Skip zero EM resident downloads

AMSS_NCKU_source/bssnEM_class.C

@@ -87,7 +87,18 @@ bool bssn_em_zero_analysis_fastpath_enabled()
   return enabled != 0;
 }
 
-bool bssn_em_analysis_zero_fastpath_ready(MyList<Patch> *PatL,
+bool bssn_em_zero_resident_download_fastpath_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH");
+    enabled = (!env || atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+bool bssn_em_resident_zero_fastpath_ready(MyList<Patch> *PatL,
 #ifdef WithShell
                                           ShellPatch *shell,
 #else
@@ -95,8 +106,6 @@ bool bssn_em_analysis_zero_fastpath_ready(MyList<Patch> *PatL,
 #endif
                                           int rank)
 {
-  if (!bssn_em_zero_analysis_fastpath_enabled())
-    return false;
   int local_ok = 1;
   int local_seen = 0;
   MyList<Patch> *Pp = PatL;
@@ -149,6 +158,19 @@ bool bssn_em_analysis_zero_fastpath_ready(MyList<Patch> *PatL,
   return global_seen && global_ok;
 }
 
+bool bssn_em_analysis_zero_fastpath_ready(MyList<Patch> *PatL,
+#ifdef WithShell
+                                          ShellPatch *shell,
+#else
+                                          ShellPatch *shell,
+#endif
+                                          int rank)
+{
+  if (!bssn_em_zero_analysis_fastpath_enabled())
+    return false;
+  return bssn_em_resident_zero_fastpath_ready(PatL, shell, rank);
+}
+
 void zero_em_analysis_outputs(MyList<Patch> *PatL,
 #ifdef WithShell
                               ShellPatch *shell,
@@ -1660,11 +1682,11 @@ void bssnEM_class::Step(int lev, int YN)
   }
 #endif
 
-    if (em_step_timing)
+  if (em_step_timing)
-      em_t0 = MPI_Wtime();
+    em_t0 = MPI_Wtime();
-    Parallel::Sync_cached(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev]);
+  Parallel::Sync_cached(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev]);
-    if (em_step_timing)
+  if (em_step_timing)
-      em_t_predictor_sync += MPI_Wtime() - em_t0;
+    em_t_predictor_sync += MPI_Wtime() - em_t0;
 
 #ifdef WithShell
   if (lev == 0)
@@ -2310,7 +2332,19 @@ void bssnEM_class::Step(int lev, int YN)
   {
     if (em_step_timing)
       em_t0 = MPI_Wtime();
-    if (!bssn_em_cuda_keep_resident_after_step(lev, trfls, a_lev))
+    const bool needs_resident_download =
+        !bssn_em_cuda_keep_resident_after_step(lev, trfls, a_lev);
+    const bool skip_zero_resident_download =
+        needs_resident_download &&
+        bssn_em_zero_resident_download_fastpath_enabled() &&
+        bssn_em_resident_zero_fastpath_ready(GH->PatL[lev],
+#ifdef WithShell
+                                             0,
+#else
+                                             0,
+#endif
+                                             myrank);
+    if (needs_resident_download && !skip_zero_resident_download)
       bssn_em_cuda_download_level_state(GH->PatL[lev], SynchList_cor, myrank, true);
     if (em_step_timing)
       em_t_resident += MPI_Wtime() - em_t0;

makefile_and_run.py

@@ -153,6 +153,7 @@ def _gpu_runtime_env():
         "AMSS_CUDA_EM_CACHE_SOURCES": "1",
         "AMSS_CUDA_EM_ZERO_FASTPATH": "1",
         "AMSS_EM_ZERO_ANALYSIS_FASTPATH": "1",
+        "AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH": "1",
         "AMSS_CUDA_AMR_HOST_STAGED": "1",
         "AMSS_CUDA_AMR_RESTRICT_DEVICE": "0",
         "AMSS_CUDA_AMR_RESTRICT_BATCH": "0",
@@ -293,6 +294,7 @@ def run_ABE():
         print(f"   AMSS_CUDA_EM_CACHE_SOURCES={mpi_env.get('AMSS_CUDA_EM_CACHE_SOURCES', '')}")
         print(f"   AMSS_CUDA_EM_ZERO_FASTPATH={mpi_env.get('AMSS_CUDA_EM_ZERO_FASTPATH', '')}")
         print(f"   AMSS_EM_ZERO_ANALYSIS_FASTPATH={mpi_env.get('AMSS_EM_ZERO_ANALYSIS_FASTPATH', '')}")
+        print(f"   AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH={mpi_env.get('AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH', '')}")
         print(f"   AMSS_CUDA_AMR_HOST_STAGED={mpi_env.get('AMSS_CUDA_AMR_HOST_STAGED', '')}")
         print(f"   AMSS_CUDA_AMR_RESTRICT_DEVICE={mpi_env.get('AMSS_CUDA_AMR_RESTRICT_DEVICE', '')}")
         print(f"   AMSS_CUDA_AMR_RESTRICT_BATCH={mpi_env.get('AMSS_CUDA_AMR_RESTRICT_BATCH', '')}")