remove eval() calls in Eigen, test=develop (#21498)

paddle/fluid/operators/batch_norm_op.cc

@@ -522,7 +522,7 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
       EigenVectorArrayMap<T> inv_var_tmp(running_inv_var_data, C);
       ConstEigenVectorArrayMap<T> var_arr(running_variance->data<T>(), C);
 
-      inv_var_tmp = (var_arr + epsilon).sqrt().inverse().eval();
+      inv_var_tmp = (var_arr + epsilon).sqrt().inverse();
       inv_var_data = running_inv_var_data;
     }
 

paddle/fluid/operators/bilinear_tensor_product_op.h

@@ -70,7 +70,7 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
     if (bias) {
       auto bias_vec = EigenMatrix<T>::From(*bias);
       Eigen::DSizes<int, 2> bcast(batch_size, 1);
-      output_mat.device(place) = bias_vec.broadcast(bcast).eval() + output_mat;
+      output_mat.device(place) = bias_vec.broadcast(bcast) + output_mat;
     }
   }
 };
@@ -143,8 +143,7 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
         if (d_x) {
           y_scale_mat.device(place) =
               output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
-                  .broadcast(bcast_for_x)
-                  .eval() *
+                  .broadcast(bcast_for_x) *
               y_mat;
           blas.GEMM(CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1,
                     y_scale.data<T>(), weight_i.data<T>(), 1, d_x->data<T>());
@@ -153,8 +152,7 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
         if (d_y || d_weight) {
           auto output_vec_y =
               output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
-                  .broadcast(bcast_for_y)
-                  .eval();
+                  .broadcast(bcast_for_y);
           x_scale_mat.device(place) = output_vec_y * x_mat;
           if (d_y) {
             blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1,

paddle/fluid/operators/clip_by_norm_op.h

@@ -75,7 +75,7 @@ class ClipByNormKernel : public framework::OpKernel<T> {
     auto& place =
         *context.template device_context<DeviceContext>().eigen_device();
 
-    auto temp = (x_norm <= max_norm).template cast<T>().eval();
+    auto temp = (x_norm <= max_norm).template cast<T>();
     auto scaling = temp + (static_cast<T>(1) - temp) * max_norm / x_norm;
     Eigen::array<int, 1> one_dim{{1}};
     Eigen::DSizes<int, 1> m_dsize(input->numel());

paddle/fluid/operators/instance_norm_op.cc

@@ -491,7 +491,7 @@ class InstanceNormDoubleGradKernel<platform::CPUDeviceContext, T>
                   sample_size * inv_var_tile_data * inv_var_tile_data *
                   (ddx_arr.colwise().sum() / sample_size - ddx_arr);
 
-        dx_arr = scale_tile_data * dx_arr.eval();
+        dx_arr = scale_tile_data * dx_arr;
       }
       if (ddScale) {
         ConstEigenVectorArrayMap<T> ddscale_arr(ddScale->data<T>(), C);
@@ -532,7 +532,7 @@ class InstanceNormDoubleGradKernel<platform::CPUDeviceContext, T>
              x_sub_mean_mul_invstd_arr *
                  (dy_arr * x_sub_mean_mul_invstd_arr).colwise().sum() /
                  sample_size);
-        first_grad_arr = first_grad_arr.eval() * ddx_arr;
+        first_grad_arr = first_grad_arr * ddx_arr;
         for (int nc = 0; nc < NxC; ++nc) {
           int c = nc % C;
           dscale_arr(c) += first_grad_arr.colwise().sum()(nc);

paddle/fluid/operators/kldiv_loss_op.h

@@ -71,7 +71,7 @@ class KLDivLossKernel : public framework::OpKernel<T> {
     if ("none" == reduction) {
       loss_t.device(place) = output;
     } else if ("batchmean" == reduction) {
-      auto output_sum = output.sum().eval();
+      auto output_sum = output.sum();
       loss_t.device(place) = output_sum / output_sum.constant(n);
     } else if ("mean" == reduction) {
       loss_t.device(place) = output.mean();

paddle/fluid/operators/reduce_ops/reduce_sum_op.h

@@ -90,7 +90,7 @@ struct SumGradFunctor {
             typename DY, typename Dim>
   void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy,
                   const Dim& dim, int size) {
-    dx->device(place) = dy->eval().broadcast(dim);
+    dx->device(place) = dy->broadcast(dim);
   }
 };