1 /* 2 Implements the Kokkos kernel 3 */ 4 #define PETSC_SKIP_CXX_COMPLEX_FIX 5 #include <petscconf.h> 6 #include <petsc/private/dmpleximpl.h> /*I "petscdmplex.h" I*/ 7 #include <petsclandau.h> 8 #include <petscts.h> 9 #include <petscveckokkos.hpp> 10 11 #include <Kokkos_Core.hpp> 12 #include <cstdio> 13 typedef Kokkos::TeamPolicy<>::member_type team_member; 14 #define PETSC_DEVICE_FUNC_DECL KOKKOS_INLINE_FUNCTION 15 #include "../land_tensors.h" 16 #define atomicAdd(e, f) Kokkos::atomic_fetch_add(e, f) 17 #include <petscaijdevice.h> 18 #undef atomicAdd 19 20 namespace landau_inner_red { // namespace helps with name resolution in reduction identity 21 template< class ScalarType > 22 struct array_type { 23 ScalarType gg2[LANDAU_DIM]; 24 ScalarType gg3[LANDAU_DIM][LANDAU_DIM]; 25 26 KOKKOS_INLINE_FUNCTION // Default constructor - Initialize to 0's 27 array_type() { 28 for (int j = 0; j < LANDAU_DIM; j++){ 29 gg2[j] = 0; 30 for (int k = 0; k < LANDAU_DIM; k++){ 31 gg3[j][k] = 0; 32 } 33 } 34 } 35 KOKKOS_INLINE_FUNCTION // Copy Constructor 36 array_type(const array_type & rhs) { 37 for (int j = 0; j < LANDAU_DIM; j++){ 38 gg2[j] = rhs.gg2[j]; 39 for (int k = 0; k < LANDAU_DIM; k++){ 40 gg3[j][k] = rhs.gg3[j][k]; 41 } 42 } 43 } 44 KOKKOS_INLINE_FUNCTION // add operator 45 array_type& operator += (const array_type& src) { 46 for (int j = 0; j < LANDAU_DIM; j++){ 47 gg2[j] += src.gg2[j]; 48 for (int k = 0; k < LANDAU_DIM; k++){ 49 gg3[j][k] += src.gg3[j][k]; 50 } 51 } 52 return *this; 53 } 54 KOKKOS_INLINE_FUNCTION // volatile add operator 55 void operator += (const volatile array_type& src) volatile { 56 for (int j = 0; j < LANDAU_DIM; j++){ 57 gg2[j] += src.gg2[j]; 58 for (int k = 0; k < LANDAU_DIM; k++){ 59 gg3[j][k] += src.gg3[j][k]; 60 } 61 } 62 } 63 }; 64 typedef array_type<PetscReal> TensorValueType; // used to simplify code below 65 } 66 67 namespace Kokkos { //reduction identity must be defined in Kokkos namespace 68 template<> 69 struct reduction_identity< landau_inner_red::TensorValueType > { 70 KOKKOS_FORCEINLINE_FUNCTION static landau_inner_red::TensorValueType sum() { 71 return landau_inner_red::TensorValueType(); 72 } 73 }; 74 } 75 76 extern "C" { 77 PetscErrorCode LandauKokkosCreateMatMaps(P4estVertexMaps *maps, pointInterpolationP4est (*pointMaps)[LANDAU_MAX_Q_FACE], PetscInt Nf, PetscInt Nq) 78 { 79 P4estVertexMaps h_maps; /* host container */ 80 const Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_points ((pointInterpolationP4est*)pointMaps, maps->num_reduced); 81 const Kokkos::View< LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_gidx ((LandauIdx*)maps->gIdx, maps->num_elements); 82 Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight> *d_points = new Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>("points", maps->num_reduced); 83 Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight> *d_gidx = new Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight>("gIdx", maps->num_elements); 84 85 PetscFunctionBegin; 86 Kokkos::deep_copy (*d_gidx, h_gidx); 87 Kokkos::deep_copy (*d_points, h_points); 88 h_maps.num_elements = maps->num_elements; 89 h_maps.num_face = maps->num_face; 90 h_maps.num_reduced = maps->num_reduced; 91 h_maps.deviceType = maps->deviceType; 92 h_maps.Nf = Nf; 93 h_maps.Nq = Nq; 94 h_maps.c_maps = (pointInterpolationP4est (*)[LANDAU_MAX_Q_FACE]) d_points->data(); 95 maps->vp1 = (void*)d_points; 96 h_maps.gIdx = (LandauIdx (*)[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ]) d_gidx->data(); 97 maps->vp2 = (void*)d_gidx; 98 { 99 Kokkos::View<P4estVertexMaps, Kokkos::HostSpace> h_maps_k(&h_maps); 100 Kokkos::View<P4estVertexMaps> *d_maps_k = new Kokkos::View<P4estVertexMaps>(Kokkos::create_mirror(DeviceMemorySpace(),h_maps_k)); 101 Kokkos::deep_copy (*d_maps_k, h_maps_k); 102 maps->data = d_maps_k->data(); 103 maps->vp3 = (void*)d_maps_k; 104 // debug 105 // PetscErrorCode ierr; 106 // PetscInt ej,q,s; 107 // ierr = PetscPrintf(PETSC_COMM_SELF,"=== c_maps size --> %D. Kokkos::View<P4estVertexMaps*>=%p maps->data=%p P4estVertexMaps *maps=%p\n",maps->num_reduced, d_maps_k, maps->data, maps);CHKERRQ(ierr); 108 // for (ej = 0; ej < maps->num_reduced; ++ej) { 109 // ierr = PetscPrintf(PETSC_COMM_SELF,"\t\tconstrant %3D) ",ej); 110 // for (q = 0; q < maps->num_face; ++q) { 111 // ierr = PetscPrintf(PETSC_COMM_SELF,"\t %3D - %13.5e ; ", maps->data->c_maps[ej][q].gid, maps->data->c_maps[ej][q].scale);CHKERRQ(ierr); 112 // } 113 // ierr = PetscPrintf(PETSC_COMM_SELF,"\n"); 114 // } 115 // ierr = PetscPrintf(PETSC_COMM_SELF,"================= gIDx \n"); 116 // for (ej = 0; ej < maps->num_elements; ++ej) { 117 // for (q = 0; q < Nq; ++q) { 118 // for (s = 0; s < Nf; ++s) { 119 // ierr = PetscPrintf(PETSC_COMM_SELF," %3D ", maps->data->gIdx[ej][s][q]); 120 // } 121 // ierr = PetscPrintf(PETSC_COMM_SELF," ; "); 122 // } 123 // ierr = PetscPrintf(PETSC_COMM_SELF,"\n"); 124 // } 125 } 126 PetscFunctionReturn(0); 127 } 128 PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps *maps) 129 { 130 Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight> *a = (Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>*)maps->vp1; 131 Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight>*b = (Kokkos::View<LandauIdx*[LANDAU_MAX_SPECIES][LANDAU_MAX_NQ], Kokkos::LayoutRight>*)maps->vp2; 132 Kokkos::View<P4estVertexMaps*> *c = (Kokkos::View<P4estVertexMaps*>*)maps->vp3; 133 delete a; delete b; delete c; 134 PetscFunctionReturn(0); 135 } 136 137 PetscErrorCode LandauKokkosJacobian(DM plex, const PetscInt Nq, PetscReal nu_alpha[], PetscReal nu_beta[], PetscReal invMass[], PetscReal Eq_m[], 138 const LandauIPData *const IPData, PetscReal a_invJ[], const PetscInt num_sub_blocks, PetscReal *mass_w, PetscReal shift, 139 const PetscLogEvent events[], Mat JacP) 140 { 141 PetscErrorCode ierr; 142 PetscInt *Nbf,Nb,cStart,cEnd,Nf,dim,numCells,totDim,ipdatasz,global_elem_mat_sz,NfJac,nip; 143 PetscTabulation *Tf; 144 PetscDS prob; 145 PetscSection section, globalSection; 146 PetscLogDouble flops; 147 PetscReal *BB,*DD; 148 LandauCtx *ctx; 149 P4estVertexMaps *d_maps=NULL; 150 PetscSplitCSRDataStructure *d_mat=NULL; 151 152 PetscFunctionBegin; 153 ierr = DMGetApplicationContext(plex, &ctx);CHKERRQ(ierr); 154 if (!ctx) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 155 ierr = DMGetDimension(plex, &dim);CHKERRQ(ierr); 156 ierr = DMPlexGetHeightStratum(plex,0,&cStart,&cEnd);CHKERRQ(ierr); 157 numCells = cEnd - cStart; 158 ierr = DMGetDS(plex, &prob);CHKERRQ(ierr); 159 ierr = PetscDSGetNumFields(prob, &Nf);CHKERRQ(ierr); 160 ierr = PetscDSGetDimensions(prob, &Nbf);CHKERRQ(ierr); Nb = Nbf[0]; 161 if (Nq != Nb) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Nq != Nb. %D %D",Nq,Nb); 162 if (LANDAU_DIM != dim) SETERRQ2(PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %D != LANDAU_DIM %d",dim,LANDAU_DIM); 163 ierr = PetscDSGetTotalDimension(prob, &totDim);CHKERRQ(ierr); 164 ierr = PetscDSGetTabulation(prob, &Tf);CHKERRQ(ierr); 165 if (ctx->gpu_assembly) { 166 PetscContainer container; 167 ierr = PetscObjectQuery((PetscObject) JacP, "assembly_maps", (PetscObject *) &container);CHKERRQ(ierr); 168 if (container) { // not here first call 169 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 170 P4estVertexMaps *h_maps=NULL; 171 ierr = PetscContainerGetPointer(container, (void **) &h_maps);CHKERRQ(ierr); 172 ierr = PetscInfo2(JacP, "Have container maps=%p maps->data=%p\n", h_maps, h_maps ? h_maps->data : NULL);CHKERRQ(ierr); 173 if (h_maps->data) { 174 d_maps = h_maps->data; 175 } else { 176 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); 177 } 178 // this does the setup the first time called 179 ierr = MatKokkosGetDeviceMatWrite(JacP,&d_mat);CHKERRQ(ierr); 180 global_elem_mat_sz = 0; 181 #else 182 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly w/o kokkos kernels -- should not be here"); 183 #endif 184 } else { // kernel output - first call assembled on device 185 global_elem_mat_sz = numCells; 186 } 187 } else { 188 global_elem_mat_sz = numCells; // no device assembly 189 } 190 BB = Tf[0]->T[0]; DD = Tf[0]->T[1]; 191 ierr = DMGetLocalSection(plex, §ion);CHKERRQ(ierr); 192 ierr = DMGetGlobalSection(plex, &globalSection);CHKERRQ(ierr); 193 if (mass_w) { 194 flops = (PetscLogDouble)numCells*Nq*(5*dim*dim*Nf*Nf); 195 ipdatasz = 0; NfJac = 0; nip = numCells*Nq; 196 } else { 197 flops = (PetscLogDouble)numCells*Nq*(5*dim*dim*Nf*Nf + 165); 198 ipdatasz = LandauGetIPDataSize(IPData); NfJac = Nf; nip = IPData->nip_; 199 } 200 ierr = PetscKokkosInitializeCheck();CHKERRQ(ierr); 201 { 202 using scr_mem_t = Kokkos::DefaultExecutionSpace::scratch_memory_space; 203 using g2_scr_t = Kokkos::View<PetscReal***, Kokkos::LayoutRight, scr_mem_t>; 204 using g3_scr_t = Kokkos::View<PetscReal****, Kokkos::LayoutRight, scr_mem_t>; 205 206 const int scr_bytes = 2*(g2_scr_t::shmem_size(dim,Nf,Nq) + g3_scr_t::shmem_size(dim,dim,Nf,Nq)); 207 int conc, team_size; 208 ierr = PetscLogEventBegin(events[3],0,0,0,0);CHKERRQ(ierr); 209 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_alpha (nu_alpha, NfJac); 210 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_alpha ("nu_alpha", NfJac); 211 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_beta (nu_beta, NfJac); 212 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_beta ("nu_beta", NfJac); 213 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invMass (invMass,NfJac); 214 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_invMass ("invMass", NfJac); 215 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_Eq_m (Eq_m,NfJac); 216 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_Eq_m ("Eq_m", NfJac); 217 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_BB (BB,Nq*Nb); 218 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_BB ("BB", Nq*Nb); 219 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_DD (DD,Nq*Nb*dim); 220 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_DD ("DD", Nq*Nb*dim); 221 const Kokkos::View<LandauIPReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_ipdata_raw (IPData->w, ipdatasz); // IPData->w==0 for mass 222 Kokkos::View<LandauIPReal*, Kokkos::LayoutLeft> d_ipdata_raw ("ipdata", ipdatasz); 223 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invJ (a_invJ, mass_w ? 0 : nip*dim*dim); 224 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_invJ ("invJ", mass_w ? 0 : nip*dim*dim); 225 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_mass_w (mass_w, mass_w ? nip : 0); 226 Kokkos::View<PetscReal*, Kokkos::LayoutLeft> d_mass_w ("mass_w", mass_w ? nip : 0); 227 Kokkos::View<PetscScalar**, Kokkos::LayoutRight> d_elem_mats("element matrices", global_elem_mat_sz, totDim*totDim); 228 Kokkos::View<PetscScalar**, Kokkos::LayoutRight> d_f("element matrices", NfJac, nip); 229 Kokkos::View<PetscScalar***, Kokkos::LayoutRight> d_df("element matrices", dim, NfJac, nip); 230 Kokkos::deep_copy (d_mass_w, h_mass_w); 231 Kokkos::deep_copy (d_ipdata_raw, h_ipdata_raw); 232 Kokkos::deep_copy (d_alpha, h_alpha); 233 Kokkos::deep_copy (d_beta, h_beta); 234 Kokkos::deep_copy (d_invMass, h_invMass); 235 Kokkos::deep_copy (d_Eq_m, h_Eq_m); 236 Kokkos::deep_copy (d_BB, h_BB); 237 Kokkos::deep_copy (d_DD, h_DD); 238 Kokkos::deep_copy (d_invJ, h_invJ); 239 240 ierr = PetscLogEventEnd(events[3],0,0,0,0);CHKERRQ(ierr); 241 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 242 ierr = PetscLogGpuFlops(flops*nip);CHKERRQ(ierr); 243 if (ctx->deviceType == LANDAU_CPU) PetscInfo(plex, "Warning: Landau selected CPU but no support for Kokkos using GPU\n"); 244 #else 245 ierr = PetscLogFlops(flops*nip);CHKERRQ(ierr); 246 #endif 247 #define KOKKOS_SHARED_LEVEL 1 248 conc = Kokkos::DefaultExecutionSpace().concurrency(), team_size = conc > Nq ? Nq : 1; 249 // PetscInfo5(plex, "shared memory size: %d bytes in level %d. mass_w.size=%d. #threads=%D team size=%D\n",scr_bytes,KOKKOS_SHARED_LEVEL,d_mass_w.size(),num_sub_blocks,team_size); 250 // get f and df 251 if (!mass_w) { 252 ierr = PetscLogEventBegin(events[8],0,0,0,0);CHKERRQ(ierr); 253 Kokkos::parallel_for("df_d_elements", Kokkos::TeamPolicy<>(numCells, team_size, num_sub_blocks), KOKKOS_LAMBDA (const team_member team) { 254 const PetscInt myelem = team.league_rank(); 255 // un pack IPData 256 LandauIPReal *IPData_coefs = &d_ipdata_raw[nip*(dim+1)]; 257 LandauIPReal *coef = &IPData_coefs[myelem*Nb*Nf]; 258 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 259 const PetscInt ipidx = myQi + myelem * Nq; 260 const PetscReal *const invJj = &d_invJ(ipidx*dim*dim); 261 const PetscReal *Bq = &d_BB[myQi*Nb], *Dq = &d_DD[myQi*Nb*dim]; 262 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nf), [=] (int f) { 263 PetscInt b, e, d; 264 PetscScalar refSpaceDer[LANDAU_DIM]; 265 d_f(f,ipidx) = 0.0; 266 for (d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0; 267 for (b = 0; b < Nb; ++b) { 268 const PetscInt cidx = b; 269 d_f(f,ipidx) += Bq[cidx]*coef[f*Nb+cidx]; 270 for (d = 0; d < dim; ++d) refSpaceDer[d] += Dq[cidx*dim+d]*coef[f*Nb+cidx]; 271 } 272 for (d = 0; d < dim; ++d) { 273 for (e = 0, d_df(d,f,ipidx) = 0.0; e < dim; ++e) { 274 d_df(d,f,ipidx) += invJj[e*dim+d]*refSpaceDer[e]; 275 } 276 } 277 }); // Nf 278 }); // Nq 279 }); // elems 280 Kokkos::fence(); 281 ierr = PetscLogEventEnd(events[8],0,0,0,0);CHKERRQ(ierr); 282 } 283 ierr = PetscLogEventBegin(events[4],0,0,0,0);CHKERRQ(ierr); 284 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 285 Kokkos::parallel_for("Landau_elements", Kokkos::TeamPolicy<>(numCells, team_size, num_sub_blocks).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes)), KOKKOS_LAMBDA (const team_member team) { 286 const PetscInt myelem = team.league_rank(); 287 g2_scr_t g2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,Nf,Nq); // we don't use these for mass matrix 288 g3_scr_t g3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,Nf,Nq); 289 if (!d_mass_w.size()) { 290 g2_scr_t gg2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,Nf,Nq); 291 g3_scr_t gg3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,Nf,Nq); 292 LandauIPData d_IPData; 293 // un pack IPData 294 d_IPData.w = &d_ipdata_raw[0]; 295 d_IPData.x = &d_ipdata_raw[1*nip]; 296 d_IPData.y = &d_ipdata_raw[2*nip]; 297 if (dim==2) d_IPData.z = NULL; 298 else d_IPData.z = &d_ipdata_raw[3*nip]; 299 // get g2[] & g3[] 300 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 301 using Kokkos::parallel_reduce; 302 const PetscInt jpidx = myQi + myelem * Nq; 303 const PetscReal* const invJj = &d_invJ(jpidx*dim*dim); 304 const PetscReal vj[3] = {d_IPData.x[jpidx], d_IPData.y[jpidx], d_IPData.z ? d_IPData.z[jpidx] : 0}, wj = d_IPData.w[jpidx]; 305 landau_inner_red::TensorValueType gg_temp; // reduce on part of gg2 and g33 for IP jpidx 306 Kokkos::parallel_reduce(Kokkos::ThreadVectorRange (team, (int)nip), [=] (const int& ipidx, landau_inner_red::TensorValueType & ggg) { 307 const PetscReal wi = d_IPData.w[ipidx], x = d_IPData.x[ipidx], y = d_IPData.y[ipidx]; 308 PetscReal temp1[3] = {0, 0, 0}, temp2 = 0; 309 PetscInt fieldA,d2,d3; 310 #if LANDAU_DIM==2 311 PetscReal Ud[2][2], Uk[2][2]; 312 LandauTensor2D(vj, x, y, Ud, Uk, (ipidx==jpidx) ? 0. : 1.); 313 #else 314 PetscReal U[3][3], z = d_IPData.z[ipidx]; 315 LandauTensor3D(vj, x, y, z, U, (ipidx==jpidx) ? 0. : 1.); 316 #endif 317 for (fieldA = 0; fieldA < Nf; ++fieldA) { 318 temp1[0] += d_df(0,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA]; 319 temp1[1] += d_df(1,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA]; 320 #if LANDAU_DIM==3 321 temp1[2] += d_df(2,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA]; 322 #endif 323 temp2 += d_f(fieldA,ipidx)*d_beta[fieldA]; 324 } 325 temp1[0] *= wi; 326 temp1[1] *= wi; 327 #if LANDAU_DIM==3 328 temp1[2] *= wi; 329 #endif 330 temp2 *= wi; 331 #if LANDAU_DIM==2 332 for (d2 = 0; d2 < 2; d2++) { 333 for (d3 = 0; d3 < 2; ++d3) { 334 /* K = U * grad(f): g2=e: i,A */ 335 ggg.gg2[d2] += Uk[d2][d3]*temp1[d3]; 336 /* D = -U * (I \kron (fx)): g3=f: i,j,A */ 337 ggg.gg3[d2][d3] += Ud[d2][d3]*temp2; 338 } 339 } 340 #else 341 for (d2 = 0; d2 < 3; ++d2) { 342 for (d3 = 0; d3 < 3; ++d3) { 343 /* K = U * grad(f): g2 = e: i,A */ 344 ggg.gg2[d2] += U[d2][d3]*temp1[d3]; 345 /* D = -U * (I \kron (fx)): g3 = f: i,j,A */ 346 ggg.gg3[d2][d3] += U[d2][d3]*temp2; 347 } 348 } 349 #endif 350 }, Kokkos::Sum<landau_inner_red::TensorValueType>(gg_temp)); 351 // if (myelem==0) printf("\t:%d.%d) temp gg3=%e %e %e %e\n",myelem,myQi,gg_temp.gg3[0][0],gg_temp.gg3[1][0],gg_temp.gg3[0][1],gg_temp.gg3[1][1]); 352 // add alpha and put in gg2/3 353 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [&] (const int& fieldA) { 354 PetscInt d2,d3; 355 for (d2 = 0; d2 < dim; d2++) { 356 gg2(d2,fieldA,myQi) = gg_temp.gg2[d2]*d_alpha[fieldA]; 357 //if (myelem==0 && fieldA==1) printf("\t\t:%d.%d) gg2[%d]=%e (+= %e)\n",myelem,myQi,d2,gg2(d2,fieldA,myQi),gg_temp.gg2[d2]*d_alpha[fieldA]); 358 //gg2[d2][myQi][fieldA] += gg_temp.gg2[d2]*d_alpha[fieldA]; 359 for (d3 = 0; d3 < dim; d3++) { 360 //gg3[d2][d3][myQi][fieldA] -= gg_temp.gg3[d2][d3]*d_alpha[fieldA]*s_invMass[fieldA]; 361 gg3(d2,d3,fieldA,myQi) = -gg_temp.gg3[d2][d3]*d_alpha[fieldA]*d_invMass[fieldA]; 362 //if (myelem==0 && fieldA==1) printf("\t\t\t:%d.%d) gg3[%d][%d]=%e\n",myelem,myQi,d2,d3,gg3(d2,d3,fieldA,myQi)); 363 } 364 } 365 }); 366 /* add electric field term once per IP */ 367 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [&] (const int& fieldA) { 368 //gg.gg2[fieldA][dim-1] += d_Eq_m[fieldA]; 369 gg2(dim-1,fieldA,myQi) += d_Eq_m[fieldA]; 370 }); 371 // Kokkos::single(Kokkos::PerThread(team), [&]() { 372 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [=] (const int& fieldA) { 373 int d,d2,d3,dp; 374 //printf("%d %d %d gg2[][1]=%18.10e\n",myelem,myQi,fieldA,gg.gg2[fieldA][dim-1]); 375 /* Jacobian transform - g2, g3 - per thread (2D) */ 376 for (d = 0; d < dim; ++d) { 377 g2(d,fieldA,myQi) = 0; 378 for (d2 = 0; d2 < dim; ++d2) { 379 g2(d,fieldA,myQi) += invJj[d*dim+d2]*gg2(d2,fieldA,myQi); 380 //if (myelem==0 && myQi==0) printf("\t:g2[%d][%d][%d]=%e. %e %e\n",(int)myQi,(int)fieldA,(int)d,g2(fieldA,myQi,d),invJj[d*dim+d2],gg.gg2[fieldA][d2]); 381 g3(d,d2,fieldA,myQi) = 0; 382 for (d3 = 0; d3 < dim; ++d3) { 383 for (dp = 0; dp < dim; ++dp) { 384 g3(d,d2,fieldA,myQi) += invJj[d*dim + d3]*gg3(d3,dp,fieldA,myQi)*invJj[d2*dim + dp]; 385 //printf("\t%d %d %d %d %d %d %d g3=%g wj=%g g3 = %g * %g * %g\n",myelem,myQi,fieldA,d,d2,d3,dp,g3(fieldA,myQi,d,d2),wj,invJj[d*dim + d3],gg.gg3[fieldA][d3][dp],invJj[d2*dim + dp]); 386 } 387 } 388 g3(d,d2,fieldA,myQi) *= wj; 389 } 390 g2(d,fieldA,myQi) *= wj; 391 } 392 }); 393 }); // Nq 394 team.team_barrier(); 395 } // Jacobian 396 /* assemble */ 397 //Kokkos::single(Kokkos::PerTeam(team), [&]() { 398 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nb), [=] (int blk_i) { 399 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nf), [=] (int fieldA) { 400 int blk_j,qj,d,d2; 401 const PetscInt i = fieldA*Nb + blk_i; /* Element matrix row */ 402 for (blk_j = 0; blk_j < Nb; ++blk_j) { 403 const PetscInt j = fieldA*Nb + blk_j; /* Element matrix column */ 404 const PetscInt fOff = i*totDim + j; 405 PetscScalar t = global_elem_mat_sz ? d_elem_mats(myelem,fOff) : 0; 406 for (qj = 0 ; qj < Nq ; qj++) { // look at others integration points 407 const PetscReal *BJq = &d_BB[qj*Nb], *DIq = &d_DD[qj*Nb*dim]; 408 if (!d_mass_w.size()) { 409 for (d = 0; d < dim; ++d) { 410 t += DIq[blk_i*dim+d]*g2(d,fieldA,qj)*BJq[blk_j]; 411 for (d2 = 0; d2 < dim; ++d2) { 412 t += DIq[blk_i*dim + d]*g3(d,d2,fieldA,qj)*DIq[blk_j*dim + d2]; 413 } 414 } 415 } else { 416 const PetscInt jpidx = qj + myelem * Nq; 417 t += BJq[blk_i] * d_mass_w(jpidx)*shift * BJq[blk_j]; 418 //printf("\tmat[%d %d %d %d %d]: B=%g w=%g shift=%g B=%g\n",myelem,fOff,fieldA,qj,d,BJq[blk_i],d_mass_w(jpidx),shift,BJq[blk_j]); 419 } 420 } 421 if (global_elem_mat_sz) d_elem_mats(myelem,fOff) = t; // can set this because local element matrix[fOff] 422 else { 423 PetscErrorCode ierr = 0; 424 PetscScalar vals[LANDAU_MAX_Q*LANDAU_MAX_Q],row_scale[LANDAU_MAX_Q],col_scale[LANDAU_MAX_Q]; 425 PetscInt q,idx,nr,nc,rows0[LANDAU_MAX_Q],cols0[LANDAU_MAX_Q],rows[LANDAU_MAX_Q],cols[LANDAU_MAX_Q]; 426 const LandauIdx *const Idxs = &d_maps->gIdx[myelem][fieldA][0]; 427 idx = Idxs[blk_i]; 428 if (idx >= 0) { 429 nr = 1; 430 rows0[0] = idx; 431 row_scale[0] = 1.; 432 } else { 433 idx = -idx - 1; 434 nr = d_maps->num_face; 435 for (q = 0; q < d_maps->num_face; q++) { 436 rows0[q] = d_maps->c_maps[idx][q].gid; 437 row_scale[q] = d_maps->c_maps[idx][q].scale; 438 } 439 } 440 idx = Idxs[blk_j]; 441 if (idx >= 0) { 442 nc = 1; 443 cols0[0] = idx; 444 col_scale[0] = 1.; 445 } else { 446 idx = -idx - 1; 447 nc = d_maps->num_face; 448 for (q = 0; q < d_maps->num_face; q++) { 449 cols0[q] = d_maps->c_maps[idx][q].gid; 450 col_scale[q] = d_maps->c_maps[idx][q].scale; 451 } 452 } 453 for (q = 0; q < nr; q++) rows[q] = rows0[q]; 454 for (q = 0; q < nc; q++) cols[q] = cols0[q]; 455 for (q = 0; q < nr; q++) { 456 for (d = 0; d < nc; d++) { 457 vals[q*nc + d] = row_scale[q]*col_scale[d]*t; 458 } 459 } 460 MatSetValuesDevice(d_mat,nr,rows,nc,cols,vals,ADD_VALUES,&ierr); 461 if (ierr) return; 462 } 463 } 464 }); 465 }); 466 }); 467 Kokkos::fence(); 468 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); 469 ierr = PetscLogEventEnd(events[4],0,0,0,0);CHKERRQ(ierr); 470 if (global_elem_mat_sz) { 471 Kokkos::View<PetscScalar**, Kokkos::LayoutRight>::HostMirror h_elem_mats = Kokkos::create_mirror_view(d_elem_mats); 472 ierr = PetscLogEventBegin(events[5],0,0,0,0);CHKERRQ(ierr); 473 Kokkos::deep_copy (h_elem_mats, d_elem_mats); 474 ierr = PetscLogEventEnd(events[5],0,0,0,0);CHKERRQ(ierr); 475 ierr = PetscLogEventBegin(events[6],0,0,0,0);CHKERRQ(ierr); 476 PetscInt ej; 477 for (ej = cStart ; ej < cEnd; ++ej) { 478 const PetscScalar *elMat = &h_elem_mats(ej-cStart,0); 479 ierr = DMPlexMatSetClosure(plex, section, globalSection, JacP, ej, elMat, ADD_VALUES);CHKERRQ(ierr); 480 if (ej==-1) { 481 int d,f; 482 PetscPrintf(PETSC_COMM_SELF,"Kokkos Element matrix %d/%d\n",1,(int)numCells); 483 for (d = 0; d < totDim; ++d){ 484 for (f = 0; f < totDim; ++f) PetscPrintf(PETSC_COMM_SELF," %12.5e", PetscRealPart(elMat[d*totDim + f])); 485 PetscPrintf(PETSC_COMM_SELF,"\n"); 486 } 487 } 488 } 489 ierr = PetscLogEventEnd(events[6],0,0,0,0);CHKERRQ(ierr); 490 } 491 } 492 PetscFunctionReturn(0); 493 } 494 } // extern "C" 495