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 ((LandauIdx*)SData_d->coo_elem_offsets, SData_d->coo_n_cellsTot+1); 233 auto coo_elem_offsets = new Kokkos::View<LandauIdx*, Kokkos::LayoutLeft> ("coo_elem_offsets",SData_d->coo_n_cellsTot+1); 234 const Kokkos::View<LandauIdx*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_coo_elem_fullNb ((LandauIdx*)SData_d->coo_elem_fullNb, SData_d->coo_n_cellsTot); 235 auto coo_elem_fullNb = new Kokkos::View<LandauIdx*, Kokkos::LayoutLeft> ("coo_elem_offsets",SData_d->coo_n_cellsTot); 236 const Kokkos::View<LandauIdx*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_coo_elem_point_offsets ((LandauIdx*)SData_d->coo_elem_point_offsets, SData_d->coo_n_cellsTot*(LANDAU_MAX_NQ+1)); 237 auto coo_elem_point_offsets = new Kokkos::View<LandauIdx*, Kokkos::LayoutLeft> ("coo_elem_point_offsets",SData_d->coo_n_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 = static_cast<void*>(coo_elem_offsets); 245 SData_d->coo_elem_fullNb = static_cast<void*>(coo_elem_fullNb); 246 SData_d->coo_elem_point_offsets = static_cast<void*>(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::LayoutLeft>*>((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[], const LandauIdx (*coo_elem_point_offsets)[LANDAU_MAX_NQ+1], 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[glb_elem_idx][f] + nr * coo_elem_point_offsets[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 for (q = 0, nc = 0; q < d_maps->num_face; q++, nc++) { 378 if (d_maps->c_maps[idx][q].gid < 0) break; 379 cols[q] = d_maps->c_maps[idx][q].gid; 380 col_scale[q] = d_maps->c_maps[idx][q].scale; 381 } 382 } 383 384 for (q = 0; q < nr; q++) rows[q] = rows[q] + moffset; 385 for (q = 0; q < nc; q++) cols[q] = cols[q] + moffset; 386 for (q = 0; q < nr; q++) { 387 for (d = 0; d < nc; d++) { 388 vals[q*nc + d] = row_scale[q]*col_scale[d]*Aij; 389 } 390 } 391 MatSetValuesDevice(d_mat,nr,rows,nc,cols,vals,ADD_VALUES); 392 } 393 return 0; 394 } 395 396 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[], 397 const PetscScalar a_xarray[], const LandauStaticData *SData_d, const PetscReal shift, 398 const PetscLogEvent events[], const PetscInt a_mat_offset[], const PetscInt a_species_offset[], Mat subJ[], Mat JacP) 399 { 400 using scr_mem_t = Kokkos::DefaultExecutionSpace::scratch_memory_space; 401 using real2_scr_t = Kokkos::View<PetscScalar**, Kokkos::LayoutRight, scr_mem_t>; 402 using g2_scr_t = Kokkos::View<PetscReal***, Kokkos::LayoutRight, scr_mem_t>; 403 using g3_scr_t = Kokkos::View<PetscReal****, Kokkos::LayoutRight, scr_mem_t>; 404 PetscErrorCode ierr; 405 PetscInt Nb=Nq,dim,num_cells_max,Nf_max,num_cells_batch; 406 int nfaces=0; 407 LandauCtx *ctx; 408 PetscReal *d_Eq_m=NULL; 409 PetscScalar *d_vertex_f=NULL; 410 P4estVertexMaps *maps[LANDAU_MAX_GRIDS]; // this gets captured 411 PetscSplitCSRDataStructure d_mat; 412 PetscContainer container; 413 const int conc = Kokkos::DefaultExecutionSpace().concurrency(), openmp = !!(conc < 1000), team_size = (openmp==0) ? Nq : 1; 414 const PetscInt coo_sz_batch = SData_d->coo_size/batch_sz; // capture 415 auto d_alpha_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->alpha); //static data 416 const PetscReal *d_alpha = d_alpha_k->data(); 417 const PetscInt Nftot = d_alpha_k->size(); // total number of species 418 auto d_beta_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->beta); 419 const PetscReal *d_beta = d_beta_k->data(); 420 auto d_invMass_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invMass); 421 const PetscReal *d_invMass = d_invMass_k->data(); 422 auto d_B_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->B); 423 const PetscReal *d_BB = d_B_k->data(); 424 auto d_D_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->D); 425 const PetscReal *d_DD = d_D_k->data(); 426 auto d_invJ_k = *static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invJ); // use Kokkos vector in kernels 427 auto d_x_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->x); //static data 428 const PetscReal *d_x = d_x_k->data(); 429 auto d_y_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->y); //static data 430 const PetscReal *d_y = d_y_k->data(); 431 auto d_z_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->z); //static data 432 const PetscReal *d_z = (LANDAU_DIM==3) ? d_z_k->data() : NULL; 433 auto d_w_k = *static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->w); //static data 434 const PetscReal *d_w = d_w_k.data(); 435 // grid offsets - single vertex grid data 436 auto d_numCells_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->NCells); 437 const PetscInt *d_numCells = d_numCells_k->data(); 438 auto d_species_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->species_offset); 439 const PetscInt *d_species_offset = d_species_offset_k->data(); 440 auto d_mat_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->mat_offset); 441 const PetscInt *d_mat_offset = d_mat_offset_k->data(); 442 auto d_ip_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->ip_offset); 443 const PetscInt *d_ip_offset = d_ip_offset_k->data(); 444 auto d_ipf_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->ipf_offset); 445 const PetscInt *d_ipf_offset = d_ipf_offset_k->data(); 446 auto d_elem_offset_k = static_cast<Kokkos::View<PetscInt*, Kokkos::LayoutLeft>*>(SData_d->elem_offset); 447 const PetscInt *d_elem_offset = d_elem_offset_k->data(); 448 #if defined(LANDAU_LAYOUT_LEFT) // preallocate dynamic data, including batched vertices 449 Kokkos::View<PetscReal***, Kokkos::LayoutLeft > d_fdf_k = *static_cast<Kokkos::View<PetscReal***, Kokkos::LayoutLeft >*>(SData_d->ipfdf_data); 450 #else 451 Kokkos::View<PetscReal***, Kokkos::LayoutRight > d_fdf_k = *static_cast<Kokkos::View<PetscReal***, Kokkos::LayoutRight >*>(SData_d->ipfdf_data); 452 #endif 453 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 454 // COO 455 auto d_coo_elem_offsets_k = static_cast<Kokkos::View<LandauIdx*, Kokkos::LayoutLeft>*>(SData_d->coo_elem_offsets); 456 LandauIdx *d_coo_elem_offsets = (SData_d->coo_size==0) ? NULL : d_coo_elem_offsets_k->data(); 457 auto d_coo_elem_fullNb_k = static_cast<Kokkos::View<LandauIdx*, Kokkos::LayoutLeft>*>(SData_d->coo_elem_fullNb); 458 LandauIdx *d_coo_elem_fullNb = (SData_d->coo_size==0) ? NULL : d_coo_elem_fullNb_k->data(); 459 auto d_coo_elem_point_offsets_k = static_cast<Kokkos::View<LandauIdx*, Kokkos::LayoutLeft>*>(SData_d->coo_elem_point_offsets); 460 LandauIdx (*d_coo_elem_point_offsets)[LANDAU_MAX_NQ+1] = (SData_d->coo_size==0) ? NULL : (LandauIdx (*)[LANDAU_MAX_NQ+1])d_coo_elem_point_offsets_k->data(); 461 462 Kokkos::View<PetscScalar*, Kokkos::LayoutRight,Kokkos::DefaultExecutionSpace> d_coo_vals_k("coo_vals", SData_d->coo_size); // device data (default space) 463 PetscScalar* d_coo_vals = (SData_d->coo_size==0) ? NULL : d_coo_vals_k.data(); 464 465 PetscFunctionBegin; 466 ierr = PetscLogEventBegin(events[3],0,0,0,0);CHKERRQ(ierr); 467 ierr = DMGetApplicationContext(plex[0], &ctx);CHKERRQ(ierr); 468 PetscCheckFalse(!ctx,PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 469 ierr = DMGetDimension(plex[0], &dim);CHKERRQ(ierr); 470 PetscCheckFalse(LANDAU_DIM != dim,PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %D != LANDAU_DIM %d",dim,LANDAU_DIM); 471 if (ctx->gpu_assembly) { 472 ierr = PetscObjectQuery((PetscObject) JacP, "assembly_maps", (PetscObject *) &container);CHKERRQ(ierr); 473 if (container) { 474 P4estVertexMaps *h_maps=NULL; 475 ierr = PetscContainerGetPointer(container, (void **) &h_maps);CHKERRQ(ierr); 476 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 477 if (h_maps[grid].d_self) { 478 maps[grid] = h_maps[grid].d_self; 479 nfaces = h_maps[grid].num_face; // nface=0 for CPU assembly 480 } else { 481 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); 482 } 483 } 484 if (!d_coo_vals) { 485 // this does the setup the first time called 486 ierr = MatKokkosGetDeviceMatWrite(JacP,&d_mat);CHKERRQ(ierr); 487 } else { 488 d_mat = NULL; 489 } 490 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); 491 } else { 492 for (PetscInt grid=0 ; grid<num_grids ; grid++) maps[grid] = NULL; 493 nfaces = 0; 494 d_mat = NULL; 495 container = NULL; 496 } 497 num_cells_batch = Nf_max = num_cells_max = 0; 498 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 499 int Nfloc = a_species_offset[grid+1] - a_species_offset[grid]; 500 if (Nfloc > Nf_max) Nf_max = Nfloc; 501 if (a_numCells[grid] > num_cells_max) num_cells_max = a_numCells[grid]; 502 num_cells_batch += a_numCells[grid]; // we don't have a host element offset here (but in ctx) 503 } 504 const PetscInt totDim_max = Nf_max*Nq, elem_mat_size_max = totDim_max*totDim_max; 505 const PetscInt elem_mat_num_cells_max_grid = container ? 0 : num_cells_max; 506 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 507 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_Eq_m_k (a_Eq_m, Nftot); 508 if (a_elem_closure || a_xarray) { 509 Kokkos::deep_copy (*d_Eq_m_k, h_Eq_m_k); 510 d_Eq_m = d_Eq_m_k->data(); 511 } else d_Eq_m = NULL; 512 ierr = PetscKokkosInitializeCheck();CHKERRQ(ierr); 513 ierr = PetscLogEventEnd(events[3],0,0,0,0);CHKERRQ(ierr); 514 if (a_elem_closure || a_xarray) { // Jacobian, create f & df 515 Kokkos::View<PetscScalar*, Kokkos::LayoutLeft> *d_vertex_f_k = NULL; 516 ierr = PetscLogEventBegin(events[1],0,0,0,0);CHKERRQ(ierr); 517 if (a_elem_closure) { 518 PetscInt closure_sz = 0; // argh, don't have this on the host!!! 519 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 520 PetscInt nfloc = a_species_offset[grid+1] - a_species_offset[grid]; 521 closure_sz += Nq*nfloc*a_numCells[grid]; 522 } 523 closure_sz *= batch_sz; 524 d_vertex_f_k = new Kokkos::View<PetscScalar*, Kokkos::LayoutLeft> ("closure",closure_sz); 525 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 526 Kokkos::deep_copy (*d_vertex_f_k, h_closure_k); 527 d_vertex_f = d_vertex_f_k->data(); 528 } else { 529 d_vertex_f = (PetscScalar*)a_xarray; 530 } 531 ierr = PetscLogEventEnd(events[1],0,0,0,0);CHKERRQ(ierr); 532 ierr = PetscLogEventBegin(events[8],0,0,0,0);CHKERRQ(ierr); 533 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 534 535 const int scr_bytes_fdf = real2_scr_t::shmem_size(Nf_max,Nb); 536 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); 537 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) { 538 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]; 539 // find my grid 540 PetscInt grid = 0; 541 while (b_elem_idx >= d_elem_offset[grid+1]) grid++; 542 { 543 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]; 544 const PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,d_mat_offset); 545 { 546 real2_scr_t s_coef_k(team.team_scratch(KOKKOS_SHARED_LEVEL),Nf_max,Nb); 547 PetscScalar *coef; 548 const PetscReal *invJe = &d_invJ_k((d_ip_offset[grid] + loc_elem*Nq)*dim*dim); 549 // un pack IPData 550 if (!maps[grid]) { 551 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 552 } else { 553 coef = s_coef_k.data(); 554 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,(int)loc_Nf), [=] (int f) { 555 //for (int f = 0; f < loc_Nf; ++f) { 556 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nb), [=] (int b) { 557 //for (int b = 0; b < Nb; ++b) { 558 LandauIdx idx = maps[grid]->gIdx[loc_elem][f][b]; 559 if (idx >= 0) { 560 coef[f*Nb+b] = d_vertex_f[idx+moffset]; // xarray data, not IP, need raw array to deal with CPU assemble case (not used) 561 } else { 562 idx = -idx - 1; 563 coef[f*Nb+b] = 0; 564 for (int q = 0; q < maps[grid]->num_face; q++) { 565 PetscInt id = maps[grid]->c_maps[idx][q].gid; 566 PetscScalar scale = maps[grid]->c_maps[idx][q].scale; 567 if (id >= 0) coef[f*Nb+b] += scale*d_vertex_f[id+moffset]; 568 } 569 } 570 }); 571 }); 572 } 573 team.team_barrier(); 574 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 575 const PetscReal *const invJ = &invJe[myQi*dim*dim]; // b_elem_idx: batch element index 576 const PetscReal *Bq = &d_BB[myQi*Nb], *Dq = &d_DD[myQi*Nb*dim]; 577 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)loc_Nf), [=] (int f) { 578 PetscInt b, e, d; 579 PetscReal refSpaceDer[LANDAU_DIM]; 580 const PetscInt idx = d_ipf_offset[grid] + f*loc_nip + loc_elem*Nq + myQi; 581 d_fdf_k(b_id,0,idx) = 0.0; 582 for (d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0; 583 for (b = 0; b < Nb; ++b) { 584 d_fdf_k(b_id,0,idx) += Bq[b]*PetscRealPart(coef[f*Nb+b]); 585 for (d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b*dim+d]*PetscRealPart(coef[f*Nb+b]); 586 } 587 for (d = 0; d < dim; ++d) { 588 for (e = 0, d_fdf_k(b_id,d+1,idx) = 0.0; e < dim; ++e) { 589 d_fdf_k(b_id,d+1,idx) += invJ[e*dim+d]*refSpaceDer[e]; 590 } 591 } 592 }); // f 593 }); // myQi 594 } 595 team.team_barrier(); 596 } // 'grid' scope 597 }); // global elems - fdf 598 Kokkos::fence(); 599 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); // is this a fence? 600 ierr = PetscLogEventEnd(events[8],0,0,0,0);CHKERRQ(ierr); 601 // Jacobian 602 auto jac_lambda = KOKKOS_LAMBDA (const team_member team) { 603 const PetscInt b_Nelem = d_elem_offset[num_grids], b_elem_idx = team.league_rank()%b_Nelem, b_id = team.league_rank()/b_Nelem; 604 // find my grid 605 PetscInt grid = 0; 606 while (b_elem_idx >= d_elem_offset[grid+1]) grid++; 607 { 608 const PetscInt loc_Nf = d_species_offset[grid+1]-d_species_offset[grid], loc_elem = b_elem_idx - d_elem_offset[grid]; 609 const PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,d_mat_offset); 610 const PetscInt f_off = d_species_offset[grid], totDim = loc_Nf*Nq; 611 g2_scr_t g2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,loc_Nf,Nq); 612 g3_scr_t g3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,loc_Nf,Nq); 613 g2_scr_t gg2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,loc_Nf,Nq); 614 g3_scr_t gg3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,loc_Nf,Nq); 615 // get g2[] & g3[] 616 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 617 using Kokkos::parallel_reduce; 618 const PetscInt jpidx_glb = d_ip_offset[grid] + loc_elem * Nq + myQi; 619 const PetscReal *invJ = &d_invJ_k(jpidx_glb*dim*dim); 620 const PetscReal vj[3] = {d_x[jpidx_glb], d_y[jpidx_glb], d_z ? d_z[jpidx_glb] : 0}, wj = d_w[jpidx_glb]; 621 landau_inner_red::TensorValueType gg_temp; // reduce on part of gg2 and g33 for IP jpidx_g 622 Kokkos::parallel_reduce(Kokkos::ThreadVectorRange (team, (int)d_ip_offset[num_grids]), [=] (const int& ipidx, landau_inner_red::TensorValueType & ggg) { 623 const PetscReal wi = d_w[ipidx], x = d_x[ipidx], y = d_y[ipidx]; 624 PetscReal temp1[3] = {0, 0, 0}, temp2 = 0; 625 PetscInt fieldA,d2,d3,f_off_r,grid_r,ipidx_g,nip_loc_r,loc_Nf_r; 626 #if LANDAU_DIM==2 627 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.; 628 LandauTensor2D(vj, x, y, Ud, Uk, mask); 629 #else 630 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.; 631 LandauTensor3D(vj, x, y, z, U, mask); 632 #endif 633 grid_r = 0; 634 while (ipidx >= d_ip_offset[grid_r+1]) grid_r++; // yuck search for grid 635 f_off_r = d_species_offset[grid_r]; 636 ipidx_g = ipidx - d_ip_offset[grid_r]; 637 nip_loc_r = d_numCells[grid_r]*Nq; 638 loc_Nf_r = d_species_offset[grid_r+1] - d_species_offset[grid_r]; 639 for (fieldA = 0; fieldA < loc_Nf_r; ++fieldA) { 640 const PetscInt idx = d_ipf_offset[grid_r] + fieldA*nip_loc_r + ipidx_g; 641 temp1[0] += d_fdf_k(b_id,1,idx)*d_beta[fieldA+f_off_r]*d_invMass[fieldA+f_off_r]; 642 temp1[1] += d_fdf_k(b_id,2,idx)*d_beta[fieldA+f_off_r]*d_invMass[fieldA+f_off_r]; 643 #if LANDAU_DIM==3 644 temp1[2] += d_fdf_k(b_id,3,idx)*d_beta[fieldA+f_off_r]*d_invMass[fieldA+f_off_r]; 645 #endif 646 temp2 += d_fdf_k(b_id,0,idx)*d_beta[fieldA+f_off_r]; 647 } 648 temp1[0] *= wi; 649 temp1[1] *= wi; 650 #if LANDAU_DIM==3 651 temp1[2] *= wi; 652 #endif 653 temp2 *= wi; 654 #if LANDAU_DIM==2 655 for (d2 = 0; d2 < 2; d2++) { 656 for (d3 = 0; d3 < 2; ++d3) { 657 /* K = U * grad(f): g2=e: i,A */ 658 ggg.gg2[d2] += Uk[d2][d3]*temp1[d3]; 659 /* D = -U * (I \kron (fx)): g3=f: i,j,A */ 660 ggg.gg3[d2][d3] += Ud[d2][d3]*temp2; 661 } 662 } 663 #else 664 for (d2 = 0; d2 < 3; ++d2) { 665 for (d3 = 0; d3 < 3; ++d3) { 666 /* K = U * grad(f): g2 = e: i,A */ 667 ggg.gg2[d2] += U[d2][d3]*temp1[d3]; 668 /* D = -U * (I \kron (fx)): g3 = f: i,j,A */ 669 ggg.gg3[d2][d3] += U[d2][d3]*temp2; 670 } 671 } 672 #endif 673 }, Kokkos::Sum<landau_inner_red::TensorValueType>(gg_temp)); 674 // add alpha and put in gg2/3 675 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)loc_Nf), [&] (const int& fieldA) { 676 PetscInt d2,d3; 677 for (d2 = 0; d2 < dim; d2++) { 678 gg2(d2,fieldA,myQi) = gg_temp.gg2[d2]*d_alpha[fieldA+f_off]; 679 for (d3 = 0; d3 < dim; d3++) { 680 gg3(d2,d3,fieldA,myQi) = -gg_temp.gg3[d2][d3]*d_alpha[fieldA+f_off]*d_invMass[fieldA+f_off]; 681 } 682 } 683 }); 684 /* add electric field term once per IP */ 685 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)loc_Nf), [&] (const int& fieldA) { 686 gg2(dim-1,fieldA,myQi) += d_Eq_m[fieldA+f_off]; 687 }); 688 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)loc_Nf), [=] (const int& fieldA) { 689 int d,d2,d3,dp; 690 /* Jacobian transform - g2, g3 - per thread (2D) */ 691 for (d = 0; d < dim; ++d) { 692 g2(d,fieldA,myQi) = 0; 693 for (d2 = 0; d2 < dim; ++d2) { 694 g2(d,fieldA,myQi) += invJ[d*dim+d2]*gg2(d2,fieldA,myQi); 695 g3(d,d2,fieldA,myQi) = 0; 696 for (d3 = 0; d3 < dim; ++d3) { 697 for (dp = 0; dp < dim; ++dp) { 698 g3(d,d2,fieldA,myQi) += invJ[d*dim + d3]*gg3(d3,dp,fieldA,myQi)*invJ[d2*dim + dp]; 699 } 700 } 701 g3(d,d2,fieldA,myQi) *= wj; 702 } 703 g2(d,fieldA,myQi) *= wj; 704 } 705 }); 706 }); // Nq 707 team.team_barrier(); 708 { /* assemble */ 709 for (PetscInt fieldA = 0; fieldA < loc_Nf; fieldA++) { 710 /* assemble */ 711 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nb), [=] (int f) { 712 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,Nb), [=] (int g) { 713 PetscScalar t = 0; 714 for (int qj = 0 ; qj < Nq ; qj++) { // look at others integration points 715 const PetscReal *BJq = &d_BB[qj*Nb], *DIq = &d_DD[qj*Nb*dim]; 716 for (int d = 0; d < dim; ++d) { 717 t += DIq[f*dim+d]*g2(d,fieldA,qj)*BJq[g]; 718 for (int d2 = 0; d2 < dim; ++d2) { 719 t += DIq[f*dim + d]*g3(d,d2,fieldA,qj)*DIq[g*dim + d2]; 720 } 721 } 722 } 723 if (elem_mat_num_cells_max_grid) { // CPU assembly 724 const PetscInt fOff = (fieldA*Nb + f)*totDim + fieldA*Nb + g; 725 d_elem_mats(b_id,grid,loc_elem,fOff) = t; 726 } else { 727 landau_mat_assemble (d_mat, 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); 728 } 729 }); 730 }); 731 } 732 } 733 } // scope with 'grid' 734 }; 735 ierr = PetscLogEventBegin(events[4],0,0,0,0);CHKERRQ(ierr); 736 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 737 const int scr_bytes = 2*(g2_scr_t::shmem_size(dim,Nf_max,Nq) + g3_scr_t::shmem_size(dim,dim,Nf_max,Nq)); 738 Kokkos::parallel_for("Jacobian", Kokkos::TeamPolicy<>(num_cells_batch*batch_sz, /*team_size*/1, 1).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes)), jac_lambda); 739 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); 740 ierr = PetscLogEventEnd(events[4],0,0,0,0);CHKERRQ(ierr); 741 if (d_vertex_f_k) delete d_vertex_f_k; 742 } else { // mass - 2*Nb is guess at max size (2D Q3 is 31 =< 2*Nb = 32) 743 using fieldMats_scr_t = Kokkos::View<PetscScalar*, Kokkos::LayoutRight, scr_mem_t>; 744 PetscInt loc_ass_sz = 1; // captured 745 ierr = PetscLogEventBegin(events[16],0,0,0,0);CHKERRQ(ierr); 746 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 747 if (loc_ass_sz) loc_ass_sz = ctx->SData_d.coo_max_fullnb*ctx->SData_d.coo_max_fullnb; 748 int scr_bytes = fieldMats_scr_t::shmem_size(loc_ass_sz); // + idx_scr_t::shmem_size(Nb,nfaces) + scale_scr_t::shmem_size(Nb,nfaces); 749 ierr = PetscInfo(plex[0], "Mass shared memory size: %d bytes in level %d conc=%D team size=%D #face=%D Nb=%D, %s assembly\n",scr_bytes,KOKKOS_SHARED_LEVEL,conc,team_size,nfaces,Nb, d_coo_vals ? (loc_ass_sz==0 ? "COO" : "optimized COO") : "CSR");CHKERRQ(ierr); 750 Kokkos::parallel_for("Mass", Kokkos::TeamPolicy<>(num_cells_batch*batch_sz, team_size, /* Kokkos::AUTO */ 16).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes)), KOKKOS_LAMBDA (const team_member team) { 751 fieldMats_scr_t s_fieldMats(team.team_scratch(KOKKOS_SHARED_LEVEL),loc_ass_sz); // Only used for GPU assembly (ie, not first pass) 752 const PetscInt b_Nelem = d_elem_offset[num_grids], b_elem_idx = team.league_rank()%b_Nelem, b_id = team.league_rank()/b_Nelem; 753 // find my grid 754 PetscInt grid = 0; 755 while (b_elem_idx >= d_elem_offset[grid+1]) grid++; 756 { 757 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; 758 const PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,d_mat_offset); 759 for (int fieldA = 0; fieldA < loc_Nf; fieldA++) { 760 /* assemble */ 761 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nb), [=] (int f) { 762 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,Nb), [=] (int g) { 763 PetscScalar t = 0; 764 for (int qj = 0 ; qj < Nq ; qj++) { // look at others integration points 765 const PetscReal *BJq = &d_BB[qj*Nb]; 766 const PetscInt jpidx_glb = jpidx_0 + qj; 767 if (dim==2) { 768 t += BJq[f] * d_w_k(jpidx_glb) * shift * BJq[g] * 2. * PETSC_PI; 769 } else { 770 t += BJq[f] * d_w_k(jpidx_glb) * shift * BJq[g]; 771 } 772 } 773 if (elem_mat_num_cells_max_grid) { 774 const PetscInt fOff = (fieldA*Nb + f)*totDim + fieldA*Nb + g; 775 d_elem_mats(b_id,grid,loc_elem,fOff) = t; 776 } else { 777 if (d_coo_vals && loc_ass_sz) { 778 PetscInt idx,q,nr,nc; 779 const LandauIdx *const Idxs = &maps[grid]->gIdx[loc_elem][fieldA][0]; 780 PetscScalar row_scale[LANDAU_MAX_Q_FACE]={0},col_scale[LANDAU_MAX_Q_FACE]={0}; 781 const int fullNb = d_coo_elem_fullNb[b_elem_idx], num_face = maps[grid]->num_face; 782 const PetscScalar Aij = t; 783 idx = Idxs[f]; 784 if (idx >= 0) { 785 nr = 1; 786 row_scale[0] = 1.; 787 } else { 788 idx = -idx - 1; 789 for (q = 0, nr = 0; q < num_face; q++, nr++) { 790 if (maps[grid]->c_maps[idx][q].gid < 0) break; 791 row_scale[q] = maps[grid]->c_maps[idx][q].scale; 792 } 793 } 794 idx = Idxs[g]; 795 if (idx >= 0) { 796 nc = 1; 797 col_scale[0] = 1.; 798 } else { 799 idx = -idx - 1; 800 for (q = 0, nc = 0; q < num_face; q++, nc++) { 801 if (maps[grid]->c_maps[idx][q].gid < 0) break; 802 col_scale[q] = maps[grid]->c_maps[idx][q].scale; 803 } 804 } 805 const int idx0 = fullNb * d_coo_elem_point_offsets[b_elem_idx][f] + nr * d_coo_elem_point_offsets[b_elem_idx][g]; 806 for (int q = 0, idx2 = idx0; q < nr; q++) { 807 for (int d = 0; d < nc; d++, idx2++) { 808 s_fieldMats(idx2) = row_scale[q]*col_scale[d]*Aij; 809 } 810 } 811 } else { 812 landau_mat_assemble (d_mat, 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); 813 } 814 } 815 }); 816 }); 817 if (!elem_mat_num_cells_max_grid) { 818 if (d_coo_vals && loc_ass_sz) { 819 const int fullNb = d_coo_elem_fullNb[b_elem_idx], fullNb2=fullNb*fullNb; 820 const int idx0 = b_id*coo_sz_batch + d_coo_elem_offsets[b_elem_idx] + fieldA*fullNb2; 821 team.team_barrier(); 822 #if 0 823 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,fullNb), [=] (int gg) { 824 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,fullNb), [=] (int ff) { 825 const int idx = fullNb * ff + gg; 826 d_coo_vals[idx0 + idx] = s_fieldMats(idx); 827 }); 828 }); 829 #else 830 Kokkos::parallel_for(Kokkos::TeamVectorRange(team,0,fullNb2), [=] (int idx) { d_coo_vals[idx0 + idx] = s_fieldMats(idx); }); 831 #endif 832 } 833 } 834 } // field 835 } // grid 836 }); 837 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); 838 ierr = PetscLogEventEnd(events[16],0,0,0,0);CHKERRQ(ierr); 839 } 840 Kokkos::fence(); 841 if (d_coo_vals) { 842 ierr = MatSetValuesCOO(JacP,d_coo_vals,ADD_VALUES);CHKERRQ(ierr); 843 } else if (elem_mat_num_cells_max_grid) { // CPU assembly 844 Kokkos::View<PetscScalar****, Kokkos::LayoutRight>::HostMirror h_elem_mats = Kokkos::create_mirror_view(d_elem_mats); 845 Kokkos::deep_copy (h_elem_mats, d_elem_mats); 846 PetscCheck(!container,PETSC_COMM_SELF, PETSC_ERR_PLIB, "?????"); 847 for (PetscInt b_id = 0 ; b_id < batch_sz ; b_id++) { // OpenMP (once) 848 for (PetscInt grid=0 ; grid<num_grids ; grid++) { 849 PetscSection section, globalSection; 850 PetscInt cStart,cEnd; 851 Mat B = subJ[ LAND_PACK_IDX(b_id,grid) ]; 852 PetscInt moffset = LAND_MOFFSET(b_id,grid,batch_sz,num_grids,a_mat_offset), nloc, nzl, colbuf[1024], row; 853 const PetscInt *cols; 854 const PetscScalar *vals; 855 ierr = PetscLogEventBegin(events[5],0,0,0,0);CHKERRQ(ierr); 856 ierr = DMPlexGetHeightStratum(plex[grid],0,&cStart,&cEnd);CHKERRQ(ierr); 857 ierr = DMGetLocalSection(plex[grid], §ion);CHKERRQ(ierr); 858 ierr = DMGetGlobalSection(plex[grid], &globalSection);CHKERRQ(ierr); 859 ierr = PetscLogEventEnd(events[5],0,0,0,0);CHKERRQ(ierr); 860 ierr = PetscLogEventBegin(events[6],0,0,0,0);CHKERRQ(ierr); 861 for (PetscInt ej = cStart ; ej < cEnd; ++ej) { 862 const PetscScalar *elMat = &h_elem_mats(b_id,grid,ej-cStart,0); 863 ierr = DMPlexMatSetClosure(plex[grid], section, globalSection, B, ej, elMat, ADD_VALUES);CHKERRQ(ierr); 864 if (grid==0 && ej==-1) { 865 const PetscInt loc_Nf = a_species_offset[grid+1]-a_species_offset[grid], totDim = loc_Nf*Nq; 866 int d,f; 867 PetscPrintf(PETSC_COMM_SELF,"Kokkos Element matrix %d/%d\n",1,(int)a_numCells[grid]); 868 for (d = 0; d < totDim; ++d){ 869 for (f = 0; f < totDim; ++f) PetscPrintf(PETSC_COMM_SELF," %12.5e", PetscRealPart(elMat[d*totDim + f])); 870 PetscPrintf(PETSC_COMM_SELF,"\n"); 871 } 872 exit(14); 873 } 874 } 875 // move nest matrix to global JacP 876 ierr = MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 877 ierr = MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 878 ierr = MatGetSize(B, &nloc, NULL);CHKERRQ(ierr); 879 for (int i=0 ; i<nloc ; i++) { 880 ierr = MatGetRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 881 PetscCheckFalse(nzl>1024,PetscObjectComm((PetscObject) B), PETSC_ERR_PLIB, "Row too big: %D",nzl); 882 for (int j=0; j<nzl; j++) colbuf[j] = cols[j] + moffset; 883 row = i + moffset; 884 ierr = MatSetValues(JacP,1,&row,nzl,colbuf,vals,ADD_VALUES);CHKERRQ(ierr); 885 ierr = MatRestoreRow(B,i,&nzl,&cols,&vals);CHKERRQ(ierr); 886 } 887 ierr = MatDestroy(&subJ[ LAND_PACK_IDX(b_id,grid) ]);CHKERRQ(ierr); 888 ierr = PetscLogEventEnd(events[6],0,0,0,0);CHKERRQ(ierr); 889 } // grids 890 } 891 } 892 PetscFunctionReturn(0); 893 } 894 } // extern "C" 895