From 133bac2b10de13d11424e52a6bbe935817cde083 Mon Sep 17 00:00:00 2001
From: minqiyang <minqiyang@baidu.com>
Date: Wed, 24 Oct 2018 17:44:02 +0800
Subject: [PATCH] Accelerate embedding op grad

test=develop
---
 paddle/fluid/operators/lookup_table_op.h | 26 ++++++++----------------
 1 file changed, 8 insertions(+), 18 deletions(-)

diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h
index 58463dc4d6f..eac6224d106 100644
--- a/paddle/fluid/operators/lookup_table_op.h
+++ b/paddle/fluid/operators/lookup_table_op.h
@@ -68,6 +68,7 @@ class LookupTableKernel : public framework::OpKernel<T> {
       const auto *table = table_t.value().data<T>();
       auto *output = output_t->mutable_data<T>(context.GetPlace());
 
+      auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
       for (int64_t i = 0; i < ids_numel; ++i) {
         if (padding_idx != kNoPadding && ids[i] == padding_idx) {
           memset(output + i * row_width, 0, row_width * sizeof(T));
@@ -75,8 +76,8 @@ class LookupTableKernel : public framework::OpKernel<T> {
           PADDLE_ENFORCE_GE(ids[i], 0);
           auto id_index = table_t.Index(ids[i]);
           PADDLE_ENFORCE_GE(id_index, 0, "the input key should be exists.");
-          memcpy(output + i * row_width, table + id_index * row_width,
-                 row_width * sizeof(T));
+          blas.VCOPY(row_width, table + id_index * row_width,
+                     output + i * row_width);
         }
       }
     }
@@ -111,27 +112,16 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
       auto *ids_data = ids->data<int64_t>();
       int64_t ids_num = ids->numel();
 
-      framework::Vector<int64_t> new_rows;
+      std::vector<int64_t> new_rows;
       new_rows.reserve(ids_num);
-      for (int64_t i = 0; i < ids_num; i++) {
-        new_rows.push_back(ids_data[i]);
-      }
+      std::memcpy(new_rows.data(), ids_data, ids_num * sizeof(int64_t));
       d_table->set_rows(new_rows);
 
       auto *d_table_value = d_table->mutable_value();
       d_table_value->Resize({ids_num, table_dim[1]});
-      d_table_value->mutable_data<T>(context.GetPlace());
-
-      d_table->set_height(table_dim[0]);
-
-      auto *d_output_data = d_output->data<T>();
-      auto *d_table_data = d_table_value->data<T>();
-
-      auto d_output_dims = d_output->dims();
-      PADDLE_ENFORCE_EQ(
-          d_table_value->dims(),
-          framework::flatten_to_2d(d_output_dims, d_output_dims.size() - 1));
-      memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel());
+      // memory optimization will NOT reuse Tensor with SelectedRows
+      // so we could just share the tensor here directly.
+      d_table_value->ShareDataWith(*d_output);
     } else {
       auto *ids = context.Input<LoDTensor>("Ids");
       auto *d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
-- 
GitLab