/* Implements the Kokkos kernel */ #include #if defined(PETSC_HAVE_CUDA_CLANG) #include #define LANDAU_NOT_IMPLEMENTED SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Not supported with CLANG") PetscErrorCode LandauKokkosJacobian(DM[], const PetscInt, const PetscInt, const PetscInt, const PetscInt, const PetscInt[], PetscReal[], PetscScalar[], const PetscScalar[], const LandauStaticData *, const PetscReal, const PetscLogEvent[], const PetscInt[], const PetscInt[], Mat[], Mat) { LANDAU_NOT_IMPLEMENTED; } PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps *, pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE], PetscInt[], PetscInt) { LANDAU_NOT_IMPLEMENTED; } PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps *, PetscInt) { LANDAU_NOT_IMPLEMENTED; } PetscErrorCode LandauKokkosStaticDataSet(DM, const PetscInt, const PetscInt, const PetscInt, const PetscInt, PetscInt[], PetscInt[], PetscInt[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], PetscReal[], LandauStaticData *) { LANDAU_NOT_IMPLEMENTED; } PetscErrorCode LandauKokkosStaticDataClear(LandauStaticData *) { LANDAU_NOT_IMPLEMENTED; } #else #include #include #include /*I "petscdmplex.h" I*/ #include #include #include #include typedef Kokkos::TeamPolicy<>::member_type team_member; #include "../land_tensors.h" namespace landau_inner_red { // namespace helps with name resolution in reduction identity template 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 TensorValueType; // used to simplify code below } // namespace landau_inner_red namespace Kokkos { //reduction identity must be defined in Kokkos namespace template <> struct reduction_identity { KOKKOS_FORCEINLINE_FUNCTION static landau_inner_red::TensorValueType sum() { return landau_inner_red::TensorValueType(); } }; } // namespace Kokkos extern "C" { PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps maps[], pointInterpolationP4est (*pointMaps)[LANDAU_MAX_Q_FACE], PetscInt Nf[], PetscInt grid) { P4estVertexMaps h_maps; /* host container */ const Kokkos::View> h_points((pointInterpolationP4est *)pointMaps, maps[grid].num_reduced); const Kokkos::View> h_gidx((LandauIdx *)maps[grid].gIdx, maps[grid].num_elements); Kokkos::View *d_points = new Kokkos::View("points", maps[grid].num_reduced); Kokkos::View *d_gidx = new Kokkos::View("gIdx", maps[grid].num_elements); PetscFunctionBegin; Kokkos::deep_copy(*d_gidx, h_gidx); Kokkos::deep_copy(*d_points, h_points); h_maps.num_elements = maps[grid].num_elements; h_maps.num_face = maps[grid].num_face; h_maps.num_reduced = maps[grid].num_reduced; h_maps.deviceType = maps[grid].deviceType; h_maps.numgrids = maps[grid].numgrids; h_maps.Nf = Nf[grid]; h_maps.c_maps = (pointInterpolationP4est(*)[LANDAU_MAX_Q_FACE])d_points->data(); maps[grid].vp1 = (void *)d_points; h_maps.gIdx = (LandauIdx(*)[LANDAU_MAX_SPECIES][LANDAU_MAX_NQND])d_gidx->data(); maps[grid].vp2 = (void *)d_gidx; { Kokkos::View h_maps_k(&h_maps); Kokkos::View *d_maps_k = new Kokkos::View(Kokkos::create_mirror(Kokkos::DefaultExecutionSpace::memory_space(), h_maps_k)); Kokkos::deep_copy(*d_maps_k, h_maps_k); maps[grid].d_self = d_maps_k->data(); maps[grid].vp3 = (void *)d_maps_k; } PetscFunctionReturn(PETSC_SUCCESS); } PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps maps[], PetscInt num_grids) { PetscFunctionBegin; for (PetscInt grid = 0; grid < num_grids; grid++) { auto a = static_cast *>(maps[grid].vp1); auto b = static_cast *>(maps[grid].vp2); auto c = static_cast *>(maps[grid].vp3); delete a; delete b; delete c; } PetscFunctionReturn(PETSC_SUCCESS); } PetscErrorCode LandauKokkosStaticDataSet(DM plex, const PetscInt Nq, const PetscInt Nb, const PetscInt batch_sz, const PetscInt num_grids, PetscInt a_numCells[], PetscInt a_species_offset[], PetscInt a_mat_offset[], PetscReal a_nu_alpha[], PetscReal a_nu_beta[], PetscReal a_invMass[], PetscReal a_lambdas[], PetscReal a_invJ[], PetscReal a_x[], PetscReal a_y[], PetscReal a_z[], PetscReal a_w[], LandauStaticData *SData_d) { PetscReal *BB, *DD; PetscTabulation *Tf; PetscInt dim; PetscInt ip_offset[LANDAU_MAX_GRIDS + 1], ipf_offset[LANDAU_MAX_GRIDS + 1], elem_offset[LANDAU_MAX_GRIDS + 1], nip, IPf_sz, Nftot; PetscDS prob; PetscFunctionBegin; PetscCall(DMGetDimension(plex, &dim)); PetscCall(DMGetDS(plex, &prob)); PetscCheck(LANDAU_DIM == dim, PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d", dim, LANDAU_DIM); PetscCall(PetscDSGetTabulation(prob, &Tf)); BB = Tf[0]->T[0]; DD = Tf[0]->T[1]; ip_offset[0] = ipf_offset[0] = elem_offset[0] = 0; nip = 0; IPf_sz = 0; for (PetscInt grid = 0; grid < num_grids; grid++) { PetscInt nfloc = a_species_offset[grid + 1] - a_species_offset[grid]; elem_offset[grid + 1] = elem_offset[grid] + a_numCells[grid]; nip += a_numCells[grid] * Nq; ip_offset[grid + 1] = nip; IPf_sz += Nq * nfloc * a_numCells[grid]; ipf_offset[grid + 1] = IPf_sz; } Nftot = a_species_offset[num_grids]; PetscCall(PetscKokkosInitializeCheck()); { const Kokkos::View> h_alpha(a_nu_alpha, Nftot); auto alpha = new Kokkos::View("alpha", Nftot); SData_d->alpha = static_cast(alpha); const Kokkos::View> h_beta(a_nu_beta, Nftot); auto beta = new Kokkos::View("beta", Nftot); SData_d->beta = static_cast(beta); const Kokkos::View> h_invMass(a_invMass, Nftot); auto invMass = new Kokkos::View("invMass", Nftot); SData_d->invMass = static_cast(invMass); const Kokkos::View> h_lambdas(a_lambdas, LANDAU_MAX_GRIDS * LANDAU_MAX_GRIDS); auto lambdas = new Kokkos::View("lambdas", LANDAU_MAX_GRIDS * LANDAU_MAX_GRIDS); SData_d->lambdas = static_cast(lambdas); const Kokkos::View> h_BB(BB, Nq * Nb); auto B = new Kokkos::View("B", Nq * Nb); SData_d->B = static_cast(B); const Kokkos::View> h_DD(DD, Nq * Nb * dim); auto D = new Kokkos::View("D", Nq * Nb * dim); SData_d->D = static_cast(D); const Kokkos::View> h_invJ(a_invJ, nip * dim * dim); auto invJ = new Kokkos::View("invJ", nip * dim * dim); SData_d->invJ = static_cast(invJ); const Kokkos::View> h_x(a_x, nip); auto x = new Kokkos::View("x", nip); SData_d->x = static_cast(x); const Kokkos::View> h_y(a_y, nip); auto y = new Kokkos::View("y", nip); SData_d->y = static_cast(y); const Kokkos::View> h_w(a_w, nip); auto w = new Kokkos::View("w", nip); SData_d->w = static_cast(w); Kokkos::deep_copy(*alpha, h_alpha); Kokkos::deep_copy(*beta, h_beta); Kokkos::deep_copy(*invMass, h_invMass); Kokkos::deep_copy(*lambdas, h_lambdas); Kokkos::deep_copy(*B, h_BB); Kokkos::deep_copy(*D, h_DD); Kokkos::deep_copy(*invJ, h_invJ); Kokkos::deep_copy(*x, h_x); Kokkos::deep_copy(*y, h_y); Kokkos::deep_copy(*w, h_w); if (dim == 3) { const Kokkos::View> h_z(a_z, nip); auto z = new Kokkos::View("z", nip); SData_d->z = static_cast(z); Kokkos::deep_copy(*z, h_z); } else SData_d->z = NULL; // const Kokkos::View> h_NCells(a_numCells, num_grids); auto nc = new Kokkos::View("NCells", num_grids); SData_d->NCells = static_cast(nc); Kokkos::deep_copy(*nc, h_NCells); const Kokkos::View> h_species_offset(a_species_offset, num_grids + 1); auto soff = new Kokkos::View("species_offset", num_grids + 1); SData_d->species_offset = static_cast(soff); Kokkos::deep_copy(*soff, h_species_offset); const Kokkos::View> h_mat_offset(a_mat_offset, num_grids + 1); auto moff = new Kokkos::View("mat_offset", num_grids + 1); SData_d->mat_offset = static_cast(moff); Kokkos::deep_copy(*moff, h_mat_offset); const Kokkos::View> h_ip_offset(ip_offset, num_grids + 1); auto ipoff = new Kokkos::View("ip_offset", num_grids + 1); SData_d->ip_offset = static_cast(ipoff); Kokkos::deep_copy(*ipoff, h_ip_offset); const Kokkos::View> h_elem_offset(elem_offset, num_grids + 1); auto eoff = new Kokkos::View("elem_offset", num_grids + 1); SData_d->elem_offset = static_cast(eoff); Kokkos::deep_copy(*eoff, h_elem_offset); const Kokkos::View> h_ipf_offset(ipf_offset, num_grids + 1); auto ipfoff = new Kokkos::View("ipf_offset", num_grids + 1); SData_d->ipf_offset = static_cast(ipfoff); Kokkos::deep_copy(*ipfoff, h_ipf_offset); #if defined(LANDAU_LAYOUT_LEFT) // preallocate dynamic data, no copy auto ipfdf_data = new Kokkos::View("fdf", batch_sz, dim + 1, IPf_sz); #else auto ipfdf_data = new Kokkos::View("fdf", batch_sz, dim + 1, IPf_sz); #endif SData_d->ipfdf_data = static_cast(ipfdf_data); auto Eq_m = new Kokkos::View("Eq_m", Nftot); // allocate but do not set, same for whole batch SData_d->Eq_m = static_cast(Eq_m); const Kokkos::View> h_coo_elem_offsets((LandauIdx *)SData_d->coo_elem_offsets, SData_d->coo_n_cellsTot + 1); auto coo_elem_offsets = new Kokkos::View("coo_elem_offsets", SData_d->coo_n_cellsTot + 1); const Kokkos::View> h_coo_elem_fullNb((LandauIdx *)SData_d->coo_elem_fullNb, SData_d->coo_n_cellsTot); auto coo_elem_fullNb = new Kokkos::View("coo_elem_offsets", SData_d->coo_n_cellsTot); const Kokkos::View> h_coo_elem_point_offsets((LandauIdx *)SData_d->coo_elem_point_offsets, SData_d->coo_n_cellsTot * (LANDAU_MAX_NQND + 1)); auto coo_elem_point_offsets = new Kokkos::View("coo_elem_point_offsets", SData_d->coo_n_cellsTot * (LANDAU_MAX_NQND + 1)); // assign & copy Kokkos::deep_copy(*coo_elem_offsets, h_coo_elem_offsets); Kokkos::deep_copy(*coo_elem_fullNb, h_coo_elem_fullNb); Kokkos::deep_copy(*coo_elem_point_offsets, h_coo_elem_point_offsets); // need to free this now and use pointer space PetscCall(PetscFree3(SData_d->coo_elem_offsets, SData_d->coo_elem_fullNb, SData_d->coo_elem_point_offsets)); SData_d->coo_elem_offsets = static_cast(coo_elem_offsets); SData_d->coo_elem_fullNb = static_cast(coo_elem_fullNb); SData_d->coo_elem_point_offsets = static_cast(coo_elem_point_offsets); auto coo_vals = new Kokkos::View("coo_vals", SData_d->coo_size); SData_d->coo_vals = static_cast(coo_vals); } SData_d->maps = NULL; // not used PetscFunctionReturn(PETSC_SUCCESS); } PetscErrorCode LandauKokkosStaticDataClear(LandauStaticData *SData_d) { PetscFunctionBegin; if (SData_d->alpha) { auto alpha = static_cast *>(SData_d->alpha); delete alpha; SData_d->alpha = NULL; auto beta = static_cast *>(SData_d->beta); delete beta; auto invMass = static_cast *>(SData_d->invMass); delete invMass; auto lambdas = static_cast *>(SData_d->lambdas); delete lambdas; auto B = static_cast *>(SData_d->B); delete B; auto D = static_cast *>(SData_d->D); delete D; auto invJ = static_cast *>(SData_d->invJ); delete invJ; auto x = static_cast *>(SData_d->x); delete x; auto y = static_cast *>(SData_d->y); delete y; if (SData_d->z) { auto z = static_cast *>(SData_d->z); delete z; } #if defined(LANDAU_LAYOUT_LEFT) // preallocate dynamic data, no copy auto z = static_cast *>(SData_d->ipfdf_data); #else auto z = static_cast *>(SData_d->ipfdf_data); #endif delete z; auto w = static_cast *>(SData_d->w); delete w; auto Eq_m = static_cast *>(SData_d->Eq_m); delete Eq_m; // offset auto nc = static_cast *>(SData_d->NCells); delete nc; auto soff = static_cast *>(SData_d->species_offset); delete soff; auto moff = static_cast *>(SData_d->mat_offset); delete moff; auto ipoff = static_cast *>(SData_d->ip_offset); delete ipoff; auto eoff = static_cast *>(SData_d->elem_offset); delete eoff; auto ipfoff = static_cast *>(SData_d->ipf_offset); delete ipfoff; auto coo_elem_offsets = static_cast *>((void *)SData_d->coo_elem_offsets); delete coo_elem_offsets; auto coo_elem_fullNb = static_cast *>((void *)SData_d->coo_elem_fullNb); delete coo_elem_fullNb; auto coo_elem_point_offsets = static_cast *>((void *)SData_d->coo_elem_point_offsets); delete coo_elem_point_offsets; SData_d->coo_elem_offsets = NULL; SData_d->coo_elem_point_offsets = NULL; SData_d->coo_elem_fullNb = NULL; auto coo_vals = static_cast *>((void *)SData_d->coo_vals); delete coo_vals; } PetscFunctionReturn(PETSC_SUCCESS); } #define KOKKOS_SHARED_LEVEL 0 // 0 is shared, 1 is global KOKKOS_INLINE_FUNCTION PetscErrorCode landau_mat_assemble(PetscScalar *coo_vals, const PetscScalar Aij, const PetscInt f, const PetscInt g, const PetscInt Nb, PetscInt moffset, const PetscInt elem, const PetscInt fieldA, const P4estVertexMaps *d_maps, const LandauIdx coo_elem_offsets[], const LandauIdx coo_elem_fullNb[], const LandauIdx (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1], const PetscInt glb_elem_idx, const PetscInt bid_coo_sz_batch) { PetscInt idx, q, nr, nc; const LandauIdx *const Idxs = &d_maps->gIdx[elem][fieldA][0]; PetscScalar row_scale[LANDAU_MAX_Q_FACE] = {0}, col_scale[LANDAU_MAX_Q_FACE] = {0}; { // mirror (i,j) in CreateStaticGPUData const int fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb; idx = Idxs[f]; if (idx >= 0) { nr = 1; row_scale[0] = 1.; } else { idx = -idx - 1; for (q = 0, nr = 0; q < d_maps->num_face; q++, nr++) { if (d_maps->c_maps[idx][q].gid < 0) break; row_scale[q] = d_maps->c_maps[idx][q].scale; } } idx = Idxs[g]; if (idx >= 0) { nc = 1; col_scale[0] = 1.; } else { idx = -idx - 1; nc = d_maps->num_face; for (q = 0, nc = 0; q < d_maps->num_face; q++, nc++) { if (d_maps->c_maps[idx][q].gid < 0) break; col_scale[q] = d_maps->c_maps[idx][q].scale; } } const int idx0 = bid_coo_sz_batch + coo_elem_offsets[glb_elem_idx] + fieldA * fullNb2 + fullNb * coo_elem_point_offsets[glb_elem_idx][f] + nr * coo_elem_point_offsets[glb_elem_idx][g]; for (int q = 0, idx2 = idx0; q < nr; q++) { for (int d = 0; d < nc; d++, idx2++) coo_vals[idx2] = row_scale[q] * col_scale[d] * Aij; } } return PETSC_SUCCESS; } PetscErrorCode LandauKokkosJacobian(DM plex[], const PetscInt Nq, const PetscInt Nb, const PetscInt batch_sz, const PetscInt num_grids, const PetscInt a_numCells[], PetscReal a_Eq_m[], PetscScalar a_elem_closure[], const PetscScalar a_xarray[], const LandauStaticData *SData_d, const PetscReal shift, const PetscLogEvent events[], const PetscInt a_mat_offset[], const PetscInt a_species_offset[], Mat subJ[], Mat JacP) { using scr_mem_t = Kokkos::DefaultExecutionSpace::scratch_memory_space; using real2_scr_t = Kokkos::View; using g2_scr_t = Kokkos::View; using g3_scr_t = Kokkos::View; PetscInt dim, num_cells_max, Nf_max, num_cells_batch; int nfaces = 0, vector_size = 256 / Nq; LandauCtx *ctx; PetscReal *d_Eq_m = NULL; PetscScalar *d_vertex_f = NULL; P4estVertexMaps *maps[LANDAU_MAX_GRIDS]; // this gets captured PetscContainer container; const int conc = Kokkos::DefaultExecutionSpace().concurrency(), openmp = !!(conc < 1000), team_size = (openmp == 0) ? Nq : 1; const PetscInt coo_sz_batch = SData_d->coo_size / batch_sz; // capture auto d_alpha_k = static_cast *>(SData_d->alpha); //static data const PetscReal *d_alpha = d_alpha_k->data(); const PetscInt Nftot = d_alpha_k->size(); // total number of species auto d_beta_k = static_cast *>(SData_d->beta); const PetscReal *d_beta = d_beta_k->data(); auto d_invMass_k = static_cast *>(SData_d->invMass); const PetscReal *d_invMass = d_invMass_k->data(); auto d_lambdas_k = static_cast *>(SData_d->lambdas); const PetscReal *d_lambdas = d_lambdas_k->data(); auto d_B_k = static_cast *>(SData_d->B); const PetscReal *d_BB = d_B_k->data(); auto d_D_k = static_cast *>(SData_d->D); const PetscReal *d_DD = d_D_k->data(); auto d_invJ_k = *static_cast *>(SData_d->invJ); // use Kokkos vector in kernels auto d_x_k = static_cast *>(SData_d->x); //static data const PetscReal *d_x = d_x_k->data(); auto d_y_k = static_cast *>(SData_d->y); //static data const PetscReal *d_y = d_y_k->data(); auto d_z_k = static_cast *>(SData_d->z); //static data const PetscReal *d_z = (LANDAU_DIM == 3) ? d_z_k->data() : NULL; auto d_w_k = *static_cast *>(SData_d->w); //static data const PetscReal *d_w = d_w_k.data(); // grid offsets - single vertex grid data auto d_numCells_k = static_cast *>(SData_d->NCells); const PetscInt *d_numCells = d_numCells_k->data(); auto d_species_offset_k = static_cast *>(SData_d->species_offset); const PetscInt *d_species_offset = d_species_offset_k->data(); auto d_mat_offset_k = static_cast *>(SData_d->mat_offset); const PetscInt *d_mat_offset = d_mat_offset_k->data(); auto d_ip_offset_k = static_cast *>(SData_d->ip_offset); const PetscInt *d_ip_offset = d_ip_offset_k->data(); auto d_ipf_offset_k = static_cast *>(SData_d->ipf_offset); const PetscInt *d_ipf_offset = d_ipf_offset_k->data(); auto d_elem_offset_k = static_cast *>(SData_d->elem_offset); const PetscInt *d_elem_offset = d_elem_offset_k->data(); #if defined(LANDAU_LAYOUT_LEFT) // preallocate dynamic data, including batched vertices Kokkos::View d_fdf_k = *static_cast *>(SData_d->ipfdf_data); #else Kokkos::View d_fdf_k = *static_cast *>(SData_d->ipfdf_data); #endif auto d_Eq_m_k = static_cast *>(SData_d->Eq_m); // static storage, dynamic data - E(t), copy later, single vertex // COO auto d_coo_elem_offsets_k = static_cast *>(SData_d->coo_elem_offsets); LandauIdx *d_coo_elem_offsets = (SData_d->coo_size == 0) ? NULL : d_coo_elem_offsets_k->data(); auto d_coo_elem_fullNb_k = static_cast *>(SData_d->coo_elem_fullNb); LandauIdx *d_coo_elem_fullNb = (SData_d->coo_size == 0) ? NULL : d_coo_elem_fullNb_k->data(); auto d_coo_elem_point_offsets_k = static_cast *>(SData_d->coo_elem_point_offsets); LandauIdx(*d_coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = (SData_d->coo_size == 0) ? NULL : (LandauIdx(*)[LANDAU_MAX_NQND + 1]) d_coo_elem_point_offsets_k->data(); auto d_coo_vals_k = static_cast *>(SData_d->coo_vals); PetscScalar *d_coo_vals = (SData_d->coo_size == 0) ? NULL : d_coo_vals_k->data(); PetscFunctionBegin; while (vector_size & (vector_size - 1)) vector_size = vector_size & (vector_size - 1); if (vector_size > 16) vector_size = 16; // printf("DEBUG\n"); PetscCall(PetscLogEventBegin(events[3], 0, 0, 0, 0)); PetscCall(DMGetApplicationContext(plex[0], &ctx)); PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); PetscCall(DMGetDimension(plex[0], &dim)); PetscCheck(LANDAU_DIM == dim, PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d", dim, LANDAU_DIM); if (ctx->gpu_assembly) { PetscCall(PetscObjectQuery((PetscObject)JacP, "assembly_maps", (PetscObject *)&container)); PetscCheck(container, PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); P4estVertexMaps *h_maps = NULL; PetscCall(PetscContainerGetPointer(container, &h_maps)); for (PetscInt grid = 0; grid < num_grids; grid++) { PetscCheck(h_maps[grid].d_self, PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); maps[grid] = h_maps[grid].d_self; nfaces = h_maps[grid].num_face; // nface=0 for CPU assembly } PetscCheck(d_coo_vals, PETSC_COMM_SELF, PETSC_ERR_PLIB, "d_coo_vals==0"); } else { for (PetscInt grid = 0; grid < num_grids; grid++) maps[grid] = NULL; nfaces = 0; container = NULL; } num_cells_batch = Nf_max = num_cells_max = 0; for (PetscInt grid = 0; grid < num_grids; grid++) { int Nfloc = a_species_offset[grid + 1] - a_species_offset[grid]; if (Nfloc > Nf_max) Nf_max = Nfloc; if (a_numCells[grid] > num_cells_max) num_cells_max = a_numCells[grid]; num_cells_batch += a_numCells[grid]; // we don't have a host element offset here (but in ctx) } const int elem_mat_num_cells_max_grid = container ? 0 : num_cells_max; #ifdef LAND_SUPPORT_CPU_ASS const int totDim_max = Nf_max * Nb, elem_mat_size_max = totDim_max * totDim_max; Kokkos::View d_elem_mats("element matrices", batch_sz, num_grids, elem_mat_num_cells_max_grid, elem_mat_size_max); // not used (cpu assembly) #endif const Kokkos::View> h_Eq_m_k(a_Eq_m, Nftot); if (a_elem_closure || a_xarray) { Kokkos::deep_copy(*d_Eq_m_k, h_Eq_m_k); d_Eq_m = d_Eq_m_k->data(); } else d_Eq_m = NULL; PetscCall(PetscKokkosInitializeCheck()); PetscCall(PetscLogEventEnd(events[3], 0, 0, 0, 0)); if (a_elem_closure || a_xarray) { // Jacobian, create f & df Kokkos::View *d_vertex_f_k = NULL; PetscCall(PetscLogEventBegin(events[1], 0, 0, 0, 0)); if (a_elem_closure) { PetscInt closure_sz = 0; // argh, don't have this on the host!!! for (PetscInt grid = 0; grid < num_grids; grid++) { PetscInt nfloc = a_species_offset[grid + 1] - a_species_offset[grid]; closure_sz += Nb * nfloc * a_numCells[grid]; } closure_sz *= batch_sz; d_vertex_f_k = new Kokkos::View("closure", closure_sz); const Kokkos::View> h_closure_k(a_elem_closure, closure_sz); // Vertex data for each element Kokkos::deep_copy(*d_vertex_f_k, h_closure_k); d_vertex_f = d_vertex_f_k->data(); } else { d_vertex_f = (PetscScalar *)a_xarray; } PetscCall(PetscLogEventEnd(events[1], 0, 0, 0, 0)); PetscCall(PetscLogEventBegin(events[8], 0, 0, 0, 0)); PetscCall(PetscLogGpuTimeBegin()); const int scr_bytes_fdf = real2_scr_t::shmem_size(Nf_max, Nb); PetscCall(PetscInfo(plex[0], "df & f shared memory size: %d bytes in level, %d num cells total=%d, team size=%d, vector size=%d, #face=%d, Nf_max=%d\n", scr_bytes_fdf, KOKKOS_SHARED_LEVEL, (int)(num_cells_batch * batch_sz), team_size, vector_size, nfaces, (int)Nf_max)); Kokkos::parallel_for( "f, df", Kokkos::TeamPolicy<>(num_cells_batch * batch_sz, team_size, vector_size).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes_fdf)), KOKKOS_LAMBDA(const team_member team) { const PetscInt b_Nelem = d_elem_offset[num_grids], b_elem_idx = team.league_rank() % b_Nelem, b_id = team.league_rank() / b_Nelem, IPf_sz_glb = d_ipf_offset[num_grids]; // find my grid PetscInt grid = 0; while (b_elem_idx >= d_elem_offset[grid + 1]) grid++; { const PetscInt loc_nip = d_numCells[grid] * Nq, loc_Nf = d_species_offset[grid + 1] - d_species_offset[grid], loc_elem = b_elem_idx - d_elem_offset[grid]; const PetscInt moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, d_mat_offset); { real2_scr_t s_coef_k(team.team_scratch(KOKKOS_SHARED_LEVEL), Nf_max, Nb); PetscScalar *coef; const PetscReal *invJe = &d_invJ_k((d_ip_offset[grid] + loc_elem * Nq) * dim * dim); // un pack IPData if (!maps[grid]) { coef = &d_vertex_f[b_id * IPf_sz_glb + d_ipf_offset[grid] + loc_elem * Nb * loc_Nf]; // closure and IP indexing are the same } else { coef = s_coef_k.data(); Kokkos::parallel_for(Kokkos::TeamThreadRange(team, 0, (int)loc_Nf), [=](int f) { //for (int f = 0; f < loc_Nf; ++f) { Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, 0, (int)Nb), [=](int b) { //for (int b = 0; b < Nb; ++b) { LandauIdx idx = maps[grid]->gIdx[loc_elem][f][b]; if (idx >= 0) { coef[f * Nb + b] = d_vertex_f[idx + moffset]; // xarray data, not IP, need raw array to deal with CPU assemble case (not used) } else { idx = -idx - 1; coef[f * Nb + b] = 0; for (int q = 0; q < maps[grid]->num_face; q++) { PetscInt id = maps[grid]->c_maps[idx][q].gid; PetscScalar scale = maps[grid]->c_maps[idx][q].scale; if (id >= 0) coef[f * Nb + b] += scale * d_vertex_f[id + moffset]; } } }); }); } team.team_barrier(); Kokkos::parallel_for(Kokkos::TeamThreadRange(team, 0, Nq), [=](int myQi) { const PetscReal *const invJ = &invJe[myQi * dim * dim]; // b_elem_idx: batch element index const PetscReal *Bq = &d_BB[myQi * Nb], *Dq = &d_DD[myQi * Nb * dim]; Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, 0, (int)loc_Nf), [=](int f) { PetscInt b, e, d; PetscReal refSpaceDer[LANDAU_DIM]; const PetscInt idx = d_ipf_offset[grid] + f * loc_nip + loc_elem * Nq + myQi; d_fdf_k(b_id, 0, idx) = 0.0; for (d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0; for (b = 0; b < Nb; ++b) { d_fdf_k(b_id, 0, idx) += Bq[b] * PetscRealPart(coef[f * Nb + b]); for (d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b * dim + d] * PetscRealPart(coef[f * Nb + b]); } for (d = 0; d < dim; ++d) { for (e = 0, d_fdf_k(b_id, d + 1, idx) = 0.0; e < dim; ++e) d_fdf_k(b_id, d + 1, idx) += invJ[e * dim + d] * refSpaceDer[e]; } }); // f }); // myQi } team.team_barrier(); } // 'grid' scope }); // global elems - fdf Kokkos::fence(); PetscCall(PetscLogGpuTimeEnd()); PetscCall(PetscLogEventEnd(events[8], 0, 0, 0, 0)); // Jacobian size_t maximum_shared_mem_size = 64000; PetscDevice device; PetscCall(PetscDeviceGetDefault_Internal(&device)); PetscCall(PetscDeviceGetAttribute(device, PETSC_DEVICE_ATTR_SIZE_T_SHARED_MEM_PER_BLOCK, &maximum_shared_mem_size)); size_t jac_scr_bytes = (size_t)2 * (g2_scr_t::shmem_size(dim, Nf_max, Nq) + g3_scr_t::shmem_size(dim, dim, Nf_max, Nq)); const int jac_shared_level = (jac_scr_bytes > maximum_shared_mem_size) ? 1 : KOKKOS_SHARED_LEVEL; // device function/lambda auto jac_lambda = KOKKOS_LAMBDA(const team_member team) { const PetscInt b_Nelem = d_elem_offset[num_grids], b_elem_idx = team.league_rank() % b_Nelem, b_id = team.league_rank() / b_Nelem; // find my grid PetscInt grid = 0; while (b_elem_idx >= d_elem_offset[grid + 1]) grid++; { const PetscInt loc_Nf = d_species_offset[grid + 1] - d_species_offset[grid], loc_elem = b_elem_idx - d_elem_offset[grid]; const PetscInt moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, d_mat_offset); const PetscInt f_off = d_species_offset[grid]; #ifdef LAND_SUPPORT_CPU_ASS const PetscInt totDim = loc_Nf * Nb; #endif g2_scr_t g2(team.team_scratch(jac_shared_level), dim, loc_Nf, Nq); g3_scr_t g3(team.team_scratch(jac_shared_level), dim, dim, loc_Nf, Nq); g2_scr_t gg2(team.team_scratch(jac_shared_level), dim, loc_Nf, Nq); g3_scr_t gg3(team.team_scratch(jac_shared_level), dim, dim, loc_Nf, Nq); // get g2[] & g3[] Kokkos::parallel_for(Kokkos::TeamThreadRange(team, 0, Nq), [=](int myQi) { using Kokkos::parallel_reduce; const PetscInt jpidx_glb = d_ip_offset[grid] + loc_elem * Nq + myQi; const PetscReal *invJ = &d_invJ_k(jpidx_glb * dim * dim); const PetscReal vj[3] = {d_x[jpidx_glb], d_y[jpidx_glb], d_z ? d_z[jpidx_glb] : 0}, wj = d_w[jpidx_glb]; landau_inner_red::TensorValueType gg_temp; // reduce on part of gg2 and g33 for IP jpidx_g Kokkos::parallel_reduce( Kokkos::ThreadVectorRange(team, (int)d_ip_offset[num_grids]), [=](const int &ipidx, landau_inner_red::TensorValueType &ggg) { const PetscReal wi = d_w[ipidx], x = d_x[ipidx], y = d_y[ipidx]; PetscReal temp1[3] = {0, 0, 0}, temp2 = 0; PetscInt fieldB, d2, d3, f_off_r, grid_r, ipidx_g, nip_loc_r, loc_Nf_r; #if LANDAU_DIM == 2 PetscReal Ud[2][2], Uk[2][2], mask = (PetscAbs(vj[0] - x) < 100 * PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[1] - y) < 100 * PETSC_SQRT_MACHINE_EPSILON) ? 0. : 1.; LandauTensor2D(vj, x, y, Ud, Uk, mask); #else PetscReal U[3][3], z = d_z[ipidx], mask = (PetscAbs(vj[0]-x) < 100*PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[1]-y) < 100*PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[2]-z) < 100*PETSC_SQRT_MACHINE_EPSILON) ? 0. : 1.; LandauTensor3D(vj, x, y, z, U, mask); #endif grid_r = 0; while (ipidx >= d_ip_offset[grid_r + 1]) grid_r++; // yuck search for grid f_off_r = d_species_offset[grid_r]; ipidx_g = ipidx - d_ip_offset[grid_r]; nip_loc_r = d_numCells[grid_r] * Nq; loc_Nf_r = d_species_offset[grid_r + 1] - d_species_offset[grid_r]; for (fieldB = 0; fieldB < loc_Nf_r; ++fieldB) { // fieldB is \beta d_lambdas[grid][grid_r] const PetscInt idx = d_ipf_offset[grid_r] + fieldB * nip_loc_r + ipidx_g; const PetscReal ff1 = d_beta[fieldB + f_off_r] * d_lambdas[LANDAU_MAX_GRIDS * grid + grid_r], ff2 = d_invMass[fieldB + f_off_r] * ff1; temp1[0] += d_fdf_k(b_id, 1, idx) * ff2; temp1[1] += d_fdf_k(b_id, 2, idx) * ff2; #if LANDAU_DIM == 3 temp1[2] += d_fdf_k(b_id, 3, idx) * ff2; #endif temp2 += d_fdf_k(b_id, 0, idx) * ff1; } 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(gg_temp)); // add alpha and put in gg2/3 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, (int)loc_Nf), [&](const int &fieldA) { // \alpha PetscInt d2, d3; for (d2 = 0; d2 < dim; d2++) { gg2(d2, fieldA, myQi) = gg_temp.gg2[d2] * d_alpha[fieldA + f_off]; for (d3 = 0; d3 < dim; d3++) gg3(d2, d3, fieldA, myQi) = -gg_temp.gg3[d2][d3] * d_alpha[fieldA + f_off] * d_invMass[fieldA + f_off]; } }); /* add electric field term once per IP */ Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, (int)loc_Nf), [&](const int &fieldA) { gg2(dim - 1, fieldA, myQi) += d_Eq_m[fieldA + f_off]; }); Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, (int)loc_Nf), [=](const int &fieldA) { int d, d2, d3, dp; /* 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) += invJ[d * dim + d2] * gg2(d2, fieldA, myQi); g3(d, d2, fieldA, myQi) = 0; for (d3 = 0; d3 < dim; ++d3) { for (dp = 0; dp < dim; ++dp) g3(d, d2, fieldA, myQi) += invJ[d * dim + d3] * gg3(d3, dp, fieldA, myQi) * invJ[d2 * dim + dp]; } g3(d, d2, fieldA, myQi) *= wj; } g2(d, fieldA, myQi) *= wj; } }); }); // Nq team.team_barrier(); { /* assemble */ for (PetscInt fieldA = 0; fieldA < loc_Nf; fieldA++) { /* assemble */ Kokkos::parallel_for(Kokkos::TeamThreadRange(team, 0, Nb), [=](int f) { Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, 0, Nb), [=](int g) { PetscScalar t = 0; for (int qj = 0; qj < Nq; qj++) { // look at others integration points const PetscReal *BJq = &d_BB[qj * Nb], *DIq = &d_DD[qj * Nb * dim]; for (int d = 0; d < dim; ++d) { t += DIq[f * dim + d] * g2(d, fieldA, qj) * BJq[g]; for (int d2 = 0; d2 < dim; ++d2) t += DIq[f * dim + d] * g3(d, d2, fieldA, qj) * DIq[g * dim + d2]; } } if (elem_mat_num_cells_max_grid) { // CPU assembly #ifdef LAND_SUPPORT_CPU_ASS const PetscInt fOff = (fieldA * Nb + f) * totDim + fieldA * Nb + g; d_elem_mats(b_id, grid, loc_elem, fOff) = t; #endif } else { static_cast(landau_mat_assemble(d_coo_vals, t, f, g, Nb, moffset, loc_elem, fieldA, maps[grid], d_coo_elem_offsets, d_coo_elem_fullNb, d_coo_elem_point_offsets, b_elem_idx, b_id * coo_sz_batch)); } }); }); } } } // scope with 'grid' }; #if defined(PETSC_HAVE_HIP) const int lbound2 = 1; #else const int lbound2 = 2; #endif PetscCall(PetscLogEventBegin(events[4], 0, 0, 0, 0)); PetscCall(PetscLogGpuTimeBegin()); PetscCall(PetscInfo(plex[0], "Jacobian shared memory size: %zu bytes in level %s (max shared=%zu), num cells total=%d, team size=%d, vector size=%d, #face=%d, Nf_max=%d\n", jac_scr_bytes, jac_shared_level == 0 ? "local" : "global", maximum_shared_mem_size, (int)(num_cells_batch * batch_sz), team_size, vector_size, nfaces, (int)Nf_max)); Kokkos::parallel_for("Jacobian", Kokkos::TeamPolicy>(num_cells_batch * batch_sz, team_size, vector_size).set_scratch_size(jac_shared_level, Kokkos::PerTeam(jac_scr_bytes)), jac_lambda); Kokkos::fence(); PetscCall(PetscLogGpuTimeEnd()); PetscCall(PetscLogEventEnd(events[4], 0, 0, 0, 0)); if (d_vertex_f_k) delete d_vertex_f_k; } else { // mass PetscCall(PetscLogEventBegin(events[16], 0, 0, 0, 0)); PetscCall(PetscLogGpuTimeBegin()); PetscCall(PetscInfo(plex[0], "Mass team size=%d vector size=%d #face=%d Nb=%" PetscInt_FMT ", %s assembly\n", team_size, vector_size, nfaces, Nb, d_coo_vals ? "COO" : "CSR")); Kokkos::parallel_for( "Mass", Kokkos::TeamPolicy<>(num_cells_batch * batch_sz, team_size, vector_size), KOKKOS_LAMBDA(const team_member team) { const PetscInt b_Nelem = d_elem_offset[num_grids], b_elem_idx = team.league_rank() % b_Nelem, b_id = team.league_rank() / b_Nelem; // find my grid PetscInt grid = 0; while (b_elem_idx >= d_elem_offset[grid + 1]) grid++; { const PetscInt loc_Nf = d_species_offset[grid + 1] - d_species_offset[grid], loc_elem = b_elem_idx - d_elem_offset[grid], jpidx_0 = d_ip_offset[grid] + loc_elem * Nq; #ifdef LAND_SUPPORT_CPU_ASS const PetscInt totDim = loc_Nf * Nb; #endif const PetscInt moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, d_mat_offset); for (int fieldA = 0; fieldA < loc_Nf; fieldA++) { /* assemble */ Kokkos::parallel_for(Kokkos::TeamThreadRange(team, 0, Nb), [=](int f) { Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, 0, Nb), [=](int g) { PetscScalar t = 0; for (int qj = 0; qj < Nq; qj++) { // look at others integration points const PetscReal *BJq = &d_BB[qj * Nb]; const PetscInt jpidx_glb = jpidx_0 + qj; if (dim == 2) { t += BJq[f] * d_w_k(jpidx_glb) * shift * BJq[g] * 2. * PETSC_PI; } else { t += BJq[f] * d_w_k(jpidx_glb) * shift * BJq[g]; } } if (elem_mat_num_cells_max_grid) { #ifdef LAND_SUPPORT_CPU_ASS const PetscInt fOff = (fieldA * Nb + f) * totDim + fieldA * Nb + g; d_elem_mats(b_id, grid, loc_elem, fOff) = t; #endif } else { static_cast(landau_mat_assemble(d_coo_vals, t, f, g, Nb, moffset, loc_elem, fieldA, maps[grid], d_coo_elem_offsets, d_coo_elem_fullNb, d_coo_elem_point_offsets, b_elem_idx, b_id * coo_sz_batch)); } }); }); } // field } // grid }); Kokkos::fence(); PetscCall(PetscLogGpuTimeEnd()); PetscCall(PetscLogEventEnd(events[16], 0, 0, 0, 0)); } if (d_coo_vals) { PetscCall(MatSetValuesCOO(JacP, d_coo_vals, ADD_VALUES)); #if !defined(LAND_SUPPORT_CPU_ASS) Kokkos::fence(); // for timers #endif } else if (elem_mat_num_cells_max_grid) { // CPU assembly #ifdef LAND_SUPPORT_CPU_ASS Kokkos::View::HostMirror h_elem_mats = Kokkos::create_mirror_view(d_elem_mats); Kokkos::deep_copy(h_elem_mats, d_elem_mats); PetscCheck(!container, PETSC_COMM_SELF, PETSC_ERR_PLIB, "?????"); for (PetscInt b_id = 0; b_id < batch_sz; b_id++) { // OpenMP (once) for (PetscInt grid = 0; grid < num_grids; grid++) { PetscSection section, globalSection; PetscInt cStart, cEnd; Mat B = subJ[LAND_PACK_IDX(b_id, grid)]; PetscInt moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, a_mat_offset), nloc, nzl, colbuf[1024], row; const PetscInt *cols; const PetscScalar *vals; PetscCall(PetscLogEventBegin(events[5], 0, 0, 0, 0)); PetscCall(DMPlexGetHeightStratum(plex[grid], 0, &cStart, &cEnd)); PetscCall(DMGetLocalSection(plex[grid], §ion)); PetscCall(DMGetGlobalSection(plex[grid], &globalSection)); PetscCall(PetscLogEventEnd(events[5], 0, 0, 0, 0)); PetscCall(PetscLogEventBegin(events[6], 0, 0, 0, 0)); for (PetscInt ej = cStart; ej < cEnd; ++ej) { const PetscScalar *elMat = &h_elem_mats(b_id, grid, ej - cStart, 0); PetscCall(DMPlexMatSetClosure(plex[grid], section, globalSection, B, ej, elMat, ADD_VALUES)); if (grid == 0 && ej == -1) { const PetscInt loc_Nf = a_species_offset[grid + 1] - a_species_offset[grid], totDim = loc_Nf * Nb; int d, f; PetscPrintf(PETSC_COMM_SELF, "Kokkos Element matrix %d/%d\n", 1, (int)a_numCells[grid]); 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"); } exit(14); } } // move nest matrix to global JacP PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY)); PetscCall(MatGetSize(B, &nloc, NULL)); for (int i = 0; i < nloc; i++) { PetscCall(MatGetRow(B, i, &nzl, &cols, &vals)); PetscCheck(nzl <= 1024, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT, nzl); for (int j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset; row = i + moffset; PetscCall(MatSetValues(JacP, 1, &row, nzl, colbuf, vals, ADD_VALUES)); PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals)); } PetscCall(MatDestroy(&subJ[LAND_PACK_IDX(b_id, grid)])); PetscCall(PetscLogEventEnd(events[6], 0, 0, 0, 0)); } // grids } #else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "CPU assembly not supported"); #endif } PetscFunctionReturn(PETSC_SUCCESS); } } // extern "C" #endif