removing dependent to fluid/framework/eigen.h in phi (#47675)

* removing dependent to fluid/framework/eigen.h in phi

* more fix according to PR-CI-Py3 fail

paddle/fluid/distributed/ps/service/communicator/communicator.h

@@ -154,7 +154,7 @@ class BlockingQueue {
 template <typename T,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+using EigenVector = phi::EigenVector<T, MajorType, IndexType>;
 
 template <typename T>
 inline void MergeVars(const std::string &var_name,

paddle/fluid/imperative/gradient_accumulator.cc

@@ -209,8 +209,8 @@ void TensorAdd(const VarType& src, VarType* dst) {
 #define TENSOR_ADD_EIGEN(T)                                \
   auto cpu_ctx = static_cast<phi::CPUContext*>(            \
       platform::DeviceContextPool::Instance().Get(place)); \
-  auto in = paddle::framework::EigenVector<T>::Flatten(src_tensor);   \
-  auto out = paddle::framework::EigenVector<T>::Flatten(*dst_tensor); \
+  auto in = phi::EigenVector<T>::Flatten(src_tensor);      \
+  auto out = phi::EigenVector<T>::Flatten(*dst_tensor);    \
   auto& p = *(cpu_ctx->eigen_device());                    \
   out.device(p) = out + in;                                \
   return;

paddle/fluid/operators/detection/anchor_generator_op.h

@@ -70,7 +70,7 @@ class AnchorGeneratorOpKernel : public framework::OpKernel<T> {
     anchors->mutable_data<T>(ctx.GetPlace());
     vars->mutable_data<T>(ctx.GetPlace());
 
-    auto e_anchors = framework::EigenTensor<T, 4>::From(*anchors);
+    auto e_anchors = phi::EigenTensor<T, 4>::From(*anchors);
     for (int h_idx = 0; h_idx < feature_height; ++h_idx) {
       for (int w_idx = 0; w_idx < feature_width; ++w_idx) {
         T x_ctr = (w_idx * stride_width) + offset * (stride_width - 1);
@@ -110,7 +110,7 @@ class AnchorGeneratorOpKernel : public framework::OpKernel<T> {
     var_t.mutable_data<T>(
         phi::make_ddim({1, static_cast<int>(variances.size())}),
         ctx.GetPlace());
-    auto var_et = framework::EigenTensor<T, 2>::From(var_t);
+    auto var_et = phi::EigenTensor<T, 2>::From(var_t);
     for (size_t i = 0; i < variances.size(); ++i) {
       var_et(0, i) = variances[i];
     }
@@ -119,7 +119,7 @@ class AnchorGeneratorOpKernel : public framework::OpKernel<T> {
     auto var_dim = vars->dims();
     vars->Resize({anchor_num, static_cast<int>(variances.size())});
 
-    auto e_vars = framework::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
+    auto e_vars = phi::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
     e_vars = var_et.broadcast(Eigen::DSizes<int, 2>(anchor_num, 1));
 
     vars->Resize(var_dim);

paddle/fluid/operators/detection/density_prior_box_op.h

@@ -66,7 +66,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
 
     auto box_dim = vars->dims();
     boxes->Resize({feature_height, feature_width, num_priors, 4});
-    auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes).setConstant(0.0);
+    auto e_boxes = phi::EigenTensor<T, 4>::From(*boxes).setConstant(0.0);
     int step_average = static_cast<int>((step_width + step_height) * 0.5);
 
     std::vector<float> sqrt_fixed_ratios;
@@ -126,7 +126,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
         phi::make_ddim({1, static_cast<int>(variances.size())}),
         ctx.GetPlace());
 
-    auto var_et = framework::EigenTensor<T, 2>::From(var_t);
+    auto var_et = phi::EigenTensor<T, 2>::From(var_t);
 
     for (size_t i = 0; i < variances.size(); ++i) {
       var_et(0, i) = variances[i];
@@ -136,7 +136,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
     auto var_dim = vars->dims();
     vars->Resize({box_num, static_cast<int>(variances.size())});
 
-    auto e_vars = framework::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
+    auto e_vars = phi::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for collapse(2)
 #endif

paddle/fluid/operators/detection/prior_box_op.h

@@ -182,7 +182,7 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
     var_t.mutable_data<K>(
         phi::make_ddim({1, static_cast<int>(variances.size())}),
         ctx.GetPlace());
-    auto var_et = framework::EigenTensor<K, 2>::From(var_t);
+    auto var_et = phi::EigenTensor<K, 2>::From(var_t);
 
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
@@ -195,7 +195,7 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
     auto var_dim = vars->dims();
     vars->Resize({box_num, static_cast<int>(variances.size())});
 
-    auto e_vars = framework::EigenMatrix<K, Eigen::RowMajor>::From(*vars);
+    auto e_vars = phi::EigenMatrix<K, Eigen::RowMajor>::From(*vars);
 
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for collapse(2)

paddle/fluid/operators/index_select_op.h

@@ -83,8 +83,8 @@ void IndexSelectInner(const framework::ExecutionContext& context,
   input->Resize(phi::make_ddim({outer_nums, input_dim[dim], slice_size}));
   output->Resize(phi::make_ddim({outer_nums, index_size, slice_size}));
 
-  auto input_tensor = framework::EigenTensor<T, 3>::From(*input);
-  auto output_tensor = framework::EigenTensor<T, 3>::From(*output);
+  auto input_tensor = phi::EigenTensor<T, 3>::From(*input);
+  auto output_tensor = phi::EigenTensor<T, 3>::From(*output);
 
   auto& place =
       *context.template device_context<DeviceContext>().eigen_device();

paddle/fluid/operators/interpolate_op.h

@@ -25,7 +25,7 @@ template <typename T,
           size_t D,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+using EigenTensor = phi::EigenTensor<T, D, MajorType, IndexType>;
 using Tensor = phi::DenseTensor;
 using DataLayout = phi::DataLayout;
 

paddle/fluid/operators/math/sequence_pooling.cc

@@ -29,11 +29,11 @@ using LoDTensor = phi::DenseTensor;
 template <typename T,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+using EigenVector = phi::EigenVector<T, MajorType, IndexType>;
 template <typename T,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+using EigenMatrix = phi::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename T, bool is_test>
 class MaxSeqPoolFunctor {

paddle/fluid/operators/sequence_ops/sequence_expand_op.h

@@ -26,7 +26,7 @@ using LoDTensor = phi::DenseTensor;
 template <typename T,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+using EigenMatrix = phi::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename DeviceContext, typename T>
 struct SequenceExpandFunctor {

paddle/phi/kernels/cpu/norm_grad_kernel.cc

@@ -45,10 +45,10 @@ void NormGradKernel(const Context& ctx,
 
   auto* place = ctx.eigen_device();
 
-  auto x_e = paddle::framework::EigenVector<T>::Flatten(*in_x);
-  auto dy_e = paddle::framework::EigenVector<T>::Flatten(*in_dy);
-  auto norm_e = paddle::framework::EigenVector<T>::Flatten(*in_norm);
-  auto dx_e = paddle::framework::EigenVector<T>::Flatten(*out_dx);
+  auto x_e = phi::EigenVector<T>::Flatten(*in_x);
+  auto dy_e = phi::EigenVector<T>::Flatten(*in_dy);
+  auto norm_e = phi::EigenVector<T>::Flatten(*in_norm);
+  auto dx_e = phi::EigenVector<T>::Flatten(*out_dx);
 
   Eigen::DSizes<int, 3> shape(pre, n, post);
   Eigen::DSizes<int, 3> rshape(pre, 1, post);
@@ -60,7 +60,7 @@ void NormGradKernel(const Context& ctx,
   DenseTensor rsum;
   rsum.Resize({pre, post});
   ctx.template Alloc<T>(&rsum);
-  auto sum = paddle::framework::EigenTensor<T, 2>::From(rsum);
+  auto sum = phi::EigenTensor<T, 2>::From(rsum);
 
   Eigen::DSizes<int, 1> rdim(1);
   Eigen::DSizes<int, 3> bcast(1, n, 1);

paddle/phi/kernels/cpu/norm_kernel.cc

@@ -55,9 +55,9 @@ void NormKernel(const Context& ctx,
   Eigen::DSizes<int, 3> shape(pre, n, post);
   Eigen::DSizes<int, 2> norm_shape(pre, post);
 
-  auto x_e = paddle::framework::EigenVector<T>::Flatten(x);
-  auto y_e = paddle::framework::EigenVector<T>::Flatten(*out);
-  auto norm_e = paddle::framework::EigenVector<T>::Flatten(*out_norm);
+  auto x_e = phi::EigenVector<T>::Flatten(x);
+  auto y_e = phi::EigenVector<T>::Flatten(*out);
+  auto norm_e = phi::EigenVector<T>::Flatten(*out_norm);
   auto x_r = x_e.reshape(shape);
   auto y = y_e.reshape(shape);
   auto norm_reshape = norm_e.reshape(norm_shape);

paddle/phi/kernels/funcs/detail/gru_cpu_kernel.h

@@ -15,9 +15,9 @@ limitations under the License. */
 #pragma once
 #include <type_traits>
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/phi/kernels/funcs/activation_functor.h"
 #include "paddle/phi/kernels/funcs/detail/activation_functions.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
 #include "paddle/phi/kernels/funcs/gru_compute.h"
 
 namespace phi {
@@ -27,7 +27,7 @@ using Array1 = Eigen::DSizes<int64_t, 1>;
 template <typename T,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenVector = paddle::framework::EigenVector<T, MajorType, IndexType>;
+using EigenVector = phi::EigenVector<T, MajorType, IndexType>;
 
 #if !defined(__NVCC__) && !defined(__HIPCC___)  // @{ Group for GRU CPU
 template <class OpResetOutput, typename T>

paddle/phi/kernels/funcs/detail/lstm_cpu_kernel.h

@@ -15,9 +15,9 @@ limitations under the License. */
 #pragma once
 #include <type_traits>
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/phi/kernels/funcs/activation_functor.h"
 #include "paddle/phi/kernels/funcs/detail/activation_functions.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
 #include "paddle/phi/kernels/funcs/lstm_compute.h"
 
 #if defined(_WIN32)
@@ -34,7 +34,7 @@ using Array1 = Eigen::DSizes<int64_t, 1>;
 template <typename T,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenVector = paddle::framework::EigenVector<T, MajorType, IndexType>;
+using EigenVector = phi::EigenVector<T, MajorType, IndexType>;
 
 #if !defined(__NVCC__) && !defined(__HIPCC___)  // @{ Group LSTM CPU
 

paddle/phi/kernels/funcs/math_function.cc

@@ -299,9 +299,9 @@ struct RowwiseAdd<phi::CPUContext, T> {
                           in_dims_cstr,
                           out_dims_cstr));
 
-    auto in = paddle::framework::EigenMatrix<T>::From(input);
-    auto vec = paddle::framework::EigenVector<T>::Flatten(vector);
-    auto out = paddle::framework::EigenMatrix<T>::From(*output);
+    auto in = phi::EigenMatrix<T>::From(input);
+    auto vec = phi::EigenVector<T>::Flatten(vector);
+    auto out = phi::EigenMatrix<T>::From(*output);
 
     for (int64_t i = 0; i < in_dims[0]; ++i) {
       out.chip(i, 0) = in.chip(i, 0) + vec;

paddle/phi/kernels/funcs/math_function.h

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/fluid/framework/convert_utils.h"
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
@@ -26,6 +25,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/utils/data_type.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
 
 namespace phi {
 namespace funcs {

paddle/phi/kernels/funcs/math_function_impl.h

@@ -22,13 +22,13 @@ limitations under the License. */
 namespace phi {
 namespace funcs {
 
-using paddle::framework::To32BitIndex;
+using phi::To32BitIndex;
 
 template <typename DeviceContext, typename T>
 void SetConstant<DeviceContext, T>::operator()(const DeviceContext& context,
                                                phi::DenseTensor* tensor,
                                                T num) {
-  auto t = paddle::framework::EigenVector<T>::Flatten(*tensor);
+  auto t = phi::EigenVector<T>::Flatten(*tensor);
   t.device(*context.eigen_device()) = t.constant(static_cast<T>(num));
 }
 
@@ -60,8 +60,8 @@ void Transpose<DeviceContext, T, Rank>::operator()(
   for (int i = 0; i < Rank; i++) {
     permute[i] = axis[i];
   }
-  auto eigen_in = paddle::framework::EigenTensor<T, Rank>::From(in);
-  auto eigen_out = paddle::framework::EigenTensor<T, Rank>::From(*out);
+  auto eigen_in = phi::EigenTensor<T, Rank>::From(in);
+  auto eigen_out = phi::EigenTensor<T, Rank>::From(*out);
   auto* dev = context.eigen_device();
   // use 32bit index to speed up computation
   bool use_32bit_index = eigen_out.size() < Eigen::NumTraits<int>::highest();
@@ -89,8 +89,8 @@ void ColwiseSum<DeviceContext, T>::operator()(const DeviceContext& context,
                         size,
                         out->numel()));
 
-  auto in = paddle::framework::EigenMatrix<T>::From(input);
-  auto vec = paddle::framework::EigenVector<T>::Flatten(*out);
+  auto in = phi::EigenMatrix<T>::From(input);
+  auto vec = phi::EigenVector<T>::Flatten(*out);
 
   vec.device(*context.eigen_device()) = in.sum(Eigen::array<int, 1>({{0}}));
 }
@@ -151,8 +151,8 @@ void RowwiseMean<DeviceContext, T>::operator()(const DeviceContext& context,
                         in_dims[0],
                         out->numel()));
 
-  auto in = paddle::framework::EigenMatrix<T>::From(input);
-  auto vec = paddle::framework::EigenVector<T>::Flatten(*out);
+  auto in = phi::EigenMatrix<T>::From(input);
+  auto vec = phi::EigenVector<T>::Flatten(*out);
 
   vec.device(*context.eigen_device()) = in.mean(Eigen::array<int, 1>({{1}}));
 }
@@ -217,8 +217,8 @@ void RowwiseSum<DeviceContext, T>::operator()(const DeviceContext& context,
                         in_dims[0],
                         out->numel()));
 
-  auto in = paddle::framework::EigenMatrix<T>::From(input);
-  auto vec = paddle::framework::EigenVector<T>::Flatten(*out);
+  auto in = phi::EigenMatrix<T>::From(input);
+  auto vec = phi::EigenVector<T>::Flatten(*out);
 
   vec.device(*context.eigen_device()) = in.sum(Eigen::array<int, 1>({{1}}));
 }

paddle/phi/kernels/funcs/sequence2batch.h

@@ -16,10 +16,10 @@ limitations under the License. */
 #include <algorithm>
 #include <vector>
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/platform/device_context.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
 
 namespace phi {
 namespace funcs {
@@ -27,7 +27,7 @@ namespace funcs {
 template <typename T,
           int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = paddle::framework::EigenMatrix<T, MajorType, IndexType>;
+using EigenMatrix = phi::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename DeviceContext, typename T>
 class CopyMatrixRowsFunctor {

paddle/phi/kernels/gpudnn/conv_grad_kernel.cu

@@ -14,7 +14,6 @@
 
 #include "paddle/phi/kernels/conv_grad_kernel.h"
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"

paddle/phi/kernels/impl/conv_cudnn_impl.h

@@ -14,7 +14,6 @@
 
 #pragma once
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"

paddle/phi/kernels/impl/segment_pool_grad_kernel_impl.h

@@ -16,7 +16,6 @@
 
 #include <string>
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/phi/common/place.h"
 #include "paddle/phi/kernels/funcs/math_function.h"
 #include "paddle/phi/kernels/funcs/segment_pooling.h"

paddle/phi/kernels/impl/segment_pool_kernel_impl.h

@@ -16,7 +16,6 @@
 
 #include <string>
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/phi/common/place.h"
 #include "paddle/phi/kernels/funcs/math_function.h"
 #include "paddle/phi/kernels/funcs/segment_pooling.h"