Merge branch 'develop' into update_cka8w8_uc

Jenkinsfile

@@ -722,6 +722,9 @@ CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCM
 
 pipeline {
     agent none
+    triggers {
+        parameterizedCron(CRON_SETTINGS)
+    }
     options {
         parallelsAlwaysFailFast()
     }

example/ck_tile/03_gemm/universal_gemm.cpp

@@ -254,22 +254,13 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
 #endif
     }
     else
-    {
-        // Tail number always Full - #PrefetchStages
-        if(tail_num == ck_tile::TailNumber::Full)
-        {
-            Run(ck_tile::bool_constant<false>{},
-                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
-        }
-        else
     {
         std::ostringstream err;
-            err << "When there's no hot loop, this tail number \"" << tail_num
-                << "\" is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
-                << "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+        err << "Num K loop must be larger than number of prefetech stages."
+            << "\n PrefetchStages: " << BaseGemmPipeline::PrefetchStages << "\n File: " << __FILE__
+            << ":" << __LINE__ << ", in function: " << __func__;
         throw std::runtime_error(err.str());
     }
-    }
 
     return ave_time;
 }

include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs.hpp

@@ -343,6 +343,8 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
         // moving k_dram_window is an in-page-block operation, so there is
         // no need to invoke k_page_block_navigator.move_tile_window() here.
         move_tile_window(k_dram_window, {0, kK0});
+        // ensure LDS access by Q is done before the over-writting by K
+        block_sync_lds();
         store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
 
         do