test_parallel_for.cpp

Go to the documentation of this file.

#include "kokkos-utils/callbacks/ConjunctionMatcher.hpp"
#include "kokkos-utils/callbacks/RecorderListener.hpp"
#include "kokkos-utils/tests/scoped/callbacks/Manager.hpp"
 
#include "kokkos-execution/graph.hpp"
#include "kokkos-execution/impl/event.hpp"
 
#include "tests/graph/events.hpp"
#include "tests/utils/callback_matchers.hpp"
#include "tests/utils/category.hpp"
#include "tests/utils/check_node_type.hpp"
#include "tests/utils/functors/load_check_add.hpp"
#include "tests/utils/functors/no_op.hpp"
#include "tests/utils/functors/sum_indices.hpp"
#include "tests/utils/functors/tag_dispatch.hpp"
#include "tests/utils/graph_context.hpp"
#include "tests/utils/just_stopped.hpp"
#include "tests/utils/sink_receiver.hpp"
#include "tests/utils/sync_wait.hpp"

 
namespace Tests::GraphImpl {
 
using namespace Kokkos::utils::callbacks;
 
class TEST_CATEGORY(ParallelForTest)
    : public Tests::Utils::GraphContextTest<TEST_EXECUTION_SPACE>
    , public Kokkos::utils::tests::scoped::callbacks::Manager {
   public:
    using recorder_listener_t = RecorderListener<
        ConjunctionMatcher<EventDiscardMatcher<TEST_EXECUTION_SPACE>, GraphEventDiscardMatcher<TEST_EXECUTION_SPACE>>,
        BeginFenceEvent,
        BeginParallelForEvent,
        AllocateDataEvent,
        DeallocateDataEvent,
        Kokkos::Execution::Impl::RecordEvent,
        Kokkos::Execution::Impl::WaitEvent,
        Kokkos::Execution::GraphImpl::GraphAddNodeEvent,
        Kokkos::Execution::GraphImpl::GraphCreateEvent,
        Kokkos::Execution::GraphImpl::GraphInstantiateEvent,
        Kokkos::Execution::GraphImpl::GraphSubmitEvent
    >;
};

template <
    template <typename...> class SndrAdptr,
    bool ExptIsDispatchingSender,
    template <typename> typename ExptTransformSenderType
>
consteval bool test_sndr_traits() {
    using schd_sndr_t = typename TEST_CATEGORY(ParallelForTest)::schedule_sender_t;

    using label_t = std::string;
    using functor_t = Tests::Utils::Functors::SumIndices<typename TEST_CATEGORY(ParallelForTest)::view_s_t>;
    using policy_t = Kokkos::RangePolicy<TEST_EXECUTION_SPACE>;
    using pfor_sndr_t = SndrAdptr<schd_sndr_t, TEST_EXECUTION_SPACE, label_t, functor_t, policy_t>;

    static_assert(Kokkos::Execution::GraphImpl::graph_completing_sender<pfor_sndr_t>);

    static_assert(Kokkos::Execution::Impl::dispatching_sender<pfor_sndr_t> == ExptIsDispatchingSender);

    static_assert(!stdexec::dependent_sender<pfor_sndr_t>);
    static_assert(
        stdexec::get_completion_signatures<pfor_sndr_t>()
        == stdexec::completion_signatures<stdexec::set_value_t(), stdexec::set_error_t(std::exception_ptr)>{});

    static_assert(std::same_as<
                  stdexec::__completion_domain_of_t<stdexec::set_value_t, pfor_sndr_t, stdexec::env<>>,
                  Kokkos::Execution::GraphImpl::Domain
    >);

    static_assert(std::same_as<
                  Kokkos::Execution::Impl::completion_scheduler_of_t<stdexec::set_value_t, pfor_sndr_t>,
                  Kokkos::Execution::GraphImpl::Scheduler<TEST_EXECUTION_SPACE>
    >);

    static_assert(stdexec::sender_to<pfor_sndr_t, Tests::Utils::SinkReceiver>);
 
    static_assert(std::same_as<
                  stdexec::transform_sender_result_t<pfor_sndr_t, stdexec::env_of_t<Tests::Utils::SinkReceiver>>,
                  typename ExptTransformSenderType<Kokkos::Execution::GraphImpl::ParallelForSender< // NOLINT(modernize-type-traits)
                      schd_sndr_t,
                      Kokkos::Execution::GraphImpl::node_props_t<TEST_EXECUTION_SPACE, label_t>,
                      functor_t,
                      policy_t
                  >>::type
    >);

    static_assert(!stdexec::__nothrow_connectable<pfor_sndr_t, Tests::Utils::SinkReceiver>);
 
    return true;
}
 
template <typename Sndr, typename Exec, typename Label, typename Functor, typename Policy>
using pfor_generic_t = Kokkos::Execution::Impl::ParallelForSender<
    Sndr,
    Kokkos::Execution::Impl::ParallelForData<Functor, Policy, Kokkos::Execution::Impl::Label<Label>>
>;
 
template <typename Sndr, typename Exec, typename Label, typename Functor, typename Policy>
using pfor_impl_t = Kokkos::Execution::GraphImpl::ParallelForSender<
    Sndr,
    Kokkos::Execution::GraphImpl::node_props_t<Exec, Label>,
    Functor,
    Policy
>;
 
static_assert(test_sndr_traits<pfor_generic_t, true, std::type_identity>());
static_assert(test_sndr_traits<pfor_impl_t, false, std::add_rvalue_reference>());

template <typename ViewType>
consteval bool test_closure_traits() {
    using functor_t = Tests::Utils::Functors::SumIndices<ViewType>;
    using policy_t = Kokkos::RangePolicy<TEST_EXECUTION_SPACE>;
    using iter_space_t = Kokkos::dims<1, int>;
    using closure_policy_t = Kokkos::Execution::GraphImpl::ParallelForClosure<
        Kokkos::Execution::GraphImpl::node_props_t<TEST_EXECUTION_SPACE, std::string>,
        functor_t,
        policy_t
    >;
    using closure_iter_space_t = Kokkos::Execution::GraphImpl::ParallelForClosure<
        Kokkos::Execution::GraphImpl::node_props_t<TEST_EXECUTION_SPACE, std::string>,
        functor_t,
        iter_space_t
    >;

    static_assert(Kokkos::Execution::GraphImpl::Closure<closure_policy_t>);
    static_assert(Kokkos::Execution::GraphImpl::Closure<closure_iter_space_t>);
 
    static_assert(std::is_nothrow_move_constructible_v<closure_policy_t>);
    static_assert(std::is_nothrow_move_constructible_v<closure_iter_space_t>);
 
    return true;
}
 
static_assert(test_closure_traits<typename TEST_CATEGORY(ParallelForTest)::view_s_t>());
static_assert(test_closure_traits<std::span<int>>());

static_assert(Tests::Utils::check_continues_on_after_just_stopped<
              typename TEST_CATEGORY(ParallelForTest)::scheduler_t,
              Kokkos::Execution::parallel_for_t,
              Kokkos::RangePolicy<TEST_EXECUTION_SPACE>,
              Tests::Utils::Functors::NoOp<false, false, false>
>());

TEST_F(TEST_CATEGORY(ParallelForTest), parallel_for_schedule) {
    constexpr size_t size = 10;
 
    const view_s_t data(Kokkos::view_alloc(exec, "data - shared space"));
 
    const context_t gctx{exec};
 
    auto sndr =
        stdexec::schedule(gctx.get_scheduler())
        | Kokkos::Execution::parallel_for(
            std::format("{}: hello from pfor", Kokkos::Execution::Impl::display_string_of<TEST_EXECUTION_SPACE>()),
            size,
            Tests::Utils::Functors::SumIndices{.data = data});
 
    const auto recorded_events = Tests::Utils::record_sync_wait<recorder_listener_t>(std::move(sndr));
 
    ASSERT_THAT(
        recorded_events,
        testing::ElementsAre(
            MATCHER_FOR_GRAPH_CREATE(device_handle),
            MATCHER_FOR_GRAPH_ADDNODE(
                recorded_events.at(0), device_handle, MATCHER_FOR_GRAPH_ROOT_NODE_OF(recorded_events.at(0))),
            MATCHER_FOR_GRAPH_SUBMIT(exec, recorded_events.at(0)),
            MATCHER_FOR_BEGIN_FENCE(exec, dispatch_label(exec, "sync_wait"))));
 
    ASSERT_EQ(data(), size / 2 * (size - 1));
}

TEST_F(TEST_CATEGORY(ParallelForTest), parallel_for_starts_on) {
    constexpr size_t size = 10;
 
    const view_s_t data(Kokkos::view_alloc(exec, "data - shared space"));
 
    auto chain =
        stdexec::just()
        | Kokkos::Execution::parallel_for(
            std::format("{}: hello from pfor", Kokkos::Execution::Impl::display_string_of<TEST_EXECUTION_SPACE>()),
            Kokkos::RangePolicy<TEST_EXECUTION_SPACE>(0, size),
            Tests::Utils::Functors::SumIndices{.data = data});
 
    using chain_t = decltype(chain);
 
    static_assert(
        stdexec::get_completion_signatures<chain_t>()
        == stdexec::completion_signatures<stdexec::set_value_t(), stdexec::set_error_t(std::exception_ptr)>{});
    static_assert(!Tests::Utils::has_completion_scheduler_for<chain_t, stdexec::set_value_t>);
    static_assert(
        std::same_as<stdexec::__completion_domain_of_t<stdexec::set_value_t, chain_t>, stdexec::indeterminate_domain<>>);
 
    const context_t gctx{exec};
    auto starts_on = stdexec::starts_on(gctx.get_scheduler(), std::move(chain));
 
    using starts_on_t = decltype(starts_on);
 
    static_assert(stdexec::__has_eptr_completion<chain_t>);
    static_assert(!stdexec::dependent_sender<starts_on_t>);
    static_assert(
        stdexec::get_completion_signatures<starts_on_t>()
        == stdexec::completion_signatures<stdexec::set_value_t(), stdexec::set_error_t(std::exception_ptr)>{});
 
    const auto recorded_events = Tests::Utils::record_sync_wait<recorder_listener_t>(std::move(starts_on));
 
    ASSERT_THAT(recorded_events, testing::SizeIs(4));
 
    ASSERT_THAT(
        recorded_events,
        testing::ElementsAre(
            MATCHER_FOR_GRAPH_CREATE(device_handle),
            MATCHER_FOR_GRAPH_ADDNODE(
                recorded_events.at(0), device_handle, MATCHER_FOR_GRAPH_ROOT_NODE_OF(recorded_events.at(0))),
            MATCHER_FOR_GRAPH_SUBMIT(exec, recorded_events.at(0)),
            MATCHER_FOR_BEGIN_FENCE(exec, dispatch_label(exec, "after dispatch"))));
 
    ASSERT_EQ(data(), size / 2 * (size - 1));
}

TEST_F(TEST_CATEGORY(ParallelForTest), with_range_slice) {
    const view_s_t data(Kokkos::view_alloc(exec, "data - shared space"));
 
    const context_t gctx{exec};
 
    using functor_t = Tests::Utils::Functors::SumIndices<view_s_t>;
 
    using expt_node_t = Kokkos::Experimental::GraphNodeRef<
        TEST_EXECUTION_SPACE,
        Kokkos::Impl::GraphNodeKernelImpl<
            TEST_EXECUTION_SPACE,
            Kokkos::RangePolicy<TEST_EXECUTION_SPACE, Kokkos::Impl::IsGraphKernelTag>,
            functor_t,
            Kokkos::ParallelForTag
        >,
        Kokkos::Experimental::GraphNodeRef<TEST_EXECUTION_SPACE, Kokkos::Experimental::GraphNodeRootTag>
    >;
 
    const auto recorded_events = Tests::Utils::record_sync_wait<recorder_listener_t>(
        stdexec::schedule(gctx.get_scheduler())
        | Kokkos::Execution::parallel_for(Kokkos::range_slice<size_t, size_t>{3, 5}, functor_t{.data = data})
        | Tests::Utils::check_node_type<expt_node_t>());
 
    ASSERT_THAT(
        recorded_events,
        testing::ElementsAre(
            MATCHER_FOR_GRAPH_CREATE(device_handle),
            MATCHER_FOR_GRAPH_ADDNODE(
                recorded_events.at(0), device_handle, MATCHER_FOR_GRAPH_ROOT_NODE_OF(recorded_events.at(0))),
            MATCHER_FOR_GRAPH_SUBMIT(exec, recorded_events.at(0)),
            MATCHER_FOR_BEGIN_FENCE(exec, dispatch_label(exec, "sync_wait"))));
 
    ASSERT_EQ(data(), 3 + 4);
}

TEST_F(TEST_CATEGORY(ParallelForTest), parallel_for_schedule_tagged_operator) {
    const view_s_t data(Kokkos::view_alloc(exec, "data - shared space"));
 
    const context_t gctx{exec};
 
    using functor_t = Tests::Utils::Functors::TagDispatch<view_s_t>;
 
    const auto recorded_events = Tests::Utils::record_sync_wait<recorder_listener_t>(
        stdexec::schedule(gctx.get_scheduler())
        | Kokkos::Execution::parallel_for(
            Kokkos::RangePolicy<typename functor_t::Tag, TEST_EXECUTION_SPACE>(0, 1), functor_t{.data = data}));
 
    ASSERT_THAT(
        recorded_events,
        testing::ElementsAre(
            MATCHER_FOR_GRAPH_CREATE(device_handle),
            MATCHER_FOR_GRAPH_ADDNODE(
                recorded_events.at(0), device_handle, MATCHER_FOR_GRAPH_ROOT_NODE_OF(recorded_events.at(0))),
            MATCHER_FOR_GRAPH_SUBMIT(exec, recorded_events.at(0)),
            MATCHER_FOR_BEGIN_FENCE(exec, dispatch_label(exec, "sync_wait"))));
 
    ASSERT_EQ(data(), 1);
}
 
} // namespace Tests::GraphImpl