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

To make CUDA_LAUNCH_KERNEL_HELPER support large size.

Merge pull request #15611 from guoshengCS/fix-beam-search-size
To make CUDA_LAUNCH_KERNEL_HELPER support large size.
430e5590 · Guo Sheng · GitHub · 46a6cac9 · b6c3b69a · 430e5590
隐藏空白更改
内联并排

Showing with 6 addition and 4 deletion

paddle/fluid/platform/cuda_device_function.h paddle/fluid/platform/cuda_device_function.h +6 -4

未找到文件。
--- 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(...)         \
+#define CUDA_LAUNCH_KERNEL_HELPER(...)          \
-  CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__); \
+  CUDA_LAUNCH_KERNEL_BASE(1024, ##__VA_ARGS__); \
-  CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__); \
+  CUDA_LAUNCH_KERNEL_BASE(512, ##__VA_ARGS__);  \
-  CUDA_LAUNCH_KERNEL_BASE(64, ##__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>