve.h"
 #include "paddle/fluid/operators/svd_helper.h"
 #include "paddle/fluid/operators/transpose_op.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 #include "paddle/pten/kernels/funcs/math_function.h"
 #define EPSILON 1e-6
 
@@ -87,18 +87,19 @@ void LapackEig(Tensor* input, Tensor* values, Tensor* vectors, int info,
   int values_stride = values->dims()[values->dims().size() - 1];
 
   Tensor rwork;
-  math::Real<T>* rwork_data = nullptr;
+  pten::funcs::Real<T>* rwork_data = nullptr;
 
   rwork.Resize(framework::make_ddim({lda * 2}));
-  rwork_data = rwork.mutable_data<math::Real<T>>(context.GetPlace());
+  rwork_data = rwork.mutable_data<pten::funcs::Real<T>>(context.GetPlace());
 
   // call lapackEig once to compute the size of work;
   T computed_work_size;
-  math::lapackEig<T, math::Real<T>>(
+  math::lapackEig<T, pten::funcs::Real<T>>(
       jobvl, jobvr, order, input_data, lda, values_data, lvector_data, ldvl,
       rvector_data, ldvr, &computed_work_size, lwork, rwork_data, &info);
 
-  lwork = std::max<int>(1, static_cast<int>(math::Real<T>(computed_work_size)));
+  lwork = std::max<int>(
+      1, static_cast<int>(pten::funcs::Real<T>(computed_work_size)));
   Tensor work;
   work.Resize(framework::make_ddim({lwork}));
   T* work_data = work.mutable_data<T>(context.GetPlace());
@@ -108,7 +109,7 @@ void LapackEig(Tensor* input, Tensor* values, Tensor* vectors, int info,
     T* current_values = &values_data[i * values_stride];
     T* current_rvectors = &rvector_data[i * matrix_stride];
 
-    math::lapackEig<T, math::Real<T>>(
+    math::lapackEig<T, pten::funcs::Real<T>>(
         jobvl, jobvr, order, current_matrix, lda, current_values, lvector_data,
         ldvl, current_rvectors, ldvr, work_data, lwork, rwork_data, &info);
     PADDLE_ENFORCE_EQ(
@@ -207,26 +208,27 @@ class EigKernel : public framework::OpKernel<T> {
       origin_dim.push_back(last_item * 2);
       framework::DDim big_dim = framework::make_ddim(origin_dim);
 
-      real_values.mutable_data<math::Real<T>>(big_dim, context.GetPlace());
-      real_vectors.mutable_data<math::Real<T>>(x->dims(), context.GetPlace());
+      real_values.mutable_data<pten::funcs::Real<T>>(big_dim,
+                                                     context.GetPlace());
+      real_vectors.mutable_data<pten::funcs::Real<T>>(x->dims(),
+                                                      context.GetPlace());
 
-      ApplyEigKernel<DeviceContext, math::Real<T>>(*x, &real_values,
-                                                   &real_vectors, context);
-      auto dito =
-          math::DeviceIndependenceTensorOperations<DeviceContext, math::Real<T>,
-                                                   Tout>(context);
+      ApplyEigKernel<DeviceContext, pten::funcs::Real<T>>(
+          *x, &real_values, &real_vectors, context);
+      auto dito = math::DeviceIndependenceTensorOperations<
+          DeviceContext, pten::funcs::Real<T>, Tout>(context);
 
       // 1. extract real part & imag part from real_values
       Tensor real_part = dito.Slice(real_values, {-1}, {0}, {order});
       Tensor imag_part = dito.Slice(real_values, {-1}, {order}, {order * 2});
 
       // 2. construct complex values
-      auto* real_part_data = real_part.data<math::Real<T>>();
-      auto* imag_part_data = imag_part.data<math::Real<T>>();
+      auto* real_part_data = real_part.data<pten::funcs::Real<T>>();
+      auto* imag_part_data = imag_part.data<pten::funcs::Real<T>>();
       int out_values_numel = out_values->numel();
       platform::ForRange<DeviceContext> for_range(
           context.template device_context<DeviceContext>(), out_values_numel);
-      math::RealImagToComplexFunctor<Tout> functor(
+      pten::funcs::RealImagToComplexFunctor<Tout> functor(
           real_part_data, imag_part_data,
           out_values->mutable_data<Tout>(context.GetPlace()), out_values_numel);
       for_range(functor);
@@ -235,7 +237,7 @@ class EigKernel : public framework::OpKernel<T> {
       Tensor real_vector_trans = dito.Transpose(real_vectors);
       Tensor out_vectors_trans;
       out_vectors_trans.mutable_data<Tout>(x->dims(), context.GetPlace());
-      ConstructComplexVectors<math::Real<T>, Tout>(
+      ConstructComplexVectors<pten::funcs::Real<T>, Tout>(
           &out_vectors_trans, *out_values, real_vector_trans, context,
           batch_count, order);
       TransposeTwoAxis<DeviceContext, Tout>(out_vectors_trans, out_vectors,
@@ -271,14 +273,14 @@ void ComputeBackwardForComplexInput(
   // turn diag_unsqueezed into complex
   auto numel = diag_unsqueezed.numel();
   Tensor diag_unsqueezed_complex;
-  auto* data_diag_un = diag_unsqueezed.data<math::Real<Tout>>();
+  auto* data_diag_un = diag_unsqueezed.data<pten::funcs::Real<Tout>>();
   auto* data_diag_un_com = diag_unsqueezed_complex.mutable_data<Tout>(
       diag_unsqueezed.dims(), context.GetPlace(),
       static_cast<size_t>(numel * sizeof(Tout)));
   auto& dev_ctx = context.template device_context<DeviceContext>();
   platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-  math::RealToComplexFunctor<Tout> functor(data_diag_un, data_diag_un_com,
-                                           numel);
+  pten::funcs::RealToComplexFunctor<Tout> functor(data_diag_un,
+                                                  data_diag_un_com, numel);
   for_range(functor);
   // real tensor multiply complex tensor in broadcast manner
   Tensor res1 = dito.RealMulComplex(V, diag_unsqueezed_complex);
diff --git a/paddle/fluid/operators/eigh_op.h b/paddle/fluid/operators/eigh_op.h
index ad9b0f598311b1f44ae3f7ec34bda8489e422fc7..77afaf681da939cdc089325d5e210e5960142f55 100644
--- a/paddle/fluid/operators/eigh_op.h
+++ b/paddle/fluid/operators/eigh_op.h
@@ -40,7 +40,7 @@ template <typename DeviceContext, typename T>
 class EighGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    using ValueType = math::Real<T>;
+    using ValueType = pten::funcs::Real<T>;
     auto& x_grad = *ctx.Output<framework::Tensor>(framework::GradVarName("X"));
     x_grad.mutable_data<T>(ctx.GetPlace());
     auto& output_w = *ctx.Input<Tensor>("Eigenvalues");
diff --git a/paddle/fluid/operators/eigvals_op.h b/paddle/fluid/operators/eigvals_op.h
index d825833b0242240cb0bdbeaf1e85057fe23dc618..a069ea164c94c5c9ec50ece0d28042347ff4777f 100644
--- a/paddle/fluid/operators/eigvals_op.h
+++ b/paddle/fluid/operators/eigvals_op.h
@@ -20,9 +20,9 @@
 #include "paddle/fluid/framework/ddim.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/memory/allocation/allocator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/math/lapack_function.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -48,7 +48,7 @@ struct PaddleComplex<
 template <typename T>
 using PaddleCType = typename PaddleComplex<T>::type;
 template <typename T>
-using Real = typename math::Real<T>;
+using Real = typename pten::funcs::Real<T>;
 
 static void SpiltBatchSquareMatrix(const Tensor& input,
                                    std::vector<Tensor>* output) {
@@ -118,7 +118,7 @@ LapackEigvals(const framework::ExecutionContext& ctx, const Tensor& input,
 
   platform::ForRange<DeviceContext> for_range(
       ctx.template device_context<DeviceContext>(), n_dim);
-  math::RealImagToComplexFunctor<PaddleCType<T>> functor(
+  pten::funcs::RealImagToComplexFunctor<PaddleCType<T>> functor(
       w_data, w_data + n_dim, output->template data<PaddleCType<T>>(), n_dim);
   for_range(functor);
 }
@@ -143,7 +143,7 @@ LapackEigvals(const framework::ExecutionContext& ctx, const Tensor& input,
           required_work_mem, work_mem));
 
   int64_t rwork_mem = rwork->memory_size();
-  int64_t required_rwork_mem = (n_dim << 1) * sizeof(Real<T>);
+  int64_t required_rwork_mem = (n_dim << 1) * sizeof(pten::funcs::Real<T>);
   PADDLE_ENFORCE_GE(
       rwork_mem, required_rwork_mem,
       platform::errors::InvalidArgument(
@@ -153,11 +153,11 @@ LapackEigvals(const framework::ExecutionContext& ctx, const Tensor& input,
           required_rwork_mem, rwork_mem));
 
   int info = 0;
-  math::lapackEig<T, Real<T>>(
+  math::lapackEig<T, pten::funcs::Real<T>>(
       'N', 'N', static_cast<int>(n_dim), a.template data<T>(),
       static_cast<int>(n_dim), output->template data<T>(), NULL, 1, NULL, 1,
       work->template data<T>(), static_cast<int>(work_mem / sizeof(T)),
-      rwork->template data<Real<T>>(), &info);
+      rwork->template data<pten::funcs::Real<T>>(), &info);
 
   std::string name = "framework::platform::dynload::cgeev_";
   if (framework::TransToProtoVarType(input.dtype()) ==
@@ -187,10 +187,10 @@ class EigvalsKernel : public framework::OpKernel<T> {
     // query workspace size
     T qwork;
     int info;
-    math::lapackEig<T, Real<T>>('N', 'N', static_cast<int>(n_dim),
-                                input_matrices[0].template data<T>(),
-                                static_cast<int>(n_dim), NULL, NULL, 1, NULL, 1,
-                                &qwork, -1, static_cast<Real<T>*>(NULL), &info);
+    math::lapackEig<T, pten::funcs::Real<T>>(
+        'N', 'N', static_cast<int>(n_dim), input_matrices[0].template data<T>(),
+        static_cast<int>(n_dim), NULL, NULL, 1, NULL, 1, &qwork, -1,
+        static_cast<pten::funcs::Real<T>*>(NULL), &info);
     int64_t lwork = static_cast<int64_t>(qwork);
 
     Tensor work, rwork;
@@ -207,8 +207,8 @@ class EigvalsKernel : public framework::OpKernel<T> {
     }
     if (framework::IsComplexType(
             framework::TransToProtoVarType(input->dtype()))) {
-      rwork.mutable_data<Real<T>>(framework::make_ddim({n_dim << 1}),
-                                  ctx.GetPlace());
+      rwork.mutable_data<pten::funcs::Real<T>>(
+          framework::make_ddim({n_dim << 1}), ctx.GetPlace());
     }
 
     for (int64_t i = 0; i < n_batch; ++i) {
diff --git a/paddle/fluid/operators/imag_op.h b/paddle/fluid/operators/imag_op.h
index 562a8dffa90623ed44c51ff1048c25550f5a7ce7..02682cfc954be57dd7900326dd98dae507fadeaa 100644
--- a/paddle/fluid/operators/imag_op.h
+++ b/paddle/fluid/operators/imag_op.h
@@ -16,8 +16,8 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -31,12 +31,13 @@ class ImagKernel : public framework::OpKernel<T> {
 
     auto numel = x->numel();
     auto* x_data = x->data<T>();
-    auto* out_data = out->mutable_data<math::Real<T>>(
-        ctx.GetPlace(), static_cast<size_t>(numel * sizeof(math::Real<T>)));
+    auto* out_data = out->mutable_data<pten::funcs::Real<T>>(
+        ctx.GetPlace(),
+        static_cast<size_t>(numel * sizeof(pten::funcs::Real<T>)));
 
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::ImagFunctor<T> functor(x_data, out_data, numel);
+    pten::funcs::ImagFunctor<T> functor(x_data, out_data, numel);
     for_range(functor);
   }
 };
@@ -51,13 +52,13 @@ class ImagGradKernel : public framework::OpKernel<T> {
         ctx.Output<framework::Tensor>(framework::GradVarName("X"));
 
     auto numel = d_out->numel();
-    auto* dout_data = d_out->data<math::Real<T>>();
+    auto* dout_data = d_out->data<pten::funcs::Real<T>>();
     auto* dx_data = d_x->mutable_data<T>(
         ctx.GetPlace(), static_cast<size_t>(numel * sizeof(T)));
 
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::ImagToComplexFunctor<T> functor(dout_data, dx_data, numel);
+    pten::funcs::ImagToComplexFunctor<T> functor(dout_data, dx_data, numel);
     for_range(functor);
   }
 };
diff --git a/paddle/fluid/operators/lstsq_op.h b/paddle/fluid/operators/lstsq_op.h
index 4819bd725183248d8711b94c546e11e6d30026ab..f39d65d681f2f8f0e18c2fae13154d76b8b2f76c 100644
--- a/paddle/fluid/operators/lstsq_op.h
+++ b/paddle/fluid/operators/lstsq_op.h
@@ -18,7 +18,6 @@
 #include <algorithm>
 #include <complex>
 #include "paddle/fluid/operators/eig_op.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/math/eigen_values_vectors.h"
 #include "paddle/fluid/operators/math/lapack_function.h"
 #include "paddle/fluid/operators/math/matrix_solve.h"
@@ -26,6 +25,7 @@
 #include "paddle/fluid/operators/transpose_op.h"
 #include "paddle/fluid/operators/triangular_solve_op.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 #include "paddle/pten/kernels/funcs/math_function.h"
 
 #define EPSILON 1e-6
@@ -46,7 +46,7 @@ template <typename DeviceContext, typename T>
 class LstsqCPUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    using ValueType = math::Real<T>;
+    using ValueType = pten::funcs::Real<T>;
 
     const Tensor& x = *context.Input<Tensor>("X");
     auto y = context.Input<Tensor>("Y");
@@ -169,7 +169,7 @@ class LstsqCPUKernel : public framework::OpKernel<T> {
                         &rwkopt, &info);
     }
 
-    lwork = std::max<int>(1, static_cast<int>(math::Real<T>(wkopt)));
+    lwork = std::max<int>(1, static_cast<int>(pten::funcs::Real<T>(wkopt)));
     Tensor work;
     work.Resize(framework::make_ddim({lwork}));
     T* work_data = work.mutable_data<T>(context.GetPlace());
diff --git a/paddle/fluid/operators/lu_op.h b/paddle/fluid/operators/lu_op.h
index 11174540cb0cd1f6e971c9fb85338b2eeb8bbfa0..0d05d766e67fb16c75d5fb0f9c798c7048a1c7f9 100644
--- a/paddle/fluid/operators/lu_op.h
+++ b/paddle/fluid/operators/lu_op.h
@@ -211,8 +211,9 @@ void Tensor_Conj(const DeviceContext& dev_ctx, const framework::Tensor& tensor,
                  framework::Tensor* out) {
   out->Resize(tensor.dims());
   platform::ForRange<DeviceContext> out_for_range(dev_ctx, tensor.numel());
-  math::ConjFunctor<T> out_functor(tensor.data<T>(), tensor.numel(),
-                                   out->mutable_data<T>(dev_ctx.GetPlace()));
+  pten::funcs::ConjFunctor<T> out_functor(
+      tensor.data<T>(), tensor.numel(),
+      out->mutable_data<T>(dev_ctx.GetPlace()));
   out_for_range(out_functor);
 }
 
diff --git a/paddle/fluid/operators/math/eigen_values_vectors.h b/paddle/fluid/operators/math/eigen_values_vectors.h
index 9ce615c949ffcb0e7ef300dfdc4f45b87604ad0c..b946d4d072ba2e276df632e5fea6960fbbe17975 100644
--- a/paddle/fluid/operators/math/eigen_values_vectors.h
+++ b/paddle/fluid/operators/math/eigen_values_vectors.h
@@ -63,7 +63,7 @@ struct MatrixEighFunctor<platform::CPUDeviceContext, T> {
   void operator()(const framework::ExecutionContext &ctx, const Tensor &input,
                   Tensor *eigen_values, Tensor *eigen_vectors, bool is_lower,
                   bool has_vectors) {
-    using ValueType = math::Real<T>;
+    using ValueType = pten::funcs::Real<T>;
     auto *out_value = eigen_values->mutable_data<ValueType>(ctx.GetPlace());
 
     auto dito =
@@ -123,9 +123,9 @@ struct MatrixEighFunctor<platform::CPUDeviceContext, T> {
     for (auto i = 0; i < batch_size; i++) {
       auto *value_data = out_value + i * values_stride;
       auto *input_data = input_vector + i * vector_stride;
-      math::lapackEigh<T, Real<T>>(jobz, uplo, n, input_data, lda, value_data,
-                                   work_data, lwork, rwork_data, lrwork,
-                                   iwork_data, liwork, &info);
+      math::lapackEigh<T, pten::funcs::Real<T>>(
+          jobz, uplo, n, input_data, lda, value_data, work_data, lwork,
+          rwork_data, lrwork, iwork_data, liwork, &info);
       CheckEighResult(i, info);
     }
     if (has_vectors) {
@@ -151,7 +151,7 @@ struct MatrixEighFunctor<platform::CUDADeviceContext, T> {
   void operator()(const framework::ExecutionContext &ctx, const Tensor &input,
                   Tensor *eigen_values, Tensor *eigen_vectors, bool is_lower,
                   bool has_vectors) {
-    using ValueType = math::Real<T>;
+    using ValueType = pten::funcs::Real<T>;
     auto *out_value = eigen_values->mutable_data<ValueType>(ctx.GetPlace());
 
     auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
@@ -233,7 +233,7 @@ struct MatrixEighFunctor<platform::CUDADeviceContext, T> {
     }
   }
 
-  using ValueType = math::Real<T>;
+  using ValueType = pten::funcs::Real<T>;
   inline void EvdBuffer(cusolverDnHandle_t handle, cusolverEigMode_t jobz,
                         cublasFillMode_t uplo, int n, const T *A, int lda,
                         const ValueType *W, int *lwork) const;
diff --git a/paddle/fluid/operators/math/inclusive_scan.h b/paddle/fluid/operators/math/inclusive_scan.h
index 5fdc2889a88858769c4bdf445367dc60265d6cbf..1c750fcb832c1ca0fae51c6c5f818fe82923897e 100644
--- a/paddle/fluid/operators/math/inclusive_scan.h
+++ b/paddle/fluid/operators/math/inclusive_scan.h
@@ -26,9 +26,9 @@ namespace cub = hipcub;
 #include <thrust/iterator/reverse_iterator.h>
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/memory/malloc.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -115,7 +115,7 @@ static __global__ void InclusiveScanInnerDimCUDAKernel(const T *x, T *y,
                                                        size_t num_rows,
                                                        size_t row_size, T init,
                                                        BinaryOp op) {
-  using RealT = math::Real<T>;
+  using RealT = pten::funcs::Real<T>;
   constexpr auto kSharedBufferSize =
       framework::IsComplex<T>::value ? 4 * kThreadNumX : 2 * kThreadNumX;
   __shared__ RealT sbuf[kThreadNumY][kSharedBufferSize];
diff --git a/paddle/fluid/operators/matmul_v2_op.h b/paddle/fluid/operators/matmul_v2_op.h
index 0e1c6b82e41922cb1a7fd8404ffae1135e7872a0..6fac2d1038334528b87c056ae0d14a366432d5bc 100644
--- a/paddle/fluid/operators/matmul_v2_op.h
+++ b/paddle/fluid/operators/matmul_v2_op.h
@@ -22,8 +22,8 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/dot_op.h"
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/reduce_ops/reduce_sum_op.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 // only can include the headers in paddle/pten/api dirs
 #include "paddle/pten/api/lib/utils/tensor_utils.h"
diff --git a/paddle/fluid/operators/matrix_rank_op.cu b/paddle/fluid/operators/matrix_rank_op.cu
index d974d7c1b78f15bb5e0f050b4e415af453e4349f..2df794fb794430910e71c0980154f682c3f4920d 100644
--- a/paddle/fluid/operators/matrix_rank_op.cu
+++ b/paddle/fluid/operators/matrix_rank_op.cu
@@ -18,11 +18,11 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/memory/memory.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/matrix_rank_op.h"
 #include "paddle/fluid/operators/svd_helper.h"
 #include "paddle/fluid/platform/dynload/cusolver.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 #include "paddle/pten/kernels/funcs/math_function.h"
 
 namespace paddle {
@@ -93,8 +93,8 @@ class MatrixRankGPUKernel : public framework::OpKernel<T> {
                    info_ptr);
       platform::ForRange<platform::CUDADeviceContext> for_range(
           dev_ctx, eigenvalue_tensor.numel());
-      math::AbsFunctor<T> functor(eigenvalue_data, eigenvalue_data,
-                                  eigenvalue_tensor.numel());
+      pten::funcs::AbsFunctor<T> functor(eigenvalue_data, eigenvalue_data,
+                                         eigenvalue_tensor.numel());
       for_range(functor);
     } else {
       Tensor U, VH;
diff --git a/paddle/fluid/operators/qr_op.cu b/paddle/fluid/operators/qr_op.cu
index c8b6404830cdac0427f99b2e1d2c642fe8aa0f38..dfeec15d9b887aa55b81004b728a7c31fc8b4be7 100644
--- a/paddle/fluid/operators/qr_op.cu
+++ b/paddle/fluid/operators/qr_op.cu
@@ -56,12 +56,13 @@ class QrGPUKernel : public framework::OpKernel<T> {
     int tau_stride = min_mn;
 
     if (compute_q) {
-      q.mutable_data<math::Real<T>>(
+      q.mutable_data<pten::funcs::Real<T>>(
           context.GetPlace(),
-          size_t(batch_size * m * k * sizeof(math::Real<T>)));
+          size_t(batch_size * m * k * sizeof(pten::funcs::Real<T>)));
     }
-    r.mutable_data<math::Real<T>>(
-        context.GetPlace(), size_t(batch_size * k * n * sizeof(math::Real<T>)));
+    r.mutable_data<pten::funcs::Real<T>>(
+        context.GetPlace(),
+        size_t(batch_size * k * n * sizeof(pten::funcs::Real<T>)));
 
     auto dito =
         math::DeviceIndependenceTensorOperations<platform::CUDADeviceContext,
@@ -70,8 +71,9 @@ class QrGPUKernel : public framework::OpKernel<T> {
     // Note: allocate temporary tensors because of lacking in-place operatios.
     // Prepare qr
     Tensor qr;
-    qr.mutable_data<math::Real<T>>(
-        context.GetPlace(), size_t(batch_size * m * n * sizeof(math::Real<T>)));
+    qr.mutable_data<pten::funcs::Real<T>>(
+        context.GetPlace(),
+        size_t(batch_size * m * n * sizeof(pten::funcs::Real<T>)));
     // BatchedGeqrf performs computation in-place and 'qr' must be a copy of
     // input
     paddle::framework::TensorCopy(x, context.GetPlace(), &qr);
@@ -124,7 +126,8 @@ class QrGPUKernel : public framework::OpKernel<T> {
           for (int i = 0; i < batch_size; ++i) {
             memory::Copy(dev_ctx.GetPlace(), (new_qr_data + i * new_qr_stride),
                          dev_ctx.GetPlace(), (qr_data + i * qr_stride),
-                         qr_stride * sizeof(math::Real<T>), dev_ctx.stream());
+                         qr_stride * sizeof(pten::funcs::Real<T>),
+                         dev_ctx.stream());
           }
           BatchedOrgqr<platform::CUDADeviceContext, T>(
               dev_ctx, batch_size, m, m, min_mn, new_qr_data, m, tau_data,
diff --git a/paddle/fluid/operators/qr_op.h b/paddle/fluid/operators/qr_op.h
index c55619a4f76e7f316c6c7bcb689e2a101e5908eb..b8308b29106be39bacbf05028809e7206ea63cec 100644
--- a/paddle/fluid/operators/qr_op.h
+++ b/paddle/fluid/operators/qr_op.h
@@ -18,9 +18,9 @@
 #include <cstdarg>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/svd_helper.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -74,17 +74,20 @@ class QrCPUKernel : public framework::OpKernel<T> {
     int q_stride = m * k;
     int r_stride = k * n;
 
-    auto* x_data = x.data<math::Real<T>>();
+    auto* x_data = x.data<pten::funcs::Real<T>>();
     T* q_data = nullptr;
     if (compute_q) {
-      q_data = q.mutable_data<math::Real<T>>(
+      q_data = q.mutable_data<pten::funcs::Real<T>>(
           context.GetPlace(),
-          size_t(batch_size * m * k * sizeof(math::Real<T>)));
-      memset(q_data, 0, size_t(batch_size * m * k * sizeof(math::Real<T>)));
+          size_t(batch_size * m * k * sizeof(pten::funcs::Real<T>)));
+      memset(q_data, 0,
+             size_t(batch_size * m * k * sizeof(pten::funcs::Real<T>)));
     }
-    auto* r_data = r.mutable_data<math::Real<T>>(
-        context.GetPlace(), size_t(batch_size * k * n * sizeof(math::Real<T>)));
-    memset(r_data, 0, size_t(batch_size * k * n * sizeof(math::Real<T>)));
+    auto* r_data = r.mutable_data<pten::funcs::Real<T>>(
+        context.GetPlace(),
+        size_t(batch_size * k * n * sizeof(pten::funcs::Real<T>)));
+    memset(r_data, 0,
+           size_t(batch_size * k * n * sizeof(pten::funcs::Real<T>)));
 
     // Implement QR by calling Eigen
     for (int i = 0; i < batch_size; ++i) {
@@ -140,7 +143,7 @@ class QrGradKernel : public framework::OpKernel<T> {
     // Use a different name dA instead of dX
     framework::Tensor& dA =
         *ctx.Output<framework::Tensor>(framework::GradVarName("X"));
-    dA.mutable_data<math::Real<T>>(ctx.GetPlace());
+    dA.mutable_data<pten::funcs::Real<T>>(ctx.GetPlace());
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     pten::funcs::SetConstant<DeviceContext, T>()(dev_ctx, &dA, T(0));
 
@@ -222,7 +225,7 @@ class QrGradKernel : public framework::OpKernel<T> {
     } else {
       // If m < n for input matrices A, we partition A = [X|Y] and R = [U|V]
       // Calculate dX and dY individually and concatenate them to get dA
-      dA.mutable_data<math::Real<T>>(ctx.GetPlace());
+      dA.mutable_data<pten::funcs::Real<T>>(ctx.GetPlace());
 
       auto Y = dito.Slice(A, {-1}, {m}, {n});
       auto U = dito.Slice(R, {-1}, {0}, {m});
diff --git a/paddle/fluid/operators/real_op.h b/paddle/fluid/operators/real_op.h
index 6cc9065269c62716b54c329d46711ff96f83f015..41549393f578ff6109b629a6036cbbef108b398c 100644
--- a/paddle/fluid/operators/real_op.h
+++ b/paddle/fluid/operators/real_op.h
@@ -16,8 +16,8 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -31,12 +31,13 @@ class RealKernel : public framework::OpKernel<T> {
 
     auto numel = x->numel();
     auto* x_data = x->data<T>();
-    auto* out_data = out->mutable_data<math::Real<T>>(
-        ctx.GetPlace(), static_cast<size_t>(numel * sizeof(math::Real<T>)));
+    auto* out_data = out->mutable_data<pten::funcs::Real<T>>(
+        ctx.GetPlace(),
+        static_cast<size_t>(numel * sizeof(pten::funcs::Real<T>)));
 
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::RealFunctor<T> functor(x_data, out_data, numel);
+    pten::funcs::RealFunctor<T> functor(x_data, out_data, numel);
     for_range(functor);
   }
 };
@@ -51,13 +52,13 @@ class RealGradKernel : public framework::OpKernel<T> {
         ctx.Output<framework::Tensor>(framework::GradVarName("X"));
 
     auto numel = d_out->numel();
-    auto* dout_data = d_out->data<math::Real<T>>();
+    auto* dout_data = d_out->data<pten::funcs::Real<T>>();
     auto* dx_data = d_x->mutable_data<T>(
         ctx.GetPlace(), static_cast<size_t>(numel * sizeof(T)));
 
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::RealToComplexFunctor<T> functor(dout_data, dx_data, numel);
+    pten::funcs::RealToComplexFunctor<T> functor(dout_data, dx_data, numel);
     for_range(functor);
   }
 };
diff --git a/paddle/fluid/operators/renorm_op.h b/paddle/fluid/operators/renorm_op.h
index 461f383ad25639fe2db9b64eb490ad1e7a769a4a..753ed9e27ac0918b7f36cd347b190b80714ccde5 100644
--- a/paddle/fluid/operators/renorm_op.h
+++ b/paddle/fluid/operators/renorm_op.h
@@ -17,8 +17,8 @@
 #include "math.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 namespace paddle {
 namespace operators {
 using Tensor = framework::Tensor;
diff --git a/paddle/fluid/operators/spectral_op.cu b/paddle/fluid/operators/spectral_op.cu
index c932834db39b30c50746aeee80fcd32b5090f58f..77703637db5cd7d34c865083bd765e1122b7fefb 100644
--- a/paddle/fluid/operators/spectral_op.cu
+++ b/paddle/fluid/operators/spectral_op.cu
@@ -20,11 +20,11 @@
 #include <vector>
 
 #include "paddle/fluid/operators/conj_op.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/spectral_helper.h"
 #include "paddle/fluid/operators/spectral_op.h"
 #include "paddle/fluid/operators/transpose_op.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -115,8 +115,8 @@ void exec_cufft_plan(const DeviceContext& ctx, const FFTConfig& config,
     framework::Tensor input_conj(input->type());
     input_conj.mutable_data<Ti>(input->dims(), ctx.GetPlace());
     platform::ForRange<DeviceContext> for_range(ctx, input->numel());
-    math::ConjFunctor<Ti> functor(input->data<Ti>(), input->numel(),
-                                  input_conj.data<Ti>());
+    pten::funcs::ConjFunctor<Ti> functor(input->data<Ti>(), input->numel(),
+                                         input_conj.data<Ti>());
     for_range(functor);
     exec_cufft_plan_raw(config, input_conj.data(), output->data(), forward);
   } else if (fft_type == FFTTransformType::R2C && !forward) {
@@ -126,8 +126,8 @@ void exec_cufft_plan(const DeviceContext& ctx, const FFTConfig& config,
     exec_cufft_plan_raw(config, input->data(), out_conj.data(), forward);
 
     platform::ForRange<DeviceContext> for_range(ctx, output->numel());
-    math::ConjFunctor<To> functor(out_conj.data<To>(), output->numel(),
-                                  output->data<To>());
+    pten::funcs::ConjFunctor<To> functor(out_conj.data<To>(), output->numel(),
+                                         output->data<To>());
     for_range(functor);
   } else {
     exec_cufft_plan_raw(config, input->data(), output->data(), forward);
@@ -227,8 +227,8 @@ void exec_hipfft_plan(const DeviceContext& ctx, const FFTConfig& config,
     framework::Tensor input_conj(input->type());
     input_conj.mutable_data<Ti>(input->dims(), ctx.GetPlace());
     platform::ForRange<DeviceContext> for_range(ctx, input->numel());
-    math::ConjFunctor<Ti> functor(input->data<Ti>(), input->numel(),
-                                  input_conj.data<Ti>());
+    pten::funcs::ConjFunctor<Ti> functor(input->data<Ti>(), input->numel(),
+                                         input_conj.data<Ti>());
     for_range(functor);
     exec_hipfft_plan_raw(config, input_conj.data(), output->data(), forward);
   } else if (fft_type == FFTTransformType::R2C && !forward) {
@@ -238,8 +238,8 @@ void exec_hipfft_plan(const DeviceContext& ctx, const FFTConfig& config,
     exec_hipfft_plan_raw(config, input->data(), out_conj.data(), forward);
 
     platform::ForRange<DeviceContext> for_range(ctx, output->numel());
-    math::ConjFunctor<To> functor(out_conj.data<To>(), output->numel(),
-                                  output->data<To>());
+    pten::funcs::ConjFunctor<To> functor(out_conj.data<To>(), output->numel(),
+                                         output->data<To>());
     for_range(functor);
   } else {
     exec_hipfft_plan_raw(config, input->data(), output->data(), forward);
diff --git a/paddle/fluid/operators/svd_helper.h b/paddle/fluid/operators/svd_helper.h
index 3a57a7b3e54cc3313654d20256b888efdb4baf5a..4384e7152fa4e56554a3effd7e82b56b03a1c585 100644
--- a/paddle/fluid/operators/svd_helper.h
+++ b/paddle/fluid/operators/svd_helper.h
@@ -25,9 +25,9 @@
 #include "paddle/fluid/operators/eigen/eigen_function.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 #include "paddle/pten/kernels/funcs/math_function.h"
 
 namespace paddle {
@@ -105,7 +105,8 @@ struct RealMulComplexFunctor {
                                         "The image part of y must to be 0"
                                         "but got [%d]",
                                         y.imag));
-    return platform::complex<Real<T>>(x.real * y.real, x.imag * y.real);
+    return platform::complex<pten::funcs::Real<T>>(x.real * y.real,
+                                                   x.imag * y.real);
   }
 };
 
@@ -390,11 +391,11 @@ struct DeviceIndependenceTensorOperations {
   // batch_diag for CPU only
   Tensor BatchDiag(const Tensor& x, int batch) {
     Tensor out;
-    auto* x_data = x.data<math::Real<T>>();
+    auto* x_data = x.data<pten::funcs::Real<T>>();
     auto numel = x.numel();
-    auto* out_data = out.mutable_data<math::Real<T>>(
+    auto* out_data = out.mutable_data<pten::funcs::Real<T>>(
         x.dims(), context.GetPlace(),
-        static_cast<size_t>(numel * sizeof(math::Real<T>)));
+        static_cast<size_t>(numel * sizeof(pten::funcs::Real<T>)));
 
     auto x_dims = x.dims();
     int num_dims = x_dims.size();
@@ -654,7 +655,7 @@ struct DeviceIndependenceTensorOperations {
     auto* out_data = out.mutable_data<T>(x.dims(), context.GetPlace());
     auto* x_data = x.data<T>();
     auto for_range = GetForRange(x.numel());
-    math::ConjFunctor<T> functor(x_data, x.numel(), out_data);
+    pten::funcs::ConjFunctor<T> functor(x_data, x.numel(), out_data);
     for_range(functor);
     return out;
   }
@@ -662,12 +663,12 @@ struct DeviceIndependenceTensorOperations {
   Tensor Real(const Tensor& x) {
     Tensor out;
     auto numel = x.numel();
-    auto* out_data = out.mutable_data<math::Real<T>>(
+    auto* out_data = out.mutable_data<pten::funcs::Real<T>>(
         x.dims(), context.GetPlace(),
-        static_cast<size_t>(numel * sizeof(math::Real<T>)));
+        static_cast<size_t>(numel * sizeof(pten::funcs::Real<T>)));
     auto* x_data = x.data<T>();
     auto for_range = GetForRange(numel);
-    math::RealFunctor<T> functor(x_data, out_data, numel);
+    pten::funcs::RealFunctor<T> functor(x_data, out_data, numel);
     for_range(functor);
     return out;
   }
diff --git a/paddle/fluid/operators/svd_op.h b/paddle/fluid/operators/svd_op.h
index f387dca7b7f9b2c4e741d8f495a58b05a46c6c6f..4042fcccf33090e11f14ec0effc1e5b9ddd95258 100644
--- a/paddle/fluid/operators/svd_op.h
+++ b/paddle/fluid/operators/svd_op.h
@@ -17,9 +17,9 @@
 #include <cstdarg>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/svd_helper.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -46,14 +46,14 @@ class SvdCPUKernel : public framework::OpKernel<T> {
     int col_u = full ? rows : k;
     int col_v = full ? cols : k;
     int batches = numel / (rows * cols);
-    auto* U_out = U->mutable_data<math::Real<T>>(
+    auto* U_out = U->mutable_data<pten::funcs::Real<T>>(
         context.GetPlace(),
-        size_t(batches * rows * col_u * sizeof(math::Real<T>)));
-    auto* VH_out = VH->mutable_data<math::Real<T>>(
+        size_t(batches * rows * col_u * sizeof(pten::funcs::Real<T>)));
+    auto* VH_out = VH->mutable_data<pten::funcs::Real<T>>(
         context.GetPlace(),
-        size_t(batches * col_v * cols * sizeof(math::Real<T>)));
-    auto* S_out = S->mutable_data<math::Real<T>>(
-        context.GetPlace(), size_t(batches * k * sizeof(math::Real<T>)));
+        size_t(batches * col_v * cols * sizeof(pten::funcs::Real<T>)));
+    auto* S_out = S->mutable_data<pten::funcs::Real<T>>(
+        context.GetPlace(), size_t(batches * k * sizeof(pten::funcs::Real<T>)));
     /*SVD Use the Eigen Library*/
     math::BatchSvd<T>(x_data, U_out, VH_out, S_out, rows, cols, batches, full);
   }
diff --git a/paddle/fluid/operators/triangular_solve_op.h b/paddle/fluid/operators/triangular_solve_op.h
index f64b016366e39b2260f4f8aebbb2e371ee2a8a7a..e892d258f3b126c0f6532f215e411837a415ee27 100644
--- a/paddle/fluid/operators/triangular_solve_op.h
+++ b/paddle/fluid/operators/triangular_solve_op.h
@@ -19,10 +19,10 @@ limitations under the License. */
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/reduce_ops/reduce_op.h"
 #include "paddle/fluid/operators/solve_op.h"
 #include "paddle/fluid/operators/tril_triu_op.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace paddle {
 namespace operators {
@@ -152,7 +152,7 @@ class TriangularSolveGradKernel : public framework::OpKernel<T> {
       // calculate x's conjugate for complex
       Tensor x_conj(x->type());
       platform::ForRange<DeviceContext> x_for_range(dev_ctx, x->numel());
-      math::ConjFunctor<T> x_functor(
+      pten::funcs::ConjFunctor<T> x_functor(
           x->data<T>(), x->numel(),
           x_conj.mutable_data<T>(x->dims(), dev_ctx.GetPlace()));
       x_for_range(x_functor);
@@ -179,7 +179,7 @@ class TriangularSolveGradKernel : public framework::OpKernel<T> {
       // calculate out's conjugate for complex
       Tensor out_conj(out->type());
       platform::ForRange<DeviceContext> out_for_range(dev_ctx, out->numel());
-      math::ConjFunctor<T> out_functor(
+      pten::funcs::ConjFunctor<T> out_functor(
           out->data<T>(), out->numel(),
           out_conj.mutable_data<T>(out->dims(), dev_ctx.GetPlace()));
       out_for_range(out_functor);
diff --git a/paddle/pten/kernels/cpu/abs_grad_kernel.cc b/paddle/pten/kernels/cpu/abs_grad_kernel.cc
index a3f3aabd16c097941ecc615ee51e62b90ef14b34..9d6675aa7b3155d1f83f4bd098a867e5f3359938 100644
--- a/paddle/pten/kernels/cpu/abs_grad_kernel.cc
+++ b/paddle/pten/kernels/cpu/abs_grad_kernel.cc
@@ -12,9 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/pten/common/complex.h"
 #include "paddle/pten/core/kernel_registry.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 #include "paddle/pten/kernels/impl/abs_grad_kernel_impl.h"
 
 using pten::dtype::complex;
diff --git a/paddle/pten/kernels/cpu/abs_kernel.cc b/paddle/pten/kernels/cpu/abs_kernel.cc
index 49094f5c64ed3762b8e27d941df2ecc8359f6bf2..ee766a18d42aba35f4d9e45d6d381beda99798d6 100644
--- a/paddle/pten/kernels/cpu/abs_kernel.cc
+++ b/paddle/pten/kernels/cpu/abs_kernel.cc
@@ -13,11 +13,11 @@
 // limitations under the License.
 
 #include "paddle/pten/kernels/abs_kernel.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
 #include "paddle/pten/backends/cpu/cpu_context.h"
 #include "paddle/pten/common/complex.h"
 #include "paddle/pten/core/kernel_registry.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace pten {
 
@@ -25,12 +25,12 @@ template <typename T, typename Context>
 void AbsKernel(const Context& ctx, const DenseTensor& x, DenseTensor* out) {
   auto numel = x.numel();
   auto* x_data = x.data<T>();
-  ctx.template Alloc<paddle::operators::math::Real<T>>(
-      out, size_t(x.numel() * sizeof(paddle::operators::math::Real<T>)));
-  auto* out_data = out->data<paddle::operators::math::Real<T>>();
+  ctx.template Alloc<pten::funcs::Real<T>>(
+      out, size_t(x.numel() * sizeof(pten::funcs::Real<T>)));
+  auto* out_data = out->data<pten::funcs::Real<T>>();
 
   paddle::platform::ForRange<Context> for_range(ctx, numel);
-  paddle::operators::math::AbsFunctor<T> functor(x_data, out_data, numel);
+  pten::funcs::AbsFunctor<T> functor(x_data, out_data, numel);
   for_range(functor);
 }
 
diff --git a/paddle/fluid/operators/math/complex_functors.h b/paddle/pten/kernels/funcs/complex_functors.h
similarity index 57%
rename from paddle/fluid/operators/math/complex_functors.h
rename to paddle/pten/kernels/funcs/complex_functors.h
index 48f16b87cbd66c6a39c74d1dbaab2349193f04ae..b0eee3ac1fdce3c9fc7de7f8162ae74f4b33daff 100644
--- a/paddle/fluid/operators/math/complex_functors.h
+++ b/paddle/pten/kernels/funcs/complex_functors.h
@@ -13,15 +13,17 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-
+#ifndef _USE_MATH_DEFINES
+#define _USE_MATH_DEFINES
+#endif
+#include <cmath>
 #include <type_traits>
 
-#include "paddle/fluid/platform/complex.h"
+#include "paddle/pten/common/complex.h"
 #include "paddle/pten/core/hostdevice.h"
 
-namespace paddle {
-namespace operators {
-namespace math {
+namespace pten {
+namespace funcs {
 
 template <bool B, typename T>
 struct cond {
@@ -64,8 +66,8 @@ using select_t = typename select<Head, Tail...>::type;
 
 template <typename T>
 using Real =
-    select_t<cond<std::is_same<T, platform::complex<float>>::value, float>,
-             cond<std::is_same<T, platform::complex<double>>::value, double>,
+    select_t<cond<std::is_same<T, pten::dtype::complex<float>>::value, float>,
+             cond<std::is_same<T, pten::dtype::complex<double>>::value, double>,
              T>;
 
 template <typename T, typename RealT>
@@ -77,13 +79,13 @@ using NoComplex = typename std::enable_if<std::is_same<T, RealT>::value>::type;
 
 template <typename T>
 using EnableComplex = typename std::enable_if<
-    std::is_same<T, platform::complex<float>>::value ||
-    std::is_same<T, platform::complex<double>>::value>::type;
+    std::is_same<T, pten::dtype::complex<float>>::value ||
+    std::is_same<T, pten::dtype::complex<double>>::value>::type;
 
 template <typename T>
 using DisableComplex = typename std::enable_if<
-    !std::is_same<T, platform::complex<float>>::value &&
-    !std::is_same<T, platform::complex<double>>::value>::type;
+    !std::is_same<T, pten::dtype::complex<float>>::value &&
+    !std::is_same<T, pten::dtype::complex<double>>::value>::type;
 
 template <typename T, typename Enable = void>
 struct RealFunctor;
@@ -154,8 +156,7 @@ struct AbsFunctor<T, NoComplex<T, Real<T>>> {
 
 template <typename T>
 struct AbsGradFunctor {
-  AbsGradFunctor(const math::Real<T>* dout, const T* x, T* output,
-                 int64_t numel)
+  AbsGradFunctor(const Real<T>* dout, const T* x, T* output, int64_t numel)
       : dout_(dout), x_(x), output_(output), numel_(numel) {}
 
   HOSTDEVICE void operator()(int64_t idx) const {
@@ -166,52 +167,55 @@ struct AbsGradFunctor {
     }
   }
 
-  const math::Real<T>* dout_;
+  const Real<T>* dout_;
   const T* x_;
   T* output_;
   int64_t numel_;
 };
 
 template <>
-struct AbsGradFunctor<paddle::platform::complex<float>> {
-  AbsGradFunctor(const float* dout, const paddle::platform::complex<float>* x,
-                 paddle::platform::complex<float>* output, int64_t numel)
+struct AbsGradFunctor<pten::dtype::complex<float>> {
+  AbsGradFunctor(const float* dout,
+                 const pten::dtype::complex<float>* x,
+                 pten::dtype::complex<float>* output,
+                 int64_t numel)
       : dout_(dout), x_(x), output_(output), numel_(numel) {}
 
   HOSTDEVICE void operator()(int64_t idx) const {
-    if (x_[idx] == paddle::platform::complex<float>(0)) {
-      output_[idx] = paddle::platform::complex<float>(0);
+    if (x_[idx] == pten::dtype::complex<float>(0)) {
+      output_[idx] = pten::dtype::complex<float>(0);
     } else {
-      output_[idx] = paddle::platform::complex<float>(dout_[idx]) *
-                     (x_[idx] / paddle::platform::complex<float>(abs(x_[idx])));
+      output_[idx] = pten::dtype::complex<float>(dout_[idx]) *
+                     (x_[idx] / pten::dtype::complex<float>(abs(x_[idx])));
     }
   }
 
   const float* dout_;
-  const paddle::platform::complex<float>* x_;
-  paddle::platform::complex<float>* output_;
+  const pten::dtype::complex<float>* x_;
+  pten::dtype::complex<float>* output_;
   int64_t numel_;
 };
 
 template <>
-struct AbsGradFunctor<paddle::platform::complex<double>> {
-  AbsGradFunctor(const double* dout, const paddle::platform::complex<double>* x,
-                 paddle::platform::complex<double>* output, int64_t numel)
+struct AbsGradFunctor<pten::dtype::complex<double>> {
+  AbsGradFunctor(const double* dout,
+                 const pten::dtype::complex<double>* x,
+                 pten::dtype::complex<double>* output,
+                 int64_t numel)
       : dout_(dout), x_(x), output_(output), numel_(numel) {}
 
   HOSTDEVICE void operator()(int64_t idx) const {
-    if (x_[idx] == paddle::platform::complex<double>(0)) {
-      output_[idx] = paddle::platform::complex<double>(0);
+    if (x_[idx] == pten::dtype::complex<double>(0)) {
+      output_[idx] = pten::dtype::complex<double>(0);
     } else {
-      output_[idx] =
-          paddle::platform::complex<double>(dout_[idx]) *
-          (x_[idx] / paddle::platform::complex<double>(abs(x_[idx])));
+      output_[idx] = pten::dtype::complex<double>(dout_[idx]) *
+                     (x_[idx] / pten::dtype::complex<double>(abs(x_[idx])));
     }
   }
 
   const double* dout_;
-  const paddle::platform::complex<double>* x_;
-  paddle::platform::complex<double>* output_;
+  const pten::dtype::complex<double>* x_;
+  pten::dtype::complex<double>* output_;
   int64_t numel_;
 };
 
@@ -235,46 +239,48 @@ struct AbsGradGradFunctor {
 };
 
 template <>
-struct AbsGradGradFunctor<paddle::platform::complex<double>> {
-  AbsGradGradFunctor(const paddle::platform::complex<double>* ddx,
-                     const paddle::platform::complex<double>* x,
-                     paddle::platform::complex<double>* output, int64_t numel)
+struct AbsGradGradFunctor<pten::dtype::complex<double>> {
+  AbsGradGradFunctor(const pten::dtype::complex<double>* ddx,
+                     const pten::dtype::complex<double>* x,
+                     pten::dtype::complex<double>* output,
+                     int64_t numel)
       : ddx_(ddx), x_(x), output_(output), numel_(numel) {}
 
   HOSTDEVICE void operator()(int64_t idx) const {
-    if (x_[idx] == paddle::platform::complex<double>(0)) {
-      output_[idx] = paddle::platform::complex<double>(0);
+    if (x_[idx] == pten::dtype::complex<double>(0)) {
+      output_[idx] = pten::dtype::complex<double>(0);
     } else {
-      output_[idx] = paddle::platform::complex<double>(ddx_[idx]) * x_[idx] /
-                     paddle::platform::complex<double>(abs(x_[idx]));
+      output_[idx] = pten::dtype::complex<double>(ddx_[idx]) * x_[idx] /
+                     pten::dtype::complex<double>(abs(x_[idx]));
     }
   }
 
-  const paddle::platform::complex<double>* ddx_;
-  const paddle::platform::complex<double>* x_;
-  paddle::platform::complex<double>* output_;
+  const pten::dtype::complex<double>* ddx_;
+  const pten::dtype::complex<double>* x_;
+  pten::dtype::complex<double>* output_;
   int64_t numel_;
 };
 
 template <>
-struct AbsGradGradFunctor<paddle::platform::complex<float>> {
-  AbsGradGradFunctor(const paddle::platform::complex<float>* ddx,
-                     const paddle::platform::complex<float>* x,
-                     paddle::platform::complex<float>* output, int64_t numel)
+struct AbsGradGradFunctor<pten::dtype::complex<float>> {
+  AbsGradGradFunctor(const pten::dtype::complex<float>* ddx,
+                     const pten::dtype::complex<float>* x,
+                     pten::dtype::complex<float>* output,
+                     int64_t numel)
       : ddx_(ddx), x_(x), output_(output), numel_(numel) {}
 
   HOSTDEVICE void operator()(int64_t idx) const {
-    if (x_[idx] == paddle::platform::complex<float>(0)) {
-      output_[idx] = paddle::platform::complex<float>(0);
+    if (x_[idx] == pten::dtype::complex<float>(0)) {
+      output_[idx] = pten::dtype::complex<float>(0);
     } else {
-      output_[idx] = paddle::platform::complex<float>(ddx_[idx]) * x_[idx] /
-                     paddle::platform::complex<float>(abs(x_[idx]));
+      output_[idx] = pten::dtype::complex<float>(ddx_[idx]) * x_[idx] /
+                     pten::dtype::complex<float>(abs(x_[idx]));
     }
   }
 
-  const paddle::platform::complex<float>* ddx_;
-  const paddle::platform::complex<float>* x_;
-  paddle::platform::complex<float>* output_;
+  const pten::dtype::complex<float>* ddx_;
+  const pten::dtype::complex<float>* x_;
+  pten::dtype::complex<float>* output_;
   int64_t numel_;
 };
 template <typename T, typename Enable = void>
@@ -318,8 +324,10 @@ struct RealImagToComplexFunctor;
 
 template <typename T>
 struct RealImagToComplexFunctor<T, Complex<T, Real<T>>> {
-  RealImagToComplexFunctor(const Real<T>* input_real, const Real<T>* input_imag,
-                           T* output, int64_t numel)
+  RealImagToComplexFunctor(const Real<T>* input_real,
+                           const Real<T>* input_imag,
+                           T* output,
+                           int64_t numel)
       : input_real_(input_real),
         input_imag_(input_imag),
         output_(output),
@@ -363,6 +371,84 @@ struct ConjFunctor<T, DisableComplex<T>> {
   T* output_;
 };
 
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle
+template <typename T, typename Enable = void>
+struct AngleFunctor;
+
+// angel function for complex
+template <typename T>
+struct AngleFunctor<T, pten::funcs::Complex<T, pten::funcs::Real<T>>> {
+  AngleFunctor(const T* input, pten::funcs::Real<T>* output, int64_t numel)
+      : input_(input), output_(output), numel_(numel) {}
+
+  HOSTDEVICE void operator()(int64_t idx) const {
+    output_[idx] = arg(input_[idx]);
+  }
+
+  const T* input_;
+  pten::funcs::Real<T>* output_;
+  int64_t numel_;
+};
+
+// angel function for real
+template <typename T>
+struct AngleFunctor<T, pten::funcs::NoComplex<T, pten::funcs::Real<T>>> {
+  AngleFunctor(const T* input, T* output, int64_t numel)
+      : input_(input), output_(output), numel_(numel) {}
+
+  HOSTDEVICE void operator()(int64_t idx) const {
+    output_[idx] = input_[idx] < static_cast<T>(0) ? M_PI : 0;
+  }
+
+  const T* input_;
+  T* output_;
+  int64_t numel_;
+};
+
+template <typename T, typename Enable = void>
+struct AngleGradFunctor;
+
+// angle grad for complex
+template <typename T>
+struct AngleGradFunctor<T, pten::funcs::Complex<T, pten::funcs::Real<T>>> {
+  AngleGradFunctor(const pten::funcs::Real<T>* dout,
+                   const T* x,
+                   T* dx,
+                   int64_t numel)
+      : dout_(dout), x_(x), dx_(dx), numel_(numel) {}
+
+  HOSTDEVICE void operator()(int64_t idx) const {
+    if (x_[idx] == T(0)) {
+      dx_[idx] = T(0);
+    } else {
+      const pten::funcs::Real<T> r_square =
+          x_[idx].real * x_[idx].real + x_[idx].imag * x_[idx].imag;
+      dx_[idx] = T(-dout_[idx] * x_[idx].imag / r_square,
+                   dout_[idx] * x_[idx].real / r_square);
+    }
+  }
+
+  const pten::funcs::Real<T>* dout_;
+  const T* x_;
+  T* dx_;
+  int64_t numel_;
+};
+
+// angle grad for real
+template <typename T>
+struct AngleGradFunctor<T, pten::funcs::NoComplex<T, pten::funcs::Real<T>>> {
+  AngleGradFunctor(const pten::funcs::Real<T>* dout,
+                   const T* x,
+                   T* dx,
+                   int64_t numel)
+      : dout_(dout), x_(x), dx_(dx), numel_(numel) {}
+
+  HOSTDEVICE void operator()(int64_t idx) const { dx_[idx] = 0; }
+
+  const pten::funcs::Real<T>* dout_;
+  const T* x_;
+  T* dx_;
+  int64_t numel_;
+};
+
+}  // namespace funcs
+}  // namespace pten
diff --git a/paddle/pten/kernels/gpu/abs_kernel.cu b/paddle/pten/kernels/gpu/abs_kernel.cu
index d97aa79105322e8678fd559f9306a60b15c39080..06eff050674c3670a2aa07cb43d0baea82fe7202 100644
--- a/paddle/pten/kernels/gpu/abs_kernel.cu
+++ b/paddle/pten/kernels/gpu/abs_kernel.cu
@@ -14,11 +14,11 @@
 
 #include <algorithm>
 #include <vector>
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/pten/backends/gpu/gpu_context.h"
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_registry.h"
 #include "paddle/pten/kernels/abs_kernel.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 #include "paddle/pten/kernels/funcs/elementwise_base.h"
 
 namespace pten {
@@ -27,19 +27,14 @@ template <typename T, typename Enable = void>
 struct CudaAbsFunctor;
 
 template <typename T>
-struct CudaAbsFunctor<
-    T,
-    paddle::operators::math::Complex<T, paddle::operators::math::Real<T>>> {
-  __device__ __forceinline__ paddle::operators::math::Real<T> operator()(
-      const T x) const {
+struct CudaAbsFunctor<T, pten::funcs::Complex<T, pten::funcs::Real<T>>> {
+  __device__ __forceinline__ pten::funcs::Real<T> operator()(const T x) const {
     return abs(x);
   }
 };
 
 template <typename T>
-struct CudaAbsFunctor<
-    T,
-    paddle::operators::math::NoComplex<T, paddle::operators::math::Real<T>>> {
+struct CudaAbsFunctor<T, pten::funcs::NoComplex<T, pten::funcs::Real<T>>> {
   __device__ __forceinline__ T operator()(const T x) const {
     return std::abs(x);
   }
@@ -47,12 +42,12 @@ struct CudaAbsFunctor<
 
 template <typename T, typename Context>
 void AbsKernel(const Context& ctx, const DenseTensor& x, DenseTensor* out) {
-  ctx.template Alloc<paddle::operators::math::Real<T>>(out);
+  ctx.template Alloc<pten::funcs::Real<T>>(out);
   std::vector<const DenseTensor*> ins = {&x};
   std::vector<DenseTensor*> outs = {out};
   auto functor = CudaAbsFunctor<T>();
 
-  funcs::LaunchSameDimsElementwiseCudaKernel<paddle::operators::math::Real<T>>(
+  funcs::LaunchSameDimsElementwiseCudaKernel<pten::funcs::Real<T>>(
       ctx, ins, &outs, functor);
 }
 
diff --git a/paddle/pten/kernels/impl/abs_grad_kernel_impl.h b/paddle/pten/kernels/impl/abs_grad_kernel_impl.h
index c702a0042dc4e2c7f75888a76c95b29dae667ec6..ff829e10b2d8bf06971173978627632bf18fa93f 100644
--- a/paddle/pten/kernels/impl/abs_grad_kernel_impl.h
+++ b/paddle/pten/kernels/impl/abs_grad_kernel_impl.h
@@ -14,9 +14,9 @@
 
 #pragma once
 
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
 #include "paddle/pten/kernels/abs_grad_kernel.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace pten {
 
@@ -26,15 +26,14 @@ void AbsGradKernel(const Context& ctx,
                    const DenseTensor& dout,
                    DenseTensor* dx) {
   auto numel = dout.numel();
-  auto* dout_data = dout.data<paddle::operators::math::Real<T>>();
+  auto* dout_data = dout.data<pten::funcs::Real<T>>();
   auto* x_data = x.data<T>();
 
   ctx.template Alloc<T>(dx, static_cast<size_t>(numel * sizeof(T)));
   auto* dx_data = dx->data<T>();
 
   paddle::platform::ForRange<Context> for_range(ctx, numel);
-  paddle::operators::math::AbsGradFunctor<T> functor(
-      dout_data, x_data, dx_data, numel);
+  pten::funcs::AbsGradFunctor<T> functor(dout_data, x_data, dx_data, numel);
   for_range(functor);
 }
 
@@ -50,7 +49,7 @@ void AbsDoubleGradKernel(const Context& ctx,
   auto* ddout_data = ddout->data<T>();
 
   paddle::platform::ForRange<Context> for_range(ctx, numel);
-  paddle::operators::math::AbsGradGradFunctor<T> functor(
+  pten::funcs::AbsGradGradFunctor<T> functor(
       ddx_data, x_data, ddout_data, numel);
   for_range(functor);
 }
diff --git a/paddle/pten/kernels/impl/complex_kernel_impl.h b/paddle/pten/kernels/impl/complex_kernel_impl.h
index 7e4c4f0d66d4fc89634eb7bde9eb24e2743d4a7c..17cfb886e57b813fa744ebc232d6cc38e6b0f951 100644
--- a/paddle/pten/kernels/impl/complex_kernel_impl.h
+++ b/paddle/pten/kernels/impl/complex_kernel_impl.h
@@ -15,8 +15,8 @@
 #pragma once
 
 // See Note [ Why still include the fluid headers? ]
-#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/platform/for_range.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace pten {
 
@@ -29,7 +29,7 @@ void ConjKernel(const Context& dev_ctx,
   auto* out_data = dev_ctx.template Alloc<T>(out);
 
   paddle::platform::ForRange<Context> for_range(dev_ctx, numel);
-  paddle::operators::math::ConjFunctor<T> functor(x_data, numel, out_data);
+  pten::funcs::ConjFunctor<T> functor(x_data, numel, out_data);
   for_range(functor);
 }
 
diff --git a/paddle/pten/kernels/impl/dot_grad_kernel_impl.h b/paddle/pten/kernels/impl/dot_grad_kernel_impl.h
index d4ea9fc944527145269fdfd1a854aca1299a6018..4ed47bd69dd5f6b37b179cc9534fde64f949b5de 100644
--- a/paddle/pten/kernels/impl/dot_grad_kernel_impl.h
+++ b/paddle/pten/kernels/impl/dot_grad_kernel_impl.h
@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/pten/kernels/complex_kernel.h"
 
 #include "paddle/fluid/operators/eigen/eigen_function.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 namespace pten {
 
@@ -35,9 +35,7 @@ struct DotGradFunction {
 };
 
 template <typename DeviceContext, typename T>
-struct DotGradFunction<DeviceContext,
-                       T,
-                       paddle::operators::math::EnableComplex<T>> {
+struct DotGradFunction<DeviceContext, T, pten::funcs::EnableComplex<T>> {
   void operator()(const DeviceContext& ctx,
                   const DenseTensor* tensor_x,
                   const DenseTensor* tensor_y,
@@ -133,9 +131,7 @@ struct DotGradFunction<DeviceContext,
 };
 
 template <typename DeviceContext, typename T>
-struct DotGradFunction<DeviceContext,
-                       T,
-                       paddle::operators::math::DisableComplex<T>> {
+struct DotGradFunction<DeviceContext, T, pten::funcs::DisableComplex<T>> {
   void operator()(const DeviceContext& ctx,
                   const DenseTensor* tensor_x,
                   const DenseTensor* tensor_y,
@@ -221,9 +217,7 @@ struct DotDoubleGradFunction {
 };
 
 template <typename DeviceContext, typename T>
-struct DotDoubleGradFunction<DeviceContext,
-                             T,
-                             paddle::operators::math::EnableComplex<T>> {
+struct DotDoubleGradFunction<DeviceContext, T, pten::funcs::EnableComplex<T>> {
   void operator()(const DeviceContext& ctx,
                   const DenseTensor* tensor_x,
                   const DenseTensor* tensor_y,
@@ -334,9 +328,7 @@ struct DotDoubleGradFunction<DeviceContext,
 };
 
 template <typename DeviceContext, typename T>
-struct DotDoubleGradFunction<DeviceContext,
-                             T,
-                             paddle::operators::math::DisableComplex<T>> {
+struct DotDoubleGradFunction<DeviceContext, T, pten::funcs::DisableComplex<T>> {
   void operator()(const DeviceContext& ctx,
                   const DenseTensor* tensor_x,
                   const DenseTensor* tensor_y,
@@ -461,9 +453,7 @@ struct DotTripleGradFunction {
 // TODO(wuweilong): enable this function when the unittests framewark for multi
 // grad is ok (dtype: complex64 or complex128).
 template <typename DeviceContext, typename T>
-struct DotTripleGradFunction<DeviceContext,
-                             T,
-                             paddle::operators::math::EnableComplex<T>> {
+struct DotTripleGradFunction<DeviceContext, T, pten::funcs::EnableComplex<T>> {
   void operator()(const DeviceContext& ctx,
                   const DenseTensor* in_tensor_x,
                   const DenseTensor* in_tensor_y,
@@ -656,9 +646,7 @@ struct DotTripleGradFunction<DeviceContext,
 };
 
 template <typename DeviceContext, typename T>
-struct DotTripleGradFunction<DeviceContext,
-                             T,
-                             paddle::operators::math::DisableComplex<T>> {
+struct DotTripleGradFunction<DeviceContext, T, pten::funcs::DisableComplex<T>> {
   void operator()(const DeviceContext& ctx,
                   const DenseTensor* in_tensor_x,
                   const DenseTensor* in_tensor_y,
diff --git a/paddle/pten/kernels/impl/matmul_kernel_impl.h b/paddle/pten/kernels/impl/matmul_kernel_impl.h
index 858807a1d4d6496d5e3091aa71f5b2dada03b92e..addea622f140210ae714da2eda775f6ce6568eca 100644
--- a/paddle/pten/kernels/impl/matmul_kernel_impl.h
+++ b/paddle/pten/kernels/impl/matmul_kernel_impl.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
+#include "paddle/pten/kernels/funcs/complex_functors.h"
 
 #include "paddle/pten/core/dense_tensor.h"