Stream-ordered wait{_any,_all}, fix wait_any implementation

Author: cwpearson

src/KokkosComm/mpi/req.hpp

@@ -64,34 +64,68 @@ class Req<Mpi> {
  private:
   std::shared_ptr<Record> record_;
 
+  template <KokkosExecutionSpace ExecSpace>
+  friend void wait(const ExecSpace &space, Req<Mpi> req);
   friend void wait(Req<Mpi> req);
+  template <KokkosExecutionSpace ExecSpace>
+  friend void wait_all(const ExecSpace &space, std::vector<Req<Mpi>> &reqs);
   friend void wait_all(std::vector<Req<Mpi>> &reqs);
+  template <KokkosExecutionSpace ExecSpace>
+  friend int wait_any(const ExecSpace &space, std::vector<Req<Mpi>> &reqs);
   friend int wait_any(std::vector<Req<Mpi>> &reqs);
 };
 
-inline void wait(Req<Mpi> req) {
+template <KokkosExecutionSpace ExecSpace>
+void wait(const ExecSpace &space, Req<Mpi> req) {
+  /* Semantically this only guarantees that `space` is waiting for request to complete. For the MPI host API, we have no
+   * choice but to fence the space before waiting on the host.*/
+  space.fence();
   MPI_Wait(&req.mpi_request(), MPI_STATUS_IGNORE);
   for (auto &f : req.record_->postWaits_) {
     f();
   }
   req.record_->postWaits_.clear();
 }
 
-inline void wait_all(std::vector<Req<Mpi>> &reqs) {
+inline void wait(Req<Mpi> req) { wait(Kokkos::DefaultExecutionSpace(), req); }
+
+template <KokkosExecutionSpace ExecSpace>
+void wait_all(const ExecSpace &space, std::vector<Req<Mpi>> &reqs) {
+  space.fence();
   for (Req<Mpi> &req : reqs) {
-    wait(req);
+    MPI_Wait(&req.mpi_request(), MPI_STATUS_IGNORE);
+    for (auto &f : req.record_->postWaits_) {
+      f();
+    }
+    req.record_->postWaits_.clear();
   }
 }
 
-inline int wait_any(std::vector<Req<Mpi>> &reqs) {
-  for (size_t i = 0; i < reqs.size(); ++i) {
-    int completed;
-    MPI_Test(&(reqs[i].mpi_request()), &completed, MPI_STATUS_IGNORE);
-    if (completed) {
-      return true;
+inline void wait_all(std::vector<Req<Mpi>> &reqs) { wait_all(Kokkos::DefaultExecutionSpace(), reqs); }
+
+template <KokkosExecutionSpace ExecSpace>
+int wait_any(const ExecSpace &space, std::vector<Req<Mpi>> &reqs) {
+  if (reqs.empty()) {
+    return -1;
+  }
+
+  space.fence();
+  while (true) {  // wait until something is done
+    for (size_t i = 0; i < reqs.size(); ++i) {
+      int completed;
+      Req<Mpi> &req = reqs[i];
+      MPI_Test(&(req.mpi_request()), &completed, MPI_STATUS_IGNORE);
+      if (completed) {
+        for (auto &f : req.record_->postWaits_) {
+          f();
+        }
+        req.record_->postWaits_.clear();
+        return i;
+      }
     }
   }
-  return false;
 }
 
+inline int wait_any(std::vector<Req<Mpi>> &reqs) { return wait_any(Kokkos::DefaultExecutionSpace(), reqs); }
+
 }  // namespace KokkosComm

unit_tests/CMakeLists.txt

@@ -86,6 +86,7 @@ target_sources(
     mpi/test_alltoall.cpp
     mpi/test_reduce.cpp
     mpi/test_allgather.cpp
+    mpi/test_waitany.cpp
 )
 target_link_libraries(
   test-main

unit_tests/mpi/test_waitany.cpp

@@ -0,0 +1,105 @@
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 4.0
+//       Copyright (2022) National Technology & Engineering
+//               Solutions of Sandia, LLC (NTESS).
+//
+// Under the terms of Contract DE-NA0003525 with NTESS,
+// the U.S. Government retains certain rights in this software.
+//
+// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions.
+// See https://kokkos.org/LICENSE for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//@HEADER
+
+#include <gtest/gtest.h>
+#include <type_traits>
+#include <algorithm>  // iota
+#include <random>
+
+#include "KokkosComm/KokkosComm.hpp"
+
+namespace {
+
+using namespace KokkosComm::mpi;
+
+template <typename T>
+class MpiWaitAny : public testing::Test {
+ public:
+  using Scalar = T;
+};
+
+using ScalarTypes = ::testing::Types<int, double, Kokkos::complex<float>>;
+TYPED_TEST_SUITE(MpiWaitAny, ScalarTypes);
+
+template <KokkosComm::KokkosExecutionSpace ExecSpace, typename Scalar>
+void wait_any() {
+  using TestView = Kokkos::View<Scalar *>;
+
+  int rank, size;
+  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+  MPI_Comm_size(MPI_COMM_WORLD, &size);
+  if (size < 2) {
+    GTEST_SKIP() << "Requires >= 2 ranks (" << size << " provided)";
+  }
+
+  constexpr size_t numMsg = 128;
+  ExecSpace space;
+  std::vector<KokkosComm::Req<>> reqs;
+  std::vector<TestView> views;
+
+  for (size_t i = 0; i < numMsg; ++i) {
+    views.push_back(TestView(std::to_string(i), i));
+  }
+
+  constexpr unsigned int SEED = 31337;
+  std::random_device rd;
+  std::mt19937 g(SEED);
+
+  // random send/recv order
+  std::vector<size_t> order(numMsg);
+  std::iota(order.begin(), order.end(), size_t(0));
+  std::shuffle(order.begin(), order.end(), g);
+
+  KokkosComm::Handle<ExecSpace, KokkosComm::Mpi> h(space, MPI_COMM_WORLD);
+
+  if (0 == rank) {
+    constexpr int dst = 1;
+
+    for (size_t i : order) {
+      reqs.push_back(KokkosComm::send(h, views[i], dst));
+    }
+
+    for (size_t i = 0; i < numMsg; ++i) {
+      reqs.erase(reqs.begin() + KokkosComm::wait_any(reqs));
+    }
+  } else if (1 == rank) {
+    constexpr int src = 0;
+
+    for (size_t i : order) {
+      reqs.push_back(KokkosComm::recv(h, views[i], src));
+    }
+
+    for (size_t i = 0; i < numMsg; ++i) {
+      reqs.erase(reqs.begin() + KokkosComm::wait_any(reqs));
+    }
+  }
+}
+
+// TODO: test call on no requests
+
+TYPED_TEST(MpiWaitAny, default_execution_space) {
+  wait_any<Kokkos::DefaultExecutionSpace, typename TestFixture::Scalar>();
+}
+
+TYPED_TEST(MpiWaitAny, default_host_execution_space) {
+  if constexpr (std::is_same_v<Kokkos::DefaultHostExecutionSpace, Kokkos::DefaultExecutionSpace>) {
+    GTEST_SKIP() << "Skipping test: DefaultHostExecSpace = DefaultExecSpace";
+  } else {
+    wait_any<Kokkos::DefaultHostExecutionSpace, typename TestFixture::Scalar>();
+  }
+}
+
+}  // namespace