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/xgemm_direct_part3.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/xgemm_direct_part3.opencl')
| -rw-r--r-- | src/kernels/level3/xgemm_direct_part3.opencl | 24 |
1 files changed, 20 insertions, 4 deletions
diff --git a/src/kernels/level3/xgemm_direct_part3.opencl b/src/kernels/level3/xgemm_direct_part3.opencl index 0822c95f..66b880e6 100644 --- a/src/kernels/level3/xgemm_direct_part3.opencl +++ b/src/kernels/level3/xgemm_direct_part3.opencl @@ -218,7 +218,11 @@ INLINE_FUNC void XgemmDirect(const int kSizeM, const int kSizeN, const int kSize // ================================================================================================= // Direct version of the GEMM kernel with [A, B] = [non-transposed, non-transposed] -__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#endif void XgemmDirectNN(const int kSizeM, const int kSizeN, const int kSizeK, const real_arg arg_alpha, const real_arg arg_beta, const __global realMD* restrict agm, const int a_offset, const int a_ld, @@ -233,7 +237,11 @@ void XgemmDirectNN(const int kSizeM, const int kSizeN, const int kSizeK, } // Direct version of the GEMM kernel with [A, B] = [non-transposed, transposed] -__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#endif void XgemmDirectNT(const int kSizeM, const int kSizeN, const int kSizeK, const real_arg arg_alpha, const real_arg arg_beta, const __global realMD* restrict agm, const int a_offset, const int a_ld, @@ -248,7 +256,11 @@ void XgemmDirectNT(const int kSizeM, const int kSizeN, const int kSizeK, } // Direct version of the GEMM kernel with [A, B] = [transposed, non-transposed] -__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#endif void XgemmDirectTN(const int kSizeM, const int kSizeN, const int kSizeK, const real_arg arg_alpha, const real_arg arg_beta, const __global realMD* restrict agm, const int a_offset, const int a_ld, @@ -263,7 +275,11 @@ void XgemmDirectTN(const int kSizeM, const int kSizeN, const int kSizeK, } // Direct version of the GEMM kernel with [A, B] = [transposed, transposed] -__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#if RELAX_WORKGROUP_SIZE == 1 + __kernel +#elif + __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +#endif void XgemmDirectTT(const int kSizeM, const int kSizeN, const int kSizeK, const real_arg arg_alpha, const real_arg arg_beta, const __global realMD* restrict agm, const int a_offset, const int a_ld, |