xref: /petsc/src/ts/utils/dmplexlandau/kokkos/landau.kokkos.cxx (revision 930e68a5a1dbab7612595fd12ba2e3af4e1d5d80)

/*
  Implements the Kokkos kernel
*/
#define PETSC_SKIP_CXX_COMPLEX_FIX
#include <petscconf.h>
#include <petsc/private/dmpleximpl.h>   /*I   "petscdmplex.h"   I*/
#include <petsclandau.h>
#include <petscts.h>
#include <petscveckokkos.hpp>

#include <Kokkos_Core.hpp>
#include <cstdio>
typedef Kokkos::TeamPolicy<>::member_type team_member;
#define PETSC_DEVICE_FUNC_DECL KOKKOS_INLINE_FUNCTION
#include "../land_tensors.h"
#define atomicAdd(e, f) Kokkos::atomic_fetch_add(e, f)
#include <petscaijdevice.h>
#undef atomicAdd

namespace landau_inner_red {  // namespace helps with name resolution in reduction identity
  template< class ScalarType >
  struct array_type {
    ScalarType gg2[LANDAU_DIM];
    ScalarType gg3[LANDAU_DIM][LANDAU_DIM];

    KOKKOS_INLINE_FUNCTION   // Default constructor - Initialize to 0's
    array_type() {
      for (int j = 0; j < LANDAU_DIM; j++){
        gg2[j] = 0;
        for (int k = 0; k < LANDAU_DIM; k++){
          gg3[j][k] = 0;
        }
      }
    }
    KOKKOS_INLINE_FUNCTION   // Copy Constructor
    array_type(const array_type & rhs) {
      for (int j = 0; j < LANDAU_DIM; j++){
        gg2[j] = rhs.gg2[j];
        for (int k = 0; k < LANDAU_DIM; k++){
          gg3[j][k] = rhs.gg3[j][k];
        }
      }
    }
    KOKKOS_INLINE_FUNCTION   // add operator
    array_type& operator += (const array_type& src) {
      for (int j = 0; j < LANDAU_DIM; j++){
        gg2[j] += src.gg2[j];
        for (int k = 0; k < LANDAU_DIM; k++){
          gg3[j][k] += src.gg3[j][k];
        }
      }
      return *this;
    }
    KOKKOS_INLINE_FUNCTION   // volatile add operator
    void operator += (const volatile array_type& src) volatile {
      for (int j = 0; j < LANDAU_DIM; j++){
        gg2[j] += src.gg2[j];
        for (int k = 0; k < LANDAU_DIM; k++){
          gg3[j][k] += src.gg3[j][k];
        }
      }
    }
  };
  typedef array_type<PetscReal> TensorValueType;  // used to simplify code below
}

namespace Kokkos { //reduction identity must be defined in Kokkos namespace
  template<>
  struct reduction_identity< landau_inner_red::TensorValueType > {
    KOKKOS_FORCEINLINE_FUNCTION static landau_inner_red::TensorValueType sum() {
      return landau_inner_red::TensorValueType();
    }
  };
}

extern "C"  {
  PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps *maps, pointInterpolationP4est (*pointMaps)[LANDAU_MAX_Q_FACE], PetscInt Nf, PetscInt Nq)
  {
    P4estVertexMaps   h_maps;  /* host container */
    const Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> >    h_points ((pointInterpolationP4est*)pointMaps, maps->num_reduced);
    const Kokkos::View< LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_gidx ((LandauIdx*)maps->gIdx, maps->num_elements);
    Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>   *d_points = new Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>("points", maps->num_reduced);
    Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight> *d_gidx = new Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight>("gIdx", maps->num_elements);

    PetscFunctionBegin;
    Kokkos::deep_copy (*d_gidx, h_gidx);
    Kokkos::deep_copy (*d_points, h_points);
    h_maps.num_elements = maps->num_elements;
    h_maps.num_face = maps->num_face;
    h_maps.num_reduced = maps->num_reduced;
    h_maps.deviceType = maps->deviceType;
    h_maps.Nf = Nf;
    h_maps.Nq = Nq;
    h_maps.c_maps = (pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE]) d_points->data();
    maps->vp1 = (void*)d_points;
    h_maps.gIdx = (LandauIdx (*)[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ]) d_gidx->data();
    maps->vp2 = (void*)d_gidx;
    {
      Kokkos::View<P4estVertexMaps, Kokkos::HostSpace> h_maps_k(&h_maps);
      Kokkos::View<P4estVertexMaps>                    *d_maps_k = new Kokkos::View<P4estVertexMaps>(Kokkos::create_mirror(DeviceMemorySpace(),h_maps_k));
      Kokkos::deep_copy (*d_maps_k, h_maps_k);
      maps->data = d_maps_k->data();
      maps->vp3 = (void*)d_maps_k;
      // debug
      // PetscErrorCode    ierr;
      // PetscInt          ej,q,s;
      // ierr = PetscPrintf(PETSC_COMM_SELF,"=== c_maps size --> %D. Kokkos::View<P4estVertexMaps*>=%p maps->data=%p P4estVertexMaps *maps=%p\n",maps->num_reduced, d_maps_k, maps->data, maps);CHKERRQ(ierr);
      // for (ej = 0; ej < maps->num_reduced; ++ej) {
      //   ierr = PetscPrintf(PETSC_COMM_SELF,"\t\tconstrant %3D) ",ej);
      //   for (q = 0; q < maps->num_face; ++q) {
      //     ierr = PetscPrintf(PETSC_COMM_SELF,"\t %3D - %13.5e ; ", maps->data->c_maps[ej][q].gid, maps->data->c_maps[ej][q].scale);CHKERRQ(ierr);
      //   }
      //   ierr = PetscPrintf(PETSC_COMM_SELF,"\n");
      // }
      // ierr = PetscPrintf(PETSC_COMM_SELF,"================= gIDx \n");
      // for (ej = 0; ej < maps->num_elements; ++ej) {
      //   for (q = 0; q < Nq; ++q) {
      //     for (s = 0; s < Nf; ++s) {
      //       ierr = PetscPrintf(PETSC_COMM_SELF," %3D ", maps->data->gIdx[ej][s][q]);
      //     }
      //     ierr = PetscPrintf(PETSC_COMM_SELF," ; ");
      //   }
      //   ierr = PetscPrintf(PETSC_COMM_SELF,"\n");
      // }
    }
    PetscFunctionReturn(0);
  }
  PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps *maps)
  {
    Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>  *a = (Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>*)maps->vp1;
    Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight>*b = (Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight>*)maps->vp2;
    Kokkos::View<P4estVertexMaps*>                                                  *c = (Kokkos::View<P4estVertexMaps*>*)maps->vp3;
    delete a;  delete b;  delete c;
    PetscFunctionReturn(0);
  }

  PetscErrorCode LandauKokkosJacobian(DM plex, const PetscInt Nq, PetscReal nu_alpha[], PetscReal nu_beta[], PetscReal invMass[], PetscReal Eq_m[],
                                      const LandauIPData *const IPData, PetscReal a_invJ[], const PetscInt num_sub_blocks, PetscReal *mass_w, PetscReal shift,
                                      const PetscLogEvent events[], Mat JacP)
  {
    PetscErrorCode    ierr;
    PetscInt          *Nbf,Nb,cStart,cEnd,Nf,dim,numCells,totDim,ipdatasz,global_elem_mat_sz,NfJac,nip;
    PetscTabulation   *Tf;
    PetscDS           prob;
    PetscSection      section, globalSection;
    PetscLogDouble    flops;
    PetscReal         *BB,*DD;
    LandauCtx         *ctx;
    P4estVertexMaps   *d_maps=NULL;
    PetscSplitCSRDataStructure *d_mat=NULL;

    PetscFunctionBegin;
    ierr = DMGetApplicationContext(plex, &ctx);CHKERRQ(ierr);
    if (!ctx) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
    ierr = DMGetDimension(plex, &dim);CHKERRQ(ierr);
    ierr = DMPlexGetHeightStratum(plex,0,&cStart,&cEnd);CHKERRQ(ierr);
    numCells = cEnd - cStart;
    ierr = DMGetDS(plex, &prob);CHKERRQ(ierr);
    ierr = PetscDSGetNumFields(prob, &Nf);CHKERRQ(ierr);
    ierr = PetscDSGetDimensions(prob, &Nbf);CHKERRQ(ierr); Nb = Nbf[0];
    if (Nq != Nb) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Nq != Nb. %D  %D",Nq,Nb);
    if (LANDAU_DIM != dim) SETERRQ2(PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %D != LANDAU_DIM %d",dim,LANDAU_DIM);
    ierr = PetscDSGetTotalDimension(prob, &totDim);CHKERRQ(ierr);
    ierr = PetscDSGetTabulation(prob, &Tf);CHKERRQ(ierr);
    if (ctx->gpu_assembly) {
      PetscContainer container;
      ierr = PetscObjectQuery((PetscObject) JacP, "assembly_maps", (PetscObject *) &container);CHKERRQ(ierr);
      if (container) { // not here first call
#if defined(PETSC_HAVE_KOKKOS_KERNELS)
        P4estVertexMaps   *h_maps=NULL;
        ierr = PetscContainerGetPointer(container, (void **) &h_maps);CHKERRQ(ierr);
        ierr = PetscInfo2(JacP, "Have container maps=%p maps->data=%p\n", h_maps, h_maps ? h_maps->data : NULL);CHKERRQ(ierr);
        if (h_maps->data) {
          d_maps = h_maps->data;
        } else {
          SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container");
        }
        // this does the setup the first time called
        ierr = MatKokkosGetDeviceMatWrite(JacP,&d_mat);CHKERRQ(ierr);
        global_elem_mat_sz = 0;
#else
        SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly w/o kokkos kernels -- should not be here");
#endif
      } else { // kernel output - first call assembled on device
        global_elem_mat_sz = numCells;
      }
    } else {
      global_elem_mat_sz = numCells; // no device assembly
    }
    BB   = Tf[0]->T[0]; DD = Tf[0]->T[1];
    ierr = DMGetLocalSection(plex, &section);CHKERRQ(ierr);
    ierr = DMGetGlobalSection(plex, &globalSection);CHKERRQ(ierr);
    if (mass_w) {
      flops = (PetscLogDouble)numCells*Nq*(5*dim*dim*Nf*Nf);
      ipdatasz = 0; NfJac = 0; nip = numCells*Nq;
    } else {
      flops = (PetscLogDouble)numCells*Nq*(5*dim*dim*Nf*Nf + 165);
      ipdatasz = LandauGetIPDataSize(IPData); NfJac = Nf; nip = IPData->nip_;
    }
    ierr = PetscKokkosInitializeCheck();CHKERRQ(ierr);
    {
      using scr_mem_t = Kokkos::DefaultExecutionSpace::scratch_memory_space;
      using g2_scr_t = Kokkos::View<PetscReal***, Kokkos::LayoutRight, scr_mem_t>;
      using g3_scr_t = Kokkos::View<PetscReal****, Kokkos::LayoutRight, scr_mem_t>;

      const int scr_bytes = 2*(g2_scr_t::shmem_size(dim,Nf,Nq) + g3_scr_t::shmem_size(dim,dim,Nf,Nq));
      int   conc, team_size;
      ierr = PetscLogEventBegin(events[3],0,0,0,0);CHKERRQ(ierr);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_alpha (nu_alpha, NfJac);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_alpha ("nu_alpha", NfJac);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_beta (nu_beta, NfJac);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_beta ("nu_beta", NfJac);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invMass (invMass,NfJac);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_invMass ("invMass", NfJac);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_Eq_m (Eq_m,NfJac);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_Eq_m ("Eq_m", NfJac);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_BB (BB,Nq*Nb);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_BB ("BB", Nq*Nb);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_DD (DD,Nq*Nb*dim);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_DD ("DD", Nq*Nb*dim);
      const Kokkos::View<LandauIPReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_ipdata_raw (IPData->w, ipdatasz); // IPData->w==0 for mass
      Kokkos::View<LandauIPReal*, Kokkos::LayoutLeft> d_ipdata_raw ("ipdata", ipdatasz);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invJ (a_invJ, mass_w ? 0 : nip*dim*dim);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_invJ ("invJ", mass_w ? 0 : nip*dim*dim);
      const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_mass_w (mass_w, mass_w ? nip : 0);
      Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_mass_w ("mass_w", mass_w ? nip : 0);
      Kokkos::View<PetscScalar**, Kokkos::LayoutRight> d_elem_mats("element matrices", global_elem_mat_sz, totDim*totDim);
      Kokkos::View<PetscScalar**, Kokkos::LayoutRight> d_f("element matrices", NfJac, nip);
      Kokkos::View<PetscScalar***, Kokkos::LayoutRight> d_df("element matrices", dim, NfJac, nip);
      Kokkos::deep_copy (d_mass_w, h_mass_w);
      Kokkos::deep_copy (d_ipdata_raw, h_ipdata_raw);
      Kokkos::deep_copy (d_alpha, h_alpha);
      Kokkos::deep_copy (d_beta, h_beta);
      Kokkos::deep_copy (d_invMass, h_invMass);
      Kokkos::deep_copy (d_Eq_m, h_Eq_m);
      Kokkos::deep_copy (d_BB, h_BB);
      Kokkos::deep_copy (d_DD, h_DD);
      Kokkos::deep_copy (d_invJ, h_invJ);

      ierr = PetscLogEventEnd(events[3],0,0,0,0);CHKERRQ(ierr);
#if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL)
      ierr = PetscLogGpuFlops(flops*nip);CHKERRQ(ierr);
      if (ctx->deviceType == LANDAU_CPU) PetscInfo(plex, "Warning: Landau selected CPU but no support for Kokkos using GPU\n");
#else
      ierr = PetscLogFlops(flops*nip);CHKERRQ(ierr);
#endif
#define KOKKOS_SHARED_LEVEL 1
      conc = Kokkos::DefaultExecutionSpace().concurrency(), team_size = conc > Nq ? Nq : 1;
      // PetscInfo5(plex, "shared memory size: %d bytes in level %d. mass_w.size=%d. #threads=%D team size=%D\n",scr_bytes,KOKKOS_SHARED_LEVEL,d_mass_w.size(),num_sub_blocks,team_size);
      // get f and df
      if (!mass_w) {
        ierr = PetscLogEventBegin(events[8],0,0,0,0);CHKERRQ(ierr);
        Kokkos::parallel_for("df_d_elements", Kokkos::TeamPolicy<>(numCells, team_size, num_sub_blocks), KOKKOS_LAMBDA (const team_member team) {
            const PetscInt  myelem = team.league_rank();
            // un pack IPData
            LandauIPReal *IPData_coefs = &d_ipdata_raw[nip*(dim+1)];
            LandauIPReal *coef = &IPData_coefs[myelem*Nb*Nf];
            Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) {
                const PetscInt          ipidx = myQi + myelem * Nq;
                const PetscReal *const  invJj = &d_invJ(ipidx*dim*dim);
                const PetscReal         *Bq = &d_BB[myQi*Nb], *Dq = &d_DD[myQi*Nb*dim];
                Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nf), [=] (int f) {
                    PetscInt     b, e, d;
                    PetscScalar  refSpaceDer[LANDAU_DIM];
                    d_f(f,ipidx) = 0.0;
                    for (d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0;
                    for (b = 0; b < Nb; ++b) {
                      const PetscInt    cidx = b;
                      d_f(f,ipidx) += Bq[cidx]*coef[f*Nb+cidx];
                      for (d = 0; d < dim; ++d) refSpaceDer[d] += Dq[cidx*dim+d]*coef[f*Nb+cidx];
                    }
                    for (d = 0; d < dim; ++d) {
                      for (e = 0, d_df(d,f,ipidx) = 0.0; e < dim; ++e) {
                        d_df(d,f,ipidx) += invJj[e*dim+d]*refSpaceDer[e];
                      }
                    }
                  }); // Nf
              }); // Nq
          }); // elems
        Kokkos::fence();
        ierr = PetscLogEventEnd(events[8],0,0,0,0);CHKERRQ(ierr);
      }
      ierr = PetscLogEventBegin(events[4],0,0,0,0);CHKERRQ(ierr);
      ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr);
      Kokkos::parallel_for("Landau_elements", Kokkos::TeamPolicy<>(numCells, team_size, num_sub_blocks).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes)), KOKKOS_LAMBDA (const team_member team) {
          const PetscInt  myelem = team.league_rank();
          g2_scr_t        g2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,Nf,Nq); // we don't use these for mass matrix
          g3_scr_t        g3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,Nf,Nq);
          if (!d_mass_w.size()) {
            g2_scr_t        gg2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,Nf,Nq);
            g3_scr_t        gg3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,Nf,Nq);
            LandauIPData    d_IPData;
            // un pack IPData
            d_IPData.w   = &d_ipdata_raw[0];
            d_IPData.x   = &d_ipdata_raw[1*nip];
            d_IPData.y   = &d_ipdata_raw[2*nip];
            if (dim==2) d_IPData.z = NULL;
            else        d_IPData.z = &d_ipdata_raw[3*nip];
            // get g2[] & g3[]
            Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) {
                using Kokkos::parallel_reduce;
                const PetscInt                    jpidx = myQi + myelem * Nq;
                const PetscReal* const            invJj = &d_invJ(jpidx*dim*dim);
                const PetscReal                   vj[3] = {d_IPData.x[jpidx], d_IPData.y[jpidx], d_IPData.z ? d_IPData.z[jpidx] : 0}, wj = d_IPData.w[jpidx];
                landau_inner_red::TensorValueType gg_temp; // reduce on part of gg2 and g33 for IP jpidx
                Kokkos::parallel_reduce(Kokkos::ThreadVectorRange (team, (int)nip), [=] (const int& ipidx, landau_inner_red::TensorValueType & ggg) {
                    const PetscReal wi = d_IPData.w[ipidx], x = d_IPData.x[ipidx], y = d_IPData.y[ipidx];
                    PetscReal       temp1[3] = {0, 0, 0}, temp2 = 0;
                    PetscInt        fieldA,d2,d3;
#if LANDAU_DIM==2
                    PetscReal Ud[2][2], Uk[2][2];
                    LandauTensor2D(vj, x, y, Ud, Uk, (ipidx==jpidx) ? 0. : 1.);
#else
                    PetscReal U[3][3], z = d_IPData.z[ipidx];
                    LandauTensor3D(vj, x, y, z, U, (ipidx==jpidx) ? 0. : 1.);
#endif
                    for (fieldA = 0; fieldA < Nf; ++fieldA) {
                      temp1[0] += d_df(0,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA];
                      temp1[1] += d_df(1,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA];
#if LANDAU_DIM==3
                      temp1[2] += d_df(2,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA];
#endif
                      temp2    += d_f(fieldA,ipidx)*d_beta[fieldA];
                    }
                    temp1[0] *= wi;
                    temp1[1] *= wi;
#if LANDAU_DIM==3
                    temp1[2] *= wi;
#endif
                    temp2    *= wi;
#if LANDAU_DIM==2
                    for (d2 = 0; d2 < 2; d2++) {
                      for (d3 = 0; d3 < 2; ++d3) {
                        /* K = U * grad(f): g2=e: i,A */
                        ggg.gg2[d2] += Uk[d2][d3]*temp1[d3];
                        /* D = -U * (I \kron (fx)): g3=f: i,j,A */
                        ggg.gg3[d2][d3] += Ud[d2][d3]*temp2;
                      }
                    }
#else
                    for (d2 = 0; d2 < 3; ++d2) {
                      for (d3 = 0; d3 < 3; ++d3) {
                        /* K = U * grad(f): g2 = e: i,A */
                        ggg.gg2[d2] += U[d2][d3]*temp1[d3];
                        /* D = -U * (I \kron (fx)): g3 = f: i,j,A */
                        ggg.gg3[d2][d3] += U[d2][d3]*temp2;
                      }
                    }
#endif
                  }, Kokkos::Sum<landau_inner_red::TensorValueType>(gg_temp));
                // if (myelem==0) printf("\t:%d.%d) temp gg3=%e %e %e %e\n",myelem,myQi,gg_temp.gg3[0][0],gg_temp.gg3[1][0],gg_temp.gg3[0][1],gg_temp.gg3[1][1]);
                // add alpha and put in gg2/3
                Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [&] (const int& fieldA) {
                    PetscInt d2,d3;
                    for (d2 = 0; d2 < dim; d2++) {
                      gg2(d2,fieldA,myQi) = gg_temp.gg2[d2]*d_alpha[fieldA];
                      //if (myelem==0 && fieldA==1) printf("\t\t:%d.%d) gg2[%d]=%e (+= %e)\n",myelem,myQi,d2,gg2(d2,fieldA,myQi),gg_temp.gg2[d2]*d_alpha[fieldA]);
                      //gg2[d2][myQi][fieldA] += gg_temp.gg2[d2]*d_alpha[fieldA];
                      for (d3 = 0; d3 < dim; d3++) {
                        //gg3[d2][d3][myQi][fieldA] -= gg_temp.gg3[d2][d3]*d_alpha[fieldA]*s_invMass[fieldA];
                        gg3(d2,d3,fieldA,myQi) = -gg_temp.gg3[d2][d3]*d_alpha[fieldA]*d_invMass[fieldA];
                        //if (myelem==0 && fieldA==1) printf("\t\t\t:%d.%d) gg3[%d][%d]=%e\n",myelem,myQi,d2,d3,gg3(d2,d3,fieldA,myQi));
                      }
                    }
                  });
                /* add electric field term once per IP */
                Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [&] (const int& fieldA) {
                    //gg.gg2[fieldA][dim-1] += d_Eq_m[fieldA];
                    gg2(dim-1,fieldA,myQi) += d_Eq_m[fieldA];
                  });
                // Kokkos::single(Kokkos::PerThread(team), [&]() {
                Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [=] (const int& fieldA) {
                    int d,d2,d3,dp;
                    //printf("%d %d %d gg2[][1]=%18.10e\n",myelem,myQi,fieldA,gg.gg2[fieldA][dim-1]);
                    /* Jacobian transform - g2, g3 - per thread (2D) */
                    for (d = 0; d < dim; ++d) {
                      g2(d,fieldA,myQi) = 0;
                      for (d2 = 0; d2 < dim; ++d2) {
                        g2(d,fieldA,myQi) += invJj[d*dim+d2]*gg2(d2,fieldA,myQi);
                        //if (myelem==0 && myQi==0) printf("\t:g2[%d][%d][%d]=%e. %e %e\n",(int)myQi,(int)fieldA,(int)d,g2(fieldA,myQi,d),invJj[d*dim+d2],gg.gg2[fieldA][d2]);
                        g3(d,d2,fieldA,myQi) = 0;
                        for (d3 = 0; d3 < dim; ++d3) {
                          for (dp = 0; dp < dim; ++dp) {
                            g3(d,d2,fieldA,myQi) += invJj[d*dim + d3]*gg3(d3,dp,fieldA,myQi)*invJj[d2*dim + dp];
                            //printf("\t%d %d %d %d %d %d %d g3=%g wj=%g g3 = %g * %g * %g\n",myelem,myQi,fieldA,d,d2,d3,dp,g3(fieldA,myQi,d,d2),wj,invJj[d*dim + d3],gg.gg3[fieldA][d3][dp],invJj[d2*dim + dp]);
                          }
                        }
                        g3(d,d2,fieldA,myQi) *= wj;
                      }
                      g2(d,fieldA,myQi) *= wj;
                    }
                  });
              }); // Nq
            team.team_barrier();
          } // Jacobian
          /* assemble */
          //Kokkos::single(Kokkos::PerTeam(team), [&]() {
          Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nb), [=] (int blk_i) {
              Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nf), [=] (int fieldA) {
                  int blk_j,qj,d,d2;
                  const PetscInt i = fieldA*Nb + blk_i; /* Element matrix row */
                  for (blk_j = 0; blk_j < Nb; ++blk_j) {
                    const PetscInt j    = fieldA*Nb + blk_j; /* Element matrix column */
                    const PetscInt fOff = i*totDim + j;
                    PetscScalar t = global_elem_mat_sz ? d_elem_mats(myelem,fOff) : 0;
                    for (qj = 0 ; qj < Nq ; qj++) { // look at others integration points
                      const PetscReal *BJq = &d_BB[qj*Nb], *DIq = &d_DD[qj*Nb*dim];
                      if (!d_mass_w.size()) {
                        for (d = 0; d < dim; ++d) {
                          t += DIq[blk_i*dim+d]*g2(d,fieldA,qj)*BJq[blk_j];
                          for (d2 = 0; d2 < dim; ++d2) {
                            t += DIq[blk_i*dim + d]*g3(d,d2,fieldA,qj)*DIq[blk_j*dim + d2];
                          }
                        }
                      } else {
                        const PetscInt jpidx = qj + myelem * Nq;
                        t += BJq[blk_i] * d_mass_w(jpidx)*shift * BJq[blk_j];
                        //printf("\tmat[%d %d %d %d %d]: B=%g w=%g shift=%g B=%g\n",myelem,fOff,fieldA,qj,d,BJq[blk_i],d_mass_w(jpidx),shift,BJq[blk_j]);
                      }
                    }
                    if (global_elem_mat_sz) d_elem_mats(myelem,fOff) = t; // can set this because local element matrix[fOff]
                    else {
                      PetscErrorCode         ierr = 0;
                      PetscScalar            vals[LANDAU_MAX_Q*LANDAU_MAX_Q],row_scale[LANDAU_MAX_Q],col_scale[LANDAU_MAX_Q];
                      PetscInt               q,idx,nr,nc,rows0[LANDAU_MAX_Q],cols0[LANDAU_MAX_Q],rows[LANDAU_MAX_Q],cols[LANDAU_MAX_Q];
                      const LandauIdx *const Idxs = &d_maps->gIdx[myelem][fieldA][0];
                      idx = Idxs[blk_i];
                      if (idx >= 0) {
                        nr = 1;
                        rows0[0] = idx;
                        row_scale[0] = 1.;
                      } else {
                        idx = -idx - 1;
                        nr = d_maps->num_face;
                        for (q = 0; q < d_maps->num_face; q++) {
                          rows0[q]     = d_maps->c_maps[idx][q].gid;
                          row_scale[q] = d_maps->c_maps[idx][q].scale;
                        }
                      }
                      idx = Idxs[blk_j];
                      if (idx >= 0) {
                        nc = 1;
                        cols0[0] = idx;
                        col_scale[0] = 1.;
                      } else {
                        idx = -idx - 1;
                        nc = d_maps->num_face;
                        for (q = 0; q < d_maps->num_face; q++) {
                          cols0[q]     = d_maps->c_maps[idx][q].gid;
                          col_scale[q] = d_maps->c_maps[idx][q].scale;
                        }
                      }
                      for (q = 0; q < nr; q++) rows[q] = rows0[q];
                      for (q = 0; q < nc; q++) cols[q] = cols0[q];
                      for (q = 0; q < nr; q++) {
                        for (d = 0; d < nc; d++) {
                          vals[q*nc + d] = row_scale[q]*col_scale[d]*t;
                        }
                      }
                      MatSetValuesDevice(d_mat,nr,rows,nc,cols,vals,ADD_VALUES,&ierr);
                      if (ierr) return;
                    }
                  }
                });
            });
        });
      Kokkos::fence();
      ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr);
      ierr = PetscLogEventEnd(events[4],0,0,0,0);CHKERRQ(ierr);
      if (global_elem_mat_sz) {
        Kokkos::View<PetscScalar**, Kokkos::LayoutRight>::HostMirror h_elem_mats = Kokkos::create_mirror_view(d_elem_mats);
        ierr = PetscLogEventBegin(events[5],0,0,0,0);CHKERRQ(ierr);
        Kokkos::deep_copy (h_elem_mats, d_elem_mats);
        ierr = PetscLogEventEnd(events[5],0,0,0,0);CHKERRQ(ierr);
        ierr = PetscLogEventBegin(events[6],0,0,0,0);CHKERRQ(ierr);
        PetscInt ej;
        for (ej = cStart ; ej < cEnd; ++ej) {
          const PetscScalar *elMat = &h_elem_mats(ej-cStart,0);
          ierr = DMPlexMatSetClosure(plex, section, globalSection, JacP, ej, elMat, ADD_VALUES);CHKERRQ(ierr);
          if (ej==-1) {
            int d,f;
            PetscPrintf(PETSC_COMM_SELF,"Kokkos Element matrix %d/%d\n",1,(int)numCells);
            for (d = 0; d < totDim; ++d){
              for (f = 0; f < totDim; ++f) PetscPrintf(PETSC_COMM_SELF," %12.5e",  PetscRealPart(elMat[d*totDim + f]));
              PetscPrintf(PETSC_COMM_SELF,"\n");
            }
          }
        }
        ierr = PetscLogEventEnd(events[6],0,0,0,0);CHKERRQ(ierr);
      }
    }
    PetscFunctionReturn(0);
  }
} // extern "C"