From 9b0eae3023e3faf6a40a69f5ff79bcc2303c674b Mon Sep 17 00:00:00 2001
From: Jacek Czaja <jacek.czaja@intel.com>
Date: Sun, 18 Nov 2018 13:27:17 +0100
Subject: [PATCH] - Removing partial specialization of sotmax for inference for
 GPU

test=develop
---
 paddle/fluid/operators/math/softmax.h      |  3 ++-
 paddle/fluid/operators/math/softmax_impl.h | 10 +++++++---
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/paddle/fluid/operators/math/softmax.h b/paddle/fluid/operators/math/softmax.h
index bf698dc2f7..089458e957 100644
--- a/paddle/fluid/operators/math/softmax.h
+++ b/paddle/fluid/operators/math/softmax.h
@@ -19,7 +19,8 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename DeviceContext, typename T, bool is_test>
+template <typename DeviceContext, typename T, bool is_test,
+          typename Enable = void>
 class SoftmaxFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor* X,
diff --git a/paddle/fluid/operators/math/softmax_impl.h b/paddle/fluid/operators/math/softmax_impl.h
index e09a243347..0f3e5b2008 100644
--- a/paddle/fluid/operators/math/softmax_impl.h
+++ b/paddle/fluid/operators/math/softmax_impl.h
@@ -33,8 +33,8 @@ struct ValueClip {
   }
 };
 
-template <typename DeviceContext, typename T, bool is_test>
-void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(
+template <typename DeviceContext, typename T, bool is_test, typename Enable>
+void SoftmaxFunctor<DeviceContext, T, is_test, Enable>::operator()(
     const DeviceContext& context, const framework::Tensor* X,
     framework::Tensor* Y) {
   auto logits = EigenMatrix<T>::From(*X);
@@ -66,8 +66,12 @@ void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(
                                                  .broadcast(one_by_class));
 }
 
+template <class DeviceContext>
+using enable_if_CPU = typename std::enable_if<
+    std::is_same<DeviceContext, platform::CPUDeviceContext>::value>::type;
+
 template <typename DeviceContext>
-class SoftmaxFunctor<DeviceContext, float, true> {
+class SoftmaxFunctor<DeviceContext, float, true, enable_if_CPU<DeviceContext>> {
   void operator()(const DeviceContext& context, const framework::Tensor* X,
                   framework::Tensor* Y) {
     auto in_dims = X->dims();
-- 
GitLab