LRCC2 seems to work with water, but hangs in the singles equations ->…

… need to run them in macrotasks to reduce communication

src/madness/chem/CC2.cc

@@ -701,12 +701,18 @@ CC2::iterate_lrcc2_pairs(World& world, const CC_vecfunction& cc2_s,
     // if no function has been computed so far use the constant part (first iteration)
     for (auto& pair : pair_vec) if (not pair.function().is_initialized()) pair.update_u(pair.constant_part);
 
+    for (const auto& p : pair_vec) p.constant_part.print_size("constant_part before iter");
     for (const auto& p : pair_vec) p.function().print_size("u before iter");
 
     // compute the coupling between the pair functions
     if (world.rank()==0) print("computing local coupling in the universe");
     Pairs<real_function_6d> coupling=compute_local_coupling(pair_vec, info);
     auto coupling_vec=Pairs<real_function_6d>::pairs2vector(coupling,triangular_map);
+    reconstruct(world,coupling_vec);
+    for (auto& p : pair_vec) {
+        p.constant_part.reconstruct();
+        p.function().reconstruct();
+    }
 
     if (info.parameters.debug()) print_size(world, coupling_vec, "couplingvector");
 
@@ -1099,7 +1105,7 @@ CC2::initialize_pairs(Pairs<CCPair>& pairs, const CCState ftype, const CalcType
                 pairs.insert(i, j, tmp);
 
             } else if (ftype == EXCITED_STATE) {
-                name = std::to_string(int(excitation)) + "_" + name;
+                // name = std::to_string(int(excitation)) + "_" + name;
                 real_function_6d utmp = real_factory_6d(world);
                 const bool found = CCOPS.load_function(utmp, name);
                 if (found) restarted = true;
@@ -1116,7 +1122,7 @@ CC2::initialize_pairs(Pairs<CCPair>& pairs, const CCState ftype, const CalcType
 
                 tmp.constant_part = const_part;
                 pairs.insert(i, j, tmp);
-                CCPotentials::compute_excited_pair_energy(world, pairs(i, j), x, info);
+                // CCPotentials::compute_excited_pair_energy(world, pairs(i, j), x, info);
             } else error("Unknown pairtype");
         }
     }

src/madness/chem/CCPotentials.h

@@ -79,8 +79,10 @@ class CCPotentials {
             f.truncate();
             f.print_size(name);
             return true;
+        } else {
+            if (world.rank()==0) print("could not find function",name);
         }
-        else return false;
+        return false;
     }
 
     /// Plotting (convenience)

src/madness/world/cloud.h

@@ -190,12 +190,17 @@ class Cloud {
     void print_size(World& universe) {
 
         std::size_t memsize=0;
-        for (auto& item : container) memsize+=item.second.size();
+        std::size_t max_record_size=0;
+        for (auto& item : container) {
+            memsize+=item.second.size();
+            max_record_size=std::max(max_record_size,item.second.size());
+        }
         std::size_t global_memsize=memsize;
         std::size_t max_memsize=memsize;
         std::size_t min_memsize=memsize;
         universe.gop.sum(global_memsize);
         universe.gop.max(max_memsize);
+        universe.gop.max(max_record_size);
         universe.gop.min(min_memsize);
 
         auto local_size=container.size();
@@ -207,13 +212,15 @@ class Cloud {
             print("Cloud memory:");
             print("  replicated:",is_replicated);
             print("size of cloud (total)");
-            print("  number of records:",global_size);
-            print("  memory in GBytes: ",global_memsize*byte2gbyte);
+            print("  number of records:        ",global_size);
+            print("  memory in GBytes:         ",global_memsize*byte2gbyte);
             print("size of cloud (average per node)");
-            print("  number of records:",double(global_size)/universe.size());
-            print("  memory in GBytes: ",global_memsize*byte2gbyte/universe.size());
+            print("  number of records:        ",double(global_size)/universe.size());
+            print("  memory in GBytes:         ",global_memsize*byte2gbyte/universe.size());
             print("min/max of node");
-            print("  memory in GBytes: ",min_memsize*byte2gbyte,max_memsize*byte2gbyte);
+            print("  memory in GBytes:         ",min_memsize*byte2gbyte,max_memsize*byte2gbyte);
+            print("  max record size in GBytes:",max_record_size*byte2gbyte);
+
         }
     }
 

src/madness/world/worlddc.h

@@ -1867,8 +1867,8 @@ namespace madness
                     }
                     world->taskq.fence(); // just need LOCAL fence
                     wall1 = wall_time();
-                    if (world->rank() == 0)
-                        printf("time in op_inspector: %8.4fs\n", wall1 - wall0);
+                    // if (world->rank() == 0)
+                        // printf("time in op_inspector: %8.4fs\n", wall1 - wall0);
                     wall0 = wall1;
 
                     // total size over all threads
@@ -1891,8 +1891,8 @@ namespace madness
                     world->taskq.fence(); // just need LOCAL fence
 
                     wall1 = wall_time();
-                    if (world->rank() == 0)
-                        printf("time in op_executor: %8.4fs\n", wall1 - wall0);
+                    // if (world->rank() == 0)
+                        // printf("time in op_executor: %8.4fs\n", wall1 - wall0);
                     wall0 = wall1;
                 }
                 // VERify that the serialization worked!!
@@ -1919,8 +1919,8 @@ namespace madness
                 for (int i = 1; i < world->size(); ++i)
                     offsets[i] = offsets[i - 1] + sizes[i - 1];
                 size_t total_size = offsets.back() + sizes.back();
-                if (world->rank() == 0)
-                    print("total_size", total_size);
+                // if (world->rank() == 0)
+                    // print("total_size", total_size);
 
                 // print("time 4",wall_time());
                 // gather the vector of data v from each process to process 0
@@ -1932,8 +1932,8 @@ namespace madness
                 MPI_Gatherv(buf, local_size, MPI_BYTE, all_data, sizes.data(), offsets.data(), MPI_BYTE, 0, world->mpi.comm().Get_mpi_comm());
 
                 wall1 = wall_time();
-                if (world->rank() == 0)
-                    printf("time in gather+gatherv: %8.4fs\n", wall1 - wall0);
+                // if (world->rank() == 0)
+                    // printf("time in gather+gatherv: %8.4fs\n", wall1 - wall0);
                 wall0 = wall1;
 
                 delete[] buf;
@@ -1949,8 +1949,8 @@ namespace madness
                     localar.store(all_data, total_size);
                     ArchivePrePostImpl<localarchiveT, dcT>::postamble_store(localar);
                     wall1 = wall_time();
-                    if (world->rank() == 0)
-                        printf("time in final copy on node 0: %8.4fs\n", wall1 - wall0);
+                    // if (world->rank() == 0)
+                        // printf("time in final copy on node 0: %8.4fs\n", wall1 - wall0);
 
                     delete[] all_data;
                 }