fix lci-pp: fix lci_progress_pool when there are multiple pus per core

STEllAR-GROUP · Jan 11, 2024 · b6d8915 · b6d8915
1 parent 975dbc2
commit b6d8915
Show file tree

Hide file tree

Showing 2 changed files with 19 additions and 16 deletions.
diff --git a/libs/full/parcelport_lci/include/hpx/parcelport_lci/parcelport_lci.hpp b/libs/full/parcelport_lci/include/hpx/parcelport_lci/parcelport_lci.hpp
@@ -210,29 +210,30 @@ namespace hpx::traits {
                     hpx::threads::policies::scheduler_mode::
                         do_background_work_only);
 
-                size_t ncores_to_add =
+                size_t npus_to_add =
                     parcelset::policies::lci::config_t::progress_thread_num;
-                std::vector<const hpx::resource::core*> cores;
+                std::vector<const hpx::resource::pu*> pus;
                 for (auto& numa_domain : rp.numa_domains())
                 {
                     for (auto& core : numa_domain.cores())
                     {
-                        cores.push_back(&core);
+                        for (auto& pu : core.pus())
+                            pus.push_back(&pu);
                     }
                 }
-                if (cores.size() <= 1)
+                if (pus.size() <= 1)
                 {
-                    fprintf(stderr, "We don't have enough cores!\n");
+                    fprintf(stderr, "We don't have enough pus!\n");
                     exit(1);
                 }
-                if ((size_t) ncores_to_add > cores.size() / 2)
+                if ((size_t) npus_to_add > pus.size() / 2)
                 {
-                    ncores_to_add = cores.size() / 2;
+                    npus_to_add = pus.size() / 2;
                 }
-                for (size_t i = 0; i < ncores_to_add; ++i)
+                for (size_t i = 0; i < npus_to_add; ++i)
                 {
-                    size_t next_core = i * cores.size() / ncores_to_add;
-                    rp.add_resource(*cores[next_core], "lci-progress-pool");
+                    size_t next_pu = i * pus.size() / npus_to_add;
+                    rp.add_resource(*pus[next_pu], "lci-progress-pool");
                 }
             }
         }

diff --git a/libs/full/parcelport_lci/src/parcelport_lci.cpp b/libs/full/parcelport_lci/src/parcelport_lci.cpp
@@ -161,12 +161,14 @@ namespace hpx::parcelset::policies::lci {
                 if (hpx::this_thread::get_pool() ==
                     &hpx::resource::get_thread_pool("lci-progress-pool"))
                 {
-                    std::size_t prg_thread_id =
-                        hpx::get_local_worker_thread_num();
-                    double rate = (double) config_t::ndevices /
-                        config_t::progress_thread_num;
-                    for (int i = prg_thread_id * rate;
-                         i < (prg_thread_id + 1) * rate; ++i)
+                    int prg_thread_id =
+                        static_cast<int>(hpx::get_local_worker_thread_num());
+                    HPX_ASSERT(prg_thread_id < config_t::progress_thread_num);
+                    for (int i = prg_thread_id * config_t::ndevices /
+                             config_t::progress_thread_num;
+                         i < (prg_thread_id + 1) * config_t::ndevices /
+                             config_t::progress_thread_num;
+                         ++i)
                     {
                         devices_to_progress.push_back(&devices[i]);
                     }