1 /* 2 Implements the Kokkos kernel 3 */ 4 #include <petscconf.h> 5 #include <petscvec_kokkos.hpp> 6 #include <petsc/private/dmpleximpl.h> /*I "petscdmplex.h" I*/ 7 #include <petsclandau.h> 8 #include <petscts.h> 9 10 #include <Kokkos_Core.hpp> 11 #include <cstdio> 12 typedef Kokkos::TeamPolicy<>::member_type team_member; 13 #include "../land_tensors.h" 14 #include <petscaijdevice.h> 15 16 namespace landau_inner_red { // namespace helps with name resolution in reduction identity 17 template< class ScalarType > 18 struct array_type { 19 ScalarType gg2[LANDAU_DIM]; 20 ScalarType gg3[LANDAU_DIM][LANDAU_DIM]; 21 22 KOKKOS_INLINE_FUNCTION // Default constructor - Initialize to 0's 23 array_type() { 24 for (int j = 0; j < LANDAU_DIM; j++) { 25 gg2[j] = 0; 26 for (int k = 0; k < LANDAU_DIM; k++) { 27 gg3[j][k] = 0; 28 } 29 } 30 } 31 KOKKOS_INLINE_FUNCTION // Copy Constructor 32 array_type(const array_type & rhs) { 33 for (int j = 0; j < LANDAU_DIM; j++) { 34 gg2[j] = rhs.gg2[j]; 35 for (int k = 0; k < LANDAU_DIM; k++) { 36 gg3[j][k] = rhs.gg3[j][k]; 37 } 38 } 39 } 40 KOKKOS_INLINE_FUNCTION // add operator 41 array_type& operator += (const array_type& src) 42 { 43 for (int j = 0; j < LANDAU_DIM; j++) { 44 gg2[j] += src.gg2[j]; 45 for (int k = 0; k < LANDAU_DIM; k++) { 46 gg3[j][k] += src.gg3[j][k]; 47 } 48 } 49 return *this; 50 } 51 KOKKOS_INLINE_FUNCTION // volatile add operator 52 void operator += (const volatile array_type& src) volatile 53 { 54 for (int j = 0; j < LANDAU_DIM; j++) { 55 gg2[j] += src.gg2[j]; 56 for (int k = 0; k < LANDAU_DIM; k++) { 57 gg3[j][k] += src.gg3[j][k]; 58 } 59 } 60 } 61 }; 62 typedef array_type<PetscReal> TensorValueType; // used to simplify code below 63 } 64 65 namespace Kokkos { //reduction identity must be defined in Kokkos namespace 66 template<> 67 struct reduction_identity< landau_inner_red::TensorValueType > { 68 KOKKOS_FORCEINLINE_FUNCTION static landau_inner_red::TensorValueType sum() { 69 return landau_inner_red::TensorValueType(); 70 } 71 }; 72 } 73 74 extern "C" { 75 PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps maps[], pointInterpolationP4est (*pointMaps)[LANDAU_MAX_Q_FACE], PetscInt Nf[], PetscInt Nq, PetscInt grid) 76 { 77 P4estVertexMaps h_maps; /* host container */ 78 const Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_points ((pointInterpolationP4est*)pointMaps, maps[grid].num_reduced); 79 const Kokkos::View< LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_gidx ((LandauIdx*)maps[grid].gIdx, maps[grid].num_elements); 80 Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight> *d_points = new Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>("points", maps[grid].num_reduced); 81 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[grid].num_elements); 82 83 PetscFunctionBegin; 84 Kokkos::deep_copy (*d_gidx, h_gidx); 85 Kokkos::deep_copy (*d_points, h_points); 86 h_maps.num_elements = maps[grid].num_elements; 87 h_maps.num_face = maps[grid].num_face; 88 h_maps.num_reduced = maps[grid].num_reduced; 89 h_maps.deviceType = maps[grid].deviceType; 90 h_maps.numgrids = maps[grid].numgrids; 91 h_maps.Nf = Nf[grid]; 92 h_maps.c_maps = (pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE]) d_points->data(); 93 maps[grid].vp1 = (void*)d_points; 94 h_maps.gIdx = (LandauIdx (*)[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ]) d_gidx->data(); 95 maps[grid].vp2 = (void*)d_gidx; 96 { 97 Kokkos::View<P4estVertexMaps, Kokkos::HostSpace> h_maps_k(&h_maps); 98 Kokkos::View<P4estVertexMaps> *d_maps_k = new Kokkos::View<P4estVertexMaps>(Kokkos::create_mirror(Kokkos::DefaultExecutionSpace::memory_space(),h_maps_k)); 99 Kokkos::deep_copy (*d_maps_k, h_maps_k); 100 maps[grid].d_self = d_maps_k->data(); 101 maps[grid].vp3 = (void*)d_maps_k; 102 } 103 PetscFunctionReturn(0); 104 } 105 PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps maps[], PetscInt num_grids) 106 { 107 PetscFunctionBegin; 108 for (PetscInt grid=0;grid<num_grids;grid++) { 109 auto a = static_cast<Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>*>(maps[grid].vp1); 110 auto b = static_cast<Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight>*>(maps[grid].vp2); 111 auto c = static_cast<Kokkos::View<P4estVertexMaps*>*>(maps[grid].vp3); 112 delete a; delete b; delete c; 113 } 114 PetscFunctionReturn(0); 115 } 116 117 PetscErrorCode LandauKokkosStaticDataSet(DM plex, const PetscInt Nq, const PetscInt batch_sz, const PetscInt num_grids, 118 PetscInt a_numCells[], PetscInt a_species_offset[], PetscInt a_mat_offset[], 119 PetscReal a_nu_alpha[], PetscReal a_nu_beta[], PetscReal a_invMass[], PetscReal a_invJ[], 120 PetscReal a_x[], PetscReal a_y[], PetscReal a_z[], PetscReal a_w[], LandauStaticData *SData_d) 121 { 122 PetscReal *BB,*DD; 123 PetscErrorCode ierr; 124 PetscTabulation *Tf; 125 PetscInt dim; 126 PetscInt Nb=Nq,ip_offset[LANDAU_MAX_GRIDS+1],ipf_offset[LANDAU_MAX_GRIDS+1],elem_offset[LANDAU_MAX_GRIDS+1],nip,IPf_sz,Nftot; 127 PetscDS prob; 128 129 PetscFunctionBegin; 130 ierr = DMGetDimension(plex, &dim);CHKERRQ(ierr); 131 ierr = DMGetDS(plex, &prob);CHKERRQ(ierr); 132 PetscCheckFalse(LANDAU_DIM != dim,PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %D != LANDAU_DIM %d",dim,LANDAU_DIM); 133 ierr = PetscDSGetTabulation(prob, &Tf);CHKERRQ(ierr); 134 BB = Tf[0]->T[0]; DD = Tf[0]->T[1]; 135 ip_offset[0] = ipf_offset[0] = elem_offset[0] = 0; 136 nip = 0; 137 IPf_sz = 0; 138 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 139 PetscInt nfloc = a_species_offset[grid+1] - a_species_offset[grid]; 140 elem_offset[grid+1] = elem_offset[grid] + a_numCells[grid]; 141 nip += a_numCells[grid]*Nq; 142 ip_offset[grid+1] = nip; 143 IPf_sz += Nq*nfloc*a_numCells[grid]; 144 ipf_offset[grid+1] = IPf_sz; 145 } 146 Nftot = a_species_offset[num_grids]; 147 ierr = PetscKokkosInitializeCheck();CHKERRQ(ierr); 148 { 149 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_alpha (a_nu_alpha, Nftot); 150 auto alpha = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("alpha", Nftot); 151 SData_d->alpha = static_cast<void*>(alpha); 152 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_beta (a_nu_beta, Nftot); 153 auto beta = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("beta", Nftot); 154 SData_d->beta = static_cast<void*>(beta); 155 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invMass (a_invMass,Nftot); 156 auto invMass = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("invMass", Nftot); 157 SData_d->invMass = static_cast<void*>(invMass); 158 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_BB (BB,Nq*Nb); 159 auto B = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("B", Nq*Nb); 160 SData_d->B = static_cast<void*>(B); 161 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_DD (DD,Nq*Nb*dim); 162 auto D = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("D", Nq*Nb*dim); 163 SData_d->D = static_cast<void*>(D); 164 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invJ (a_invJ, nip*dim*dim); 165 auto invJ = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("invJ", nip*dim*dim); 166 SData_d->invJ = static_cast<void*>(invJ); 167 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_x (a_x, nip); 168 auto x = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("x", nip); 169 SData_d->x = static_cast<void*>(x); 170 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_y (a_y, nip); 171 auto y = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("y", nip); 172 SData_d->y = static_cast<void*>(y); 173 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_w (a_w, nip); 174 auto w = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("w", nip); 175 SData_d->w = static_cast<void*>(w); 176 177 Kokkos::deep_copy (*alpha, h_alpha); 178 Kokkos::deep_copy (*beta, h_beta); 179 Kokkos::deep_copy (*invMass, h_invMass); 180 Kokkos::deep_copy (*B, h_BB); 181 Kokkos::deep_copy (*D, h_DD); 182 Kokkos::deep_copy (*invJ, h_invJ); 183 Kokkos::deep_copy (*x, h_x); 184 Kokkos::deep_copy (*y, h_y); 185 Kokkos::deep_copy (*w, h_w); 186 187 if (dim==3) { 188 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_z (a_z , dim==3 ? nip : 0); 189 auto z = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("z", nip); 190 SData_d->z = static_cast<void*>(z); 191 Kokkos::deep_copy (*z, h_z); 192 } else SData_d->z = NULL; 193 194 // 195 const Kokkos::View<PetscInt*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_NCells (a_numCells, num_grids); 196 auto nc = new Kokkos::View<PetscInt*, Kokkos::LayoutLeft> ("NCells",num_grids); 197 SData_d->NCells = static_cast<void*>(nc); 198 Kokkos::deep_copy (*nc, h_NCells); 199 200 const Kokkos::View<PetscInt*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_species_offset (a_species_offset, num_grids+1); 201 auto soff = new Kokkos::View<PetscInt*, Kokkos::LayoutLeft> ("species_offset",num_grids+1); 202 SData_d->species_offset = static_cast<void*>(soff); 203 Kokkos::deep_copy (*soff, h_species_offset); 204 205 const Kokkos::View<PetscInt*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_mat_offset (a_mat_offset, num_grids+1); 206 auto moff = new Kokkos::View<PetscInt*, Kokkos::LayoutLeft> ("mat_offset",num_grids+1); 207 SData_d->mat_offset = static_cast<void*>(moff); 208 Kokkos::deep_copy (*moff, h_mat_offset); 209 210 const Kokkos::View<PetscInt*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_ip_offset (ip_offset, num_grids+1); 211 auto ipoff = new Kokkos::View<PetscInt*, Kokkos::LayoutLeft> ("ip_offset",num_grids+1); 212 SData_d->ip_offset = static_cast<void*>(ipoff); 213 Kokkos::deep_copy (*ipoff, h_ip_offset); 214 215 const Kokkos::View<PetscInt*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_elem_offset (elem_offset, num_grids+1); 216 auto eoff = new Kokkos::View<PetscInt*, Kokkos::LayoutLeft> ("elem_offset",num_grids+1); 217 SData_d->elem_offset = static_cast<void*>(eoff); 218 Kokkos::deep_copy (*eoff, h_elem_offset); 219 220 const Kokkos::View<PetscInt*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_ipf_offset (ipf_offset, num_grids+1); 221 auto ipfoff = new Kokkos::View<PetscInt*, Kokkos::LayoutLeft> ("ipf_offset",num_grids+1); 222 SData_d->ipf_offset = static_cast<void*>(ipfoff); 223 Kokkos::deep_copy (*ipfoff, h_ipf_offset); 224 #if defined(LANDAU_LAYOUT_LEFT) // preallocate dynamic data, no copy 225 auto ipfdf_data = new Kokkos::View<PetscReal***, Kokkos::LayoutLeft > ("fdf", batch_sz, dim+1, IPf_sz); 226 #else 227 auto ipfdf_data = new Kokkos::View<PetscReal***, Kokkos::LayoutRight > ("fdf",batch_sz, dim+1, IPf_sz); 228 #endif 229 SData_d->ipfdf_data = static_cast<void*>(ipfdf_data); 230 auto Eq_m = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("Eq_m",Nftot); // allocate but do not set, same for whole batch 231 SData_d->Eq_m = static_cast<void*>(Eq_m); 232 const Kokkos::View<LandauIdx*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_coo_elem_offsets (SData_d->coo_elem_offsets, SData_d->n_coo_cellsTot+1); 233 auto coo_elem_offsets = new Kokkos::View<LandauIdx*, Kokkos::LayoutLeft> ("coo_elem_offsets",SData_d->n_coo_cellsTot+1); 234 const Kokkos::View<LandauIdx*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_coo_elem_fullNb (SData_d->coo_elem_fullNb, SData_d->n_coo_cellsTot); 235 auto coo_elem_fullNb = new Kokkos::View<LandauIdx*, Kokkos::LayoutLeft> ("coo_elem_offsets",SData_d->n_coo_cellsTot); 236 const Kokkos::View<LandauIdx**, Kokkos::LayoutRight, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_coo_elem_point_offsets ((LandauIdx*)SData_d->coo_elem_point_offsets, SData_d->n_coo_cellsTot, LANDAU_MAX_NQ+1); 237 auto coo_elem_point_offsets = new Kokkos::View<LandauIdx**, Kokkos::LayoutRight> ("coo_elem_point_offsets",SData_d->n_coo_cellsTot,LANDAU_MAX_NQ+1); 238 // assign & copy 239 Kokkos::deep_copy (*coo_elem_offsets, h_coo_elem_offsets); 240 Kokkos::deep_copy (*coo_elem_fullNb, h_coo_elem_fullNb); 241 Kokkos::deep_copy (*coo_elem_point_offsets, h_coo_elem_point_offsets); 242 // need to free this now and use pointer space 243 ierr = PetscFree3(SData_d->coo_elem_offsets,SData_d->coo_elem_fullNb,SData_d->coo_elem_point_offsets);CHKERRQ(ierr); 244 SData_d->coo_elem_offsets = reinterpret_cast<LandauIdx*>(coo_elem_offsets); 245 SData_d->coo_elem_fullNb = reinterpret_cast<LandauIdx*>(coo_elem_fullNb); 246 SData_d->coo_elem_point_offsets = reinterpret_cast<LandauIdx (*)[LANDAU_MAX_NQ+1]>(coo_elem_point_offsets); 247 } 248 SData_d->maps = NULL; // not used 249 PetscFunctionReturn(0); 250 } 251 252 PetscErrorCode LandauKokkosStaticDataClear(LandauStaticData *SData_d) 253 { 254 PetscFunctionBegin; 255 if (SData_d->alpha) { 256 auto alpha = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->alpha); 257 delete alpha; 258 SData_d->alpha = NULL; 259 auto beta = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->beta); 260 delete beta; 261 auto invMass = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invMass); 262 delete invMass; 263 auto B = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->B); 264 delete B; 265 auto D = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->D); 266 delete D; 267 auto invJ = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invJ); 268 delete invJ; 269 auto x = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->x); 270 delete x; 271 auto y = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->y); 272 delete y; 273 if (SData_d->z) { 274 auto z = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->z); 275 delete z; 276 } 277 #if defined(LANDAU_LAYOUT_LEFT) // preallocate dynamic data, no copy 278 auto z = static_cast<Kokkos::View<PetscReal***, Kokkos::LayoutLeft>*>(SData_d->ipfdf_data); 279 #else 280 auto z = static_cast<Kokkos::View<PetscReal***, Kokkos::LayoutRight>*>(SData_d->ipfdf_data); 281 #endif 282 delete z; 283 auto w = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->w); 284 delete w; 285 auto Eq_m = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->Eq_m); 286 delete Eq_m; 287 // offset 288 auto nc = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->NCells); 289 delete nc; 290 auto soff = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->species_offset); 291 delete soff; 292 auto moff = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->mat_offset); 293 delete moff; 294 auto ipoff = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->ip_offset); 295 delete ipoff; 296 auto eoff = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->elem_offset); 297 delete eoff; 298 auto ipfoff = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->ipf_offset); 299 delete ipfoff; 300 auto coo_elem_offsets = static_cast<Kokkos::View<LandauIdx*, Kokkos::LayoutLeft>*>((void*)SData_d->coo_elem_offsets); 301 delete coo_elem_offsets; 302 auto coo_elem_fullNb = static_cast<Kokkos::View<LandauIdx*, Kokkos::LayoutLeft>*>((void*)SData_d->coo_elem_fullNb); 303 delete coo_elem_fullNb; 304 auto coo_elem_point_offsets = static_cast<Kokkos::View<LandauIdx**, Kokkos::LayoutRight>*>((void*)SData_d->coo_elem_point_offsets); 305 delete coo_elem_point_offsets; 306 SData_d->coo_elem_offsets = NULL; 307 SData_d->coo_elem_point_offsets = NULL; 308 SData_d->coo_elem_fullNb = NULL; 309 } 310 PetscFunctionReturn(0); 311 } 312 313 #define KOKKOS_SHARED_LEVEL 1 314 KOKKOS_INLINE_FUNCTION 315 PetscErrorCode landau_mat_assemble(PetscSplitCSRDataStructure d_mat, PetscScalar *coo_vals, const PetscScalar Aij, const PetscInt f, const PetscInt g, const PetscInt Nb, 316 PetscInt moffset, const PetscInt elem, const PetscInt fieldA, const P4estVertexMaps *d_maps, 317 const LandauIdx coo_elem_offsets[], const LandauIdx coo_elem_fullNb[], Kokkos::View<LandauIdx**, Kokkos::LayoutRight> &coo_elem_point_offsets_k, 318 const PetscInt glb_elem_idx, const PetscInt bid_coo_sz_batch) 319 { 320 PetscInt idx,q,nr,nc,d; 321 const LandauIdx *const Idxs = &d_maps->gIdx[elem][fieldA][0]; 322 PetscScalar row_scale[LANDAU_MAX_Q_FACE]={0},col_scale[LANDAU_MAX_Q_FACE]={0}; 323 if (coo_vals) { // mirror (i,j) in CreateStaticGPUData 324 const int fullNb = coo_elem_fullNb[glb_elem_idx],fullNb2=fullNb*fullNb; 325 idx = Idxs[f]; 326 if (idx >= 0) { 327 nr = 1; 328 row_scale[0] = 1.; 329 } else { 330 idx = -idx - 1; 331 for (q = 0, nr = 0; q < d_maps->num_face; q++, nr++) { 332 if (d_maps->c_maps[idx][q].gid < 0) break; 333 row_scale[q] = d_maps->c_maps[idx][q].scale; 334 } 335 } 336 idx = Idxs[g]; 337 if (idx >= 0) { 338 nc = 1; 339 col_scale[0] = 1.; 340 } else { 341 idx = -idx - 1; 342 nc = d_maps->num_face; 343 for (q = 0, nc = 0; q < d_maps->num_face; q++, nc++) { 344 if (d_maps->c_maps[idx][q].gid < 0) break; 345 col_scale[q] = d_maps->c_maps[idx][q].scale; 346 } 347 } 348 const int idx0 = bid_coo_sz_batch + coo_elem_offsets[glb_elem_idx] + fieldA*fullNb2 + fullNb * coo_elem_point_offsets_k(glb_elem_idx,f) + nr * coo_elem_point_offsets_k(glb_elem_idx,g); 349 for (int q = 0, idx2 = idx0; q < nr; q++) { 350 for (int d = 0; d < nc; d++, idx2++) { 351 coo_vals[idx2] = row_scale[q]*col_scale[d]*Aij; 352 } 353 } 354 } else { 355 PetscScalar vals[LANDAU_MAX_Q_FACE*LANDAU_MAX_Q_FACE]={0}; 356 PetscInt rows[LANDAU_MAX_Q_FACE],cols[LANDAU_MAX_Q_FACE]; 357 idx = Idxs[f]; 358 if (idx >= 0) { 359 nr = 1; 360 rows[0] = idx; 361 row_scale[0] = 1.; 362 } else { 363 idx = -idx - 1; 364 for (q = 0, nr = 0; q < d_maps->num_face; q++, nr++) { 365 if (d_maps->c_maps[idx][q].gid < 0) break; 366 rows[q] = d_maps->c_maps[idx][q].gid; 367 row_scale[q] = d_maps->c_maps[idx][q].scale; 368 } 369 } 370 idx = Idxs[g]; 371 if (idx >= 0) { 372 nc = 1; 373 cols[0] = idx; 374 col_scale[0] = 1.; 375 } else { 376 idx = -idx - 1; 377 nc = d_maps->num_face; 378 for (q = 0, nc = 0; q < d_maps->num_face; q++, nc++) { 379 if (d_maps->c_maps[idx][q].gid < 0) break; 380 cols[q] = d_maps->c_maps[idx][q].gid; 381 col_scale[q] = d_maps->c_maps[idx][q].scale; 382 } 383 } 384 385 for (q = 0; q < nr; q++) rows[q] = rows[q] + moffset; 386 for (q = 0; q < nc; q++) cols[q] = cols[q] + moffset; 387 for (q = 0; q < nr; q++) { 388 for (d = 0; d < nc; d++) { 389 vals[q*nc + d] = row_scale[q]*col_scale[d]*Aij; 390 } 391 } 392 MatSetValuesDevice(d_mat,nr,rows,nc,cols,vals,ADD_VALUES); 393 } 394 return 0; 395 } 396 397 PetscErrorCode LandauKokkosJacobian(DM plex[], const PetscInt Nq, const PetscInt batch_sz, const PetscInt num_grids, const PetscInt a_numCells[], PetscReal a_Eq_m[], PetscScalar a_elem_closure[], 398 const PetscScalar a_xarray[], const LandauStaticData *SData_d, const PetscReal shift, 399 const PetscLogEvent events[], const PetscInt a_mat_offset[], const PetscInt a_species_offset[], Mat subJ[], Mat JacP) 400 { 401 using scr_mem_t = Kokkos::DefaultExecutionSpace::scratch_memory_space; 402 using real2_scr_t = Kokkos::View<PetscScalar**, Kokkos::LayoutRight, scr_mem_t>; 403 using g2_scr_t = Kokkos::View<PetscReal***, Kokkos::LayoutRight, scr_mem_t>; 404 using g3_scr_t = Kokkos::View<PetscReal****, Kokkos::LayoutRight, scr_mem_t>; 405 PetscErrorCode ierr; 406 PetscInt Nb=Nq,dim,num_cells_max,Nf_max,num_cells_batch; 407 int nfaces=0; 408 LandauCtx *ctx; 409 PetscReal *d_Eq_m=NULL; 410 PetscScalar *d_vertex_f=NULL; 411 P4estVertexMaps *maps[LANDAU_MAX_GRIDS]; // this gets captured 412 PetscSplitCSRDataStructure d_mat; 413 PetscContainer container; 414 const int conc = Kokkos::DefaultExecutionSpace().concurrency(), openmp = !!(conc < 1000), team_size = (openmp==0) ? Nq : 1; 415 const PetscInt coo_sz_batch = SData_d->coo_size/batch_sz; // capture 416 auto d_alpha_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->alpha); //static data 417 const PetscReal *d_alpha = d_alpha_k->data(); 418 const PetscInt Nftot = d_alpha_k->size(); // total number of species 419 auto d_beta_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->beta); 420 const PetscReal *d_beta = d_beta_k->data(); 421 auto d_invMass_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invMass); 422 const PetscReal *d_invMass = d_invMass_k->data(); 423 auto d_B_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->B); 424 const PetscReal *d_BB = d_B_k->data(); 425 auto d_D_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->D); 426 const PetscReal *d_DD = d_D_k->data(); 427 auto d_invJ_k = *static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invJ); // use Kokkos vector in kernels 428 auto d_x_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->x); //static data 429 const PetscReal *d_x = d_x_k->data(); 430 auto d_y_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->y); //static data 431 const PetscReal *d_y = d_y_k->data(); 432 auto d_z_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->z); //static data 433 const PetscReal *d_z = (LANDAU_DIM==3) ? d_z_k->data() : NULL; 434 auto d_w_k = *static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->w); //static data 435 const PetscReal *d_w = d_w_k.data(); 436 // grid offsets - single vertex grid data 437 auto d_numCells_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->NCells); 438 const PetscInt *d_numCells = d_numCells_k->data(); 439 auto d_species_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->species_offset); 440 const PetscInt *d_species_offset = d_species_offset_k->data(); 441 auto d_mat_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->mat_offset); 442 const PetscInt *d_mat_offset = d_mat_offset_k->data(); 443 auto d_ip_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->ip_offset); 444 const PetscInt *d_ip_offset = d_ip_offset_k->data(); 445 auto d_ipf_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->ipf_offset); 446 const PetscInt *d_ipf_offset = d_ipf_offset_k->data(); 447 auto d_elem_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->elem_offset); 448 const PetscInt *d_elem_offset = d_elem_offset_k->data(); 449 #if defined(LANDAU_LAYOUT_LEFT) // preallocate dynamic data, including batched vertices 450 Kokkos::View<PetscReal***, Kokkos::LayoutLeft > d_fdf_k = *static_cast<Kokkos::View<PetscReal***, Kokkos::LayoutLeft >*>(SData_d->ipfdf_data); 451 #else 452 Kokkos::View<PetscReal***, Kokkos::LayoutRight > d_fdf_k = *static_cast<Kokkos::View<PetscReal***, Kokkos::LayoutRight >*>(SData_d->ipfdf_data); 453 #endif 454 auto d_Eq_m_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->Eq_m); // static storage, dynamci data - E(t), copy later, single vertex 455 // COO 456 auto d_coo_elem_offsets_k = static_cast<Kokkos::View<LandauIdx*, Kokkos::LayoutLeft>*>((void*)SData_d->coo_elem_offsets); 457 auto d_coo_elem_fullNb_k = static_cast<Kokkos::View<LandauIdx*, Kokkos::LayoutLeft>*>((void*)SData_d->coo_elem_fullNb); 458 auto d_coo_elem_point_offsets_k = static_cast<Kokkos::View<LandauIdx**, Kokkos::LayoutRight>*>((void*)SData_d->coo_elem_point_offsets); 459 Kokkos::View<PetscScalar*, Kokkos::LayoutRight,Kokkos::DefaultExecutionSpace> d_coo_vals_k("coo_vals", SData_d->coo_size); // device data (default space) 460 PetscScalar* d_coo_vals = (SData_d->coo_size==0) ? NULL : d_coo_vals_k.data(); 461 462 PetscFunctionBegin; 463 ierr = PetscLogEventBegin(events[3],0,0,0,0);CHKERRQ(ierr); 464 ierr = DMGetApplicationContext(plex[0], &ctx);CHKERRQ(ierr); 465 PetscCheckFalse(!ctx,PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 466 ierr = DMGetDimension(plex[0], &dim);CHKERRQ(ierr); 467 PetscCheckFalse(LANDAU_DIM != dim,PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %D != LANDAU_DIM %d",dim,LANDAU_DIM); 468 if (ctx->gpu_assembly) { 469 ierr = PetscObjectQuery((PetscObject) JacP, "assembly_maps", (PetscObject *) &container);CHKERRQ(ierr); 470 if (container) { 471 P4estVertexMaps *h_maps=NULL; 472 ierr = PetscContainerGetPointer(container, (void **) &h_maps);CHKERRQ(ierr); 473 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 474 if (h_maps[grid].d_self) { 475 maps[grid] = h_maps[grid].d_self; 476 nfaces = h_maps[grid].num_face; // nface=0 for CPU assembly 477 } else { 478 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); 479 } 480 } 481 if (!d_coo_vals) { 482 // this does the setup the first time called 483 ierr = MatKokkosGetDeviceMatWrite(JacP,&d_mat);CHKERRQ(ierr); 484 } else { 485 d_mat = NULL; 486 } 487 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); 488 } else { 489 for (PetscInt grid=0 ; grid<num_grids ; grid++) maps[grid] = NULL; 490 nfaces = 0; 491 d_mat = NULL; 492 container = NULL; 493 } 494 num_cells_batch = Nf_max = num_cells_max = 0; 495 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 496 int Nfloc = a_species_offset[grid+1] - a_species_offset[grid]; 497 if (Nfloc > Nf_max) Nf_max = Nfloc; 498 if (a_numCells[grid] > num_cells_max) num_cells_max = a_numCells[grid]; 499 num_cells_batch += a_numCells[grid]; // we don't have a host element offset here (but in ctx) 500 } 501 const PetscInt totDim_max = Nf_max*Nq, elem_mat_size_max = totDim_max*totDim_max; 502 const PetscInt elem_mat_num_cells_max_grid = container ? 0 : num_cells_max; 503 Kokkos::View<PetscScalar****, Kokkos::LayoutRight> d_elem_mats("element matrices", batch_sz, num_grids, elem_mat_num_cells_max_grid, elem_mat_size_max); // first call have large set of global (Jac) element matrices 504 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_Eq_m_k (a_Eq_m, Nftot); 505 if (a_elem_closure || a_xarray) { 506 Kokkos::deep_copy (*d_Eq_m_k, h_Eq_m_k); 507 d_Eq_m = d_Eq_m_k->data(); 508 } else d_Eq_m = NULL; 509 ierr = PetscKokkosInitializeCheck();CHKERRQ(ierr); 510 ierr = PetscLogEventEnd(events[3],0,0,0,0);CHKERRQ(ierr); 511 if (a_elem_closure || a_xarray) { // Jacobian, create f & df 512 Kokkos::View<PetscScalar*, Kokkos::LayoutLeft> *d_vertex_f_k = NULL; 513 ierr = PetscLogEventBegin(events[1],0,0,0,0);CHKERRQ(ierr); 514 if (a_elem_closure) { 515 PetscInt closure_sz = 0; // argh, don't have this on the host!!! 516 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 517 PetscInt nfloc = a_species_offset[grid+1] - a_species_offset[grid]; 518 closure_sz += Nq*nfloc*a_numCells[grid]; 519 } 520 closure_sz *= batch_sz; 521 d_vertex_f_k = new Kokkos::View<PetscScalar*, Kokkos::LayoutLeft> ("closure",closure_sz); 522 const Kokkos::View<PetscScalar*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_closure_k (a_elem_closure, closure_sz); // Vertex data for each element 523 Kokkos::deep_copy (*d_vertex_f_k, h_closure_k); 524 d_vertex_f = d_vertex_f_k->data(); 525 } else { 526 d_vertex_f = (PetscScalar*)a_xarray; 527 } 528 ierr = PetscLogEventEnd(events[1],0,0,0,0);CHKERRQ(ierr); 529 ierr = PetscLogEventBegin(events[8],0,0,0,0);CHKERRQ(ierr); 530 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 531 532 const int scr_bytes_fdf = real2_scr_t::shmem_size(Nf_max,Nb); 533 ierr = PetscInfo(plex[0], "Jacobian shared memory size: %d bytes in level %d num cells total=%D team size=%D #face=%D Nf_max=%D\n",scr_bytes_fdf,KOKKOS_SHARED_LEVEL,num_cells_batch*batch_sz,team_size,nfaces,Nf_max);CHKERRQ(ierr); 534 Kokkos::parallel_for("f, df", Kokkos::TeamPolicy<>(num_cells_batch*batch_sz, team_size, /* Kokkos::AUTO */ 16).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes_fdf)), KOKKOS_LAMBDA (const team_member team) { 535 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]; 536 // find my grid 537 PetscInt grid = 0; 538 while (b_elem_idx >= d_elem_offset[grid+1]) grid++; 539 { 540 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]; 541 const PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,d_mat_offset); 542 { 543 real2_scr_t s_coef_k(team.team_scratch(KOKKOS_SHARED_LEVEL),Nf_max,Nb); 544 PetscScalar *coef; 545 const PetscReal *invJe = &d_invJ_k((d_ip_offset[grid] + loc_elem*Nq)*dim*dim); 546 // un pack IPData 547 if (!maps[grid]) { 548 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 549 } else { 550 coef = s_coef_k.data(); 551 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,(int)loc_Nf), [=] (int f) { 552 //for (int f = 0; f < loc_Nf; ++f) { 553 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nb), [=] (int b) { 554 //for (int b = 0; b < Nb; ++b) { 555 LandauIdx idx = maps[grid]->gIdx[loc_elem][f][b]; 556 if (idx >= 0) { 557 coef[f*Nb+b] = d_vertex_f[idx+moffset]; // xarray data, not IP, need raw array to deal with CPU assemble case (not used) 558 } else { 559 idx = -idx - 1; 560 coef[f*Nb+b] = 0; 561 for (int q = 0; q < maps[grid]->num_face; q++) { 562 PetscInt id = maps[grid]->c_maps[idx][q].gid; 563 PetscScalar scale = maps[grid]->c_maps[idx][q].scale; 564 if (id >= 0) coef[f*Nb+b] += scale*d_vertex_f[id+moffset]; 565 } 566 } 567 }); 568 }); 569 } 570 team.team_barrier(); 571 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 572 const PetscReal *const invJ = &invJe[myQi*dim*dim]; // b_elem_idx: batch element index 573 const PetscReal *Bq = &d_BB[myQi*Nb], *Dq = &d_DD[myQi*Nb*dim]; 574 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)loc_Nf), [=] (int f) { 575 PetscInt b, e, d; 576 PetscReal refSpaceDer[LANDAU_DIM]; 577 const PetscInt idx = d_ipf_offset[grid] + f*loc_nip + loc_elem*Nq + myQi; 578 d_fdf_k(b_id,0,idx) = 0.0; 579 for (d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0; 580 for (b = 0; b < Nb; ++b) { 581 d_fdf_k(b_id,0,idx) += Bq[b]*PetscRealPart(coef[f*Nb+b]); 582 for (d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b*dim+d]*PetscRealPart(coef[f*Nb+b]); 583 } 584 for (d = 0; d < dim; ++d) { 585 for (e = 0, d_fdf_k(b_id,d+1,idx) = 0.0; e < dim; ++e) { 586 d_fdf_k(b_id,d+1,idx) += invJ[e*dim+d]*refSpaceDer[e]; 587 } 588 } 589 }); // f 590 }); // myQi 591 } 592 team.team_barrier(); 593 } // 'grid' scope 594 }); // global elems - fdf 595 Kokkos::fence(); 596 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); // is this a fence? 597 ierr = PetscLogEventEnd(events[8],0,0,0,0);CHKERRQ(ierr); 598 // Jacobian 599 auto jac_lambda = KOKKOS_LAMBDA (const team_member team) { 600 const PetscInt b_Nelem = d_elem_offset[num_grids], b_elem_idx = team.league_rank()%b_Nelem, b_id = team.league_rank()/b_Nelem; 601 // find my grid 602 PetscInt grid = 0; 603 while (b_elem_idx >= d_elem_offset[grid+1]) grid++; 604 { 605 const PetscInt loc_Nf = d_species_offset[grid+1]-d_species_offset[grid], loc_elem = b_elem_idx - d_elem_offset[grid]; 606 const PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,d_mat_offset); 607 const PetscInt f_off = d_species_offset[grid], totDim = loc_Nf*Nq; 608 g2_scr_t g2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,loc_Nf,Nq); 609 g3_scr_t g3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,loc_Nf,Nq); 610 g2_scr_t gg2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,loc_Nf,Nq); 611 g3_scr_t gg3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,loc_Nf,Nq); 612 // get g2[] & g3[] 613 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 614 using Kokkos::parallel_reduce; 615 const PetscInt jpidx_glb = d_ip_offset[grid] + loc_elem * Nq + myQi; 616 const PetscReal *invJ = &d_invJ_k(jpidx_glb*dim*dim); 617 const PetscReal vj[3] = {d_x[jpidx_glb], d_y[jpidx_glb], d_z ? d_z[jpidx_glb] : 0}, wj = d_w[jpidx_glb]; 618 landau_inner_red::TensorValueType gg_temp; // reduce on part of gg2 and g33 for IP jpidx_g 619 Kokkos::parallel_reduce(Kokkos::ThreadVectorRange (team, (int)d_ip_offset[num_grids]), [=] (const int& ipidx, landau_inner_red::TensorValueType & ggg) { 620 const PetscReal wi = d_w[ipidx], x = d_x[ipidx], y = d_y[ipidx]; 621 PetscReal temp1[3] = {0, 0, 0}, temp2 = 0; 622 PetscInt fieldA,d2,d3,f_off_r,grid_r,ipidx_g,nip_loc_r,loc_Nf_r; 623 #if LANDAU_DIM==2 624 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.; 625 LandauTensor2D(vj, x, y, Ud, Uk, mask); 626 #else 627 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.; 628 LandauTensor3D(vj, x, y, z, U, mask); 629 #endif 630 grid_r = 0; 631 while (ipidx >= d_ip_offset[grid_r+1]) grid_r++; // yuck search for grid 632 f_off_r = d_species_offset[grid_r]; 633 ipidx_g = ipidx - d_ip_offset[grid_r]; 634 nip_loc_r = d_numCells[grid_r]*Nq; 635 loc_Nf_r = d_species_offset[grid_r+1] - d_species_offset[grid_r]; 636 for (fieldA = 0; fieldA < loc_Nf_r; ++fieldA) { 637 const PetscInt idx = d_ipf_offset[grid_r] + fieldA*nip_loc_r + ipidx_g; 638 temp1[0] += d_fdf_k(b_id,1,idx)*d_beta[fieldA+f_off_r]*d_invMass[fieldA+f_off_r]; 639 temp1[1] += d_fdf_k(b_id,2,idx)*d_beta[fieldA+f_off_r]*d_invMass[fieldA+f_off_r]; 640 #if LANDAU_DIM==3 641 temp1[2] += d_fdf_k(b_id,3,idx)*d_beta[fieldA+f_off_r]*d_invMass[fieldA+f_off_r]; 642 #endif 643 temp2 += d_fdf_k(b_id,0,idx)*d_beta[fieldA+f_off_r]; 644 } 645 temp1[0] *= wi; 646 temp1[1] *= wi; 647 #if LANDAU_DIM==3 648 temp1[2] *= wi; 649 #endif 650 temp2 *= wi; 651 #if LANDAU_DIM==2 652 for (d2 = 0; d2 < 2; d2++) { 653 for (d3 = 0; d3 < 2; ++d3) { 654 /* K = U * grad(f): g2=e: i,A */ 655 ggg.gg2[d2] += Uk[d2][d3]*temp1[d3]; 656 /* D = -U * (I \kron (fx)): g3=f: i,j,A */ 657 ggg.gg3[d2][d3] += Ud[d2][d3]*temp2; 658 } 659 } 660 #else 661 for (d2 = 0; d2 < 3; ++d2) { 662 for (d3 = 0; d3 < 3; ++d3) { 663 /* K = U * grad(f): g2 = e: i,A */ 664 ggg.gg2[d2] += U[d2][d3]*temp1[d3]; 665 /* D = -U * (I \kron (fx)): g3 = f: i,j,A */ 666 ggg.gg3[d2][d3] += U[d2][d3]*temp2; 667 } 668 } 669 #endif 670 }, Kokkos::Sum<landau_inner_red::TensorValueType>(gg_temp)); 671 // add alpha and put in gg2/3 672 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)loc_Nf), [&] (const int& fieldA) { 673 PetscInt d2,d3; 674 for (d2 = 0; d2 < dim; d2++) { 675 gg2(d2,fieldA,myQi) = gg_temp.gg2[d2]*d_alpha[fieldA+f_off]; 676 for (d3 = 0; d3 < dim; d3++) { 677 gg3(d2,d3,fieldA,myQi) = -gg_temp.gg3[d2][d3]*d_alpha[fieldA+f_off]*d_invMass[fieldA+f_off]; 678 } 679 } 680 }); 681 /* add electric field term once per IP */ 682 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)loc_Nf), [&] (const int& fieldA) { 683 gg2(dim-1,fieldA,myQi) += d_Eq_m[fieldA+f_off]; 684 }); 685 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)loc_Nf), [=] (const int& fieldA) { 686 int d,d2,d3,dp; 687 /* Jacobian transform - g2, g3 - per thread (2D) */ 688 for (d = 0; d < dim; ++d) { 689 g2(d,fieldA,myQi) = 0; 690 for (d2 = 0; d2 < dim; ++d2) { 691 g2(d,fieldA,myQi) += invJ[d*dim+d2]*gg2(d2,fieldA,myQi); 692 g3(d,d2,fieldA,myQi) = 0; 693 for (d3 = 0; d3 < dim; ++d3) { 694 for (dp = 0; dp < dim; ++dp) { 695 g3(d,d2,fieldA,myQi) += invJ[d*dim + d3]*gg3(d3,dp,fieldA,myQi)*invJ[d2*dim + dp]; 696 } 697 } 698 g3(d,d2,fieldA,myQi) *= wj; 699 } 700 g2(d,fieldA,myQi) *= wj; 701 } 702 }); 703 }); // Nq 704 team.team_barrier(); 705 { /* assemble */ 706 for (PetscInt fieldA = 0; fieldA < loc_Nf; fieldA++) { 707 /* assemble */ 708 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nb), [=] (int f) { 709 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,Nb), [=] (int g) { 710 PetscScalar t = 0; 711 for (int qj = 0 ; qj < Nq ; qj++) { // look at others integration points 712 const PetscReal *BJq = &d_BB[qj*Nb], *DIq = &d_DD[qj*Nb*dim]; 713 for (int d = 0; d < dim; ++d) { 714 t += DIq[f*dim+d]*g2(d,fieldA,qj)*BJq[g]; 715 for (int d2 = 0; d2 < dim; ++d2) { 716 t += DIq[f*dim + d]*g3(d,d2,fieldA,qj)*DIq[g*dim + d2]; 717 } 718 } 719 } 720 if (elem_mat_num_cells_max_grid) { // CPU assembly 721 const PetscInt fOff = (fieldA*Nb + f)*totDim + fieldA*Nb + g; 722 d_elem_mats(b_id,grid,loc_elem,fOff) = t; 723 } else { 724 landau_mat_assemble (d_mat, d_coo_vals, t, f, g, Nb, moffset, loc_elem, fieldA, maps[grid], d_coo_elem_offsets_k->data(), d_coo_elem_fullNb_k->data(), *d_coo_elem_point_offsets_k, b_elem_idx, b_id*coo_sz_batch); 725 } 726 }); 727 }); 728 } 729 } 730 } // scope with 'grid' 731 }; 732 ierr = PetscLogEventBegin(events[4],0,0,0,0);CHKERRQ(ierr); 733 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 734 const int scr_bytes = 2*(g2_scr_t::shmem_size(dim,Nf_max,Nq) + g3_scr_t::shmem_size(dim,dim,Nf_max,Nq)); 735 Kokkos::parallel_for("Jacobian", Kokkos::TeamPolicy<Kokkos::LaunchBounds<512,1>>(num_cells_batch*batch_sz, team_size, /* Kokkos::AUTO */ 16).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes)), jac_lambda); 736 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); 737 ierr = PetscLogEventEnd(events[4],0,0,0,0);CHKERRQ(ierr); 738 if (d_vertex_f_k) delete d_vertex_f_k; 739 } else { // mass 740 ierr = PetscLogEventBegin(events[16],0,0,0,0);CHKERRQ(ierr); 741 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 742 ierr = PetscInfo(plex[0], "Mass conc=%" PetscInt_FMT " team size=%" PetscInt_FMT " #face=%" PetscInt_FMT " Nb=%" PetscInt_FMT "\n",conc,team_size,nfaces,Nb);CHKERRQ(ierr); 743 Kokkos::parallel_for("Mass", Kokkos::TeamPolicy<>(num_cells_batch*batch_sz, team_size, /* Kokkos::AUTO */ 16), KOKKOS_LAMBDA (const team_member team) { 744 const PetscInt b_Nelem = d_elem_offset[num_grids], b_elem_idx = team.league_rank()%b_Nelem, b_id = team.league_rank()/b_Nelem; 745 // find my grid 746 PetscInt grid = 0; 747 while (b_elem_idx >= d_elem_offset[grid+1]) grid++; 748 { 749 const PetscInt loc_Nf = d_species_offset[grid+1]-d_species_offset[grid], loc_elem = b_elem_idx - d_elem_offset[grid], totDim = loc_Nf*Nq, jpidx_0 = d_ip_offset[grid] + loc_elem * Nq; 750 const PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,d_mat_offset); 751 for (int fieldA = 0; fieldA < loc_Nf; fieldA++) { 752 /* assemble */ 753 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nb), [=] (int f) { 754 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,Nb), [=] (int g) { 755 PetscScalar t = 0; 756 for (int qj = 0 ; qj < Nq ; qj++) { // look at others integration points 757 const PetscReal *BJq = &d_BB[qj*Nb]; 758 const PetscInt jpidx_glb = jpidx_0 + qj; 759 if (dim==2) { 760 t += BJq[f] * d_w_k(jpidx_glb) * shift * BJq[g] * 2. * PETSC_PI; 761 } else { 762 t += BJq[f] * d_w_k(jpidx_glb) * shift * BJq[g]; 763 } 764 } 765 if (elem_mat_num_cells_max_grid) { 766 const PetscInt fOff = (fieldA*Nb + f)*totDim + fieldA*Nb + g; 767 d_elem_mats(b_id,grid,loc_elem,fOff) = t; 768 } else { 769 landau_mat_assemble (d_mat, d_coo_vals, t, f, g, Nb, moffset, loc_elem, fieldA, maps[grid], d_coo_elem_offsets_k->data(), d_coo_elem_fullNb_k->data(), *d_coo_elem_point_offsets_k, b_elem_idx, b_id*coo_sz_batch); 770 } 771 }); 772 }); 773 } 774 } 775 }); 776 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); 777 ierr = PetscLogEventEnd(events[16],0,0,0,0);CHKERRQ(ierr); 778 } 779 Kokkos::fence(); 780 if (d_coo_vals) { 781 ierr = MatSetValuesCOO(JacP,d_coo_vals,ADD_VALUES);CHKERRQ(ierr); 782 } else if (elem_mat_num_cells_max_grid) { // CPU assembly 783 Kokkos::View<PetscScalar****, Kokkos::LayoutRight>::HostMirror h_elem_mats = Kokkos::create_mirror_view(d_elem_mats); 784 Kokkos::deep_copy (h_elem_mats, d_elem_mats); 785 PetscCheck(!container,PETSC_COMM_SELF, PETSC_ERR_PLIB, "?????"); 786 for (PetscInt b_id = 0 ; b_id < batch_sz ; b_id++) { // OpenMP (once) 787 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 788 PetscSection section, globalSection; 789 PetscInt cStart,cEnd; 790 Mat B = subJ[ LAND_PACK_IDX(b_id,grid) ]; 791 PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,a_mat_offset), nloc, nzl, colbuf[1024], row; 792 const PetscInt *cols; 793 const PetscScalar *vals; 794 ierr = PetscLogEventBegin(events[5],0,0,0,0);CHKERRQ(ierr); 795 ierr = DMPlexGetHeightStratum(plex[grid],0,&cStart,&cEnd);CHKERRQ(ierr); 796 ierr = DMGetLocalSection(plex[grid], §ion);CHKERRQ(ierr); 797 ierr = DMGetGlobalSection(plex[grid], &globalSection);CHKERRQ(ierr); 798 ierr = PetscLogEventEnd(events[5],0,0,0,0);CHKERRQ(ierr); 799 ierr = PetscLogEventBegin(events[6],0,0,0,0);CHKERRQ(ierr); 800 for (PetscInt ej = cStart ; ej < cEnd; ++ej) { 801 const PetscScalar *elMat = &h_elem_mats(b_id,grid,ej-cStart,0); 802 ierr = DMPlexMatSetClosure(plex[grid], section, globalSection, B, ej, elMat, ADD_VALUES);CHKERRQ(ierr); 803 if (grid==0 && ej==-1) { 804 const PetscInt loc_Nf = a_species_offset[grid+1]-a_species_offset[grid], totDim = loc_Nf*Nq; 805 int d,f; 806 PetscPrintf(PETSC_COMM_SELF,"Kokkos Element matrix %d/%d\n",1,(int)a_numCells[grid]); 807 for (d = 0; d < totDim; ++d){ 808 for (f = 0; f < totDim; ++f) PetscPrintf(PETSC_COMM_SELF," %12.5e", PetscRealPart(elMat[d*totDim + f])); 809 PetscPrintf(PETSC_COMM_SELF,"\n"); 810 } 811 exit(14); 812 } 813 } 814 // move nest matrix to global JacP 815 ierr = MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 816 ierr = MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 817 ierr = MatGetSize(B, &nloc, NULL);CHKERRQ(ierr); 818 for (int i=0 ; i<nloc ; i++) { 819 ierr = MatGetRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 820 PetscCheckFalse(nzl>1024,PetscObjectComm((PetscObject) B), PETSC_ERR_PLIB, "Row too big: %D",nzl); 821 for (int j=0; j<nzl; j++) colbuf[j] = cols[j] + moffset; 822 row = i + moffset; 823 ierr = MatSetValues(JacP,1,&row,nzl,colbuf,vals,ADD_VALUES);CHKERRQ(ierr); 824 ierr = MatRestoreRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 825 } 826 ierr = MatDestroy(&subJ[ LAND_PACK_IDX(b_id,grid) ]);CHKERRQ(ierr); 827 ierr = PetscLogEventEnd(events[6],0,0,0,0);CHKERRQ(ierr); 828 } // grids 829 } 830 } 831 PetscFunctionReturn(0); 832 } 833 } // extern "C" 834