From 67308822f85a387433867ea330624b9c16ae029c Mon Sep 17 00:00:00 2001
From: minqiyang <minqiyang@baidu.com>
Date: Sun, 30 Sep 2018 19:45:05 +0800
Subject: [PATCH] Add selected_rows merge for clip_by_norm op

test=develop
---
 paddle/fluid/operators/CMakeLists.txt    |  3 ++-
 paddle/fluid/operators/clip_by_norm_op.h | 24 +++++++++++++++++++++++-
 2 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index b61bca8c3d1..e10fc422fac 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -229,7 +229,7 @@ if(WITH_DISTRIBUTE)
         op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS})
         set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     endforeach()
-    
+
     #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
     #        listen_and_serv_op sum_op executor SERIAL)
@@ -267,6 +267,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
 else()
     set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op)
 endif()
+op_library(clip_by_norm_op DEPS selected_rows_functor)
 op_library(sum_op DEPS selected_rows_functor)
 op_library(sgd_op DEPS selected_rows_functor)
 op_library(print_op DEPS lod_tensor)
diff --git a/paddle/fluid/operators/clip_by_norm_op.h b/paddle/fluid/operators/clip_by_norm_op.h
index 5af0eb0b2ad..83461159134 100644
--- a/paddle/fluid/operators/clip_by_norm_op.h
+++ b/paddle/fluid/operators/clip_by_norm_op.h
@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/platform/transform.h"
 
 namespace paddle {
@@ -31,10 +32,31 @@ class ClipByNormKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto max_norm = context.Attr<T>("max_norm");
-    auto* input = context.Input<Tensor>("X");
+    auto in_var = context.InputVar("X");
     auto* output = context.Output<Tensor>("Out");
     output->mutable_data<T>(context.GetPlace());
 
+    const Tensor* input = nullptr;
+    if (in_var->IsType<framework::LoDTensor>()) {
+      input = context.Input<Tensor>("X");
+    } else if (in_var->IsType<framework::SelectedRows>()) {
+      auto* x = context.Input<framework::SelectedRows>("X");
+
+      // merge ids in selected rows first
+      math::scatter::MergeAdd<DeviceContext, T> merge_func;
+      auto* merged_input = const_cast<framework::Scope&>(context.scope())
+                               .Var()
+                               ->GetMutable<framework::SelectedRows>();
+      merge_func(context.template device_context<DeviceContext>(), *x,
+                 merged_input);
+      input = &(merged_input->value());
+    } else {
+      PADDLE_THROW("Unexpected branch, input variable type is %s",
+                   in_var->Type().name());
+    }
+
+    PADDLE_ENFORCE_NOT_NULL(input);
+
     auto x = EigenVector<T>::Flatten(*input);
     auto out = EigenVector<T>::Flatten(*output);
     auto x_norm = x.square().sum().sqrt();
-- 
GitLab