Merge pull request #15611 from guoshengCS/fix-beam-search-size

To make CUDA_LAUNCH_KERNEL_HELPER support large size.
6 years ago · 430e5590e4
parent 46a6cac91f b6c3b69af8
commit 430e5590e4
1 changed files with 6 additions and 4 deletions
--- a/paddle/fluid/platform/cuda_device_function.h
+++ b/paddle/fluid/platform/cuda_device_function.h
@ -53,10 +53,12 @@ inline static int RoundToPowerOfTwo(int dim) {
    __VA_ARGS__;                           \
  } break

-#define CUDA_LAUNCH_KERNEL_HELPER(...)         \
-  CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__); \
-  CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__); \
-  CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__);  \
+#define CUDA_LAUNCH_KERNEL_HELPER(...)          \
+  CUDA_LAUNCH_KERNEL_BASE(1024, ##__VA_ARGS__); \
+  CUDA_LAUNCH_KERNEL_BASE(512, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__);   \
  CUDA_LAUNCH_KERNEL_BASE(32, ##__VA_ARGS__);

 template <typename T>