diff options
| author | Tyler Sorensen <tylersorensen3221@hotmail.com> | 2018-07-11 15:12:22 -0400 |
|---|---|---|
| committer | Tyler Sorensen <tylersorensen3221@hotmail.com> | 2018-07-11 15:12:22 -0400 |
| commit | 7f2e98a1406da6c5293f0c988df95edc246ef88d (patch) | |
| tree | 5fdb94393d54cff495f7f2e6a2f0edc6df1eebd8 /src/kernels/level3/xgemm_part1.opencl | |
| parent | 7bae54f61f8a2b589421afd57c9da6c8775155ef (diff) | |
added inline ptx to support shuffle on Nvidia GPUs
Diffstat (limited to 'src/kernels/level3/xgemm_part1.opencl')
| -rw-r--r-- | src/kernels/level3/xgemm_part1.opencl | 24 |
1 files changed, 20 insertions, 4 deletions
diff --git a/src/kernels/level3/xgemm_part1.opencl b/src/kernels/level3/xgemm_part1.opencl index 99d64c91..9e483b3e 100644 --- a/src/kernels/level3/xgemm_part1.opencl +++ b/src/kernels/level3/xgemm_part1.opencl @@ -114,13 +114,29 @@ R"( #define GLOBAL_MEM_FENCE 0 // Global synchronisation barrier for potential better performance #endif -// Intel subgroups (https://www.khronos.org/registry/OpenCL/extensions/intel/cl_intel_subgroups.txt) -#ifndef USE_SUBGROUP_SHUFFLING - #define USE_SUBGROUP_SHUFFLING 0 // Optionally enables subgroup shuffling for Intel GPUs +#ifndef NVIDIA_WARPS_AS_SUBGROUPS + #define NVIDIA_WARPS_AS_SUBGROUPS 0 +#endif +#ifndef NVIDIA_POST_VOLTA + #define NVIDIA_POST_VOLTA 0 #endif -#if USE_SUBGROUP_SHUFFLING == 1 +#ifndef INTEL_SUBGROUP_EXTENSION + #define INTEL_SUBGROUP_EXTENSION 0 +#endif +//#ifndef USE_SUBGROUP_SHUFFLING + #define USE_SUBGROUP_SHUFFLING 0 // Optionally enables subgroup shuffling for Intel GPUs +//#endif + +// Intel subgroups (https://www.khronos.org/registry/OpenCL/extensions/intel/cl_intel_subgroups.txt) +#if USE_SUBGROUP_SHUFFLING == 1 && INTEL_SUBGROUP_EXTENSION #define SUBGROUP_SIZE 8 // Assumes subgroup size is always 8 on Intel GPUs #endif + +// NVIDIA warps as subgroups using inline PTX (https://docs.nvidia.com/cuda/inline-ptx-assembly/index.html) +#if USE_SUBGROUP_SHUFFLING == 1 && NVIDIA_WARPS_AS_SUBGROUPS + #define SUBGROUP_SIZE 32 // Assumes subgroup size is always 32 on NVIDIA GPUs +#endif + #if NWI != SUBGROUP_SIZE || MDIMC < SUBGROUP_SIZE #undef USE_SUBGROUP_SHUFFLING #define USE_SUBGROUP_SHUFFLING 0 // Disables subgroups in case the assumptions don't hold |