From 14cfb8c262c1f16c8916087c8dc4ce2d16500c7e Mon Sep 17 00:00:00 2001
From: qijun <qijun1994@hotmail.com>
Date: Wed, 19 Jul 2017 08:22:21 +0000
Subject: [PATCH] fix gpu build error

---
 cmake/flags.cmake                 |  1 +
 paddle/operators/mul_op.h         |  7 ++++---
 paddle/operators/rowwise_add_op.h |  1 +
 paddle/operators/softmax_op.h     | 21 ++++++++++++---------
 4 files changed, 18 insertions(+), 12 deletions(-)

diff --git a/cmake/flags.cmake b/cmake/flags.cmake
index c31e62fc08..34fd348893 100644
--- a/cmake/flags.cmake
+++ b/cmake/flags.cmake
@@ -124,6 +124,7 @@ set(GPU_COMMON_FLAGS
     -Wno-error=literal-suffix
     -Wno-error=unused-local-typedefs
     -Wno-error=unused-function  # Warnings in Numpy Header.
+    -Wno-error=array-bounds # Warnings in Eigen::array
 )
 
 if (APPLE)
diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h
index 13e5b6a950..81d5953cf0 100644
--- a/paddle/operators/mul_op.h
+++ b/paddle/operators/mul_op.h
@@ -24,9 +24,10 @@ template <typename Place, typename T>
 class MulKernel : public framework::OpKernel {
 public:
   void Compute(const framework::KernelContext& context) const override {
-    Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair;
-    dim_pair[0].first = 1;
-    dim_pair[0].second = 0;
+    Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair = {
+        Eigen::IndexPair<Eigen::DenseIndex>(1, 0)};
+    // dim_pair[0].first = 1;
+    // dim_pair[0].second = 0;
 
     auto input0 = context.Input(0)->Get<framework::Tensor>();
     auto input1 = context.Input(1)->Get<framework::Tensor>();
diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h
index f1d43002dc..dd5cde0c5d 100644
--- a/paddle/operators/rowwise_add_op.h
+++ b/paddle/operators/rowwise_add_op.h
@@ -26,6 +26,7 @@ public:
     auto in0 = context.Input(0)->Get<framework::Tensor>();
     auto in1 = context.Input(1)->Get<framework::Tensor>();
     auto* out = context.Output(0)->GetMutable<framework::Tensor>();
+    out->mutable_data<T>(context.GetPlace());
 
     auto input = in0.matrix<T>();
     auto bias = in1.vec<T>();
diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h
index 34a6c299bb..6d675ea5f6 100644
--- a/paddle/operators/softmax_op.h
+++ b/paddle/operators/softmax_op.h
@@ -26,6 +26,7 @@ public:
   void Compute(const framework::KernelContext& context) const override {
     auto input = context.Input(0)->Get<framework::Tensor>();
     auto* output = context.Output(0)->GetMutable<framework::Tensor>();
+    output->mutable_data<T>(context.GetPlace());
 
     auto logits = input.matrix<T>();
     auto softmax = output->matrix<T>();
@@ -40,19 +41,21 @@ public:
     Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
     Eigen::DSizes<int, 2> one_by_class(1, num_classes);
 
-    auto shifted_logits = (logits - logits.maximum(along_class)
-                                        .eval()
-                                        .reshape(batch_by_one)
-                                        .broadcast(one_by_class));
+    auto shifted_logits = (logits -
+                           logits.maximum(along_class)
+                               .eval()
+                               .reshape(batch_by_one)
+                               .broadcast(one_by_class));
 
     softmax.device(*(context.GetEigenDevice<Place>())) = shifted_logits.exp();
 
     softmax.device(*(context.GetEigenDevice<Place>())) =
-        (softmax * softmax.sum(along_class)
-                       .inverse()
-                       .eval()
-                       .reshape(batch_by_one)
-                       .broadcast(one_by_class));
+        (softmax *
+         softmax.sum(along_class)
+             .inverse()
+             .eval()
+             .reshape(batch_by_one)
+             .broadcast(one_by_class));
   }
 };
 }  // namespace operators
-- 
GitLab