diff options
| author | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2016-10-02 15:48:39 +0300 |
|---|---|---|
| committer | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2016-10-03 23:15:25 +0300 |
| commit | a3abb020e37a072eb71fd30de9ab125d1c16623a (patch) | |
| tree | b525be7f8a0792eedecb2b95802ede88dc3f330e /intern/cycles/util | |
| parent | 49ad4215baf16d850d0e367f003ab688e4a3d08e (diff) | |
Fix Cycles CUDA performance on CUDA 8.0.
Mostly this is making inlining match CUDA 7.5 in a few performance critical
places. The end result is that performance is now better than before, possibly
due to less register spilling or other CUDA 8.0 compiler improvements.
On benchmarks scenes, there are 3% to 35% render time reductions. Stack memory
usage is reduced a little too.
Reviewed By: sergey
Differential Revision: https://developer.blender.org/D2269
Diffstat (limited to 'intern/cycles/util')
| -rw-r--r-- | intern/cycles/util/util_types.h | 2 |
1 files changed, 2 insertions, 0 deletions
diff --git a/intern/cycles/util/util_types.h b/intern/cycles/util/util_types.h index 257c6ad7491..6af65f88a02 100644 --- a/intern/cycles/util/util_types.h +++ b/intern/cycles/util/util_types.h @@ -42,6 +42,7 @@ #if defined(_WIN32) && !defined(FREE_WINDOWS) #define ccl_device_inline static __forceinline +#define ccl_device_forceinline static __forceinline #define ccl_align(...) __declspec(align(__VA_ARGS__)) #ifdef __KERNEL_64_BIT__ #define ccl_try_align(...) __declspec(align(__VA_ARGS__)) @@ -56,6 +57,7 @@ #else #define ccl_device_inline static inline __attribute__((always_inline)) +#define ccl_device_forceinline static inline __attribute__((always_inline)) #define ccl_align(...) __attribute__((aligned(__VA_ARGS__))) #ifndef FREE_WINDOWS64 #define __forceinline inline __attribute__((always_inline)) |