diff options
author | Angus, Alexander <aangus@qti.qualcomm.com> | 2023-01-03 10:56:04 -0800 |
---|---|---|
committer | Angus, Alexander <aangus@qti.qualcomm.com> | 2023-01-03 10:56:04 -0800 |
commit | 4f394608a28f419dfd6091c704148d9e638a26f0 (patch) | |
tree | 4c0e042109c4d249ff5b700fc49a862169edec5a /src/kernels/level3/transpose_pad.opencl | |
parent | 03cffa83c5f7742f8ec0c5e762bb7048e38952f3 (diff) |
implemented changes to boost Adreno performance according to https://jira-dc.qualcomm.com/jira/browse/OSR-8731
Diffstat (limited to 'src/kernels/level3/transpose_pad.opencl')
-rw-r--r-- | src/kernels/level3/transpose_pad.opencl | 36 |
1 files changed, 30 insertions, 6 deletions
diff --git a/src/kernels/level3/transpose_pad.opencl b/src/kernels/level3/transpose_pad.opencl index e55a8b7c..31de0e62 100644 --- a/src/kernels/level3/transpose_pad.opencl +++ b/src/kernels/level3/transpose_pad.opencl @@ -84,7 +84,11 @@ INLINE_FUNC void _TransposePadMatrix(LOCAL_PTR real* tile, } // Interface to the above function -__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#endif void TransposePadMatrix(const int src_one, const int src_two, const int src_ld, const int src_offset, __global const real* restrict src, @@ -172,7 +176,11 @@ INLINE_FUNC void _TransposeMatrix(LOCAL_PTR real* tile, } // Interface to the above function -__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#endif void TransposeMatrix(const int src_one, const int src_two, const int src_ld, const int src_offset, __global const real* restrict src, @@ -193,7 +201,11 @@ void TransposeMatrix(const int src_one, const int src_two, #if defined(ROUTINE_GEMMBATCHED) // Batched version of the above -__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#endif void TransposePadMatrixBatched(const int src_one, const int src_two, const int src_ld, const __constant int* src_offsets, __global const real* restrict src, @@ -212,7 +224,11 @@ void TransposePadMatrixBatched(const int src_one, const int src_two, } // Batched version of the above -__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#endif void TransposeMatrixBatched(const int src_one, const int src_two, const int src_ld, const __constant int* src_offsets, __global const real* restrict src, @@ -234,7 +250,11 @@ void TransposeMatrixBatched(const int src_one, const int src_two, #if defined(ROUTINE_GEMMSTRIDEDBATCHED) // Strided-batched version of the above -__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#endif void TransposePadMatrixStridedBatched(const int src_one, const int src_two, const int src_ld, const int src_offset, const int src_stride, __global const real* restrict src, @@ -253,7 +273,11 @@ void TransposePadMatrixStridedBatched(const int src_one, const int src_two, } // Strided-batched version of the above -__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +#endif void TransposeMatrixStridedBatched(const int src_one, const int src_two, const int src_ld, const int src_offset, const int src_stride, __global const real* restrict src, |