refactor!: rename and rewrite Handle impls

.clang-format

@@ -4,4 +4,6 @@ SortIncludes: false
 AlignConsecutiveAssignments: true
 AllowShortCaseLabelsOnASingleLine: true
 AllowShortIfStatementsOnASingleLine: true
+AlignAfterOpenBracket: BlockIndent
+BinPackParameters: false
 ColumnLimit: 120

docs/api/concepts.rst

@@ -15,45 +15,51 @@ Kokkos Comm-specific concepts
 
     .. cpp:type:: T::communication_space
 
-    .. cpp:type:: T::handle_type
+    .. cpp:type:: T::communicator_type
 
     .. cpp:type:: T::request_type
 
     .. cpp:type:: T::datatype_type
 
     .. cpp:type:: T::reduction_op_type
 
+    .. cpp:type:: T::size_type
+
     .. cpp:type:: T::rank_type
 
 Types implementing the ``CommunicationSpace`` concept
------------------------------------------------------
+"""""""""""""""""""""""""""""""""""""""""""""""""""""
 
 .. cpp:class:: MpiSpace
 
     .. cpp:type:: communication_space = MpiSpace
 
-    .. cpp:type:: handle_type = MPI_Comm
+    .. cpp:type:: communicator_type = MPI_Comm
 
     .. cpp:type:: request_type = MPI_Request
 
     .. cpp:type:: datatype_type = MPI_Datatype
 
     .. cpp:type:: reduction_op_type = MPI_Op
 
+    .. cpp:type:: size_type = int
+
     .. cpp:type:: rank_type = int
 
 .. cpp:class:: Experimental::NcclSpace
 
     .. cpp:type:: communication_space = NcclSpace
 
-    .. cpp:type:: handle_type = ncclComm_t
+    .. cpp:type:: communicator_type = ncclComm_t
 
     .. cpp:type:: request_type = cudaEvent_t
 
     .. cpp:type:: datatype_type = ncclDataType_t
 
     .. cpp:type:: reduction_op_type = ncclRedOp_t
 
+    .. cpp:type:: size_type = int
+
     .. cpp:type:: rank_type = int
 
 
@@ -62,7 +68,7 @@ Types implementing the ``CommunicationSpace`` concept
     Specifies that a type ``T`` is a Kokkos Comm reduction operator.
 
 Types implementing the ``ReductionOperator`` concept
-----------------------------------------------------
+""""""""""""""""""""""""""""""""""""""""""""""""""""
 
 .. cpp:class:: BAnd
 

docs/api/core_recv.cpp

@@ -1,24 +1,22 @@
-#include "KokkosComm/KokkosComm.hpp"
+#include <Kokkos_Core.hpp>
+#include <KokkosComm/KokkosComm.hpp>
 
-// Define the execution space and transport
-using ExecSpace = Kokkos::DefaultExecutionSpace;
-using CommSpace = DefaultCommunicationSpace;
-
-// Source rank
-int src = 1;
-
-// Create a handle
-KokkosComm::Handle<> handle; // Same as Handle<Execspace, CommSpace>
+// Create an execution space instance
+auto exec = Kokkos::DefaultExecutionSpace();
+// Create a communicator
+auto comm = KokkosComm::Communicator<>::duplicate_from_raw(raw_comm_handle, exec).value();
 
 // Allocate a view to receive the data
 Kokkos::View<double*> data("recv_view", 100);
 
+// Source rank
+int src_rank = 1;
+
 // Initiate a non-blocking receive with a handle
-auto req1 = recv(handle, data, src);
+auto req1 = KokkosComm::recv(comm, data, src_rank);
 
-// Initiate a non-blocking receive with a default handle
-auto req2 = recv(data, src);
+// Simulate a blocking receive by waiting immediately
+KokkosComm::recv(comm, data, src_rank).wait();
 
-// Wait for the requests to complete (assuming a wait function exists)
+// Wait for a requests to complete
 KokkosComm::wait(req1);
-KokkosComm::wait(req2);

docs/api/core_send.cpp

@@ -1,29 +1,28 @@
-#include "KokkosComm/KokkosComm.hpp"
+#include <Kokkos_Core.hpp>
+#include <KokkosComm/KokkosComm.hpp>
 
-// Define the execution space and transport
-using ExecSpace = Kokkos::DefaultExecutionSpace;
-using CommSpace = DefaultCommunicationSpace;
+// Create an execution space instance
+auto exec = Kokkos::DefaultExecutionSpace();
+// Create a communicator
+auto comm = KokkosComm::Communicator<>::duplicate_from_raw(raw_comm_handle, exec).value();
 
 // Create a Kokkos view
-Kokkos::View<double*> data("data", 100);
+Kokkos::View<double*> data("send_data", 100);
 
 // Fill the view with some data
-Kokkos::parallel_for("fill_data", Kokkos::RangePolicy<ExecSpace>(0, 100), KOKKOS_LAMBDA(int i) {
-  data(i) = static_cast<double>(i);
-});
+Kokkos::parallel_for(
+    "fill_data", Kokkos::RangePolicy(exec, 0, 100), KOKKOS_LAMBDA(int i) { data(i) = static_cast<double>(i); }
+);
+exec.fence();
 
 // Destination rank
-int dest = 1;
-
-// Create a handle
-KokkosComm::Handle<> handle; // Same as Handle<Execspace, CommSpace>
+int dst_rank = 1;
 
 // Initiate a non-blocking send with a handle
-auto req1 = send(handle, data, dest);
+auto req1 = KokkosComm::send(comm, data, dst_rank);
 
-// Initiate a non-blocking send with a default handle
-auto req2 = send(data, dest);
+// Simulate a blocking send by waiting immediately
+KokkosComm::send(comm, data, dst_rank).wait();
 
-// Wait for the requests to complete (assuming a wait function exists)
+// Wait for a request to complete
 KokkosComm::wait(req1);
-KokkosComm::wait(req2);

docs/api/mpi.rst

@@ -79,7 +79,7 @@ Point-to-point
 
 
 .. cpp:function:: template <CommMode SendMode, KokkosExecutionSpace ExecSpace, KokkosView SendView> \
-                  auto isend(Handle<ExecSpace, Mpi> &h, const SendView &sv, int dest, int tag) -> Request<MpiSpace>
+                  auto isend(Communicator<MpiSpace, ExecSpace> &h, const SendView &sv, int dest, int tag) -> Request<MpiSpace>
 
     Initiates a non-blocking send operation.
 

docs/dev/impl_comm_space.rst

@@ -33,8 +33,8 @@ For example, for the MPI communication space, we define the following:
 To let core API functions know that your communication space is something KokkosComm can use to dispatch messages, you also need to declare the ``Impl::is_communication_space`` specialization using the ``CommunicationSpace`` concept.
 
 
-Partial specialization of ``Handle``
-=====================================
+Partial specialization of ``Communicator``
+==========================================
 
 .. attention:: Section in construction...
 
@@ -49,13 +49,13 @@ For example, for the MPI communication space handle, we define the following:
     namespace KokkosComm {
 
     template <KokkosExecutionSpace ExecSpace>
-    class Handle<ExecSpace, Mpi> { /* ... */ };
+    class Communicator<MpiSpace, ExecSpace> { /* ... */ };
 
     } // end KokkosComm
 
 
-Partial specialization of ``Req``
-===================================
+Partial specialization of ``Request``
+=====================================
 
 .. attention:: Section in construction...
 
@@ -92,7 +92,7 @@ The core API functions are actually implemented by partial specializations of st
 
 
 In the above, ``CommSpace`` is a type that represents the communication space implementation.
-For example, for the MPI communication space, we create a partial specialization of that struct template (notice fewer template parameters and the use of the ``Mpi`` "tag" struct):
+For example, for the MPI communication space, we create a partial specialization of that struct template (notice fewer template parameters and the use of the ``MpiSpace`` "tag" struct):
 
 .. code-block:: cpp
 
@@ -126,7 +126,7 @@ An asynchronous/non-blocking message send:
 
     template <KokkosView SendView, KokkosExecutionSpace ExecSpace>
     struct Send<SendView, ExecSpace, MyCommSpace> {
-      static auto execute(Handle<ExecSpace, MyCommSpace> &h, const SendView &sv, int dest) -> Request<MyCommSpace> {
+      static auto execute(Communicator<MyCommSpace, ExecSpace> &h, const SendView &sv, int dest) -> Request<MyCommSpace> {
         // actual implementation of `send` with your communication backend
       }
     };
@@ -148,31 +148,9 @@ An asynchronous/non-blocking message receive.
 
     template <KokkosView RecvView, KokkosExecutionSpace ExecSpace>
     struct Recv<RecvView, ExecSpace, MyCommSpace> {
-      static auto execute(Handle<ExecSpace, MyCommSpace> &h, const RecvView &sv, int src) -> Request<MyCommSpace> {
+      static auto execute(Communicator<MyCommSpace, ExecSpace> &h, const RecvView &sv, int src) -> Request<MyCommSpace> {
         // actual implementation of `recv` with your communication backend
       }
     };
 
     } // end KokkosComm::Impl
-
-
-``Barrier`` concept
-^^^^^^^^^^^^^^^^^^^
-
-A global barrier.
-
-.. code-block:: cpp
-
-    #include "KokkosComm/concepts.hpp"
-    #include "my_comm_space.hpp"
-
-    namespace KokkosComm::Impl {
-
-    template <KokkosExecutionSpace ExecSpace>
-    struct Recv<ExecSpace, MyCommSpace> {
-      static auto execute(Handle<ExecSpace, MyCommSpace> &&h) -> Request<MyCommSpace> {
-        // actual implementation of `barrier` with your communication backend
-      }
-    };
-
-    } // end KokkosComm::Impl

perf_tests/mpi/test_2d_halo.cpp

@@ -1,6 +1,8 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 // SPDX-FileCopyrightText: Copyright Contributors to the Kokkos project
 
+#include "KokkosComm/fwd.hpp"
+#include "KokkosComm/mpi/mpi_space.hpp"
 #include "test_utils.hpp"
 
 #include <iostream>
@@ -10,9 +12,10 @@
 void noop(benchmark::State, MPI_Comm) {}
 
 template <typename Space, typename View>
-void send_recv(benchmark::State&, MPI_Comm comm, const Space& space, int nx, int ny, int rx, int ry, int rs,
-               const View& v) {
-  KokkosComm::Handle<> h{space, comm};
+void send_recv(
+    benchmark::State&, MPI_Comm comm, const Space& space, int nx, int ny, int rx, int ry, int rs, const View& v
+) {
+  auto h = KokkosComm::Communicator<KokkosComm::MpiSpace, Space>::from_raw(comm, space);
 
   // 2D index of nbrs in minus and plus direction (periodic)
   const int xm1 = (rx + rs - 1) % rs;
@@ -73,8 +76,9 @@ void benchmark_2dhalo(benchmark::State& state) {
     // grid of elements, each with 3 properties, and a radius-1 halo
     grid_type grid("", nx + 2, ny + 2, nprops);
     while (state.KeepRunning()) {
-      do_iteration(state, MPI_COMM_WORLD, send_recv<Kokkos::DefaultExecutionSpace, grid_type>, space, nx, ny, rx, ry,
-                   rs, grid);
+      do_iteration(
+          state, MPI_COMM_WORLD, send_recv<Kokkos::DefaultExecutionSpace, grid_type>, space, nx, ny, rx, ry, rs, grid
+      );
     }
   } else {
     while (state.KeepRunning()) {

perf_tests/mpi/test_osu_latency.cpp

@@ -5,12 +5,15 @@
 // Copyright (c) 2002-2024 the Network-Based Computing Laboratory
 // (NBCL), The Ohio State University.
 
+#include "KokkosComm/mpi/mpi_space.hpp"
 #include "test_utils.hpp"
 
 #include <KokkosComm/KokkosComm.hpp>
 
 template <typename Space, typename View>
-void osu_latency_Kokkos_Comm_sendrecv(benchmark::State &, MPI_Comm, KokkosComm::Handle<> &h, const View &v) {
+void osu_latency_Kokkos_Comm_sendrecv(
+    benchmark::State &, MPI_Comm, KokkosComm::Communicator<KokkosComm::MpiSpace, Space> &h, const View &v
+) {
   if (h.rank() == 0) {
     KokkosComm::wait(KokkosComm::send(h, v, 1));
   } else if (h.rank() == 1) {
@@ -19,7 +22,9 @@ void osu_latency_Kokkos_Comm_sendrecv(benchmark::State &, MPI_Comm, KokkosComm::
 }
 
 void benchmark_osu_latency_KokkosComm_sendrecv(benchmark::State &state) {
-  KokkosComm::Handle<> h;
+  auto h = KokkosComm::Communicator<KokkosComm::MpiSpace, Kokkos::DefaultExecutionSpace>::from_raw(
+      MPI_COMM_WORLD, Kokkos::DefaultExecutionSpace()
+  );
   if (h.size() != 2) {
     state.SkipWithError("benchmark_osu_latency_KokkosComm needs exactly 2 ranks");
   }
@@ -28,14 +33,15 @@ void benchmark_osu_latency_KokkosComm_sendrecv(benchmark::State &state) {
   view_type a("A", state.range(0));
 
   while (state.KeepRunning()) {
-    do_iteration(state, h.mpi_comm(), osu_latency_Kokkos_Comm_sendrecv<Kokkos::DefaultExecutionSpace, view_type>, h, a);
+    do_iteration(state, h.comm(), osu_latency_Kokkos_Comm_sendrecv<Kokkos::DefaultExecutionSpace, view_type>, h, a);
   }
   state.counters["bytes"] = a.size() * 2;
 }
 
 template <typename Space, typename View>
-void osu_latency_Kokkos_Comm_mpi_sendrecv(benchmark::State &, MPI_Comm comm, const Space &space, int rank,
-                                          const View &v) {
+void osu_latency_Kokkos_Comm_mpi_sendrecv(
+    benchmark::State &, MPI_Comm comm, const Space &space, int rank, const View &v
+) {
   if (rank == 0) {
     KokkosComm::mpi::send(space, v, 1, 0, comm);
   } else if (rank == 1) {
@@ -56,8 +62,10 @@ void benchmark_osu_latency_Kokkos_Comm_mpi_sendrecv(benchmark::State &state) {
   view_type a("A", state.range(0));
 
   while (state.KeepRunning()) {
-    do_iteration(state, MPI_COMM_WORLD, osu_latency_Kokkos_Comm_mpi_sendrecv<Kokkos::DefaultExecutionSpace, view_type>,
-                 space, rank, a);
+    do_iteration(
+        state, MPI_COMM_WORLD, osu_latency_Kokkos_Comm_mpi_sendrecv<Kokkos::DefaultExecutionSpace, view_type>, space,
+        rank, a
+    );
   }
   state.counters["bytes"] = a.size() * 2;
 }
@@ -66,12 +74,16 @@ template <typename View>
 void osu_latency_MPI_isendirecv(benchmark::State &, MPI_Comm comm, int rank, const View &v) {
   MPI_Request sendreq, recvreq;
   if (rank == 0) {
-    MPI_Irecv(v.data(), v.size(), KokkosComm::datatype<KokkosComm::MpiSpace, typename View::value_type>(), 1, 0, comm,
-              &recvreq);
+    MPI_Irecv(
+        v.data(), v.size(), KokkosComm::datatype<KokkosComm::MpiSpace, typename View::value_type>(), 1, 0, comm,
+        &recvreq
+    );
     MPI_Wait(&recvreq, MPI_STATUS_IGNORE);
   } else if (rank == 1) {
-    MPI_Isend(v.data(), v.size(), KokkosComm::datatype<KokkosComm::MpiSpace, typename View::value_type>(), 0, 0, comm,
-              &sendreq);
+    MPI_Isend(
+        v.data(), v.size(), KokkosComm::datatype<KokkosComm::MpiSpace, typename View::value_type>(), 0, 0, comm,
+        &sendreq
+    );
     MPI_Wait(&sendreq, MPI_STATUS_IGNORE);
   }
 }
@@ -96,8 +108,10 @@ void benchmark_osu_latency_MPI_isendirecv(benchmark::State &state) {
 template <typename View>
 void osu_latency_MPI_sendrecv(benchmark::State &, MPI_Comm comm, int rank, const View &v) {
   if (rank == 0) {
-    MPI_Recv(v.data(), v.size(), KokkosComm::datatype<KokkosComm::MpiSpace, typename View::value_type>(), 1, 0, comm,
-             MPI_STATUS_IGNORE);
+    MPI_Recv(
+        v.data(), v.size(), KokkosComm::datatype<KokkosComm::MpiSpace, typename View::value_type>(), 1, 0, comm,
+        MPI_STATUS_IGNORE
+    );
   } else if (rank == 1) {
     MPI_Send(v.data(), v.size(), KokkosComm::datatype<KokkosComm::MpiSpace, typename View::value_type>(), 0, 0, comm);
   }

perf_tests/mpi/test_utils.hpp

@@ -10,12 +10,12 @@
 
 // F is a function that takes (state, MPI_Comm, args...)
 template <typename F, typename... Args>
-void do_iteration(benchmark::State &state, MPI_Comm comm, F &&func, Args... args) {
+void do_iteration(benchmark::State& state, MPI_Comm comm, F&& func, Args&&... args) {
   using Clock    = std::chrono::steady_clock;
   using Duration = std::chrono::duration<double>;
 
   auto start = Clock::now();
-  func(state, comm, args...);
+  func(state, comm, std::forward<Args>(args)...);
   Duration elapsed = Clock::now() - start;
 
   double max_elapsed_second;