diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2023-01-21 20:28:32 +0100 |
---|---|---|
committer | GitHub <noreply@github.com> | 2023-01-21 20:28:32 +0100 |
commit | e72f87ae5eca5e2ea8aea4f2ce49408c1faa0521 (patch) | |
tree | 588a426b6350a5c982d89d98749ae78667fd23b4 /src/kernels/common.opencl | |
parent | 03cffa83c5f7742f8ec0c5e762bb7048e38952f3 (diff) | |
parent | 73f49e9b3d4abc4214122e4b8c07a736e01626ee (diff) |
Merge pull request #451 from CodeLinaro/master
CLBlast modifications to address Qualcomm Adreno performance
Diffstat (limited to 'src/kernels/common.opencl')
-rw-r--r-- | src/kernels/common.opencl | 6 |
1 files changed, 6 insertions, 0 deletions
diff --git a/src/kernels/common.opencl b/src/kernels/common.opencl index 56c1dae4..0ec741ad 100644 --- a/src/kernels/common.opencl +++ b/src/kernels/common.opencl @@ -132,6 +132,12 @@ R"( #define USE_CL_MAD 0 #endif +// By default the workgroup size requirement is enabled. For Qualcomm devices the workgroup size +// requirement results in worse performance and is disabled (src/utilities/compile.cpp) +#ifndef RELAX_WORKGROUP_SIZE + #define RELAX_WORKGROUP_SIZE 0 +#endif + // Sets a variable to zero #if PRECISION == 3232 || PRECISION == 6464 #define SetToZero(a) a.x = ZERO; a.y = ZERO |