diff options
author | Gard Spreemann <gspr@nonempty.org> | 2023-06-08 11:52:00 +0200 |
---|---|---|
committer | Gard Spreemann <gspr@nonempty.org> | 2023-06-08 11:52:00 +0200 |
commit | 63870a2e60c1bc8bfa7e3672457b551a8e51ffaf (patch) | |
tree | fe2c0cd5f62e3fbd17e58d3903ec6bb37983f620 /src/kernels/level3/copy_pad.opencl | |
parent | d31fb141cb597aaf405674621aa25f263aa375e1 (diff) | |
parent | b0b302889cc786907efb080c4e1beea30d2fa39f (diff) |
Merge tag '1.6.0' into gspr/post-bookworm
Diffstat (limited to 'src/kernels/level3/copy_pad.opencl')
-rw-r--r-- | src/kernels/level3/copy_pad.opencl | 36 |
1 files changed, 30 insertions, 6 deletions
diff --git a/src/kernels/level3/copy_pad.opencl b/src/kernels/level3/copy_pad.opencl index 3d389b74..dca93b76 100644 --- a/src/kernels/level3/copy_pad.opencl +++ b/src/kernels/level3/copy_pad.opencl @@ -59,7 +59,11 @@ INLINE_FUNC void _CopyPadMatrix(const int src_one, const int src_two, } // Interface to the above function -__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#endif void CopyPadMatrix(const int src_one, const int src_two, const int src_ld, const int src_offset, __global const real* restrict src, @@ -118,7 +122,11 @@ INLINE_FUNC void _CopyMatrix(const int src_one, const int src_two, } // Interface to the above function -__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#endif void CopyMatrix(const int src_one, const int src_two, const int src_ld, const int src_offset, __global const real* restrict src, @@ -138,7 +146,11 @@ void CopyMatrix(const int src_one, const int src_two, #if defined(ROUTINE_GEMMBATCHED) // Batched version of the above -__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#endif void CopyPadMatrixBatched(const int src_one, const int src_two, const int src_ld, const __constant int* src_offsets, __global const real* restrict src, @@ -156,7 +168,11 @@ void CopyPadMatrixBatched(const int src_one, const int src_two, } // Batched version of the above -__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#endif void CopyMatrixBatched(const int src_one, const int src_two, const int src_ld, const __constant int* src_offsets, __global const real* restrict src, @@ -177,7 +193,11 @@ void CopyMatrixBatched(const int src_one, const int src_two, #if defined(ROUTINE_GEMMSTRIDEDBATCHED) // Strided-batched version of the above -__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#endif void CopyPadMatrixStridedBatched(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, @@ -195,7 +215,11 @@ void CopyPadMatrixStridedBatched(const int src_one, const int src_two, } // Strided-batched version of the above -__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +#endif void CopyMatrixStridedBatched(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, |