diff options
Diffstat (limited to 'src/kernels/common.opencl')
| -rw-r--r-- | src/kernels/common.opencl | 19 |
1 files changed, 14 insertions, 5 deletions
diff --git a/src/kernels/common.opencl b/src/kernels/common.opencl index db4c8ec4..9481881e 100644 --- a/src/kernels/common.opencl +++ b/src/kernels/common.opencl @@ -235,6 +235,15 @@ R"( // ================================================================================================= +// Force inlining functions or not: some compilers don't support the inline keyword +#ifdef USE_INLINE_KEYWORD + #define INLINE_FUNC inline +#else + #define INLINE_FUNC +#endif + +// ================================================================================================= + // Shuffled workgroup indices to avoid partition camping, see below. For specific devices, this is // enabled (see src/routine.cc). #ifndef USE_STAGGERED_INDICES @@ -245,18 +254,18 @@ R"( // http://docs.nvidia.com/cuda/samples/6_Advanced/transpose/doc/MatrixTranspose.pdf // More details: https://github.com/CNugteren/CLBlast/issues/53 #if USE_STAGGERED_INDICES == 1 - inline size_t GetGroupIDFlat() { + INLINE_FUNC size_t GetGroupIDFlat() { return get_group_id(0) + get_num_groups(0) * get_group_id(1); } - inline size_t GetGroupID1() { + INLINE_FUNC size_t GetGroupID1() { return (GetGroupIDFlat()) % get_num_groups(1); } - inline size_t GetGroupID0() { + INLINE_FUNC size_t GetGroupID0() { return ((GetGroupIDFlat() / get_num_groups(1)) + GetGroupID1()) % get_num_groups(0); } #else - inline size_t GetGroupID1() { return get_group_id(1); } - inline size_t GetGroupID0() { return get_group_id(0); } + INLINE_FUNC size_t GetGroupID1() { return get_group_id(1); } + INLINE_FUNC size_t GetGroupID0() { return get_group_id(0); } #endif // ================================================================================================= |