AMReX_ParticleTile.H

Go to the documentation of this file.

#ifndef AMREX_PARTICLETILE_H_
#define AMREX_PARTICLETILE_H_
#include <AMReX_Config.H>
 
#include <AMReX_Extension.H>
#include <AMReX_Particle.H>
#include <AMReX_ArrayOfStructs.H>
#include <AMReX_StructOfArrays.H>
#include <AMReX_Vector.H>
#include <AMReX_REAL.H>
#include <AMReX_RealVect.H>
 
#include <array>
#include <atomic>
#include <string>
#include <type_traits>
#include <vector>
 
 
namespace amrex {
 
// Forward Declaration
template <int NArrayReal, int NArrayInt>
struct ConstSoAParticle;
template <int NArrayReal, int NArrayInt>
struct SoAParticle;
 
template <typename T_ParticleType, int NArrayReal, int NArrayInt>
struct ConstParticleTileData;
 
template <typename T_ParticleType, int NArrayReal, int NArrayInt>
struct ParticleTileData
{
    static constexpr int NAR = NArrayReal;
    static constexpr int NAI = NArrayInt;
 
    using ParticleType = T_ParticleType;
    using ParticleRefType = T_ParticleType&;
    using Self = ParticleTileData<ParticleType, NAR, NAI>;
    using RealType = ParticleReal;
    using IntType = int;
 
    static constexpr int NStructReal = ParticleType::NReal;
    static constexpr int NStructInt = ParticleType::NInt;
 
    using SuperParticleType = Particle<NStructReal+NAR, NStructInt+NAI>;
 
    static constexpr bool is_particle_tile_data = true;
 
    Long m_size;
 
    using AOS_PTR = std::conditional_t<T_ParticleType::is_soa_particle,
                                       void * AMREX_RESTRICT, ParticleType * AMREX_RESTRICT>;
    AOS_PTR m_aos;
 
    uint64_t* m_idcpu;
    GpuArray<ParticleReal*, NAR> m_rdata;
    GpuArray<int*, NAI> m_idata;
 
    int m_num_runtime_real;
    int m_num_runtime_int;
    ParticleReal* AMREX_RESTRICT * AMREX_RESTRICT m_runtime_rdata;
    int* AMREX_RESTRICT * AMREX_RESTRICT m_runtime_idata;
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) pos (const int dir, const int index) const &
    {
        if constexpr(!ParticleType::is_soa_particle) {
            return this->m_aos[index].pos(dir);
        } else {
            return this->m_rdata[dir][index];
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) id (const int index) const &
    {
        if constexpr(!ParticleType::is_soa_particle) {
            return this->m_aos[index].id();
        } else {
            return ParticleIDWrapper(this->m_idcpu[index]);
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) cpu (const int index) const &
    {
        if constexpr(!ParticleType::is_soa_particle) {
            return this->m_aos[index].cpu();
        } else {
            return ParticleCPUWrapper(this->m_idcpu[index]);
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) idcpu (const int index) const &
    {
        if constexpr(ParticleType::is_soa_particle) {
            return this->m_idcpu[index];
        } else {
            return this->m_aos[index].idcpu();
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ParticleReal * rdata (const int attribute_index) const
    {
        return this->m_rdata[attribute_index];
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    int * idata (const int attribute_index) const
    {
        return this->m_idata[attribute_index];
    }
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) operator[] (const int index) const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos[index];
        } else {
            return SoAParticle<NAR, NAI>(*this, index);
        }
    }
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    void packParticleData (char* buffer, int src_index, std::size_t dst_offset,
                           const int* comm_real, const int * comm_int) const noexcept
    {
        AMREX_ASSERT(src_index < m_size);
        auto* dst = buffer + dst_offset;
        if constexpr (!ParticleType::is_soa_particle) {
            memcpy(dst, m_aos + src_index, sizeof(ParticleType));
            dst += sizeof(ParticleType);
        } else {
            memcpy(dst, m_idcpu + src_index, sizeof(uint64_t));
            dst += sizeof(uint64_t);
        }
        int array_start_index = 0;
        if constexpr (!ParticleType::is_soa_particle) {
            array_start_index = AMREX_SPACEDIM + NStructReal;
        }
        for (int i = 0; i < NAR; ++i)
        {
            if (comm_real[array_start_index + i])
            {
                memcpy(dst, m_rdata[i] + src_index, sizeof(ParticleReal));
                dst += sizeof(ParticleReal);
            }
        }
        int runtime_start_index  = array_start_index + NAR;
        for (int i = 0; i < m_num_runtime_real; ++i)
        {
            if (comm_real[runtime_start_index + i])
            {
                memcpy(dst, m_runtime_rdata[i] + src_index, sizeof(ParticleReal));
                dst += sizeof(ParticleReal);
            }
        }
        array_start_index  = 2 + NStructInt;
        for (int i = 0; i < NAI; ++i)
        {
            if (comm_int[array_start_index + i])
            {
                memcpy(dst, m_idata[i] + src_index, sizeof(int));
                dst += sizeof(int);
            }
        }
        runtime_start_index  = 2 + NStructInt + NAI;
        for (int i = 0; i < m_num_runtime_int; ++i)
        {
            if (comm_int[runtime_start_index + i])
            {
                memcpy(dst, m_runtime_idata[i] + src_index, sizeof(int));
                dst += sizeof(int);
            }
        }
    }
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    void unpackParticleData (const char* buffer, Long src_offset, int dst_index,
                             const int* comm_real, const int* comm_int) const noexcept
    {
        AMREX_ASSERT(dst_index < m_size);
        const auto* src = buffer + src_offset;
        if constexpr (!ParticleType::is_soa_particle) {
            memcpy(m_aos + dst_index, src, sizeof(ParticleType));
            src += sizeof(ParticleType);
        } else {
            memcpy(m_idcpu + dst_index, src, sizeof(uint64_t));
            src += sizeof(uint64_t);
        }
        int array_start_index = 0;
        if constexpr (!ParticleType::is_soa_particle) {
            array_start_index = AMREX_SPACEDIM + NStructReal;
        }
        if constexpr (NAR > 0) {
            for (int i = 0; i < NAR; ++i)
            {
                if (comm_real[array_start_index + i])
                {
                    memcpy(m_rdata[i] + dst_index, src, sizeof(ParticleReal));
                    src += sizeof(ParticleReal);
                }
            }
        }
        int runtime_start_index  = array_start_index + NAR;
        for (int i = 0; i < m_num_runtime_real; ++i)
        {
            if (comm_real[runtime_start_index + i])
            {
                memcpy(m_runtime_rdata[i] + dst_index, src, sizeof(ParticleReal));
                src += sizeof(ParticleReal);
            }
        }
        array_start_index  = 2 + NStructInt;
        if constexpr (NAI > 0) {
            for (int i = 0; i < NAI; ++i)
            {
                if (comm_int[array_start_index + i])
                {
                    memcpy(m_idata[i] + dst_index, src, sizeof(int));
                    src += sizeof(int);
                }
            }
        }
        runtime_start_index  = 2 + NStructInt + NAI;
        for (int i = 0; i < m_num_runtime_int; ++i)
        {
            if (comm_int[runtime_start_index + i])
            {
                memcpy(m_runtime_idata[i] + dst_index, src, sizeof(int));
                src += sizeof(int);
            }
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    SuperParticleType getSuperParticle (int index) const noexcept
        requires (!ParticleType::is_soa_particle)
    {
        AMREX_ASSERT(index < m_size);
        SuperParticleType sp;
        for (int i = 0; i < AMREX_SPACEDIM; ++i) {
            sp.pos(i) = m_aos[index].pos(i);
        }
        for (int i = 0; i < NStructReal; ++i) {
            sp.rdata(i) = m_aos[index].rdata(i);
        }
        if constexpr (NAR >0) {
            for (int i = 0; i < NAR; ++i) {
                sp.rdata(NStructReal+i) = m_rdata[i][index];
            }
        }
        sp.id() = m_aos[index].id();
        sp.cpu() = m_aos[index].cpu();
        for (int i = 0; i < NStructInt; ++i) {
            sp.idata(i) = m_aos[index].idata(i);
        }
        if constexpr (NAI > 0) {
            for (int i = 0; i < NAI; ++i) {
                sp.idata(NStructInt+i) = m_idata[i][index];
            }
        }
        return sp;
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    SuperParticleType getSuperParticle (int index) const noexcept
        requires (ParticleType::is_soa_particle)
    {
        AMREX_ASSERT(index < m_size);
        SuperParticleType sp;
        sp.m_idcpu = m_idcpu[index];
        for (int i = 0; i < AMREX_SPACEDIM; ++i) {sp.pos(i) = m_rdata[i][index];}
        for (int i = 0; i < NAR; ++i) {
            sp.rdata(i) = m_rdata[i][index];
        }
        for (int i = 0; i < NAI; ++i) {
            sp.idata(i) = m_idata[i][index];
        }
        return sp;
    }
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    void setSuperParticle (const SuperParticleType& sp, int index) const noexcept
        requires (!ParticleType::is_soa_particle)
    {
        for (int i = 0; i < AMREX_SPACEDIM; ++i) {
            m_aos[index].pos(i) = sp.pos(i);
        }
        for (int i = 0; i < NStructReal; ++i) {
            m_aos[index].rdata(i) = sp.rdata(i);
        }
        for (int i = 0; i < NAR; ++i) {
            m_rdata[i][index] = sp.rdata(NStructReal+i);
        }
        m_aos[index].id() = sp.id();
        m_aos[index].cpu() = sp.cpu();
        for (int i = 0; i < NStructInt; ++i) {
            m_aos[index].idata(i) = sp.idata(i);
        }
        for (int i = 0; i < NAI; ++i) {
            m_idata[i][index] = sp.idata(NStructInt+i);
        }
    }
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    void setSuperParticle (const SuperParticleType& sp, int index) const noexcept
        requires (ParticleType::is_soa_particle)
    {
        m_idcpu[index] = sp.m_idcpu;
        for (int i = 0; i < NAR; ++i) {
            m_rdata[i][index] = sp.rdata(i);
        }
        for (int i = 0; i < NAI; ++i) {
            m_idata[i][index] = sp.idata(i);
        }
    }
};
 
// SOA Particle Structure
template <int T_NArrayReal, int T_NArrayInt>
struct alignas(sizeof(double)) ConstSoAParticle : SoAParticleBase
{
    static constexpr int NArrayReal = T_NArrayReal;
    static constexpr int NArrayInt = T_NArrayInt;
    using StorageParticleType = SoAParticleBase;
    using ConstPTD = ConstParticleTileData<SoAParticleBase, NArrayReal, NArrayInt>;
    static constexpr bool is_soa_particle = true;
    static constexpr bool is_rtsoa_particle = false;
    static constexpr bool is_constsoa_particle = true;
 
    using RealType = ParticleReal;
    using IntType = int;
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ConstSoAParticle (ConstPTD const& ptd, long i) : // Note: should this be int instead?
        m_constparticle_tile_data(ptd), m_index(int(i))
    {
    }
 
    //static Long the_next_id;
 
    //functions to get id and cpu in the SOA data
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ConstParticleCPUWrapper cpu () const { return this->m_constparticle_tile_data.m_idcpu[m_index]; }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ConstParticleIDWrapper id () const { return this->m_constparticle_tile_data.m_idcpu[m_index]; }
 
    //functions to get positions of the particle in the SOA data
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    RealVect pos () const & {return RealVect(AMREX_D_DECL(this->m_constparticle_tile_data.m_rdata[0][m_index], this->m_constparticle_tile_data.m_rdata[1][m_index], this->m_constparticle_tile_data.m_rdata[2][m_index]));}
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    const RealType&  pos (int position_index) const &
    {
        AMREX_ASSERT(position_index < AMREX_SPACEDIM);
        return this->m_constparticle_tile_data.m_rdata[position_index][m_index];
    }
 
    static Long NextID ();
 
    static Long UnprotectedNextID ();
 
    static void NextID (Long nextid);
 
    private :
 
    static_assert(std::is_trivially_copyable<ConstPTD>(), "ParticleTileData is not trivially copyable");
 
    ConstPTD m_constparticle_tile_data;
    int m_index;
};
 
template <int T_NArrayReal, int T_NArrayInt>
struct alignas(sizeof(double)) SoAParticle : SoAParticleBase
{
    static constexpr int NArrayReal = T_NArrayReal;
    static constexpr int NArrayInt = T_NArrayInt;
    using StorageParticleType = SoAParticleBase;
    using PTD = ParticleTileData<SoAParticleBase, NArrayReal, NArrayInt>;
    static constexpr bool is_soa_particle = true;
    static constexpr bool is_rtsoa_particle = false;
    static constexpr bool is_constsoa_particle = false;
 
    using ConstType = ConstSoAParticle<T_NArrayReal, T_NArrayInt>;
    using RealType = ParticleReal;
    using IntType = int;
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    SoAParticle (PTD const& ptd, long i) : // Note: should this be int instead?
        m_particle_tile_data(ptd), m_index(int(i))
    {
    }
 
    static Long the_next_id;
 
    //functions to get id and cpu in the SOA data
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ParticleCPUWrapper cpu () & { return this->m_particle_tile_data.m_idcpu[m_index]; }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ParticleIDWrapper<> id () & { return this->m_particle_tile_data.m_idcpu[m_index]; }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    uint64_t& idcpu () & { return this->m_particle_tile_data.m_idcpu[m_index]; }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ConstParticleCPUWrapper cpu () const & { return this->m_particle_tile_data.m_idcpu[m_index]; }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    ConstParticleIDWrapper id () const & { return this->m_particle_tile_data.m_idcpu[m_index]; }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    const uint64_t& idcpu () const & { return this->m_particle_tile_data.m_idcpu[m_index]; }
 
    //functions to get positions of the particle in the SOA data
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    RealVect pos () const & {return RealVect(AMREX_D_DECL(this->m_particle_tile_data.m_rdata[0][m_index], this->m_particle_tile_data.m_rdata[1][m_index], this->m_particle_tile_data.m_rdata[2][m_index]));}
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    RealType& pos (int position_index) &
    {
        AMREX_ASSERT(position_index < AMREX_SPACEDIM);
        return this->m_particle_tile_data.m_rdata[position_index][m_index];
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    RealType pos (int position_index) const &
    {
        AMREX_ASSERT(position_index < AMREX_SPACEDIM);
        return this->m_particle_tile_data.m_rdata[position_index][m_index];
    }
 
    static Long NextID ();
 
    static Long UnprotectedNextID ();
 
    static void NextID (Long nextid);
 
private :
 
    static_assert(std::is_trivially_copyable<PTD>(), "ParticleTileData is not trivially copyable");
 
    PTD m_particle_tile_data;
    int m_index;
};
 
//template <int NArrayReal, int NArrayInt> Long ConstSoAParticle<NArrayReal, NArrayInt>::the_next_id = 1;
template <int NArrayReal, int NArrayInt> Long SoAParticle<NArrayReal, NArrayInt>::the_next_id = 1;
 
template <int NArrayReal, int NArrayInt>
Long
SoAParticle<NArrayReal, NArrayInt>::NextID ()
{
    Long next;
// we should be able to test on _OPENMP < 201107 for capture (version 3.1)
// but we must work around a bug in gcc < 4.9
#if defined(AMREX_USE_OMP) && defined(_OPENMP) && _OPENMP < 201307
#pragma omp critical (amrex_particle_nextid)
#elif defined(AMREX_USE_OMP)
#pragma omp atomic capture
#endif
    next = the_next_id++;
 
    if (next > LongParticleIds::LastParticleID) {
        amrex::Abort("SoAParticle<NArrayReal, NArrayInt>::NextID() -- too many particles");
    }
 
    return next;
}
 
template <int NArrayReal, int NArrayInt>
Long
SoAParticle<NArrayReal, NArrayInt>::UnprotectedNextID ()
{
    Long next = the_next_id++;
    if (next > LongParticleIds::LastParticleID) {
        amrex::Abort("SoAParticle<NArrayReal, NArrayInt>::NextID() -- too many particles");
    }
    return next;
}
 
template <int NArrayReal, int NArrayInt>
void
SoAParticle<NArrayReal, NArrayInt>::NextID (Long nextid)
{
    the_next_id = nextid;
}
 
template <typename T_ParticleType, int NArrayReal, int NArrayInt>
struct ConstParticleTileData
{
    static constexpr int NAR = NArrayReal;
    static constexpr int NAI = NArrayInt;
    using ParticleType = T_ParticleType;
    using ParticleRefType = T_ParticleType const&;
    using RealType = ParticleReal;
    using IntType = int;
 
    static constexpr int NStructReal = ParticleType::NReal;
    static constexpr int NStructInt = ParticleType::NInt;
 
    using SuperParticleType = Particle<NStructReal+NArrayReal, NStructInt+NArrayInt>;
 
    static constexpr bool is_particle_tile_data = true;
 
    Long m_size;
 
    using AOS_PTR = std::conditional_t<T_ParticleType::is_soa_particle,
                                       void const * AMREX_RESTRICT, ParticleType const * AMREX_RESTRICT>;
    AOS_PTR m_aos;
 
    const uint64_t* m_idcpu;
    GpuArray<const ParticleReal*, NArrayReal> m_rdata;
    GpuArray<const int*, NArrayInt > m_idata;
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) pos (const int dir, const int index) const &
    {
        if constexpr(!ParticleType::is_soa_particle) {
            return this->m_aos[index].pos(dir);
        } else {
            return this->m_rdata[dir][index];
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) id (const int index) const &
    {
        if constexpr(!ParticleType::is_soa_particle) {
            return this->m_aos[index].id();
        } else {
            return ConstParticleIDWrapper(this->m_idcpu[index]);
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) cpu (const int index) const &
    {
        if constexpr(!ParticleType::is_soa_particle) {
            return this->m_aos[index].cpu();
        } else {
            return ConstParticleCPUWrapper(this->m_idcpu[index]);
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) idcpu (const int index) const &
    {
        if constexpr(ParticleType::is_soa_particle) {
            return this->m_idcpu[index];
        } else {
            return this->m_aos[index].idcpu();
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    const ParticleReal * rdata (const int attribute_index) const
    {
        return this->m_rdata[attribute_index];
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    const int * idata (const int attribute_index) const
    {
        return this->m_idata[attribute_index];
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    decltype(auto) operator[] (const int index) const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos[index];
        } else {
            return ConstSoAParticle<NAR, NAI>(*this, index);
        }
    }
 
    int m_num_runtime_real;
    int m_num_runtime_int;
    const ParticleReal* AMREX_RESTRICT * AMREX_RESTRICT m_runtime_rdata;
    const int* AMREX_RESTRICT * AMREX_RESTRICT m_runtime_idata;
 
    AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    void packParticleData(char* buffer, int src_index, Long dst_offset,
                          const int* comm_real, const int * comm_int) const noexcept
    {
        AMREX_ASSERT(src_index < m_size);
        auto* dst = buffer + dst_offset;
        if constexpr (!ParticleType::is_soa_particle) {
            memcpy(dst, m_aos + src_index, sizeof(ParticleType));
            dst += sizeof(ParticleType);
        } else {
            memcpy(dst, m_idcpu + src_index, sizeof(uint64_t));
            dst += sizeof(uint64_t);
        }
        int array_start_index = 0;
        if constexpr (!ParticleType::is_soa_particle) {
            array_start_index = AMREX_SPACEDIM + NStructReal;
        }
        if constexpr (NArrayReal > 0) {
            for (int i = 0; i < NArrayReal; ++i)
            {
                if (comm_real[array_start_index + i])
                {
                    memcpy(dst, m_rdata[i] + src_index, sizeof(ParticleReal));
                    dst += sizeof(ParticleReal);
                }
            }
        }
        int runtime_start_index  = array_start_index + NArrayReal;
        for (int i = 0; i < m_num_runtime_real; ++i)
        {
            if (comm_real[runtime_start_index + i])
            {
                memcpy(dst, m_runtime_rdata[i] + src_index, sizeof(ParticleReal));
                dst += sizeof(ParticleReal);
            }
        }
        array_start_index  = 2 + NStructInt;
        if constexpr (NArrayInt > 0) {
            for (int i = 0; i < NArrayInt; ++i)
            {
                if (comm_int[array_start_index + i])
                {
                    memcpy(dst, m_idata[i] + src_index, sizeof(int));
                    dst += sizeof(int);
                }
            }
        }
        runtime_start_index  = 2 + NStructInt + NArrayInt;
        for (int i = 0; i < m_num_runtime_int; ++i)
        {
            if (comm_int[runtime_start_index + i])
            {
                memcpy(dst, m_runtime_idata[i] + src_index, sizeof(int));
                dst += sizeof(int);
            }
        }
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    SuperParticleType getSuperParticle (int index) const noexcept
        requires (!ParticleType::is_soa_particle)
    {
        AMREX_ASSERT(index < m_size);
        SuperParticleType sp;
        for (int i = 0; i < AMREX_SPACEDIM; ++i) {
            sp.pos(i) = m_aos[index].pos(i);
        }
        for (int i = 0; i < NStructReal; ++i) {
            sp.rdata(i) = m_aos[index].rdata(i);
        }
        if constexpr(NArrayReal > 0) {
            for (int i = 0; i < NArrayReal; ++i) {
                sp.rdata(NStructReal+i) = m_rdata[i][index];
            }
        }
        sp.id() = m_aos[index].id();
        sp.cpu() = m_aos[index].cpu();
        for (int i = 0; i < NStructInt; ++i) {
            sp.idata(i) = m_aos[index].idata(i);
        }
        if constexpr(NArrayInt > 0) {
            for (int i = 0; i < NArrayInt; ++i) {
                sp.idata(NStructInt+i) = m_idata[i][index];
            }
        }
        return sp;
    }
 
    [[nodiscard]] AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
    SuperParticleType getSuperParticle (int index) const noexcept
        requires (ParticleType::is_soa_particle)
    {
        AMREX_ASSERT(index < m_size);
        SuperParticleType sp;
        for (int i = 0; i < AMREX_SPACEDIM; ++i) {sp.pos(i) = m_rdata[i][index];}
        sp.m_idcpu = m_idcpu[index];
        for (int i = 0; i < NAR; ++i) {
            sp.rdata(i) = m_rdata[i][index];
        }
        for (int i = 0; i < NAI; ++i) {
            sp.idata(i) = m_idata[i][index];
        }
        return sp;
    }
};
 
struct ThisParticleTileHasNoParticleVector {};
 
struct ThisParticleTileHasNoAoS {
    using ParticleVector = ThisParticleTileHasNoParticleVector;
};
 
struct RuntimePtrCacheDirtyFlag
{
    RuntimePtrCacheDirtyFlag () noexcept = default;
 
    ~RuntimePtrCacheDirtyFlag () = default;
 
    RuntimePtrCacheDirtyFlag (RuntimePtrCacheDirtyFlag const&) noexcept
        : m_value(true)
    {}
 
    RuntimePtrCacheDirtyFlag (RuntimePtrCacheDirtyFlag&&) noexcept
        : m_value(true)
    {}
 
    RuntimePtrCacheDirtyFlag& operator= (RuntimePtrCacheDirtyFlag const&) noexcept
    {
        m_value.store(true, std::memory_order_relaxed);
        return *this;
    }
 
    RuntimePtrCacheDirtyFlag& operator= (RuntimePtrCacheDirtyFlag&&) noexcept
    {
        m_value.store(true, std::memory_order_relaxed);
        return *this;
    }
 
    void store (bool value, std::memory_order order = std::memory_order_seq_cst) noexcept
    {
        m_value.store(value, order);
    }
 
    [[nodiscard]] bool load (std::memory_order order = std::memory_order_seq_cst) const noexcept
    {
        return m_value.load(order);
    }
 
private:
    std::atomic<bool> m_value {true};
};
 
template <typename T_ParticleType, int NArrayReal, int NArrayInt,
          template<class> class Allocator=DefaultAllocator>
struct ParticleTile
{
    template <typename T>
    using AllocatorType = Allocator<T>;
 
    using ParticleType = T_ParticleType;
    static constexpr int NAR = NArrayReal;
    static constexpr int NAI = NArrayInt;
    using RealType = typename ParticleType::RealType;
 
    static constexpr int NStructReal = ParticleType::NReal;
    static constexpr int NStructInt = ParticleType::NInt;
 
    using SuperParticleType = Particle<NStructReal + NArrayReal, NStructInt + NArrayInt>;
 
    using AoS = std::conditional_t<
        ParticleType::is_soa_particle,
        ThisParticleTileHasNoAoS,
        ArrayOfStructs<ParticleType, Allocator>>;
    using ParticleVector = typename AoS::ParticleVector;
 
    using SoA = std::conditional_t<
        ParticleType::is_soa_particle,
        StructOfArrays<NArrayReal, NArrayInt, Allocator, true>,
        StructOfArrays<NArrayReal, NArrayInt, Allocator, false>>;
    using RealVector = typename SoA::RealVector;
    using IntVector = typename SoA::IntVector;
    using StorageParticleType = typename ParticleType::StorageParticleType;
 
    using ParticleTileDataType = ParticleTileData<StorageParticleType, NArrayReal, NArrayInt>;
    using ConstParticleTileDataType = ConstParticleTileData<StorageParticleType, NArrayReal, NArrayInt>;
 
    static constexpr bool has_polymorphic_allocator =
        IsPolymorphicArenaAllocator<Allocator<RealType>>::value;
 
    ParticleTile () = default;
 
#ifndef _WIN32  // workaround windows compiler bug
    ~ParticleTile () = default;
 
    ParticleTile (ParticleTile const&) = delete;
    ParticleTile (ParticleTile &&) noexcept = default;
 
    ParticleTile& operator= (ParticleTile const&) = delete;
    ParticleTile& operator= (ParticleTile &&) noexcept = default;
#endif
 
    void define (
        int a_num_runtime_real,
        int a_num_runtime_int,
        std::vector<std::string>* soa_rdata_names=nullptr,
        std::vector<std::string>* soa_idata_names=nullptr,
        Arena* a_arena=nullptr
    )
    {
        GetStructOfArrays().define(a_num_runtime_real, a_num_runtime_int, soa_rdata_names, soa_idata_names);
        m_runtime_r_ptrs.resize(a_num_runtime_real);
        m_runtime_i_ptrs.resize(a_num_runtime_int);
        m_runtime_r_cptrs.resize(a_num_runtime_real);
        m_runtime_i_cptrs.resize(a_num_runtime_int);
 
        if constexpr (has_polymorphic_allocator) {
            if (m_defined) {
                // it is not allowed to change the arena after the tile has been defined
                if constexpr (ParticleType::is_soa_particle) {
                    AMREX_ALWAYS_ASSERT_WITH_MESSAGE(
                        a_arena == GetStructOfArrays().GetIdCPUData().arena(),
                        "ParticleTile with PolymorphicArenaAllocator redefined with "
                        "different memory arena");
                } else {
                    AMREX_ALWAYS_ASSERT_WITH_MESSAGE(
                        a_arena == m_aos_tile().arena(),
                        "ParticleTile with PolymorphicArenaAllocator redefined with "
                        "different memory arena");
                }
            }
 
            AMREX_ALWAYS_ASSERT_WITH_MESSAGE(a_arena != nullptr,
                "ParticleTile with PolymorphicArenaAllocator defined with no memory arena! "
                "Make sure to call setArena() on the ParticleContainer before initialization or "
                "to pass an Arena to ParticleTile::define()");
 
            if constexpr (ParticleType::is_soa_particle) {
                GetStructOfArrays().GetIdCPUData().setArena(a_arena);
            } else {
                m_aos_tile().setArena(a_arena);
            }
            for (int j = 0; j < NumRealComps(); ++j) {
                GetStructOfArrays().GetRealData(j).setArena(a_arena);
            }
            for (int j = 0; j < NumIntComps(); ++j) {
                GetStructOfArrays().GetIntData(j).setArena(a_arena);
            }
        }
 
        m_defined = true;
        invalidateRuntimePtrCaches();
    }
 
    // Get id data
    decltype(auto) id (int index) & {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile[index].id();
        } else {
            return ParticleIDWrapper(m_soa_tile.GetIdCPUData()[index]);
        }
    }
 
    // const
    decltype(auto) id (int index) const & {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile[index].id();
        } else {
            return ConstParticleIDWrapper(m_soa_tile.GetIdCPUData()[index]);
        }
    }
 
    // Get cpu data
    decltype(auto) cpu (int index) & {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile[index].cpu();
        } else {
            return ParticleCPUWrapper(m_soa_tile.GetIdCPUData()[index]);
        }
    }
 
    // const
    decltype(auto) cpu (int index) const & {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile[index].cpu();
        } else {
            return ConstParticleCPUWrapper(m_soa_tile.GetIdCPUData()[index]);
        }
    }
 
    // Get positions data
    RealType& pos (int index, int position_index) & {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile[index].pos(position_index);
        } else {
            static_assert(NArrayReal == ParticleType::PTD::NAR, "ParticleTile mismatch in R");
            static_assert(NArrayInt == ParticleType::PTD::NAI, "ParticleTile mismatch in I");
            static_assert(0 == ParticleType::StorageParticleType::NReal, "ParticleTile 2 mismatch in R");
            static_assert(0 == ParticleType::StorageParticleType::NInt, "ParticleTile 2 mismatch in I");
 
            return m_soa_tile.GetRealData(position_index)[index];
        }
    }
 
    // const
    RealType  pos (int index, int position_index) const &
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile[index].pos(position_index);
        } else {
            return m_soa_tile.GetRealData(position_index)[index];
        }
    }
 
    AoS&       GetArrayOfStructs ()       { invalidateRuntimePtrCaches(); return m_aos_tile; }
    const AoS& GetArrayOfStructs () const { return m_aos_tile; }
 
    SoA&       GetStructOfArrays ()       { invalidateRuntimePtrCaches(); return m_soa_tile; }
    const SoA& GetStructOfArrays () const { return m_soa_tile; }
 
    bool empty () const { return size() == 0; }
 
    std::size_t size () const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile.size();
        } else {
            return m_soa_tile.size();
        }
    }
 
    int numParticles () const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile.numParticles();
        } else {
            return m_soa_tile.numParticles();
        }
    }
 
    int numRealParticles () const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile.numRealParticles();
        } else {
            return m_soa_tile.numRealParticles();
        }
    }
 
    int numNeighborParticles () const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile.numNeighborParticles();
        } else {
            return m_soa_tile.numNeighborParticles();
        }
    }
 
    int numTotalParticles () const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            return m_aos_tile.numTotalParticles();
        } else {
            return m_soa_tile.numTotalParticles();
        }
    }
 
    void setNumNeighbors (int num_neighbors)
    {
        invalidateRuntimePtrCaches();
        if constexpr(!ParticleType::is_soa_particle) {
            m_aos_tile.setNumNeighbors(num_neighbors);
        }
        m_soa_tile.setNumNeighbors(num_neighbors);
    }
 
    int getNumNeighbors () const
    {
        if constexpr (!ParticleType::is_soa_particle) {
            AMREX_ASSERT( m_soa_tile.getNumNeighbors() == m_aos_tile.getNumNeighbors() );
            return m_aos_tile.getNumNeighbors();
        } else {
            return m_soa_tile.getNumNeighbors();
        }
    }
 
    void resize (std::size_t count, GrowthStrategy strategy = GrowthStrategy::Poisson)
    {
        invalidateRuntimePtrCaches();
        if constexpr (!ParticleType::is_soa_particle) {
            m_aos_tile.resize(count, strategy);
        }
        m_soa_tile.resize(count, strategy);
    }
 
    void reserve (std::size_t capacity, GrowthStrategy strategy = GrowthStrategy::Poisson)
    {
        invalidateRuntimePtrCaches();
        if constexpr (!ParticleType::is_soa_particle) {
            m_aos_tile.reserve(capacity, strategy);
        }
        m_soa_tile.reserve(capacity, strategy);
    }
 
    void push_back (const ParticleType& p)
        requires (!ParticleType::is_soa_particle)
    {
        invalidateRuntimePtrCaches();
        m_aos_tile().push_back(p);
    }
 
    void push_back (const SuperParticleType& sp)
        requires (NArrayReal != 0 || NArrayInt != 0)
    {
        invalidateRuntimePtrCaches();
        auto np = numParticles();
 
        if constexpr (!ParticleType::is_soa_particle) {
            m_aos_tile.resize(np+1);
            for (int i = 0; i < AMREX_SPACEDIM; ++i) {
                m_aos_tile[np].pos(i) = sp.pos(i);
            }
            for (int i = 0; i < NStructReal; ++i) {
                m_aos_tile[np].rdata(i) = sp.rdata(i);
            }
            m_aos_tile[np].id() = sp.id();
            m_aos_tile[np].cpu() = sp.cpu();
            for (int i = 0; i < NStructInt; ++i) {
                m_aos_tile[np].idata(i) = sp.idata(i);
            }
        }
 
        m_soa_tile.resize(np+1);
        if constexpr (ParticleType::is_soa_particle) {
            m_soa_tile.GetIdCPUData()[np] = sp.m_idcpu;
        }
        auto& arr_rdata = m_soa_tile.GetRealData();
        auto& arr_idata = m_soa_tile.GetIntData();
        for (int i = 0; i < NArrayReal; ++i) {
            arr_rdata[i][np] = sp.rdata(NStructReal+i);
        }
        for (int i = 0; i < NArrayInt; ++i) {
            arr_idata[i][np] = sp.idata(NStructInt+i);
        }
    }
 
    void push_back_real (int comp, ParticleReal v) {
        invalidateRuntimePtrCaches();
        m_soa_tile.GetRealData(comp).push_back(v);
    }
 
    void push_back_real (const std::array<ParticleReal, NArrayReal>& v) {
        for (int i = 0; i < NArrayReal; ++i) {
            m_soa_tile.GetRealData(i).push_back(v[i]);
        }
    }
 
    void push_back_real (int comp, const ParticleReal* beg, const ParticleReal* end) {
        invalidateRuntimePtrCaches();
        auto it = m_soa_tile.GetRealData(comp).end();
        m_soa_tile.GetRealData(comp).insert(it, beg, end);
    }
 
    void push_back_real (int comp, amrex::Vector<amrex::ParticleReal>::const_iterator beg, amrex::Vector<amrex::ParticleReal>::const_iterator end) {
        push_back_real(comp, &(*beg), &(*end));
    }
 
    void push_back_real (int comp, amrex::Vector<amrex::ParticleReal> const & vec) {
        push_back_real(comp, vec.cbegin(), vec.cend());
    }
 
    void push_back_real (int comp, std::size_t npar, ParticleReal v) {
        invalidateRuntimePtrCaches();
        auto new_size = m_soa_tile.GetRealData(comp).size() + npar;
        m_soa_tile.GetRealData(comp).resize(new_size, v);
    }
 
    void push_back_int (int comp, int v) {
        invalidateRuntimePtrCaches();
        m_soa_tile.GetIntData(comp).push_back(v);
    }
 
    void push_back_int (const std::array<int, NArrayInt>& v) {
        for (int i = 0; i < NArrayInt; ++i) {
            m_soa_tile.GetIntData(i).push_back(v[i]);
        }
    }
 
    void push_back_int (int comp, const int* beg, const int* end) {
        invalidateRuntimePtrCaches();
        auto it = m_soa_tile.GetIntData(comp).end();
        m_soa_tile.GetIntData(comp).insert(it, beg, end);
    }
 
    void push_back_int (int comp, amrex::Vector<int>::const_iterator beg, amrex::Vector<int>::const_iterator end) {
        push_back_int(comp, &(*beg), &(*end));
    }
 
    void push_back_int (int comp, amrex::Vector<int> const & vec) {
        push_back_int(comp, vec.cbegin(), vec.cend());
    }
 
    void push_back_int (int comp, std::size_t npar, int v) {
        invalidateRuntimePtrCaches();
        auto new_size = m_soa_tile.GetIntData(comp).size() + npar;
        m_soa_tile.GetIntData(comp).resize(new_size, v);
    }
 
    int NumRealComps () const noexcept { return m_soa_tile.NumRealComps(); }
 
    int NumIntComps () const noexcept { return m_soa_tile.NumIntComps(); }
 
    int NumRuntimeRealComps () const noexcept { return m_runtime_r_ptrs.size(); }
 
    int NumRuntimeIntComps () const noexcept { return m_runtime_i_ptrs.size(); }
 
    void shrink_to_fit ()
    {
        invalidateRuntimePtrCaches();
        if constexpr (ParticleType::is_soa_particle) {
            GetStructOfArrays().GetIdCPUData().shrink_to_fit();
        } else {
            m_aos_tile().shrink_to_fit();
        }
        for (int j = 0; j < NumRealComps(); ++j)
        {
            auto& rdata = GetStructOfArrays().GetRealData(j);
            rdata.shrink_to_fit();
        }
 
        for (int j = 0; j < NumIntComps(); ++j)
        {
            auto& idata = GetStructOfArrays().GetIntData(j);
            idata.shrink_to_fit();
        }
    }
 
    Long capacity () const
    {
        Long nbytes = 0;
        if constexpr (ParticleType::is_soa_particle) {
            nbytes += GetStructOfArrays().GetIdCPUData().capacity() * sizeof(uint64_t);
        } else {
            nbytes += m_aos_tile().capacity() * sizeof(ParticleType);
        }
        for (int j = 0; j < NumRealComps(); ++j)
        {
            auto& rdata = GetStructOfArrays().GetRealData(j);
            nbytes += rdata.capacity() * sizeof(ParticleReal);
        }
 
        for (int j = 0; j < NumIntComps(); ++j)
        {
            auto& idata = GetStructOfArrays().GetIntData(j);
            nbytes += idata.capacity()*sizeof(int);
        }
        return nbytes;
    }
 
    void swap (ParticleTile<ParticleType, NArrayReal, NArrayInt, Allocator>& other) noexcept
    {
        if constexpr (ParticleType::is_soa_particle) {
            GetStructOfArrays().GetIdCPUData().swap(other.GetStructOfArrays().GetIdCPUData());
        } else {
            m_aos_tile().swap(other.GetArrayOfStructs()());
        }
        for (int j = 0; j < NumRealComps(); ++j)
        {
            auto& rdata = GetStructOfArrays().GetRealData(j);
            rdata.swap(other.GetStructOfArrays().GetRealData(j));
        }
 
        for (int j = 0; j < NumIntComps(); ++j)
        {
            auto& idata = GetStructOfArrays().GetIntData(j);
            idata.swap(other.GetStructOfArrays().GetIntData(j));
        }
        invalidateRuntimePtrCaches();
        other.invalidateRuntimePtrCaches();
    }
 
    ParticleTileDataType getParticleTileData ()
    {
        refreshRuntimePtrCaches();
 
        ParticleTileDataType ptd;
        if constexpr (!ParticleType::is_soa_particle) {
            ptd.m_aos = m_aos_tile().dataPtr();
        } else {
            ptd.m_aos = nullptr;
        }
        if constexpr (ParticleType::is_soa_particle) {
            ptd.m_idcpu = m_soa_tile.GetIdCPUData().dataPtr();
        } else {
            ptd.m_idcpu = nullptr;
        }
        if constexpr(NArrayReal > 0) {
            for (int i = 0; i < NArrayReal; ++i) {
                ptd.m_rdata[i] = m_soa_tile.GetRealData(i).dataPtr();
            }
        }
        if constexpr(NArrayInt > 0) {
            for (int i = 0; i < NArrayInt; ++i) {
                ptd.m_idata[i] = m_soa_tile.GetIntData(i).dataPtr();
            }
        }
        ptd.m_size = size();
        ptd.m_num_runtime_real = m_runtime_r_ptrs.size();
        ptd.m_num_runtime_int = m_runtime_i_ptrs.size();
        ptd.m_runtime_rdata = m_runtime_r_ptrs.dataPtr();
        ptd.m_runtime_idata = m_runtime_i_ptrs.dataPtr();
 
        return ptd;
    }
 
    ConstParticleTileDataType getConstParticleTileData () const
    {
        refreshConstRuntimePtrCaches();
 
        ConstParticleTileDataType ptd;
        if constexpr (!ParticleType::is_soa_particle) {
            ptd.m_aos = m_aos_tile().dataPtr();
        } else {
            ptd.m_aos = nullptr;
        }
        if constexpr (ParticleType::is_soa_particle) {
            ptd.m_idcpu = m_soa_tile.GetIdCPUData().dataPtr();
        } else {
            ptd.m_idcpu = nullptr;
        }
        if constexpr(NArrayReal > 0) {
            for (int i = 0; i < NArrayReal; ++i) {
                ptd.m_rdata[i] = m_soa_tile.GetRealData(i).dataPtr();
            }
        }
        if constexpr(NArrayInt > 0) {
            for (int i = 0; i < NArrayInt; ++i) {
                ptd.m_idata[i] = m_soa_tile.GetIntData(i).dataPtr();
            }
        }
        ptd.m_size = size();
        ptd.m_num_runtime_real = m_runtime_r_cptrs.size();
        ptd.m_num_runtime_int = m_runtime_i_cptrs.size();
        ptd.m_runtime_rdata = m_runtime_r_cptrs.dataPtr();
        ptd.m_runtime_idata = m_runtime_i_cptrs.dataPtr();
 
        return ptd;
    }
 
    void collectVectors (Vector<ParticleVector*>& pv,
                         Vector<typename SoA::IdCPU*>& idcpuv,
                         Vector<RealVector*>& rv, Vector<IntVector*>& iv)
    {
        invalidateRuntimePtrCaches();
        if constexpr (!ParticleType::is_soa_particle) {
            m_aos_tile.collectVectors(pv);
        } else {
            amrex::ignore_unused(pv);
        }
        m_soa_tile.collectVectors(idcpuv, rv, iv);
    }
 
    static void reserve (std::map<ParticleTile<T_ParticleType,NArrayReal,NArrayInt,Allocator>*, int> const& addsizes)
    {
        if constexpr (!IsArenaAllocator<Allocator<int>>::value) {
            for (auto [p,s] : addsizes) {
                p->reserve(p->size()+s);
            }
        } else {
            using PV = std::conditional_t<ParticleType::is_soa_particle,
                                          IntVector, ParticleVector>;
            Vector<std::pair<PV*,int>> pvs;
            Vector<std::pair<typename SoA::IdCPU*,int>> ids;
            Vector<std::pair<RealVector*,int>> rvs;
            Vector<std::pair<IntVector*,int>> ivs;
            for (auto [p,s] : addsizes) {
                if (s > 0) {
                    Vector<ParticleVector*> pv;
                    Vector<typename SoA::IdCPU*> idcpuv;
                    Vector<RealVector*> rv;
                    Vector<IntVector*> iv;
                    p->collectVectors(pv, idcpuv, rv, iv);
                    if constexpr (!ParticleType::is_soa_particle) {
                        for (auto* v : pv) {
                            pvs.emplace_back(v, s);
                        }
                    }
                    for (auto* v : idcpuv) {
                        ids.emplace_back(v, s);
                    }
                    for (auto* v : rv) {
                        rvs.emplace_back(v, s);
                    }
                    for (auto* v : iv) {
                        ivs.emplace_back(v, s);
                    }
                }
            }
 
            std::sort(pvs.begin(), pvs.end(), [] (auto const& a, auto const& b) {
                return (a.first->size() + a.second) >
                       (b.first->size() + b.second);
            });
            std::sort(ids.begin(), ids.end(), [] (auto const& a, auto const& b) {
                return (a.first->size() + a.second) >
                       (b.first->size() + b.second);
            });
            std::sort(rvs.begin(), rvs.end(), [] (auto const& a, auto const& b) {
                return (a.first->size() + a.second) >
                       (b.first->size() + b.second);
            });
            std::sort(ivs.begin(), ivs.end(), [] (auto const& a, auto const& b) {
                return (a.first->size() + a.second) >
                       (b.first->size() + b.second);
            });
 
            // Handle big vectors first
            {
                int i_pvs = 0, i_ids = 0, i_rvs = 0, i_ivs = 0;
                auto n_pvs = int(pvs.size());
                auto n_ids = int(ids.size());
                auto n_rvs = int(rvs.size());
                auto n_ivs = int(ivs.size());
                while ((i_pvs < n_pvs) || (i_ids < n_ids) || (i_rvs < n_rvs) ||
                       (i_ivs < n_ivs)) {
                    std::size_t nbytes = 0;
                    int ii = -1;
                    if (i_pvs < n_pvs) {
                        std::size_t my_bytes = (pvs[i_pvs].first->size()
                                              + pvs[i_pvs].second) * sizeof(typename PV::value_type);
                        if (my_bytes > nbytes) {
                            nbytes = my_bytes;
                            ii = 0;
                        }
                    }
                    if (i_ids < n_ids) {
                        std::size_t my_bytes = (ids[i_ids].first->size()
                                              + ids[i_ids].second) * sizeof(typename SoA::IdCPU::value_type);
                        if (my_bytes > nbytes) {
                            nbytes = my_bytes;
                            ii = 1;
                        }
                    }
                    if (i_rvs < n_rvs) {
                        std::size_t my_bytes = (rvs[i_rvs].first->size()
                                              + rvs[i_rvs].second) * sizeof(typename RealVector::value_type);
                        if (my_bytes > nbytes) {
                            nbytes = my_bytes;
                            ii = 2;
                        }
                    }
                    if (i_ivs < n_ivs) {
                        std::size_t my_bytes = (ivs[i_ivs].first->size()
                                              + ivs[i_ivs].second) * sizeof(typename IntVector::value_type);
                        if (my_bytes > nbytes) {
                            nbytes = my_bytes;
                            ii = 3;
                        }
                    }
                    if (ii == 0) {
                        auto [p,s] = pvs[i_pvs++];
                        p->reserve(p->size() + s);
                    } else if (ii == 1) {
                        auto [p,s] = ids[i_ids++];
                        p->reserve(p->size() + s);
                    } else if (ii == 2) {
                        auto [p,s] = rvs[i_rvs++];
                        p->reserve(p->size() + s);
                    } else {
                        auto [p,s] = ivs[i_ivs++];
                        p->reserve(p->size() + s);
                    }
                }
            }
        }
    }
 
private:
 
    void invalidateRuntimePtrCaches () noexcept
    {
        m_runtime_ptrs_dirty.store(true, std::memory_order_release);
        m_runtime_cptrs_dirty.store(true, std::memory_order_release);
    }
 
    void refreshRuntimePtrCaches ()
    {
        if (!m_runtime_ptrs_dirty.load(std::memory_order_acquire)) {
            return;
        }
#if defined(AMREX_USE_OMP)
#pragma omp critical (amrex_particle_tile_data_cache)
#endif
        {
            if (m_runtime_ptrs_dirty.load(std::memory_order_relaxed)) {
                m_runtime_r_ptrs.resize(m_soa_tile.NumRealComps() - NArrayReal);
                m_runtime_i_ptrs.resize(m_soa_tile.NumIntComps() - NArrayInt);
#ifdef AMREX_USE_GPU
                bool copy_real = false;
                m_h_runtime_r_ptrs.resize(m_soa_tile.NumRealComps() - NArrayReal, nullptr);
                for (std::size_t i = 0; i < m_h_runtime_r_ptrs.size(); ++i) {
                    if (m_h_runtime_r_ptrs[i] != m_soa_tile.GetRealData(i + NArrayReal).dataPtr()) {
                        m_h_runtime_r_ptrs[i] = m_soa_tile.GetRealData(i + NArrayReal).dataPtr();
                        copy_real = true;
                    }
                }
                if (copy_real) {
                    Gpu::htod_memcpy_async(m_runtime_r_ptrs.data(), m_h_runtime_r_ptrs.data(),
                                           m_h_runtime_r_ptrs.size()*sizeof(ParticleReal*));
                }
 
                bool copy_int = false;
                m_h_runtime_i_ptrs.resize(m_soa_tile.NumIntComps() - NArrayInt, nullptr);
                for (std::size_t i = 0; i < m_h_runtime_i_ptrs.size(); ++i) {
                    if (m_h_runtime_i_ptrs[i] != m_soa_tile.GetIntData(i + NArrayInt).dataPtr()) {
                        m_h_runtime_i_ptrs[i] = m_soa_tile.GetIntData(i + NArrayInt).dataPtr();
                        copy_int = true;
                    }
                }
                if (copy_int) {
                    Gpu::htod_memcpy_async(m_runtime_i_ptrs.data(), m_h_runtime_i_ptrs.data(),
                                           m_h_runtime_i_ptrs.size()*sizeof(int*));
                }
 
                if (copy_real || copy_int) {
                    Gpu::streamSynchronize();
                }
#else
                for (std::size_t i = 0; i < m_runtime_r_ptrs.size(); ++i) {
                    m_runtime_r_ptrs[i] = m_soa_tile.GetRealData(i + NArrayReal).dataPtr();
                }
 
                for (std::size_t i = 0; i < m_runtime_i_ptrs.size(); ++i) {
                    m_runtime_i_ptrs[i] = m_soa_tile.GetIntData(i + NArrayInt).dataPtr();
                }
#endif
                m_runtime_ptrs_dirty.store(false, std::memory_order_release);
            }
        }
    }
 
    void refreshConstRuntimePtrCaches () const
    {
        if (!m_runtime_cptrs_dirty.load(std::memory_order_acquire)) {
            return;
        }
#if defined(AMREX_USE_OMP)
#pragma omp critical (amrex_const_particle_tile_data_cache)
#endif
        {
            if (m_runtime_cptrs_dirty.load(std::memory_order_relaxed)) {
                m_runtime_r_cptrs.resize(m_soa_tile.NumRealComps() - NArrayReal);
                m_runtime_i_cptrs.resize(m_soa_tile.NumIntComps() - NArrayInt);
#ifdef AMREX_USE_GPU
                bool copy_real = false;
                m_h_runtime_r_cptrs.resize(m_soa_tile.NumRealComps() - NArrayReal, nullptr);
                for (std::size_t i = 0; i < m_h_runtime_r_cptrs.size(); ++i) {
                    if (m_h_runtime_r_cptrs[i] != m_soa_tile.GetRealData(i + NArrayReal).dataPtr()) {
                        m_h_runtime_r_cptrs[i] = m_soa_tile.GetRealData(i + NArrayReal).dataPtr();
                        copy_real = true;
                    }
                }
                if (copy_real) {
                    Gpu::htod_memcpy_async(m_runtime_r_cptrs.data(), m_h_runtime_r_cptrs.data(),
                                           m_h_runtime_r_cptrs.size()*sizeof(ParticleReal*));
                }
 
                bool copy_int = false;
                m_h_runtime_i_cptrs.resize(m_soa_tile.NumIntComps() - NArrayInt, nullptr);
                for (std::size_t i = 0; i < m_h_runtime_i_cptrs.size(); ++i) {
                    if (m_h_runtime_i_cptrs[i] != m_soa_tile.GetIntData(i + NArrayInt).dataPtr()) {
                        m_h_runtime_i_cptrs[i] = m_soa_tile.GetIntData(i + NArrayInt).dataPtr();
                        copy_int = true;
                    }
                }
                if (copy_int) {
                    Gpu::htod_memcpy_async(m_runtime_i_cptrs.data(), m_h_runtime_i_cptrs.data(),
                                           m_h_runtime_i_cptrs.size()*sizeof(int*));
                }
 
                if (copy_real || copy_int) {
                    Gpu::streamSynchronize();
                }
#else
                for (std::size_t i = 0; i < m_runtime_r_cptrs.size(); ++i) {
                    m_runtime_r_cptrs[i] = m_soa_tile.GetRealData(i + NArrayReal).dataPtr();
                }
 
                for (std::size_t i = 0; i < m_runtime_i_cptrs.size(); ++i) {
                    m_runtime_i_cptrs[i] = m_soa_tile.GetIntData(i + NArrayInt).dataPtr();
                }
#endif
                m_runtime_cptrs_dirty.store(false, std::memory_order_release);
            }
        }
    }
 
    AoS m_aos_tile;
    SoA m_soa_tile;
 
    bool m_defined = false;
 
    amrex::PODVector<ParticleReal*, Allocator<ParticleReal*> > m_runtime_r_ptrs;
    amrex::PODVector<int*, Allocator<int*> > m_runtime_i_ptrs;
 
    mutable amrex::PODVector<const ParticleReal*, Allocator<const ParticleReal*> > m_runtime_r_cptrs;
    mutable amrex::PODVector<const int*, Allocator<const int*> >m_runtime_i_cptrs;
 
    amrex::Gpu::HostVector<ParticleReal*> m_h_runtime_r_ptrs;
    amrex::Gpu::HostVector<int*> m_h_runtime_i_ptrs;
 
    mutable amrex::Gpu::HostVector<const ParticleReal*> m_h_runtime_r_cptrs;
    mutable amrex::Gpu::HostVector<const int*> m_h_runtime_i_cptrs;
 
    RuntimePtrCacheDirtyFlag m_runtime_ptrs_dirty;
    mutable RuntimePtrCacheDirtyFlag m_runtime_cptrs_dirty;
};
 
} // namespace amrex
 
#endif // AMREX_PARTICLETILE_H_