diff options
author | Tyler Sorensen <tylersorensen3221@hotmail.com> | 2018-07-14 19:50:47 -0400 |
---|---|---|
committer | Tyler Sorensen <tylersorensen3221@hotmail.com> | 2018-07-14 19:50:47 -0400 |
commit | 7709a7308bce5492e06d8867a4dd9dff5b2ba950 (patch) | |
tree | ed35acf41257752ec165480c2298edf17080da4c /src/kernels | |
parent | 36093429fd444d0a1fc7de25dfaf7f2f775cfabc (diff) |
Applied feedback from Cedric from first pull request
Diffstat (limited to 'src/kernels')
-rw-r--r-- | src/kernels/level3/xgemm_part1.opencl | 21 | ||||
-rw-r--r-- | src/kernels/level3/xgemm_part3.opencl | 10 |
2 files changed, 16 insertions, 15 deletions
diff --git a/src/kernels/level3/xgemm_part1.opencl b/src/kernels/level3/xgemm_part1.opencl index 9e483b3e..32386312 100644 --- a/src/kernels/level3/xgemm_part1.opencl +++ b/src/kernels/level3/xgemm_part1.opencl @@ -114,26 +114,27 @@ R"( #define GLOBAL_MEM_FENCE 0 // Global synchronisation barrier for potential better performance #endif -#ifndef NVIDIA_WARPS_AS_SUBGROUPS - #define NVIDIA_WARPS_AS_SUBGROUPS 0 +#ifndef SUBGROUP_SHUFFLING_NVIDIA_PRE_VOLTA + #define SUBGROUP_SHUFFLING_NVIDIA_PRE_VOLTA 0 #endif -#ifndef NVIDIA_POST_VOLTA - #define NVIDIA_POST_VOLTA 0 +#ifndef SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA + #define SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA 0 #endif -#ifndef INTEL_SUBGROUP_EXTENSION - #define INTEL_SUBGROUP_EXTENSION 0 +#ifndef SUBGROUP_SHUFFLING_INTEL + #define SUBGROUP_SHUFFLING_INTEL 0 #endif -//#ifndef USE_SUBGROUP_SHUFFLING +#ifndef USE_SUBGROUP_SHUFFLING #define USE_SUBGROUP_SHUFFLING 0 // Optionally enables subgroup shuffling for Intel GPUs -//#endif +#endif // Intel subgroups (https://www.khronos.org/registry/OpenCL/extensions/intel/cl_intel_subgroups.txt) -#if USE_SUBGROUP_SHUFFLING == 1 && INTEL_SUBGROUP_EXTENSION +#if USE_SUBGROUP_SHUFFLING == 1 && SUBGROUP_SHUFFLING_INTEL #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 +#if USE_SUBGROUP_SHUFFLING == 1 && (SUBGROUP_SHUFFLING_NVIDIA_PRE_VOLTA || \ + SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA) #define SUBGROUP_SIZE 32 // Assumes subgroup size is always 32 on NVIDIA GPUs #endif diff --git a/src/kernels/level3/xgemm_part3.opencl b/src/kernels/level3/xgemm_part3.opencl index 8e20b1b8..35ec735c 100644 --- a/src/kernels/level3/xgemm_part3.opencl +++ b/src/kernels/level3/xgemm_part3.opencl @@ -24,11 +24,11 @@ R"( INLINE_FUNC int clblast_get_sub_group_local_id() { // Intel extension - #if INTEL_SUBGROUP_EXTENSION == 1 + #if SUBGROUP_SHUFFLING_INTEL == 1 return get_sub_group_local_id(); // Nvidia inline PTX - #elif NVIDIA_WARPS_AS_SUBGROUPS == 1 + #elif SUBGROUP_SHUFFLING_NVIDIA_PRE_VOLTA == 1 || SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA == 1 int ret; asm volatile("mov.u32 %0, %%laneid;" : "=r"(ret) ); return ret; @@ -38,14 +38,14 @@ INLINE_FUNC int clblast_get_sub_group_local_id() { INLINE_FUNC realN clblast_sub_group_shuffle(realN reg, int src) { // Intel extension - #if INTEL_SUBGROUP_EXTENSION == 1 + #if SUBGROUP_SHUFFLING_INTEL == 1 return intel_sub_group_shuffle(reg, src); // Nvidia inline PTX // Volta and later requires .sync shuffle instructions with an extra mask arg - #elif NVIDIA_WARPS_AS_SUBGROUPS == 1 + #elif SUBGROUP_SHUFFLING_NVIDIA_PRE_VOLTA == 1 || SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA == 1 realN ret; - #if NVIDIA_POST_VOLTA == 1 + #if SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA == 1 asm volatile("shfl.sync.idx.b32 %0, %1, %2, 0x1f, 0xffffffff;" : "=f"(ret): "f"(reg), "r"(src)); #else asm volatile("shfl.idx.b32 %0, %1, %2, 0x1f;" : "=f"(ret): "f"(reg), "r"(src)); |