diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp index 0f2a8512fe..4dc5c4cece 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp @@ -533,7 +533,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR __builtin_amdgcn_sched_barrier(0); // Hot loop - do + while(i_total_loops < (num_total_loop - 1)) { // STAGE 1, Q@K Gemm0 auto st_acc = SPTBlockTileType{}; @@ -799,7 +799,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR i_total_loops += 1; seqlen_q_step += kM0; - } while(i_total_loops < (num_total_loop - 1)); + } __builtin_amdgcn_sched_barrier(0); // Tail