From e97c1a8ca04bdbfe8906e74f9433afad58fa2d7f Mon Sep 17 00:00:00 2001
From: chengduoZH <zhaochengduo@163.com>
Date: Thu, 3 May 2018 12:58:32 +0800
Subject: [PATCH] fix __shfl

---
 paddle/fluid/platform/cuda_device_function.h | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h
index 7cfeaab35b..2405f33d4f 100644
--- a/paddle/fluid/platform/cuda_device_function.h
+++ b/paddle/fluid/platform/cuda_device_function.h
@@ -35,6 +35,16 @@ __forceinline__ __device__ T __shfl_sync(unsigned, T val, int src_line,
 #define FULL_WARP_MASK 0xFFFFFFFF
 #define CREATE_SHFL_MASK(mask, predicate) \
   mask = __ballot_sync(FULL_WARP_MASK, (predicate))
+template <typename T>
+__forceinline__ __device__ T __shfl_down_sync(unsigned mask, T val, int delta) {
+  return __shfl_down_sync(mask, val, delta);
+}
+
+template <typename T>
+__forceinline__ __device__ T __shfl_sync(unsigned mask, T val, int src_line,
+                                         int width) {
+  return __shfl_sync(mask, val, src_line, width);
+}
 #endif
 
 template <typename T>
-- 
GitLab