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/levelx/xconvgemm_part2.opencl | |
parent | d31fb141cb597aaf405674621aa25f263aa375e1 (diff) | |
parent | b0b302889cc786907efb080c4e1beea30d2fa39f (diff) |
Merge tag '1.6.0' into gspr/post-bookworm
Diffstat (limited to 'src/kernels/levelx/xconvgemm_part2.opencl')
-rw-r--r-- | src/kernels/levelx/xconvgemm_part2.opencl | 18 |
1 files changed, 15 insertions, 3 deletions
diff --git a/src/kernels/levelx/xconvgemm_part2.opencl b/src/kernels/levelx/xconvgemm_part2.opencl index 693cb120..38ddd7eb 100644 --- a/src/kernels/levelx/xconvgemm_part2.opencl +++ b/src/kernels/levelx/xconvgemm_part2.opencl @@ -23,7 +23,11 @@ R"( // ConvGEMM kernel #if defined(CONVGEMM_WITH_IM2COL) -__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#endif void Xconvgemm(const int num_patches, const int num_kernels, const int patch_size, const __global realND* restrict kernelgm, const int kernel_offset, __global real* resultgm, const int result_offset, const int result_stride, @@ -285,7 +289,11 @@ INLINE_FUNC void Xconvgemm(const int num_patches, const int num_kernels, const i } #if !defined(CONVGEMM_WITH_IM2COL) -__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#endif void XconvgemmFlip(const int num_patches, const int num_kernels, const int patch_size, const __global realND* restrict kernelgm, const int kernel_offset, __global real* resultgm, const int result_offset, const int result_stride, @@ -306,7 +314,11 @@ void XconvgemmFlip(const int num_patches, const int num_kernels, const int patch output_h, output_w, alm, blm, kernel_flip); } -__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#else + __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#endif void XconvgemmNormal(const int num_patches, const int num_kernels, const int patch_size, const __global realND* restrict kernelgm, const int kernel_offset, __global real* resultgm, const int result_offset, const int result_stride, |