Merge branch 'develop' of github.com:khuck/xpress-apex into develop

src/apex/dependency_tree.cpp

@@ -339,7 +339,7 @@ double Node::writeNodeCSV(std::stringstream& outfile, double total, int node_id)
     outfile << ((parent == nullptr) ? 0 : parent->index) << ",";
     outfile << depth << ",\"";
     outfile << data->get_tree_name() << "\",";
-    // write out the inclusive
+    // write out the accumulated
     double acc = (data == task_identifier::get_main_task_id() || getAccumulated() == 0.0) ?
         total : getAccumulated();
     // write the number of calls
@@ -350,6 +350,7 @@ double Node::writeNodeCSV(std::stringstream& outfile, double total, int node_id)
     double mean = acc / ncalls;
     outfile << std::setprecision(9);
     outfile << acc << ",";
+    outfile << inclusive << ",";
     outfile << getMinimum() << ",";
     outfile << mean << ",";
     outfile << getMaximum() << ",";
@@ -398,18 +399,18 @@ double Node::writeNodeCSV(std::stringstream& outfile, double total, int node_id)
     // end the line
     outfile << std::endl;
 
-    // sort the children by accumulated time
-    std::vector<std::pair<task_identifier, Node*> > sorted;
+    // sort the children by name to make tree merging easier (I hope)
+    std::vector<Node*> sorted;
     for (auto& it : children) {
-        sorted.push_back(it);
+        sorted.push_back(it.second);
     }
-    sort(sorted.begin(), sorted.end(), cmp);
+    sort(sorted.begin(), sorted.end(), Node::compareNodeByParentName);
 
     // do all the children
     double remainder = acc;
     depth++;
     for (auto c : sorted) {
-        double tmp = c.second->writeNodeCSV(outfile, total, node_id);
+        double tmp = c->writeNodeCSV(outfile, total, node_id);
         remainder = remainder - tmp;
     }
     depth--;

src/apex/dependency_tree.hpp

@@ -94,7 +94,7 @@ class Node {
         inline double& getSumSquares() { return prof.sum_squares; }
         void addAccumulated(double value, double incl, bool is_resume, uint64_t thread_id);
         size_t getIndex() { return index; };
-        std::string getName() { return data->get_name(); };
+        std::string getName() const { return data->get_name(); };
         void writeNode(std::ofstream& outfile, double total);
         double writeNodeASCII(std::ofstream& outfile, double total, size_t indent);
         double writeNodeCSV(std::stringstream& outfile, double total, int node_id);
@@ -107,6 +107,17 @@ class Node {
         static std::set<std::string>& getKnownMetrics() {
             return known_metrics;
         }
+        // required for using this class as a key in a map, vector, etc.
+        static bool compareNodeByParentName (const Node* lhs, const Node* rhs) {
+            if (lhs->parent < rhs->parent) {
+                return true;
+            }
+            if (lhs->getName().compare(lhs->getName()) < 0) {
+                return true;
+            }
+            return false;
+        }
+
 };
 
 } // dependency_tree

src/apex/proc_read.cpp

@@ -1020,9 +1020,11 @@ std::array<double,2> getAvailableMemory() {
             }
             fclose(f);
         }
+#ifdef APEX_WITH_HIP
         if (global_rsmi_reader != nullptr) {
             values[1] = global_rsmi_reader->getAvailableMemory();
         }
+#endif
     return values;
 }
 

src/apex/profile_reducer.cpp

@@ -21,7 +21,8 @@
  * 8 values (up to) when PAPI enabled */
 constexpr size_t num_fields{23};
 
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
 #include "mpi.h"
 #endif
 
@@ -46,7 +47,8 @@ namespace apex {
 std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
     int commrank = 0;
     int commsize = 1;
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
     int mpi_initialized = 0;
     MPI_CALL(MPI_Initialized( &mpi_initialized ));
     if (mpi_initialized) {
@@ -94,7 +96,8 @@ std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
     length[1] = length[1] + 1;
 
     /* AllReduce all profile name counts */
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
     if (mpi_initialized && commsize > 1) {
         MPI_CALL(PMPI_Allreduce(&length, &max_length, 2,
             MPI_UINT64_T, MPI_MAX, MPI_COMM_WORLD));
@@ -119,7 +122,8 @@ std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
         strncpy(ptr, name.c_str(), max_length[1]);
         ptr = ptr + max_length[1];
     }
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
     if (mpi_initialized && commsize > 1) {
         MPI_CALL(PMPI_Allgather(sbuf, sbuf_length, MPI_CHAR,
             rbuf, sbuf_length, MPI_CHAR, MPI_COMM_WORLD));
@@ -192,7 +196,8 @@ std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
     }
 
     /* Reduce the data */
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
     if (mpi_initialized && commsize > 1) {
         MPI_CALL(PMPI_Gather(s_pdata, sbuf_length, MPI_DOUBLE,
             r_pdata, sbuf_length, MPI_DOUBLE, 0, MPI_COMM_WORLD));
@@ -256,7 +261,8 @@ std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
     }
 
     }
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
     if (mpi_initialized && commsize > 1) {
         MPI_CALL(PMPI_Barrier(MPI_COMM_WORLD));
     }
@@ -267,7 +273,8 @@ std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
     void reduce_profiles(std::stringstream& csv_output, std::string filename) {
         int commrank = 0;
         int commsize = 1;
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
         int mpi_initialized = 0;
         MPI_CALL(MPI_Initialized( &mpi_initialized ));
         if (mpi_initialized) {
@@ -291,7 +298,8 @@ std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
         size_t length{csv_output.str().size()};
         size_t max_length{length};
         // get the longest string from all ranks
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
         if (mpi_initialized && commsize > 1) {
             MPI_CALL(PMPI_Allreduce(&length, &max_length, 1,
                 MPI_UINT64_T, MPI_MAX, MPI_COMM_WORLD));
@@ -307,14 +315,16 @@ std::map<std::string, apex_profile*> reduce_profiles_for_screen() {
         // allocate the memory to hold all output
         char * rbuf = nullptr;
         if (commrank == 0) {
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
             rbuf = (char*)calloc(max_length * commsize, sizeof(char));
 #else
             rbuf = sbuf;
 #endif
         }
 
-#if !defined(HPX_HAVE_NETWORKING) && defined(APEX_HAVE_MPI)
+#if defined(APEX_HAVE_MPI) || \
+    (defined(HPX_HAVE_NETWORKING) && defined(HPX_HAVE_PARCELPORT_MPI))
         MPI_Gather(sbuf, max_length, MPI_CHAR, rbuf, max_length, MPI_CHAR, 0, MPI_COMM_WORLD);
 #endif
 

src/apex/profiler_listener.cpp

@@ -1175,9 +1175,9 @@ std::unordered_set<profile*> free_profiles;
     stringstream tree_stream;
     if (node_id == 0) {
         tree_stream << "\"process rank\",\"node index\",\"parent index\",\"depth\",";
-        tree_stream << "\"name\",\"calls\",\"threads\",\"total time(ns)\",";
-        tree_stream << "\"minimum time(ns)\",\"mean time(ns)\",\"maximum time(ns)\",";
-        tree_stream << "\"stddev time(ns)\"";
+        tree_stream << "\"name\",\"calls\",\"threads\",\"total time(s)\",\"inclusive time(s)\",";
+        tree_stream << "\"minimum time(s)\",\"mean time(s)\",\"maximum time(s)\",";
+        tree_stream << "\"stddev time(s)\"";
         for (auto& x : dependency::Node::getKnownMetrics()) {
             tree_stream << ",\"total " << x << "\"";
             tree_stream << ",\"minimum " << x << "\"";