ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 02:27:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix the calculation of number of instructions used by sched_group_barrier

Browse Source
This commit is contained in:
Qianfeng Zhang
2025-07-15 08:59:48 +00:00
parent 0206b3442a
commit c01695597e
1 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp
View File

@@ -363,9 +363,10 @@ struct HstuAttentionFwdPipelineQRKSVS
if constexpr(HstuMask::kUseLocal)
{
constexpr index_t V_VMEM_LOAD_INST = (kN1 * kK1) / kBlockSize / kAlignmentV;
constexpr index_t K_VMEM_LOAD_INST = (kN0 * kK0) / kBlockSize / kAlignmentV;
constexpr index_t K_LDS_WRITE_INST =
(kN0 * kK0) / kBlockSize / Policy::template GetSmemKPackK<Problem>();
constexpr index_t K_VMEM_LOAD_INST =
(kK1 * kSubQKHeaddim) / kBlockSize / kAlignmentK;
constexpr index_t K_LDS_WRITE_INST = (kK1 * kSubQKHeaddim) / kBlockSize /
Policy::template GetSmemKPackK<Problem>();
constexpr index_t MFMA_INST = (kM0 * kSubQKHeaddim) / kBlockSize / 4;
constexpr index_t K_LDS_READ_INST = MFMA_INST / kGemmNumRepM;
@@ -401,7 +402,7 @@ struct HstuAttentionFwdPipelineQRKSVS
__builtin_amdgcn_sched_group_barrier(0x00000008, kGemmNumRepM, 0);
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_sched_barrier(0x00000001);
}
else
{