xref: /petsc/src/ksp/pc/impls/bjacobi/bjkokkos/bjkokkoskernels.kokkos.cxx (revision d0e6bf2ad94dcc89b258ce16c7987200a4714786)
1 #include <petsc/private/pcbjkokkosimpl.h>
2 
3 #if defined(PETSC_HAVE_KOKKOS_KERNELS_BATCH)
4   #include <fstream>
5 
6   #include "Kokkos_Timer.hpp"
7   #include "Kokkos_Random.hpp"
8   #include "Kokkos_UnorderedMap.hpp"
9   #include "Kokkos_Sort.hpp"
10 
11   /// KokkosKernels headers
12   #include "KokkosBatched_Util.hpp"
13   #include "KokkosBatched_Vector.hpp"
14 
15   #include <Kokkos_ArithTraits.hpp>
16   #include <KokkosBatched_Util.hpp>
17   #include <KokkosBatched_Vector.hpp>
18   #include <KokkosBatched_Copy_Decl.hpp>
19   #include <KokkosBatched_Copy_Impl.hpp>
20   #include <KokkosBatched_AddRadial_Decl.hpp>
21   #include <KokkosBatched_AddRadial_Impl.hpp>
22   #include <KokkosBatched_Gemm_Decl.hpp>
23   #include <KokkosBatched_Gemm_Serial_Impl.hpp>
24   #include <KokkosBatched_Gemm_Team_Impl.hpp>
25   #include <KokkosBatched_Gemv_Decl.hpp>
26   // #include <KokkosBatched_Gemv_Serial_Impl.hpp>
27   #include <KokkosBatched_Gemv_Team_Impl.hpp>
28   #include <KokkosBatched_Trsm_Decl.hpp>
29   #include <KokkosBatched_Trsm_Serial_Impl.hpp>
30   #include <KokkosBatched_Trsm_Team_Impl.hpp>
31   #include <KokkosBatched_Trsv_Decl.hpp>
32   #include <KokkosBatched_Trsv_Serial_Impl.hpp>
33   #include <KokkosBatched_Trsv_Team_Impl.hpp>
34   #include <KokkosBatched_LU_Decl.hpp>
35   #include <KokkosBatched_LU_Serial_Impl.hpp>
36   #include <KokkosBatched_LU_Team_Impl.hpp>
37   #include <KokkosSparse_CrsMatrix.hpp>
38   #include "KokkosBatched_Spmv.hpp"
39   #include "KokkosBatched_CrsMatrix.hpp"
40   #include "KokkosBatched_Krylov_Handle.hpp"
41 
42   #include "KokkosBatched_GMRES.hpp"
43   #include "KokkosBatched_JacobiPrec.hpp"
44 
45 template <typename DeviceType, typename ValuesViewType, typename IntView, typename VectorViewType, typename KrylovHandleType>
46 struct Functor_TestBatchedTeamVectorGMRES {
47   const ValuesViewType _D;
48   const ValuesViewType _diag;
49   const IntView        _r;
50   const IntView        _c;
51   const VectorViewType _X;
52   const VectorViewType _B;
53   const int            _N_team, _team_size, _vector_length;
54   const int            _N_iteration;
55   const double         _tol;
56   const int            _ortho_strategy;
57   const int            _scratch_pad_level;
58   KrylovHandleType     _handle;
59 
60   KOKKOS_INLINE_FUNCTION
Functor_TestBatchedTeamVectorGMRESFunctor_TestBatchedTeamVectorGMRES61   Functor_TestBatchedTeamVectorGMRES(const ValuesViewType &D, const IntView &r, const IntView &c, const VectorViewType &X, const VectorViewType &B, const int N_team, const int team_size, const int vector_length, const int N_iteration, const double tol, const int ortho_strategy, const int scratch_pad_level, KrylovHandleType &handle) :
62     _D(D), _r(r), _c(c), _X(X), _B(B), _N_team(N_team), _team_size(team_size), _vector_length(vector_length), _N_iteration(N_iteration), _tol(tol), _ortho_strategy(ortho_strategy), _scratch_pad_level(scratch_pad_level), _handle(handle)
63   {
64   }
65 
66   KOKKOS_INLINE_FUNCTION
Functor_TestBatchedTeamVectorGMRESFunctor_TestBatchedTeamVectorGMRES67   Functor_TestBatchedTeamVectorGMRES(const ValuesViewType &D, const ValuesViewType &diag, const IntView &r, const IntView &c, const VectorViewType &X, const VectorViewType &B, const int N_team, const int team_size, const int vector_length, const int N_iteration, const double tol, int ortho_strategy, const int scratch_pad_level, KrylovHandleType &handle) :
68     _D(D), _diag(diag), _r(r), _c(c), _X(X), _B(B), _N_team(N_team), _team_size(team_size), _vector_length(vector_length), _N_iteration(N_iteration), _tol(tol), _ortho_strategy(ortho_strategy), _scratch_pad_level(scratch_pad_level), _handle(handle)
69   {
70   }
71 
72   template <typename MemberType>
operator ()Functor_TestBatchedTeamVectorGMRES73   KOKKOS_INLINE_FUNCTION void operator()(const MemberType &member) const
74   {
75     const int first_matrix = static_cast<int>(member.league_rank()) * _N_team;
76     const int N            = _D.extent(0);
77     const int last_matrix  = (static_cast<int>(member.league_rank() + 1) * _N_team < N ? static_cast<int>(member.league_rank() + 1) * _N_team : N);
78     const int graphID      = static_cast<int>(member.league_rank());
79     using TeamVectorCopy1D = KokkosBatched::TeamVectorCopy<MemberType, KokkosBatched::Trans::NoTranspose, 1>;
80 
81     auto d                         = Kokkos::subview(_D, Kokkos::make_pair(first_matrix, last_matrix), Kokkos::ALL);
82     auto x                         = Kokkos::subview(_X, Kokkos::make_pair(first_matrix, last_matrix), Kokkos::ALL);
83     auto b                         = Kokkos::subview(_B, Kokkos::make_pair(first_matrix, last_matrix), Kokkos::ALL);
84     using ScratchPadIntViewType    = Kokkos::View<typename IntView::non_const_value_type *, typename IntView::array_layout, typename IntView::execution_space::scratch_memory_space>;
85     using ScratchPadValuesViewType = Kokkos::View<typename ValuesViewType::non_const_value_type **, typename ValuesViewType::array_layout, typename ValuesViewType::execution_space::scratch_memory_space>;
86 
87     using Operator = KokkosBatched::CrsMatrix<ValuesViewType, ScratchPadIntViewType>;
88     ScratchPadIntViewType r(member.team_scratch(1), _r.extent(1));
89     ScratchPadIntViewType c(member.team_scratch(1), _c.extent(1));
90 
91     TeamVectorCopy1D::invoke(member, Kokkos::subview(_r, graphID, Kokkos::ALL), r);
92     TeamVectorCopy1D::invoke(member, Kokkos::subview(_c, graphID, Kokkos::ALL), c);
93     Operator A(d, r, c);
94 
95     ScratchPadValuesViewType diag(member.team_scratch(1), last_matrix - first_matrix, _diag.extent(1));
96     using PrecOperator = KokkosBatched::JacobiPrec<ScratchPadValuesViewType>;
97 
98     KokkosBatched::TeamVectorCopy<MemberType>::invoke(member, Kokkos::subview(_diag, Kokkos::make_pair(first_matrix, last_matrix), Kokkos::ALL), diag);
99     PrecOperator P(diag);
100     P.setComputedInverse();
101 
102     KokkosBatched::TeamVectorGMRES<MemberType>::template invoke<Operator, VectorViewType, PrecOperator, KrylovHandleType>(member, A, b, x, P, _handle);
103   }
runFunctor_TestBatchedTeamVectorGMRES104   inline double run(PC pc)
105   {
106     //typedef typename ValuesViewType::value_type value_type;
107     std::string   name("KokkosBatched::Test::TeamVectorGMRES");
108     Kokkos::Timer timer;
109     Kokkos::Profiling::pushRegion(name.c_str());
110 
111     Kokkos::TeamPolicy<DeviceType> auto_policy(ceil(1. * _D.extent(0) / _N_team), Kokkos::AUTO(), Kokkos::AUTO());
112     Kokkos::TeamPolicy<DeviceType> tuned_policy(ceil(1. * _D.extent(0) / _N_team), _team_size, _vector_length);
113     Kokkos::TeamPolicy<DeviceType> policy;
114 
115     if (_team_size < 1) policy = auto_policy;
116     else policy = tuned_policy;
117 
118     _handle.set_max_iteration(_N_iteration);
119     _handle.set_tolerance(_tol);
120     _handle.set_ortho_strategy(_ortho_strategy);
121     _handle.set_scratch_pad_level(_scratch_pad_level);
122     _handle.set_compute_last_residual(true);
123 
124     int maximum_iteration = _handle.get_max_iteration();
125 
126     using ScalarType = typename ValuesViewType::non_const_value_type;
127     using Layout     = typename ValuesViewType::array_layout;
128     using EXSP       = typename ValuesViewType::execution_space;
129 
130     using ViewType2D    = Kokkos::View<ScalarType **, Layout, EXSP>;
131     using IntViewType1D = Kokkos::View<PetscInt *, Layout, EXSP>;
132 
133     size_t bytes_1D      = ViewType2D::shmem_size(_N_team, 1);
134     size_t bytes_row_ptr = IntViewType1D::shmem_size(_r.extent(1));
135     size_t bytes_col_idc = IntViewType1D::shmem_size(_c.extent(1));
136     size_t bytes_2D_1    = ViewType2D::shmem_size(_N_team, _X.extent(1));
137     size_t bytes_2D_2    = ViewType2D::shmem_size(_N_team, maximum_iteration + 1);
138 
139     size_t bytes_diag = bytes_2D_1;
140     size_t bytes_tmp  = 2 * bytes_2D_1 + 2 * bytes_1D + bytes_2D_2;
141 
142     policy.set_scratch_size(0, Kokkos::PerTeam(bytes_tmp));
143     policy.set_scratch_size(1, Kokkos::PerTeam(bytes_col_idc + bytes_row_ptr + bytes_diag));
144     PetscCall(PetscInfo(pc, "%d scratch memory(0) = %d + %d + %d bytes_diag=%d; %d scratch memory(1); %d maximum_iterations\n", (int)bytes_tmp, 2 * (int)bytes_2D_1, 2 * (int)bytes_1D, (int)bytes_2D_2, (int)bytes_diag, (int)(bytes_row_ptr + bytes_col_idc + bytes_diag), (int)maximum_iteration));
145     exec_space().fence();
146     timer.reset();
147     Kokkos::parallel_for(name.c_str(), policy, *this);
148     exec_space().fence();
149     double sec = timer.seconds();
150 
151     return sec;
152   }
153 };
154 
PCApply_BJKOKKOSKERNELS(PC pc,const PetscScalar * glb_bdata,PetscScalar * glb_xdata,const PetscInt * glb_Aai,const PetscInt * glb_Aaj,const PetscScalar * glb_Aaa,const PetscInt team_size,MatInfo info,const PetscInt batch_sz,PCFailedReason * pcreason)155 PETSC_INTERN PetscErrorCode PCApply_BJKOKKOSKERNELS(PC pc, const PetscScalar *glb_bdata, PetscScalar *glb_xdata, const PetscInt *glb_Aai, const PetscInt *glb_Aaj, const PetscScalar *glb_Aaa, const PetscInt team_size, MatInfo info, const PetscInt batch_sz, PCFailedReason *pcreason)
156 {
157   PC_PCBJKOKKOS     *jac   = (PC_PCBJKOKKOS *)pc->data;
158   Mat                A     = pc->pmat;
159   const PetscInt     maxit = jac->ksp->max_it, nBlk = jac->nBlocks;
160   const int          Nsolves      = nBlk;
161   int                Nsolves_team = jac->nsolves_team, fill_idx = 0;
162   int                Nloc           = jac->const_block_size;       // same grids
163   const int          nnz            = (int)info.nz_used / Nsolves; // fix for variable grid size
164   PetscReal          rtol           = jac->ksp->rtol;
165   const PetscScalar *glb_idiag      = jac->d_idiag_k->data();
166   const PetscInt    *d_bid_eqOffset = jac->d_bid_eqOffset_k->data();
167   const PetscInt    *d_isicol = jac->d_isicol_k->data(), *d_isrow = jac->d_isrow_k->data();
168 
169   PetscFunctionBegin;
170   if (Nsolves_team > batch_sz) Nsolves_team = batch_sz; // silently fix this
171   PetscCheck(jac->const_block_size, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "Kokkos (GMRES) solver requires constant block size (but can be made to work with species ordering or N_team==1)");
172   PetscCheck(Nsolves % Nsolves_team == 0, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "Nsolves.mod(Nsolves_team) != 0: Nsolves = %d, Nsolves_team = %d", Nsolves, Nsolves_team);
173   PetscCheck(((int)info.nz_used) % Nsolves == 0, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "info.nz_used.mod(Nsolves) != 0: info.nz_used = %g, Nsolves = %d", info.nz_used, Nsolves);
174   #if defined(PETSC_HAVE_CUDA)
175   nvtxRangePushA("gmres-kk");
176   #endif
177   Kokkos::View<PetscScalar **, layout, exec_space, Kokkos::MemoryTraits<Kokkos::Unmanaged>> inv_diag((PetscScalar *)glb_idiag, Nsolves, Nloc); // in correct order
178   if (!jac->rowOffsets) {
179     jac->rowOffsets   = new IntView("rowOffsets", Nsolves / Nsolves_team, Nloc + 1); // same grids
180     jac->colIndices   = new IntView("colIndices", Nsolves / Nsolves_team, nnz);
181     jac->batch_b      = new XYType("batch rhs", Nsolves, Nloc);
182     jac->batch_x      = new XYType("batch sol", Nsolves, Nloc);
183     jac->batch_values = new AMatrixValueView("batch values", Nsolves, nnz);
184     fill_idx          = 1;
185     PetscCall(PetscInfo(pc, "Setup indices Nloc=%d, nnz=%d\n", Nloc, nnz));
186   }
187   IntView          &rowOffsets   = *jac->rowOffsets;
188   IntView          &colIndices   = *jac->colIndices;
189   XYType           &batch_b      = *jac->batch_b;
190   XYType           &batch_x      = *jac->batch_x;
191   AMatrixValueView &batch_values = *jac->batch_values;
192 
193   Kokkos::deep_copy(batch_x, 0.);
194   PetscCall(PetscInfo(pc, "\tjac->n = %d, Nloc = %d, Nsolves = %d, nnz = %d, Nsolves_team = %d, league size = %d, maxit = %d\n", (int)jac->n, Nloc, Nsolves, nnz, Nsolves_team, Nsolves / Nsolves_team, (int)maxit));
195   Kokkos::parallel_for(
196     "rowOffsets+map", Kokkos::TeamPolicy<>(Nsolves, team_size, PCBJKOKKOS_VEC_SIZE), KOKKOS_LAMBDA(const team_member team) {
197       const int blkID = team.league_rank(), start = d_bid_eqOffset[blkID], end = d_bid_eqOffset[blkID + 1];
198       if (fill_idx) {
199         if (blkID % Nsolves_team == 0) {                                                        // first matrix on this member
200           Kokkos::parallel_for(Kokkos::TeamVectorRange(team, start, end), [=](const int rowb) { // Nloc
201             int rowa                                           = d_isicol[rowb];
202             int n                                              = glb_Aai[rowa + 1] - glb_Aai[rowa];
203             rowOffsets(blkID / Nsolves_team, rowb + 1 - start) = n; // save sizes
204           });
205         }
206       }
207       // map b into field major space
208       Kokkos::parallel_for(Kokkos::TeamVectorRange(team, start, end), [=](int rowb) {
209         int rowa                     = d_isicol[rowb];
210         batch_b(blkID, rowb - start) = glb_bdata[rowa];
211       });
212     });
213   Kokkos::fence();
214   if (fill_idx) {
215     Kokkos::parallel_for(
216       "prefix sum", Kokkos::TeamPolicy<>(Nsolves / Nsolves_team, 1, 1), KOKKOS_LAMBDA(const team_member team) {
217         const int graphID      = team.league_rank();
218         rowOffsets(graphID, 0) = 0;
219         for (int i = 0; i < Nloc; ++i) rowOffsets(graphID, i + 1) += rowOffsets(graphID, i);
220       });
221     Kokkos::fence();
222   }
223   Kokkos::parallel_for(
224     "copy matrix", Kokkos::TeamPolicy<>(Nsolves /* /batch_sz */, team_size, PCBJKOKKOS_VEC_SIZE), KOKKOS_LAMBDA(const team_member team) {
225       const int blkID = team.league_rank(), start = d_bid_eqOffset[blkID], end = d_bid_eqOffset[blkID + 1], graphID = blkID / Nsolves_team;
226       Kokkos::parallel_for(Kokkos::TeamThreadRange(team, start, end), [=](const int rowb) {
227         int                rowa = d_isicol[rowb];
228         int                n    = glb_Aai[rowa + 1] - glb_Aai[rowa];
229         const PetscInt    *aj   = glb_Aaj + glb_Aai[rowa]; // global index
230         const PetscScalar *aa   = glb_Aaa + glb_Aai[rowa];
231         Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, n), [=](const int &i) {
232           PetscScalar val = aa[i];
233           if (fill_idx && blkID % Nsolves_team == 0) colIndices(graphID, rowOffsets(graphID, rowb - start) + i) = d_isrow[aj[i]] - blkID * Nloc; // local" global - block start
234           batch_values(blkID, rowOffsets(graphID, rowb - start) + i) = val;
235         });
236       });
237     });
238   Kokkos::fence();
239   // setup solver
240   using ScalarType    = typename AMatrixValueView::non_const_value_type;
241   using MagnitudeType = typename Kokkos::Details::ArithTraits<ScalarType>::mag_type;
242   //using NormViewType              = Kokkos::View<MagnitudeType *, layout, exec_space>;
243   using Norm2DViewType   = Kokkos::View<MagnitudeType **, layout, exec_space>;
244   using Scalar3DViewType = Kokkos::View<ScalarType ***, layout, exec_space>;
245   using IntViewType      = Kokkos::View<int *, layout, exec_space>;
246   using KrylovHandleType = KokkosBatched::KrylovHandle<Norm2DViewType, IntViewType, Scalar3DViewType>;
247   const int n_iterations = maxit;
248   //const int        team_size      = -1;
249   const int        vector_length  = -1;
250   const double     tol            = rtol;
251   const int        ortho_strategy = 0;
252   KrylovHandleType handle(Nsolves, Nsolves_team, n_iterations, true);
253   handle.Arnoldi_view = Scalar3DViewType("", Nsolves, n_iterations, Nloc + n_iterations + 3);
254   // solve
255   Functor_TestBatchedTeamVectorGMRES<exec_space, AMatrixValueView, IntView, XYType, KrylovHandleType>(batch_values, inv_diag, rowOffsets, colIndices, batch_x, batch_b, Nsolves_team, -1, vector_length, n_iterations, tol, ortho_strategy, 0, handle).run(pc);
256   Kokkos::fence();
257   // get data back
258   Kokkos::parallel_for(
259     "map", Kokkos::TeamPolicy<>(Nsolves /* /batch_sz */, -1, PCBJKOKKOS_VEC_SIZE), KOKKOS_LAMBDA(const team_member team) {
260       const int blkID = team.league_rank(), start = d_bid_eqOffset[blkID], end = d_bid_eqOffset[blkID + 1]; // 0
261       // map x into Plex/PETSc
262       Kokkos::parallel_for(Kokkos::TeamVectorRange(team, start, end), [=](int rowb) {
263         int rowa        = d_isicol[rowb];
264         glb_xdata[rowa] = batch_x(blkID, rowb - start);
265       });
266     });
267   // output assume species major - clone from Kokkos solvers
268   #if PCBJKOKKOS_VERBOSE_LEVEL >= 3
269     #if PCBJKOKKOS_VERBOSE_LEVEL >= 4
270   PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "Iterations\n"));
271     #else
272   PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "max iterations per species (gmres) :"));
273     #endif
274   for (PetscInt dmIdx = 0, s = 0, head = 0; dmIdx < jac->num_dms; dmIdx += batch_sz) {
275     for (PetscInt f = 0, idx = head; f < jac->dm_Nf[dmIdx]; f++, s++, idx++) {
276     #if PCBJKOKKOS_VERBOSE_LEVEL >= 4
277       PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "%2D:", s));
278       for (int bid = 0; bid < batch_sz; bid++) PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "%3D ", handle.get_iteration_host(idx + bid * jac->dm_Nf[dmIdx])));
279       PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "\n"));
280     #else
281       int count = 0, ii;
282       for (int bid = 0; bid < batch_sz; bid++) {
283         if ((ii = handle.get_iteration_host(idx + bid * jac->dm_Nf[dmIdx])) > count) count = ii;
284       }
285       PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "%3d", count));
286     #endif
287     }
288     head += batch_sz * jac->dm_Nf[dmIdx];
289   }
290     #if PCBJKOKKOS_VERBOSE_LEVEL == 3
291   PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "\n"));
292     #endif
293   #endif
294   // return error code, get max it
295   PetscInt count = 0, mbid = 0;
296   if (handle.is_converged_host()) {
297     *pcreason = PC_NOERROR;
298     if (!jac->max_nits) {
299       for (int blkID = 0; blkID < nBlk; blkID++) {
300         if (handle.get_iteration_host(blkID) > jac->max_nits) {
301           jac->max_nits = handle.get_iteration_host(blkID);
302           mbid          = blkID;
303         }
304       }
305     }
306   } else {
307     PetscCall(PetscPrintf(PETSC_COMM_SELF, "There is at least one system that did not converge."));
308     *pcreason = PC_SUBPC_ERROR;
309   }
310   // output - assume species major order
311   for (int blkID = 0; blkID < nBlk; blkID++) {
312     if (jac->reason) { // -pc_bjkokkos_ksp_converged_reason
313       if (jac->batch_target == blkID) {
314         if (batch_sz != 1)
315           PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "    Linear solve %s in %d iterations, batch %" PetscInt_FMT ", species %" PetscInt_FMT "\n", handle.is_converged_host(blkID) ? "converged" : "diverged", handle.get_iteration_host(blkID), blkID % batch_sz, blkID / batch_sz));
316         else PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "    Linear solve %s in %d iterations, block %" PetscInt_FMT "\n", handle.is_converged_host(blkID) ? "converged" : "diverged", handle.get_iteration_host(blkID), blkID));
317       } else if (jac->batch_target == -1 && handle.get_iteration_host(blkID) >= count) {
318         jac->max_nits = count = handle.get_iteration_host(blkID);
319         mbid                  = blkID;
320       }
321       if (!handle.is_converged_host(blkID)) PetscCall(PetscPrintf(PETSC_COMM_SELF, "ERROR species %d, batch %d did not converge with %d iterations\n", (int)(blkID / batch_sz), (int)blkID % batch_sz, handle.get_iteration_host(blkID)));
322     }
323   }
324   if (jac->batch_target == -1 && jac->reason) {
325     if (batch_sz != 1)
326       PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "    Linear solve %s in %d iteration, batch %" PetscInt_FMT ", specie %" PetscInt_FMT "\n", handle.is_converged_host(mbid) ? "converged" : "diverged", jac->max_nits, mbid % batch_sz, mbid / batch_sz));
327     else PetscCall(PetscPrintf(PetscObjectComm((PetscObject)A), "    Linear solve %s in %d iteration, block %" PetscInt_FMT "\n", handle.is_converged_host(mbid) ? "converged" : "diverged", jac->max_nits, mbid));
328   }
329   #if defined(PETSC_HAVE_CUDA)
330   nvtxRangePop();
331   #endif
332 
333   return PETSC_SUCCESS;
334 }
335 #endif
336