1 /* 2 Implements the Kokkos kernel 3 */ 4 #define PETSC_SKIP_CXX_COMPLEX_FIX 5 #include <petscconf.h> 6 #include <petscvec_kokkos.hpp> 7 #include <petsc/private/dmpleximpl.h> /*I "petscdmplex.h" I*/ 8 #include <petsclandau.h> 9 #include <petscts.h> 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(Kokkos::DefaultExecutionSpace::memory_space(),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 } 105 PetscFunctionReturn(0); 106 } 107 PetscErrorCode LandauKokkosDestroyMatMaps(P4estVertexMaps *maps) 108 { 109 Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight> *a = (Kokkos::View<pointInterpolationP4est*[LANDAU_MAX_Q_FACE], Kokkos::LayoutRight>*)maps->vp1; 110 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; 111 Kokkos::View<P4estVertexMaps*> *c = (Kokkos::View<P4estVertexMaps*>*)maps->vp3; 112 delete a; delete b; delete c; 113 PetscFunctionReturn(0); 114 } 115 116 PetscErrorCode LandauKokkosStaticDataSet(DM plex, const PetscInt Nq, PetscReal nu_alpha[], PetscReal nu_beta[], PetscReal a_invMass[], PetscReal a_invJ[], PetscReal a_mass_w[], 117 PetscReal a_x[], PetscReal a_y[], PetscReal a_z[], PetscReal a_w[], LandauGeomData *SData_d) 118 { 119 PetscReal *BB,*DD; 120 PetscErrorCode ierr; 121 PetscTabulation *Tf; 122 LandauCtx *ctx; 123 PetscInt *Nbf,dim,Nf,Nb,nip,cStart,cEnd; 124 PetscDS prob; 125 126 PetscFunctionBegin; 127 ierr = DMGetApplicationContext(plex, &ctx);CHKERRQ(ierr); 128 if (!ctx) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 129 ierr = DMGetDimension(plex, &dim);CHKERRQ(ierr); 130 ierr = DMPlexGetHeightStratum(plex,0,&cStart,&cEnd);CHKERRQ(ierr); 131 nip = (cEnd - cStart)*Nq; 132 ierr = DMGetDS(plex, &prob);CHKERRQ(ierr); 133 ierr = PetscDSGetNumFields(prob, &Nf);CHKERRQ(ierr); 134 ierr = PetscDSGetDimensions(prob, &Nbf);CHKERRQ(ierr); Nb = Nbf[0]; 135 if (Nq != Nb) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Nq != Nb. %D %D",Nq,Nb); 136 if (LANDAU_DIM != dim) SETERRQ2(PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %D != LANDAU_DIM %d",dim,LANDAU_DIM); 137 ierr = PetscDSGetTabulation(prob, &Tf);CHKERRQ(ierr); 138 BB = Tf[0]->T[0]; DD = Tf[0]->T[1]; 139 ierr = PetscKokkosInitializeCheck();CHKERRQ(ierr); 140 { 141 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_alpha (nu_alpha, Nf); 142 auto alpha = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("alpha", Nf); 143 SData_d->alpha = static_cast<void*>(alpha); 144 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_beta (nu_beta, Nf); 145 auto beta = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("beta", Nf); 146 SData_d->beta = static_cast<void*>(beta); 147 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invMass (a_invMass,Nf); 148 auto invMass = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("invMass", Nf); 149 SData_d->invMass = static_cast<void*>(invMass); 150 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_BB (BB,Nq*Nb); 151 auto B = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("B", Nq*Nb); 152 SData_d->B = static_cast<void*>(B); 153 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_DD (DD,Nq*Nb*dim); 154 auto D = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("D", Nq*Nb*dim); 155 SData_d->D = static_cast<void*>(D); 156 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_mass_w (a_mass_w, nip); 157 auto mass_w = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("mass_w", nip); 158 SData_d->mass_w = static_cast<void*>(mass_w); 159 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_invJ (a_invJ, nip*dim*dim); 160 auto invJ = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("invJ", nip*dim*dim); 161 SData_d->invJ = static_cast<void*>(invJ); 162 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_x (a_x, nip); 163 auto x = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("x", nip); 164 SData_d->x = static_cast<void*>(x); 165 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_y (a_y, nip); 166 auto y = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("y", nip); 167 SData_d->y = static_cast<void*>(y); 168 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_w (a_w, nip); 169 auto w = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("w", nip); 170 SData_d->w = static_cast<void*>(w); 171 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_z (a_z , dim==3 ? nip : 0); 172 auto z = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("z", dim==3 ? nip : 0); 173 SData_d->z = static_cast<void*>(z); 174 175 Kokkos::deep_copy (*mass_w, h_mass_w); 176 Kokkos::deep_copy (*alpha, h_alpha); 177 Kokkos::deep_copy (*beta, h_beta); 178 Kokkos::deep_copy (*invMass, h_invMass); 179 Kokkos::deep_copy (*B, h_BB); 180 Kokkos::deep_copy (*D, h_DD); 181 Kokkos::deep_copy (*invJ, h_invJ); 182 Kokkos::deep_copy (*x, h_x); 183 Kokkos::deep_copy (*y, h_y); 184 Kokkos::deep_copy (*z, h_z); 185 Kokkos::deep_copy (*w, h_w); 186 187 auto Eq_m = new Kokkos::View<PetscReal*, Kokkos::LayoutLeft> ("Eq_m",Nf); 188 SData_d->Eq_m = static_cast<void*>(Eq_m); 189 auto IPf = new Kokkos::View<PetscScalar*, Kokkos::LayoutLeft> ("IPf",nip*Nf); // Nq==Nb 190 SData_d->IPf = static_cast<void*>(IPf); 191 } 192 PetscFunctionReturn(0); 193 } 194 195 PetscErrorCode LandauKokkosStaticDataClear(LandauGeomData *SData_d) 196 { 197 PetscFunctionBegin; 198 if (SData_d->alpha) { 199 auto alpha = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->alpha); 200 delete alpha; 201 auto beta = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->beta); 202 delete beta; 203 auto invMass = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invMass); 204 delete invMass; 205 auto B = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->B); 206 delete B; 207 auto D = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->D); 208 delete D; 209 auto mass_w = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->mass_w); 210 delete mass_w; 211 auto invJ = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invJ); 212 delete invJ; 213 auto x = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->x); 214 delete x; 215 auto y = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->y); 216 delete y; 217 auto z = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->z); 218 delete z; 219 auto w = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->w); 220 delete w; 221 auto Eq_m = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->Eq_m); 222 delete Eq_m; 223 if (SData_d->IPf) { 224 auto IPf = static_cast<Kokkos::View<PetscScalar*, Kokkos::LayoutLeft>*>(SData_d->IPf); 225 delete IPf; 226 } 227 } 228 PetscFunctionReturn(0); 229 } 230 231 PetscErrorCode LandauKokkosJacobian(DM plex, const PetscInt Nq, PetscReal a_Eq_m[], PetscScalar a_IPf[], const PetscInt N, const PetscScalar a_xarray[], LandauGeomData *SData_d, 232 const PetscInt num_sub_blocks, PetscReal shift, const PetscLogEvent events[], Mat JacP) 233 { 234 using scr_mem_t = Kokkos::DefaultExecutionSpace::scratch_memory_space; 235 using g2_scr_t = Kokkos::View<PetscReal***, Kokkos::LayoutRight, scr_mem_t>; 236 using g3_scr_t = Kokkos::View<PetscReal****, Kokkos::LayoutRight, scr_mem_t>; 237 PetscErrorCode ierr; 238 PetscInt *Nbf,Nb,cStart,cEnd,Nf,dim,numCells,totDim,global_elem_mat_sz,nip; 239 PetscDS prob; 240 LandauCtx *ctx; 241 PetscReal *d_Eq_m=NULL; 242 PetscScalar *d_IPf=NULL; 243 P4estVertexMaps *d_maps=NULL; 244 PetscSplitCSRDataStructure *d_mat=NULL; 245 const int conc = Kokkos::DefaultExecutionSpace().concurrency(), openmp = !!(conc < 1000), team_size = (openmp==0) ? Nq : 1; 246 int scr_bytes; 247 auto d_alpha_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->alpha); //static data 248 const PetscReal *d_alpha = d_alpha_k->data(); 249 auto d_beta_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->beta); 250 const PetscReal *d_beta = d_beta_k->data(); 251 auto d_invMass_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invMass); 252 const PetscReal *d_invMass = d_invMass_k->data(); 253 auto d_B_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->B); 254 const PetscReal *d_BB = d_B_k->data(); 255 auto d_D_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->D); 256 const PetscReal *d_DD = d_D_k->data(); 257 auto d_mass_w_k = *static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->mass_w); 258 auto d_invJ_k = *static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->invJ); // use Kokkos vector in kernels 259 auto d_x_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->x); //static data 260 const PetscReal *d_x = d_x_k->data(); 261 auto d_y_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->y); //static data 262 const PetscReal *d_y = d_y_k->data(); 263 auto d_z_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->z); //static data 264 const PetscReal *d_z = d_z_k->data(); 265 auto d_w_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->w); //static data 266 const PetscReal *d_w = d_w_k->data(); 267 // dynamic data pointers 268 auto d_Eq_m_k = static_cast<Kokkos::View<PetscReal*, Kokkos::LayoutLeft>*>(SData_d->Eq_m); //static data 269 auto d_IPf_k = static_cast<Kokkos::View<PetscScalar*, Kokkos::LayoutLeft>*>(SData_d->IPf); //static data 270 271 PetscFunctionBegin; 272 ierr = PetscLogEventBegin(events[3],0,0,0,0);CHKERRQ(ierr); 273 ierr = DMGetApplicationContext(plex, &ctx);CHKERRQ(ierr); 274 if (!ctx) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context"); 275 ierr = DMGetDimension(plex, &dim);CHKERRQ(ierr); 276 ierr = DMPlexGetHeightStratum(plex,0,&cStart,&cEnd);CHKERRQ(ierr); 277 numCells = cEnd - cStart; 278 nip = numCells*Nq; 279 ierr = DMGetDS(plex, &prob);CHKERRQ(ierr); 280 ierr = PetscDSGetNumFields(prob, &Nf);CHKERRQ(ierr); 281 ierr = PetscDSGetDimensions(prob, &Nbf);CHKERRQ(ierr); Nb = Nbf[0]; 282 if (Nq != Nb) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Nq != Nb. %D %D",Nq,Nb); 283 if (LANDAU_DIM != dim) SETERRQ2(PETSC_COMM_WORLD, PETSC_ERR_PLIB, "dim %D != LANDAU_DIM %d",dim,LANDAU_DIM); 284 scr_bytes = 2*(g2_scr_t::shmem_size(dim,Nf,Nq) + g3_scr_t::shmem_size(dim,dim,Nf,Nq)); 285 ierr = PetscDSGetTotalDimension(prob, &totDim);CHKERRQ(ierr); 286 if (ctx->gpu_assembly) { 287 PetscContainer container; 288 ierr = PetscObjectQuery((PetscObject) JacP, "assembly_maps", (PetscObject *) &container);CHKERRQ(ierr); 289 if (container) { // not here first call 290 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 291 P4estVertexMaps *h_maps=NULL; 292 ierr = PetscContainerGetPointer(container, (void **) &h_maps);CHKERRQ(ierr); 293 if (h_maps->data) { 294 d_maps = h_maps->data; 295 } else { 296 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly but no metadata in container"); 297 } 298 // this does the setup the first time called 299 ierr = MatKokkosGetDeviceMatWrite(JacP,&d_mat);CHKERRQ(ierr); 300 global_elem_mat_sz = 0; 301 #else 302 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "GPU assembly w/o kokkos kernels -- should not be here"); 303 #endif 304 } else { // kernel output - first call assembled on device 305 global_elem_mat_sz = numCells; 306 } 307 } else { 308 global_elem_mat_sz = numCells; // no device assembly 309 } 310 ierr = PetscKokkosInitializeCheck();CHKERRQ(ierr); 311 ierr = PetscLogEventEnd(events[3],0,0,0,0);CHKERRQ(ierr); 312 { 313 #if defined(LANDAU_LAYOUT_LEFT) 314 Kokkos::View<PetscReal***, Kokkos::LayoutLeft > d_fdf_k("df", dim+1, Nf, (a_IPf || a_xarray) ? nip : 0); 315 #else 316 Kokkos::View<PetscReal***, Kokkos::LayoutRight > d_fdf_k("df", dim+1, Nf, (a_IPf || a_xarray) ? nip : 0); 317 #endif 318 const Kokkos::View<PetscScalar*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> >h_IPf_k (a_IPf, (a_IPf || a_xarray) ? nip*Nf : 0); 319 const Kokkos::View<PetscReal*, Kokkos::LayoutLeft, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> > h_Eq_m_k (a_Eq_m, (a_IPf || a_xarray) ? Nf : 0); 320 Kokkos::View<PetscScalar**, Kokkos::LayoutRight> d_elem_mats("element matrices", global_elem_mat_sz, totDim*totDim); 321 // copy dynamic data to device 322 if (a_IPf || a_xarray) { // form f and df 323 static int cc=0; 324 ierr = PetscLogEventBegin(events[1],0,0,0,0);CHKERRQ(ierr); 325 Kokkos::deep_copy (*d_Eq_m_k, h_Eq_m_k); 326 d_Eq_m = d_Eq_m_k->data(); 327 if (a_IPf) { 328 Kokkos::deep_copy (*d_IPf_k, h_IPf_k); 329 d_IPf = d_IPf_k->data(); 330 } else { 331 d_IPf = (PetscScalar*)a_xarray; 332 } 333 ierr = PetscLogEventEnd(events[1],0,0,0,0);CHKERRQ(ierr); 334 #define KOKKOS_SHARED_LEVEL 1 335 if (cc++ == 0) { 336 ierr = PetscInfo4(plex, "shared memory size: %d bytes in level %d conc=%D team size=%D\n",scr_bytes,KOKKOS_SHARED_LEVEL,conc,team_size);CHKERRQ(ierr); 337 } 338 // get f and df 339 ierr = PetscLogEventBegin(events[8],0,0,0,0);CHKERRQ(ierr); 340 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 341 Kokkos::parallel_for("f, df", Kokkos::TeamPolicy<>(numCells, team_size, /* Kokkos::AUTO */ 16), KOKKOS_LAMBDA (const team_member team) { 342 const PetscInt myelem = team.league_rank(); 343 const PetscScalar *coef; 344 PetscScalar coef_buff[LANDAU_MAX_SPECIES*LANDAU_MAX_NQ]; 345 // un pack IPData 346 if (!d_maps) { 347 coef = &d_IPf[myelem*Nb*Nf]; 348 } else { 349 coef = coef_buff; 350 for (int f = 0; f < Nf; ++f) { 351 LandauIdx *const Idxs = &d_maps->gIdx[myelem][f][0]; 352 for (int b = 0; b < Nb; ++b) { 353 PetscInt idx = Idxs[b]; 354 if (idx >= 0) { 355 coef_buff[f*Nb+b] = d_IPf[idx]; 356 } else { 357 idx = -idx - 1; 358 coef_buff[f*Nb+b] = 0; 359 for (int q = 0; q < d_maps->num_face; q++) { 360 PetscInt id = d_maps->c_maps[idx][q].gid; 361 PetscScalar scale = d_maps->c_maps[idx][q].scale; 362 coef_buff[f*Nb+b] += scale*d_IPf[id]; 363 } 364 } 365 //printf("%f ",coef[f*Nb+b]); 366 } 367 //printf("\n"); 368 } 369 //printf("\n"); 370 } 371 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 372 const PetscInt ipidx = myQi + myelem * Nq; 373 const PetscReal *const invJj = &d_invJ_k(ipidx*dim*dim); 374 const PetscReal *Bq = &d_BB[myQi*Nb], *Dq = &d_DD[myQi*Nb*dim]; 375 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nf), [=] (int f) { 376 PetscInt b, e, d; 377 PetscReal refSpaceDer[LANDAU_DIM]; 378 d_fdf_k(0,f,ipidx) = 0.0; 379 for (d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0; 380 for (b = 0; b < Nb; ++b) { 381 const PetscInt cidx = b; 382 d_fdf_k(0,f,ipidx) += Bq[cidx]*PetscRealPart(coef[f*Nb+cidx]); 383 for (d = 0; d < dim; ++d) refSpaceDer[d] += Dq[cidx*dim+d]*PetscRealPart(coef[f*Nb+cidx]); 384 } 385 for (d = 0; d < dim; ++d) { 386 for (e = 0, d_fdf_k(d+1,f,ipidx) = 0.0; e < dim; ++e) { 387 d_fdf_k(d+1,f,ipidx) += invJj[e*dim+d]*refSpaceDer[e]; 388 } 389 } 390 }); // Nf 391 }); // Nq 392 }); // elems 393 Kokkos::fence(); 394 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) 395 ierr = PetscLogGpuFlops(nip*(PetscLogDouble)(2*Nb*(1+dim)));CHKERRQ(ierr); 396 #else 397 ierr = PetscLogFlops(nip*(PetscLogDouble)(2*Nb*(1+dim)));CHKERRQ(ierr); 398 #endif 399 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); 400 ierr = PetscLogEventEnd(events[8],0,0,0,0);CHKERRQ(ierr); 401 } 402 ierr = PetscLogEventBegin(events[4],0,0,0,0);CHKERRQ(ierr); 403 ierr = PetscLogGpuTimeBegin();CHKERRQ(ierr); 404 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) 405 ierr = PetscLogGpuFlops(nip*(PetscLogDouble)(!(a_IPf || a_xarray) ? (nip*(11*Nf+ 4*dim*dim) + 6*Nf*dim*dim*dim + 10*Nf*dim*dim + 4*Nf*dim + Nb*Nf*Nb*Nq*dim*dim*5) : Nb*Nf*Nb*Nq*4));CHKERRQ(ierr); 406 if (ctx->deviceType == LANDAU_CPU) PetscInfo(plex, "Warning: Landau selected CPU but no support for Kokkos using CPU\n"); 407 #else 408 ierr = PetscLogFlops(nip*(PetscLogDouble)(!(a_IPf || a_xarray) ? (nip*(11*Nf+ 4*dim*dim) + 6*Nf*dim*dim*dim + 10*Nf*dim*dim + 4*Nf*dim + Nb*Nf*Nb*Nq*dim*dim*5) : Nb*Nf*Nb*Nq*4));CHKERRQ(ierr); 409 #endif 410 Kokkos::parallel_for("Landau elements", Kokkos::TeamPolicy<>(numCells, team_size, /*Kokkos::AUTO*/ 16).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes)), KOKKOS_LAMBDA (const team_member team) { 411 const PetscInt myelem = team.league_rank(); 412 g2_scr_t g2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,Nf,Nq); // we don't use these for mass matrix 413 g3_scr_t g3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,Nf,Nq); 414 if (d_IPf) { 415 g2_scr_t gg2(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,Nf,Nq); 416 g3_scr_t gg3(team.team_scratch(KOKKOS_SHARED_LEVEL),dim,dim,Nf,Nq); 417 // get g2[] & g3[] 418 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nq), [=] (int myQi) { 419 using Kokkos::parallel_reduce; 420 const PetscInt jpidx = myQi + myelem * Nq; 421 const PetscReal* const invJj = &d_invJ_k(jpidx*dim*dim); 422 const PetscReal vj[3] = {d_x[jpidx], d_y[jpidx], d_z ? d_z[jpidx] : 0}, wj = d_w[jpidx]; 423 landau_inner_red::TensorValueType gg_temp; // reduce on part of gg2 and g33 for IP jpidx 424 Kokkos::parallel_reduce(Kokkos::ThreadVectorRange (team, (int)nip), [=] (const int& ipidx, landau_inner_red::TensorValueType & ggg) { 425 const PetscReal wi = d_w[ipidx], x = d_x[ipidx], y = d_y[ipidx]; 426 PetscReal temp1[3] = {0, 0, 0}, temp2 = 0; 427 PetscInt fieldA,d2,d3; 428 #if LANDAU_DIM==2 429 PetscReal Ud[2][2], Uk[2][2]; 430 LandauTensor2D(vj, x, y, Ud, Uk, (ipidx==jpidx) ? 0. : 1.); 431 #else 432 PetscReal U[3][3], z = d_z[ipidx]; 433 LandauTensor3D(vj, x, y, z, U, (ipidx==jpidx) ? 0. : 1.); 434 #endif 435 for (fieldA = 0; fieldA < Nf; ++fieldA) { 436 temp1[0] += d_fdf_k(1,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA]; 437 temp1[1] += d_fdf_k(2,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA]; 438 #if LANDAU_DIM==3 439 temp1[2] += d_fdf_k(3,fieldA,ipidx)*d_beta[fieldA]*d_invMass[fieldA]; 440 #endif 441 temp2 += d_fdf_k(0,fieldA,ipidx)*d_beta[fieldA]; 442 } 443 temp1[0] *= wi; 444 temp1[1] *= wi; 445 #if LANDAU_DIM==3 446 temp1[2] *= wi; 447 #endif 448 temp2 *= wi; 449 #if LANDAU_DIM==2 450 for (d2 = 0; d2 < 2; d2++) { 451 for (d3 = 0; d3 < 2; ++d3) { 452 /* K = U * grad(f): g2=e: i,A */ 453 ggg.gg2[d2] += Uk[d2][d3]*temp1[d3]; 454 /* D = -U * (I \kron (fx)): g3=f: i,j,A */ 455 ggg.gg3[d2][d3] += Ud[d2][d3]*temp2; 456 } 457 } 458 #else 459 for (d2 = 0; d2 < 3; ++d2) { 460 for (d3 = 0; d3 < 3; ++d3) { 461 /* K = U * grad(f): g2 = e: i,A */ 462 ggg.gg2[d2] += U[d2][d3]*temp1[d3]; 463 /* D = -U * (I \kron (fx)): g3 = f: i,j,A */ 464 ggg.gg3[d2][d3] += U[d2][d3]*temp2; 465 } 466 } 467 #endif 468 }, Kokkos::Sum<landau_inner_red::TensorValueType>(gg_temp)); 469 // 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]); 470 // add alpha and put in gg2/3 471 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [&] (const int& fieldA) { 472 PetscInt d2,d3; 473 for (d2 = 0; d2 < dim; d2++) { 474 gg2(d2,fieldA,myQi) = gg_temp.gg2[d2]*d_alpha[fieldA]; 475 //if (myelem==0 && fieldA==1) printf("\t\t:%d.%d) gg2[%d]=%e = %e %e\n",myelem,myQi,d2,gg2(d2,fieldA,myQi),gg_temp.gg2[d2],d_alpha[fieldA]); 476 //gg2[d2][myQi][fieldA] += gg_temp.gg2[d2]*d_alpha[fieldA]; 477 for (d3 = 0; d3 < dim; d3++) { 478 //gg3[d2][d3][myQi][fieldA] -= gg_temp.gg3[d2][d3]*d_alpha[fieldA]*s_invMass[fieldA]; 479 gg3(d2,d3,fieldA,myQi) = -gg_temp.gg3[d2][d3]*d_alpha[fieldA]*d_invMass[fieldA]; 480 //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)); 481 } 482 } 483 }); 484 /* add electric field term once per IP */ 485 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [&] (const int& fieldA) { 486 //gg.gg2[fieldA][dim-1] += d_Eq_m[fieldA]; 487 gg2(dim-1,fieldA,myQi) += d_Eq_m[fieldA]; 488 }); 489 // Kokkos::single(Kokkos::PerThread(team), [&]() { 490 Kokkos::parallel_for(Kokkos::ThreadVectorRange (team, (int)Nf), [=] (const int& fieldA) { 491 int d,d2,d3,dp; 492 //printf("%d %d %d gg2[][1]=%18.10e\n",myelem,myQi,fieldA,gg.gg2[fieldA][dim-1]); 493 /* Jacobian transform - g2, g3 - per thread (2D) */ 494 for (d = 0; d < dim; ++d) { 495 g2(d,fieldA,myQi) = 0; 496 for (d2 = 0; d2 < dim; ++d2) { 497 g2(d,fieldA,myQi) += invJj[d*dim+d2]*gg2(d2,fieldA,myQi); 498 //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]); 499 g3(d,d2,fieldA,myQi) = 0; 500 for (d3 = 0; d3 < dim; ++d3) { 501 for (dp = 0; dp < dim; ++dp) { 502 g3(d,d2,fieldA,myQi) += invJj[d*dim + d3]*gg3(d3,dp,fieldA,myQi)*invJj[d2*dim + dp]; 503 //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]); 504 } 505 } 506 g3(d,d2,fieldA,myQi) *= wj; 507 } 508 g2(d,fieldA,myQi) *= wj; 509 } 510 }); 511 }); // Nq 512 team.team_barrier(); 513 } // Jacobian 514 /* assemble */ 515 //Kokkos::single(Kokkos::PerTeam(team), [&]() { 516 Kokkos::parallel_for(Kokkos::TeamThreadRange(team,0,Nb), [=] (int blk_i) { 517 Kokkos::parallel_for(Kokkos::ThreadVectorRange(team,0,(int)Nf), [=] (int fieldA) { 518 int blk_j,qj,d,d2; 519 const PetscInt i = fieldA*Nb + blk_i; /* Element matrix row */ 520 for (blk_j = 0; blk_j < Nb; ++blk_j) { 521 const PetscInt j = fieldA*Nb + blk_j; /* Element matrix column */ 522 const PetscInt fOff = i*totDim + j; 523 PetscScalar t = global_elem_mat_sz ? d_elem_mats(myelem,fOff) : 0; 524 for (qj = 0 ; qj < Nq ; qj++) { // look at others integration points 525 const PetscReal *BJq = &d_BB[qj*Nb], *DIq = &d_DD[qj*Nb*dim]; 526 if (d_IPf) { 527 for (d = 0; d < dim; ++d) { 528 t += DIq[blk_i*dim+d]*g2(d,fieldA,qj)*BJq[blk_j]; 529 for (d2 = 0; d2 < dim; ++d2) { 530 t += DIq[blk_i*dim + d]*g3(d,d2,fieldA,qj)*DIq[blk_j*dim + d2]; 531 } 532 } 533 } else { 534 const PetscInt jpidx = qj + myelem * Nq; 535 t += BJq[blk_i] * d_mass_w_k(jpidx) * shift * BJq[blk_j]; 536 } 537 } 538 if (global_elem_mat_sz) d_elem_mats(myelem,fOff) = t; // can set this because local element matrix[fOff] 539 else { 540 PetscErrorCode ierr = 0; 541 PetscScalar vals[LANDAU_MAX_Q_FACE*LANDAU_MAX_Q_FACE],row_scale[LANDAU_MAX_Q_FACE],col_scale[LANDAU_MAX_Q_FACE]; 542 PetscInt q,idx,nr,nc,rows0[LANDAU_MAX_Q_FACE],cols0[LANDAU_MAX_Q_FACE],rows[LANDAU_MAX_Q_FACE],cols[LANDAU_MAX_Q_FACE]; 543 const LandauIdx *const Idxs = &d_maps->gIdx[myelem][fieldA][0]; 544 idx = Idxs[blk_i]; 545 if (idx >= 0) { 546 nr = 1; 547 rows0[0] = idx; 548 row_scale[0] = 1.; 549 } else { 550 idx = -idx - 1; 551 nr = d_maps->num_face; 552 for (q = 0; q < d_maps->num_face; q++) { 553 rows0[q] = d_maps->c_maps[idx][q].gid; 554 row_scale[q] = d_maps->c_maps[idx][q].scale; 555 } 556 } 557 idx = Idxs[blk_j]; 558 if (idx >= 0) { 559 nc = 1; 560 cols0[0] = idx; 561 col_scale[0] = 1.; 562 } else { 563 idx = -idx - 1; 564 nc = d_maps->num_face; 565 for (q = 0; q < d_maps->num_face; q++) { 566 cols0[q] = d_maps->c_maps[idx][q].gid; 567 col_scale[q] = d_maps->c_maps[idx][q].scale; 568 } 569 } 570 for (q = 0; q < nr; q++) rows[q] = rows0[q]; 571 for (q = 0; q < nc; q++) cols[q] = cols0[q]; 572 for (q = 0; q < nr; q++) { 573 for (d = 0; d < nc; d++) { 574 vals[q*nc + d] = row_scale[q]*col_scale[d]*t; 575 } 576 } 577 MatSetValuesDevice(d_mat,nr,rows,nc,cols,vals,ADD_VALUES,&ierr); 578 if (ierr) return; 579 } 580 } 581 }); 582 }); 583 }); 584 ierr = PetscLogGpuTimeEnd();CHKERRQ(ierr); 585 Kokkos::fence(); 586 ierr = PetscLogEventEnd(events[4],0,0,0,0);CHKERRQ(ierr); 587 if (global_elem_mat_sz) { 588 PetscSection section, globalSection; 589 Kokkos::View<PetscScalar**, Kokkos::LayoutRight>::HostMirror h_elem_mats = Kokkos::create_mirror_view(d_elem_mats); 590 ierr = PetscLogEventBegin(events[5],0,0,0,0);CHKERRQ(ierr); 591 ierr = DMGetLocalSection(plex, §ion);CHKERRQ(ierr); 592 ierr = DMGetGlobalSection(plex, &globalSection);CHKERRQ(ierr); 593 Kokkos::deep_copy (h_elem_mats, d_elem_mats); 594 ierr = PetscLogEventEnd(events[5],0,0,0,0);CHKERRQ(ierr); 595 ierr = PetscLogEventBegin(events[6],0,0,0,0);CHKERRQ(ierr); 596 PetscInt ej; 597 for (ej = cStart ; ej < cEnd; ++ej) { 598 const PetscScalar *elMat = &h_elem_mats(ej-cStart,0); 599 ierr = DMPlexMatSetClosure(plex, section, globalSection, JacP, ej, elMat, ADD_VALUES);CHKERRQ(ierr); 600 if (ej==-1) { 601 int d,f; 602 PetscPrintf(PETSC_COMM_SELF,"Kokkos Element matrix %d/%d\n",1,(int)numCells); 603 for (d = 0; d < totDim; ++d){ 604 for (f = 0; f < totDim; ++f) PetscPrintf(PETSC_COMM_SELF," %12.5e", PetscRealPart(elMat[d*totDim + f])); 605 PetscPrintf(PETSC_COMM_SELF,"\n"); 606 } 607 exit(14); 608 } 609 } 610 ierr = PetscLogEventEnd(events[6],0,0,0,0);CHKERRQ(ierr); 611 // transition to use of maps for VecGetClosure 612 { 613 auto IPf = static_cast<Kokkos::View<PetscScalar*, Kokkos::LayoutLeft>*>(SData_d->IPf); 614 delete IPf; 615 SData_d->IPf = NULL; 616 } 617 if (!(a_IPf || a_xarray)) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "transition without Jacobian"); 618 } 619 } 620 PetscFunctionReturn(0); 621 } 622 } // extern "C" 623