diff --git a/paddle/phi/kernels/cpu/dirichlet_kernel.cc b/paddle/phi/kernels/cpu/dirichlet_kernel.cc
index 76ef231344199e542afcc85d2aaebffa8b017c76..c124920dfa0db8af9cb85e5b4b5889b664dfe989 100644
--- a/paddle/phi/kernels/cpu/dirichlet_kernel.cc
+++ b/paddle/phi/kernels/cpu/dirichlet_kernel.cc
@@ -16,9 +16,9 @@
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/cpu/elementwise.h"
-#include "paddle/phi/kernels/cpu/reduce.h"
 #include "paddle/phi/kernels/funcs/elementwise_functor.h"
 #include "paddle/phi/kernels/funcs/for_range.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/funcs/reduce_functor.h"
 #include "paddle/phi/kernels/impl/dirichlet_kernel_impl.h"
 
@@ -83,7 +83,7 @@ struct DirichletSampler<CPUContext, T> {
     gamma_sum.Resize(new_shape);
     dev_ctx.template Alloc<T>(&gamma_sum);
 
-    ReduceKernelImpl<CPUContext, T, T, funcs::SumFunctor>(
+    funcs::ReduceKernelImpl<CPUContext, T, T, funcs::SumFunctor>(
         dev_ctx,
         gamma_samples,
         &gamma_sum,
diff --git a/paddle/phi/kernels/cpu/reduce.h b/paddle/phi/kernels/cpu/reduce.h
index dad288cff2c1a4565a7b67d688936d231a6947c9..e5f610b9554097676afe26f94207b7fcd8cb06a2 100644
--- a/paddle/phi/kernels/cpu/reduce.h
+++ b/paddle/phi/kernels/cpu/reduce.h
@@ -16,181 +16,11 @@
 
 #include <set>
 
-#include "paddle/phi/backends/cpu/cpu_context.h"
-#include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/visit_type.h"
 #include "paddle/phi/kernels/cast_kernel.h"
-#include "paddle/phi/kernels/funcs/eigen/common.h"
-#include "paddle/phi/kernels/funcs/math_function.h"
-// See Note [ Why still include the fluid headers? ]
-#include "paddle/fluid/operators/eigen/eigen_function.h"
-namespace phi {
-
-template <typename DeviceContext,
-          typename T,
-          size_t D,
-          size_t R_D,
-          typename Functor>
-void ReduceFunctor(const DeviceContext& context,
-                   const phi::DenseTensor& input,
-                   phi::DenseTensor* output,
-                   const std::vector<int64_t>& dims,
-                   bool keep_dim) {
-  auto x = EigenTensor<T, D>::From(input);
-  auto x_rank = static_cast<int>(x.dimensions().size());
-  auto reduce_dim = Eigen::array<int, R_D>();
-  std::vector<int64_t> dims_ref = dims;
-  for (size_t i = 0; i < dims_ref.size(); ++i) {
-    if (dims_ref[i] < 0) dims_ref[i] = x_rank + dims_ref[i];
-    reduce_dim[i] = dims_ref[i];
-  }
-  // construct the squeezed output tensor
-  DDim out_dims = output->dims();
-  if (keep_dim && x_rank > 1) {
-    const int kDelFlag = -2;
-    auto dims_vector = phi::vectorize(out_dims);
-    for (size_t i = 0; i < dims_ref.size(); ++i) {
-      dims_vector[dims_ref[i]] = kDelFlag;
-    }
-    dims_vector.erase(remove(dims_vector.begin(), dims_vector.end(), kDelFlag),
-                      dims_vector.end());
-    out_dims = phi::make_ddim(dims_vector);
-  }
-  auto& place = *context.eigen_device();
-  Functor functor;
-
-  if (D == 1) {
-    auto out = EigenScalar<T>::From(*output);
-    functor(place, &x, &out, reduce_dim);
-  } else {
-    auto out = EigenTensor<T, (D - R_D)>::From(*output, out_dims);
-    functor(place, &x, &out, reduce_dim);
-  }
-}
-
-#define HANDLE_REDUCE_DIM(NDIM, RDIM)                        \
-  if (ndim == NDIM && rdim == RDIM) {                        \
-    ReduceFunctor<DeviceContext, OutT, NDIM, RDIM, Functor>( \
-        dev_ctx, input, output, dims, keep_dim);             \
-  }
-//////////////// HandleLargeDim
-
-inline void GetShuffledDim(const DDim& src_dims,
-                           DDim* dst_dims,
-                           const std::vector<int64_t>& reduced_dims,
-                           std::vector<int>* perm_axis) {
-  // check if it's a reduced dim
-  std::vector<bool> src_dims_check(src_dims.size(), false);
-  size_t src_size = src_dims.size();
-  size_t reduce_size = reduced_dims.size();
-  std::vector<int64_t> regular_reduced_dims = reduced_dims;
-  for (size_t i = 0; i < regular_reduced_dims.size(); i++) {
-    if (regular_reduced_dims[i] < 0) {
-      regular_reduced_dims[i] = src_size + regular_reduced_dims[i];
-    }
-  }
-
-  for (size_t i = 0; i < reduce_size; ++i) {
-    dst_dims->at(src_size - reduce_size + i) =
-        src_dims[regular_reduced_dims[i]];
-    (*perm_axis)[src_size - reduce_size + i] = regular_reduced_dims[i];
-    src_dims_check[regular_reduced_dims[i]] = true;
-  }
-
-  size_t offset = 0;
-  for (size_t i = 0; i < src_dims_check.size(); ++i) {
-    bool is_reduced = src_dims_check[i];
-    if (!is_reduced) {
-      (*perm_axis)[offset] = i;
-      dst_dims->at(offset++) = src_dims[i];
-    }
-  }
-}
-
-template <typename DeviceContext, typename OutT>
-void GetShuffledInput(const DeviceContext& dev_ctx,
-                      const phi::DenseTensor& input,
-                      phi::DenseTensor* shuffled_input,
-                      const std::vector<int64_t>& dims) {
-  DDim shuffled_dims(input.dims());
-  std::vector<int> perm_axis(input.dims().size());
-  GetShuffledDim(input.dims(), &shuffled_dims, dims, &perm_axis);
-
-  shuffled_input->Resize(shuffled_dims);
-  dev_ctx.template Alloc<OutT>(shuffled_input);
-
-  phi::funcs::TransposeNormal<DeviceContext, OutT> trans;
-  trans(dev_ctx, input, shuffled_input, perm_axis);
-}
-
-template <typename DeviceContext, typename OutT, typename Functor>
-void HandleLargeDim(const DeviceContext& dev_ctx,
-                    const phi::DenseTensor& input,
-                    phi::DenseTensor* output,
-                    const std::vector<int64_t>& dims,
-                    bool keep_dim) {
-  //  shuffle the reduced dim to the end
-  phi::DenseTensor shuffled_input;
-  GetShuffledInput<DeviceContext, OutT>(dev_ctx, input, &shuffled_input, dims);
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 
-  // transpose to 2D tensor whose shape is {unreduced, reduced}.
-  const int64_t unreduced = output->numel();
-  const int64_t reduced = shuffled_input.numel() / unreduced;
-  shuffled_input.ResizeAndAllocate({unreduced, reduced});
-  DDim output_dim = output->dims();
-  output->ResizeAndAllocate({unreduced});
-  ReduceFunctor<DeviceContext, OutT, 2, 1, Functor>(
-      dev_ctx, shuffled_input, output, {1}, keep_dim);
-  output->ResizeAndAllocate(output_dim);
-}
-
-////////////// ReduceKernel
-
-template <typename DeviceContext, typename T, typename OutT, typename Functor>
-void ReduceKernelImpl(const DeviceContext& dev_ctx,
-                      const phi::DenseTensor& input,
-                      phi::DenseTensor* output,
-                      const std::vector<int64_t>& dims,
-                      bool keep_dim,
-                      bool reduce_all) {
-  dev_ctx.template Alloc<OutT>(output);
-
-  if (reduce_all) {
-    // Flatten and reduce 1-D tensor
-    auto x = EigenVector<OutT>::Flatten(input);
-    auto out = EigenScalar<OutT>::From(*output);
-    auto& dev = *dev_ctx.eigen_device();
-    auto reduce_dim = Eigen::array<int, 1>({{0}});
-
-    Functor functor;
-    functor(dev, &x, &out, reduce_dim);
-  } else {
-    int ndim = input.dims().size();
-    int rdim = dims.size();
-    if (ndim > 6) {
-      HandleLargeDim<DeviceContext, OutT, Functor>(
-          dev_ctx, input, output, dims, keep_dim);
-
-    } else {
-      HANDLE_REDUCE_DIM(6, 5);
-      HANDLE_REDUCE_DIM(6, 4);
-      HANDLE_REDUCE_DIM(6, 3);
-      HANDLE_REDUCE_DIM(6, 2);
-      HANDLE_REDUCE_DIM(6, 1);
-      HANDLE_REDUCE_DIM(5, 4);
-      HANDLE_REDUCE_DIM(5, 3);
-      HANDLE_REDUCE_DIM(5, 2);
-      HANDLE_REDUCE_DIM(5, 1);
-      HANDLE_REDUCE_DIM(4, 3);
-      HANDLE_REDUCE_DIM(4, 2);
-      HANDLE_REDUCE_DIM(4, 1);
-      HANDLE_REDUCE_DIM(3, 2);
-      HANDLE_REDUCE_DIM(3, 1);
-      HANDLE_REDUCE_DIM(2, 1);
-      HANDLE_REDUCE_DIM(1, 1);
-    }
-  }
-}
+namespace phi {
 
 template <typename DeviceContext, typename T, typename Functor>
 void Reduce(const DeviceContext& dev_ctx,
@@ -218,7 +48,7 @@ void Reduce(const DeviceContext& dev_ctx,
     // do reduce sum
     PD_VISIT_ALL_TYPES(
         x.dtype(), "ReduceKernelImpl", ([&] {
-          phi::ReduceKernelImpl<DeviceContext, T, data_t, Functor>(
+          phi::funcs::ReduceKernelImpl<DeviceContext, T, data_t, Functor>(
               dev_ctx, x, out, dims, keep_dim, reduce_all);
         }));
   } else {
@@ -228,7 +58,7 @@ void Reduce(const DeviceContext& dev_ctx,
     // do reduce sum
     PD_VISIT_ALL_TYPES(
         out_dtype, "ReduceKernelImpl", ([&] {
-          phi::ReduceKernelImpl<DeviceContext, T, data_t, Functor>(
+          phi::funcs::ReduceKernelImpl<DeviceContext, T, data_t, Functor>(
               dev_ctx, tmp_tensor, out, dims, keep_dim, reduce_all);
         }));
   }
@@ -255,7 +85,7 @@ void BoolReduceKernel(const DeviceContext& dev_ctx,
   }
   reduce_all = (reduce_all || full_dim);
 
-  ReduceKernelImpl<DeviceContext, bool, OutT, Functor>(
+  funcs::ReduceKernelImpl<DeviceContext, bool, OutT, Functor>(
       dev_ctx, input, output, dims, keep_dim, reduce_all);
 }
 
diff --git a/paddle/phi/kernels/funcs/matrix_reduce.cc b/paddle/phi/kernels/funcs/matrix_reduce.cc
index 849fd7a0075a89cedeab4d87c779931f2a14f115..82396185e94521c0447d4c9ca9d82f9b19e933e1 100644
--- a/paddle/phi/kernels/funcs/matrix_reduce.cc
+++ b/paddle/phi/kernels/funcs/matrix_reduce.cc
@@ -14,7 +14,7 @@
 
 #include "paddle/phi/kernels/funcs/matrix_reduce.h"
 
-#include "paddle/phi/kernels/cpu/reduce.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/funcs/reduce_functor.h"
 
 namespace phi {
@@ -47,7 +47,7 @@ class MatrixReduceSumFunctor<T, CPUContext> {
         out_reduce_dims.push_back(idx);
       }
     }
-    phi::ReduceKernelImpl<CPUContext, T, T, phi::funcs::SumFunctor>(
+    ReduceKernelImpl<CPUContext, T, T, phi::funcs::SumFunctor>(
         dev_ctx, in, out, out_reduce_dims, true, false);
   }
 };
diff --git a/paddle/phi/kernels/funcs/reduce_function.h b/paddle/phi/kernels/funcs/reduce_function.h
index 446dfc73d5bd692ba15eabc5ffc3a61d55a3809b..92fe3885b42f0a5a6c23bf6f0ea9445658b09a3d 100644
--- a/paddle/phi/kernels/funcs/reduce_function.h
+++ b/paddle/phi/kernels/funcs/reduce_function.h
@@ -15,8 +15,7 @@
 #pragma once
 
 // CUDA, XPU and HIP use same api
-#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
-    defined(PADDLE_WITH_XPU_KP)
+#if defined(__NVCC__) || defined(__HIPCC__) || defined(__xpu__)
 
 #include <algorithm>
 #include <cmath>
@@ -40,10 +39,6 @@ namespace cub = hipcub;
 #include "paddle/phi/backends/gpu/gpu_info.h"
 #endif
 
-#include "paddle/phi/api/ext/dispatch.h"
-#include "paddle/phi/core/dense_tensor.h"
-#include "paddle/phi/core/enforce.h"
-#include "paddle/phi/core/utils/array.h"
 #include "paddle/phi/kernels/cast_kernel.h"
 #include "paddle/phi/kernels/empty_kernel.h"
 #include "paddle/phi/kernels/funcs/elementwise_base.h"
@@ -58,9 +53,19 @@ namespace kps = phi::kps;
 #ifdef PADDLE_WITH_XPU_KP
 using dim3 = phi::kps::dim3;
 #endif
+
+#endif
+
+#include "paddle/phi/core/dense_tensor.h"
+#include "paddle/phi/core/enforce.h"
+#include "paddle/phi/core/utils/array.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
+#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
+#include "paddle/phi/kernels/funcs/math_function.h"
 namespace phi {
 namespace funcs {
 
+#if defined(__NVCC__) || defined(__HIPCC__) || defined(__xpu__)
 namespace details {
 
 static inline int GetLastPow2(int n) {
@@ -1190,8 +1195,174 @@ void ReduceKernel(const KPDevice& dev_ctx,
       is_mean);
 }
 
+#endif
+
+template <typename DeviceContext,
+          typename T,
+          size_t D,
+          size_t R_D,
+          typename Functor>
+void ReduceFunctor(const DeviceContext& context,
+                   const phi::DenseTensor& input,
+                   phi::DenseTensor* output,
+                   const std::vector<int64_t>& dims,
+                   bool keep_dim) {
+  auto x = EigenTensor<T, D>::From(input);
+  auto x_rank = static_cast<int>(x.dimensions().size());
+  auto reduce_dim = Eigen::array<int, R_D>();
+  std::vector<int64_t> dims_ref = dims;
+  for (size_t i = 0; i < dims_ref.size(); ++i) {
+    if (dims_ref[i] < 0) dims_ref[i] = x_rank + dims_ref[i];
+    reduce_dim[i] = dims_ref[i];
+  }
+  // construct the squeezed output tensor
+  DDim out_dims = output->dims();
+  if (keep_dim && x_rank > 1) {
+    const int kDelFlag = -2;
+    auto dims_vector = phi::vectorize(out_dims);
+    for (size_t i = 0; i < dims_ref.size(); ++i) {
+      dims_vector[dims_ref[i]] = kDelFlag;
+    }
+    dims_vector.erase(remove(dims_vector.begin(), dims_vector.end(), kDelFlag),
+                      dims_vector.end());
+    out_dims = phi::make_ddim(dims_vector);
+  }
+  auto& place = *context.eigen_device();
+  Functor functor;
+
+  if (D == 1) {
+    auto out = EigenScalar<T>::From(*output);
+    functor(place, &x, &out, reduce_dim);
+  } else {
+    auto out = EigenTensor<T, (D - R_D)>::From(*output, out_dims);
+    functor(place, &x, &out, reduce_dim);
+  }
+}
+
+#define HANDLE_REDUCE_DIM(NDIM, RDIM)                        \
+  if (ndim == NDIM && rdim == RDIM) {                        \
+    ReduceFunctor<DeviceContext, OutT, NDIM, RDIM, Functor>( \
+        dev_ctx, input, output, dims, keep_dim);             \
+  }
+//////////////// HandleLargeDim
+
+inline void GetShuffledDim(const DDim& src_dims,
+                           DDim* dst_dims,
+                           const std::vector<int64_t>& reduced_dims,
+                           std::vector<int>* perm_axis) {
+  // check if it's a reduced dim
+  std::vector<bool> src_dims_check(src_dims.size(), false);
+  size_t src_size = src_dims.size();
+  size_t reduce_size = reduced_dims.size();
+  std::vector<int64_t> regular_reduced_dims = reduced_dims;
+  for (size_t i = 0; i < regular_reduced_dims.size(); i++) {
+    if (regular_reduced_dims[i] < 0) {
+      regular_reduced_dims[i] = src_size + regular_reduced_dims[i];
+    }
+  }
+
+  for (size_t i = 0; i < reduce_size; ++i) {
+    dst_dims->at(src_size - reduce_size + i) =
+        src_dims[regular_reduced_dims[i]];
+    (*perm_axis)[src_size - reduce_size + i] = regular_reduced_dims[i];
+    src_dims_check[regular_reduced_dims[i]] = true;
+  }
+
+  size_t offset = 0;
+  for (size_t i = 0; i < src_dims_check.size(); ++i) {
+    bool is_reduced = src_dims_check[i];
+    if (!is_reduced) {
+      (*perm_axis)[offset] = i;
+      dst_dims->at(offset++) = src_dims[i];
+    }
+  }
+}
+
+template <typename DeviceContext, typename OutT>
+void GetShuffledInput(const DeviceContext& dev_ctx,
+                      const phi::DenseTensor& input,
+                      phi::DenseTensor* shuffled_input,
+                      const std::vector<int64_t>& dims) {
+  DDim shuffled_dims(input.dims());
+  std::vector<int> perm_axis(input.dims().size());
+  GetShuffledDim(input.dims(), &shuffled_dims, dims, &perm_axis);
+
+  shuffled_input->Resize(shuffled_dims);
+  dev_ctx.template Alloc<OutT>(shuffled_input);
+
+  phi::funcs::TransposeNormal<DeviceContext, OutT> trans;
+  trans(dev_ctx, input, shuffled_input, perm_axis);
+}
+
+template <typename DeviceContext, typename OutT, typename Functor>
+void HandleLargeDim(const DeviceContext& dev_ctx,
+                    const phi::DenseTensor& input,
+                    phi::DenseTensor* output,
+                    const std::vector<int64_t>& dims,
+                    bool keep_dim) {
+  //  shuffle the reduced dim to the end
+  phi::DenseTensor shuffled_input;
+  GetShuffledInput<DeviceContext, OutT>(dev_ctx, input, &shuffled_input, dims);
+
+  // transpose to 2D tensor whose shape is {unreduced, reduced}.
+  const int64_t unreduced = output->numel();
+  const int64_t reduced = shuffled_input.numel() / unreduced;
+  shuffled_input.ResizeAndAllocate({unreduced, reduced});
+  DDim output_dim = output->dims();
+  output->ResizeAndAllocate({unreduced});
+  ReduceFunctor<DeviceContext, OutT, 2, 1, Functor>(
+      dev_ctx, shuffled_input, output, {1}, keep_dim);
+  output->ResizeAndAllocate(output_dim);
+}
+
+////////////// ReduceKernel
+
+template <typename DeviceContext, typename T, typename OutT, typename Functor>
+void ReduceKernelImpl(const DeviceContext& dev_ctx,
+                      const phi::DenseTensor& input,
+                      phi::DenseTensor* output,
+                      const std::vector<int64_t>& dims,
+                      bool keep_dim,
+                      bool reduce_all) {
+  dev_ctx.template Alloc<OutT>(output);
+
+  if (reduce_all) {
+    // Flatten and reduce 1-D tensor
+    auto x = EigenVector<OutT>::Flatten(input);
+    auto out = EigenScalar<OutT>::From(*output);
+    auto& dev = *dev_ctx.eigen_device();
+    auto reduce_dim = Eigen::array<int, 1>({{0}});
+
+    Functor functor;
+    functor(dev, &x, &out, reduce_dim);
+  } else {
+    int ndim = input.dims().size();
+    int rdim = dims.size();
+    if (ndim > 6) {
+      HandleLargeDim<DeviceContext, OutT, Functor>(
+          dev_ctx, input, output, dims, keep_dim);
+
+    } else {
+      HANDLE_REDUCE_DIM(6, 5);
+      HANDLE_REDUCE_DIM(6, 4);
+      HANDLE_REDUCE_DIM(6, 3);
+      HANDLE_REDUCE_DIM(6, 2);
+      HANDLE_REDUCE_DIM(6, 1);
+      HANDLE_REDUCE_DIM(5, 4);
+      HANDLE_REDUCE_DIM(5, 3);
+      HANDLE_REDUCE_DIM(5, 2);
+      HANDLE_REDUCE_DIM(5, 1);
+      HANDLE_REDUCE_DIM(4, 3);
+      HANDLE_REDUCE_DIM(4, 2);
+      HANDLE_REDUCE_DIM(4, 1);
+      HANDLE_REDUCE_DIM(3, 2);
+      HANDLE_REDUCE_DIM(3, 1);
+      HANDLE_REDUCE_DIM(2, 1);
+      HANDLE_REDUCE_DIM(1, 1);
+    }
+  }
+}
+
 }  // namespace funcs
 
 }  // namespace phi
-
-#endif
diff --git a/paddle/phi/kernels/gpu/dirichlet_kernel.cu b/paddle/phi/kernels/gpu/dirichlet_kernel.cu
index eb34df90f0864952ec1177a2d91d41411090587d..5a2077c8a592e058f27aeb9e3ac2de8c521c45e7 100644
--- a/paddle/phi/kernels/gpu/dirichlet_kernel.cu
+++ b/paddle/phi/kernels/gpu/dirichlet_kernel.cu
@@ -16,10 +16,10 @@
 
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/cpu/reduce.h"
 #include "paddle/phi/kernels/funcs/broadcast_function.h"
 #include "paddle/phi/kernels/funcs/elementwise_functor.h"
 #include "paddle/phi/kernels/funcs/for_range.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/funcs/reduce_functor.h"
 #include "paddle/phi/kernels/impl/dirichlet_kernel_impl.h"
 
@@ -99,7 +99,7 @@ struct DirichletSampler<GPUContext, T> {
     gamma_sum.Resize(new_shape);
     dev_ctx.template Alloc<T>(&gamma_sum);
 
-    ReduceKernelImpl<GPUContext, T, T, funcs::SumFunctor>(
+    funcs::ReduceKernelImpl<GPUContext, T, T, funcs::SumFunctor>(
         dev_ctx,
         gamma_samples,
         &gamma_sum,
diff --git a/paddle/phi/kernels/impl/logsumexp_kernel_impl.h b/paddle/phi/kernels/impl/logsumexp_kernel_impl.h
index 7a9573ff522b0a2f3c9cc62e39054c434b55282d..7c2eadcb7df7825e52b411d4e1eb9154b204641a 100644
--- a/paddle/phi/kernels/impl/logsumexp_kernel_impl.h
+++ b/paddle/phi/kernels/impl/logsumexp_kernel_impl.h
@@ -16,17 +16,17 @@
 #include <type_traits>
 #include <vector>
 
-#include "paddle/phi/kernels/cpu/reduce.h"
 #include "paddle/phi/kernels/funcs/eigen/common.h"
 #include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/logsumexp_kernel.h"
 
 namespace phi {
 
-#define HANDLE_DIM(NDIM, RDIM)                               \
-  if (ndim == NDIM && rdim == RDIM) {                        \
-    ReduceFunctor<Context, T, NDIM, RDIM, LogsumexpFunctor>( \
-        dev_ctx, x, out, axis, keepdim);                     \
+#define HANDLE_DIM(NDIM, RDIM)                                      \
+  if (ndim == NDIM && rdim == RDIM) {                               \
+    funcs::ReduceFunctor<Context, T, NDIM, RDIM, LogsumexpFunctor>( \
+        dev_ctx, x, out, axis, keepdim);                            \
   }
 
 struct LogsumexpFunctor {
diff --git a/paddle/phi/kernels/impl/matmul_grad_kernel_impl.h b/paddle/phi/kernels/impl/matmul_grad_kernel_impl.h
index be32f85fe99a431d837fa9cb467ff5efd6a18c4d..f499e59c307291f91de5b25b72fd4a12b20d68a6 100644
--- a/paddle/phi/kernels/impl/matmul_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/matmul_grad_kernel_impl.h
@@ -17,8 +17,8 @@ limitations under the License. */
 #include "paddle/phi/backends/cpu/cpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/kernels/complex_kernel.h"
-#include "paddle/phi/kernels/cpu/reduce.h"
 #include "paddle/phi/kernels/empty_kernel.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/funcs/reduce_functor.h"
 #include "paddle/phi/kernels/impl/dot_grad_kernel_impl.h"
 #include "paddle/phi/kernels/impl/matmul_kernel_impl.h"
@@ -45,7 +45,7 @@ struct ReduceSumForMatmulGrad<CPUContext, T> {
                   const std::vector<int>& reduce_dims) {
     std::vector<int64_t> reduce_dims_tmp(reduce_dims.begin(),
                                          reduce_dims.end());
-    ReduceKernelImpl<CPUContext, T, T, phi::funcs::SumFunctor>(
+    funcs::ReduceKernelImpl<CPUContext, T, T, phi::funcs::SumFunctor>(
         dev_ctx, input, output, reduce_dims_tmp, true, false);
   }
 };
diff --git a/paddle/phi/kernels/impl/solve_grad_kernel_impl.h b/paddle/phi/kernels/impl/solve_grad_kernel_impl.h
index 214db79383800a6e463a160ef360b3e13a9c2927..a661035ab5b749b81369d9b1dfc6454a14145f98 100644
--- a/paddle/phi/kernels/impl/solve_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/solve_grad_kernel_impl.h
@@ -16,11 +16,11 @@ limitations under the License. */
 
 #include "paddle/phi/backends/cpu/cpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
-#include "paddle/phi/kernels/cpu/reduce.h"
 #include "paddle/phi/kernels/expand_as_kernel.h"
 #include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/math_function.h"
 #include "paddle/phi/kernels/funcs/matrix_solve.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/funcs/reduce_functor.h"
 #include "paddle/phi/kernels/impl/solve_kernel_impl.h"
 #include "paddle/phi/kernels/squeeze_kernel.h"
@@ -50,7 +50,7 @@ struct ReduceSumForSolvelGrad<CPUContext, T> {
                   bool keep_dims) {
     std::vector<int64_t> reduce_dims_tmp(reduce_dims.begin(),
                                          reduce_dims.end());
-    phi::ReduceKernelImpl<CPUContext, T, T, phi::funcs::SumFunctor>(
+    funcs::ReduceKernelImpl<CPUContext, T, T, phi::funcs::SumFunctor>(
         dev_ctx, input, output, reduce_dims_tmp, keep_dims, false);
   }
 };