From b2d0ee5159bb824e07b30dde12edaf941f8036cb Mon Sep 17 00:00:00 2001
From: sneaxiy <sneaxiy@126.com>
Date: Mon, 6 Aug 2018 09:42:56 +0000
Subject: [PATCH] refine elementwise_add op

---
 paddle/fluid/operators/elementwise_add_op.cu | 54 ++++++++++++++++++++
 paddle/fluid/operators/elementwise_add_op.h  | 29 +++++++++--
 2 files changed, 78 insertions(+), 5 deletions(-)

diff --git a/paddle/fluid/operators/elementwise_add_op.cu b/paddle/fluid/operators/elementwise_add_op.cu
index dfff518f170..6cbf6066c92 100644
--- a/paddle/fluid/operators/elementwise_add_op.cu
+++ b/paddle/fluid/operators/elementwise_add_op.cu
@@ -16,6 +16,60 @@ limitations under the License. */
 #include "paddle/fluid/operators/elementwise_add_op.h"
 #include "paddle/fluid/platform/float16.h"
 
+namespace paddle {
+namespace operators {
+
+template <typename T>
+__global__ void ElementwiseAddCUDAKernel(const T *x, const T *y, T *z, int n,
+                                         int post, int size) {
+  int idx_x = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx_x < size) {
+    int idx_y = idx_x / post - (idx_x / (n * post)) * n;
+    z[idx_x] = x[idx_x] + y[idx_y];
+  }
+}
+
+template <typename T>
+class ElementwiseAddKernel<platform::CUDADeviceContext, T>
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    using Tensor = framework::Tensor;
+
+    const auto x = ctx.Input<Tensor>("X");
+    const auto y = ctx.Input<Tensor>("Y");
+    auto z = ctx.Output<Tensor>("Out");
+    auto *z_data = z->mutable_data<T>(ctx.GetPlace());
+
+    auto &device = *(ctx.cuda_device_context().eigen_device());
+    const framework::DDim &x_dim = x->dims();
+    framework::DDim y_dim = y->dims();
+    int size = x->numel();
+    if (x_dim == y_dim) {
+      auto dim = framework::make_ddim({size});
+      auto z_eigen = framework::EigenTensor<T, 1>::From(*z, dim);
+      auto x_eigen = framework::EigenTensor<T, 1>::From(*x, dim);
+      auto y_eigen = framework::EigenTensor<T, 1>::From(*y, dim);
+      z_eigen.device(device) = x_eigen + y_eigen;
+    } else {
+      int axis = ctx.Attr<int>("axis");
+      axis = (axis == -1 ? x_dim.size() - y_dim.size() : axis);
+      y_dim = trim_trailing_singular_dims(y_dim);
+      axis = (y_dim.size() == 0) ? x_dim.size() : axis;
+      int pre, n, post;
+      get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);
+      int threads = 512;
+      int grids = (size + threads - 1) / threads;
+      auto stream = ctx.cuda_device_context().stream();
+      ElementwiseAddCUDAKernel<T><<<grids, threads, 0, stream>>>(
+          x->data<T>(), y->data<T>(), z_data, n, post, size);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
diff --git a/paddle/fluid/operators/elementwise_add_op.h b/paddle/fluid/operators/elementwise_add_op.h
index baf04c30b17..ccda870618f 100644
--- a/paddle/fluid/operators/elementwise_add_op.h
+++ b/paddle/fluid/operators/elementwise_add_op.h
@@ -142,16 +142,35 @@ class ElementwiseAddGradKernel : public framework::OpKernel<T> {
 
     auto* x = ctx.Input<Tensor>("X");
     auto* y = ctx.Input<Tensor>("Y");
-    auto* out = ctx.Input<Tensor>("Out");
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
     auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
 
-    if (platform::is_cpu_place(ctx.GetPlace()) && (x->dims() == y->dims())) {
-      elementwise_add_grad<DeviceContext, T>(ctx, x, y, out, dout, dx, dy);
+    if (dx != nullptr) dx->ShareDataWith(*dout);
+    if (dy == nullptr) return;
+
+    if (x->dims() == y->dims()) {
+      dy->ShareDataWith(*dout);
     } else {
-      default_elementwise_add_grad<DeviceContext, T>(ctx, x, y, out, dout, dx,
-                                                     dy);
+      dy->mutable_data<T>(ctx.GetPlace());
+      // Perform reduction to dout to calculate dy
+      const framework::DDim& x_dim = x->dims();
+      framework::DDim y_dim = y->dims();
+      int axis = ctx.Attr<int>("axis");
+      axis = (axis == -1 ? x_dim.size() - y_dim.size() : axis);
+      y_dim = trim_trailing_singular_dims(y_dim);
+      axis = (y_dim.size() == 0) ? x_dim.size() : axis;
+
+      auto* device =
+          ctx.template device_context<DeviceContext>().eigen_device();
+      int pre, n, post;
+      get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);
+      auto eigen_dout = framework::EigenTensor<T, 3>::From(
+          *dout, framework::make_ddim({pre, n, post}));
+      auto eigen_dy =
+          framework::EigenTensor<T, 1>::From(*dy, framework::make_ddim({n}));
+      eigen_dy.device(*device) = eigen_dout.sum(
+          framework::EigenDim<2>::From(framework::make_ddim({0, 2})));
     }
   }
 };
-- 
GitLab