Skip to content

Commit

Permalink
Finalize implementation of AMReX-Codes#1 ParallelCopy
Browse files Browse the repository at this point in the history
  • Loading branch information
@maikel
maikel committed Mar 5, 2021
1 parent 8dcc1ff commit b952187
Showing 1 changed file with 48 additions and 27 deletions.
75 changes: 48 additions & 27 deletions Src/Base/AMReX_NonLocalBCImpl.H
View file
Original file line number Diff line number Diff line change
Expand Up @@ -363,13 +363,13 @@ AMREX_STATIC_ASSERT_NO_MESSAGE((SwapComponents{0, 1}(2) == 2));

template <class FAB, class DTOS = Identity, class Proj = Identity>
EnableIf_t<IsBaseFab<FAB>() && IsCallableR<Dim3, DTOS, Dim3>() && IsFabProjection<Proj, FAB>()>
local_copy_cpu(FabArray<FAB>& dest, const FabArray<FAB>& src, int dcomp, int scomp, int ncomp,
FabArrayBase::CopyComTagsContainer const& local_tags, DTOS dtos = DTOS{},
Proj proj = Proj{}) noexcept {
local_copy_cpu (FabArray<FAB>& dest, const FabArray<FAB>& src, int dcomp, int scomp, int ncomp,
FabArrayBase::CopyComTagsContainer const& local_tags, DTOS dtos = DTOS{},
Proj proj = Proj{}) noexcept {
#ifdef AMREX_USE_OMP
#pragma omp parallel for
#endif
for (const FabArrayBase::CopyComTag& tag : local_tags) {
for (auto const& tag : local_tags) {
auto const& sfab = src.const_array(tag.srcIndex);
auto const& dfab = dest.array(tag.dstIndex);
amrex::LoopConcurrentOnCpu(tag.dbox, ncomp, [=](int i, int j, int k, int n) noexcept {
Expand All @@ -381,10 +381,10 @@ local_copy_cpu(FabArray<FAB>& dest, const FabArray<FAB>& src, int dcomp, int sco

template <class FAB, class DTOS = Identity, class Proj = Identity>
EnableIf_t<IsBaseFab<FAB>() && IsCallableR<Dim3, DTOS, Dim3>() && IsFabProjection<Proj, FAB>()>
unpack_recv_buffer_cpu(FabArray<FAB>& mf, int scomp, int ncomp, Vector<char*> const& recv_data,
Vector<std::size_t> const& recv_size,
Vector<FabArrayBase::CopyComTagsContainer const*> const& recv_cctc,
DTOS dtos = DTOS{}, Proj proj = Proj{}) noexcept {
unpack_recv_buffer_cpu (FabArray<FAB>& mf, int scomp, int ncomp, Vector<char*> const& recv_data,
Vector<std::size_t> const& recv_size,
Vector<FabArrayBase::CopyComTagsContainer const*> const& recv_cctc,
DTOS dtos = DTOS{}, Proj proj = Proj{}) noexcept {
amrex::ignore_unused(recv_size);

const int N_rcvs = recv_cctc.size();
Expand Down Expand Up @@ -527,6 +527,11 @@ struct CommHandler {
#endif
};

#ifdef AMREX_USE_MPI
void PostRecvs(CommData& comm, int mpi_tag);
void PostSends(CommData& comm, int mpi_tag);
#endif

////////////////////////////////////////////////////////////////////////////////////
// [concept.DataPacking]
//
Expand Down Expand Up @@ -632,37 +637,36 @@ static_assert(IsDataPacking<PackComponents, FArrayBox>(),
//
template <typename DTOS = Identity, typename FabProj = Identity>
struct ApplyDtosAndProjectionOnReciever : PackComponents {
constexpr ApplyDtosAndProjectionOnReciever() = default;
constexpr ApplyDtosAndProjectionOnReciever(const PackComponents& components, DTOS dtos_, FabProj proj_)
: PackComponents(components), dtos(std::move(dtos_)), proj(std::move(proj_)) {}

AMREX_NO_UNIQUE_ADDRESS DTOS dtos;
AMREX_NO_UNIQUE_ADDRESS FabProj proj;

static_assert(IsIndexMapping<DTOS>(), "DTOS needs to be an index map");
static_assert(IsIndexMapping<DTOS>(), "DTOS needs to be an index map");
};

template <typename FAB, typename DTOS, typename FabProj>
amrex::EnableIf_t<IsBaseFab<FAB>() && IsIndexMapping<DTOS>() && IsFabProjection<FabProj, FAB>()>
LocalCopy(const ApplyDtosAndProjectionOnReciever<DTOS, FabProj>& packing, FabArray<FAB>& dest,
const FabArray<FAB>& src, const FabArrayBase::CopyComTagsContainer& local_tags) {
LocalCopy (const ApplyDtosAndProjectionOnReciever<DTOS, FabProj>& packing, FabArray<FAB>& dest,
const FabArray<FAB>& src, const FabArrayBase::CopyComTagsContainer& local_tags) {
local_copy_cpu(dest, src, packing.dest_component, packing.src_component, packing.n_components,
local_tags, packing.dtos, packing.proj);
}

#ifdef AMREX_USE_MPI
template <typename FAB, typename DTOS, typename FabProj>
amrex::EnableIf_t<IsBaseFab<FAB>() && IsIndexMapping<DTOS>() && IsFabProjection<FabProj, FAB>()>
UnpackRecvBuffers(const ApplyDtosAndProjectionOnReciever<DTOS, FabProj>& packing,
FabArray<FAB>& dest, const CommData& comm) {
UnpackRecvBuffers (const ApplyDtosAndProjectionOnReciever<DTOS, FabProj>& packing,
FabArray<FAB>& dest, const CommData& comm) {
unpack_recv_buffer_cpu(dest, packing.dest_component, packing.n_components, comm.data, comm.size,
comm.cctc, packing.dtos, packing.proj);
}
#endif

static_assert(IsDataPacking<ApplyDtosAndProjectionOnReciever<>, FArrayBox>(),
"ApplyDtosAndProjectionOnReciever is expected to satisfy the DataPacking concept.");

#ifdef AMREX_USE_MPI
void PostRecvs(CommData& comm, int mpi_tag);
void PostSends(CommData& comm, int mpi_tag);
#endif
"ApplyDtosAndProjectionOnReciever<> is expected to satisfy the DataPacking concept.");

/// Initiate recv and send calls for MPI and immediately return without doing any work.
///
Expand All @@ -678,9 +682,9 @@ void PostSends(CommData& comm, int mpi_tag);
template <typename FAB, typename DataPacking,
typename = EnableIf_t<IsBaseFab<FAB>::value>,
typename = EnableIf_t<IsDataPacking<DataPacking, FAB>::value>>
AMREX_NODISCARD CommHandler ParallelCopy_nowait(FabArray<FAB>& dest, const FabArray<FAB>& src,
const FabArrayBase::CommMetaData& cmd,
const DataPacking& data_packing) {
AMREX_NODISCARD CommHandler
ParallelCopy_nowait (FabArray<FAB>& dest, const FabArray<FAB>& src,
const FabArrayBase::CommMetaData& cmd, const DataPacking& data_packing) {
CommHandler handler{};
#ifdef AMREX_USE_MPI
if (ParallelContext::NProcsSub() == 1) {
Expand Down Expand Up @@ -710,8 +714,8 @@ AMREX_NODISCARD CommHandler ParallelCopy_nowait(FabArray<FAB>& dest, const FabAr

template <typename FAB, typename DataPacking>
EnableIf_t<IsBaseFab<FAB>() && IsDataPacking<DataPacking, FAB>()>
ParallelCopy_finish(FabArray<FAB>& dest, const FabArray<FAB>& src, CommHandler handler,
const FabArrayBase::CommMetaData& cmd, const DataPacking& data_packing) {
ParallelCopy_finish (FabArray<FAB>& dest, const FabArray<FAB>& src, CommHandler handler,
const FabArrayBase::CommMetaData& cmd, const DataPacking& data_packing) {
// If any FabArray is empty we have nothing to do.
if (dest.size() == 0) {
return;
Expand Down Expand Up @@ -803,11 +807,24 @@ struct MultiBlockCommMetaData : FabArrayBase::CommMetaData {
//! @}
};

template <typename FAB, typename DTOS, typename Proj>
template <typename FAB, typename DTOS = Identity, typename Proj = Identity>
EnableIf_t<IsBaseFab<FAB>() && IsIndexMapping<DTOS>() && IsFabProjection<Proj, FAB>()>
ParallelCopy(FabArray<FAB>& dest, const Box& destbox, const FabArray<FAB>& src, int destcomp, int srccomp, int numcomp, DTOS dtos = DTOS{}, Proj proj = Proj{})
{
ParallelCopy (FabArray<FAB>& dest, const Box& destbox, const FabArray<FAB>& src,
const MultiBlockCommMetaData& cmd, int destcomp, int srccomp, int numcomp,
DTOS dtos = DTOS{}, Proj proj = Proj{}) {
ApplyDtosAndProjectionOnReciever<DTOS, Proj> packing{PackComponents{destcomp, srccomp, numcomp},
std::move(dtos), std::move(proj)};
CommHandler handler = ParallelCopy_nowait(dest, src, cmd, packing);
ParallelCopy_finish(dest, src, std::move(handler), cmd, packing);
}

template <typename FAB, typename DTOS = Identity, typename Proj = Identity>
EnableIf_t<IsBaseFab<FAB>() && IsIndexMapping<DTOS>() && IsFabProjection<Proj, FAB>()>
ParallelCopy (FabArray<FAB>& dest, const Box& destbox, const FabArray<FAB>& src, int destcomp,
int srccomp, int numcomp, DTOS dtos = DTOS{}, Proj proj = Proj{}) {
MultiBlockCommMetaData cmd(dest, destbox, src, dtos);
ParallelCopy(dest, destbox, src, cmd, destcomp, srccomp, numcomp, std::move(dtos),
std::move(proj));
}

template <class FAB, class DTOS>
Expand Down Expand Up @@ -1070,6 +1087,10 @@ void FillPolar(FabArray<FArrayBox>& mf, int scomp, int ncomp, IntVect const& ngh
extern template
void FillPolar (FabArray<FArrayBox>& mf, Box const& domain);

extern template
void ParallelCopy(FabArray<FArrayBox>& dest, const Box& destbox, const FabArray<FArrayBox>& src, int destcomp,
int srccomp, int numcomp, Identity, Identity);

}}


Expand Down

0 comments on commit b952187

Please sign in to comment.