xref: /petsc/src/ksp/pc/impls/pbjacobi/kokkos/pbjacobi_kok.kokkos.cxx (revision 3ba1676111f5c958fe6c2729b46ca4d523958bb3)
112facf1bSJunchao Zhang #include <petscvec_kokkos.hpp>
212facf1bSJunchao Zhang #include <../src/vec/vec/impls/seq/kokkos/veckokkosimpl.hpp>
312facf1bSJunchao Zhang #include <petscdevice.h>
412facf1bSJunchao Zhang #include <../src/ksp/pc/impls/pbjacobi/pbjacobi.h>
512facf1bSJunchao Zhang 
612facf1bSJunchao Zhang struct PC_PBJacobi_Kokkos {
712facf1bSJunchao Zhang   PetscScalarKokkosDualView diag_dual;
812facf1bSJunchao Zhang 
912facf1bSJunchao Zhang   PC_PBJacobi_Kokkos(PetscInt len, PetscScalar *diag_ptr_h)
1012facf1bSJunchao Zhang   {
1112facf1bSJunchao Zhang     PetscScalarKokkosViewHost diag_h(diag_ptr_h, len);
1212facf1bSJunchao Zhang     auto                      diag_d = Kokkos::create_mirror_view_and_copy(DefaultMemorySpace(), diag_h);
1312facf1bSJunchao Zhang     diag_dual                        = PetscScalarKokkosDualView(diag_d, diag_h);
1412facf1bSJunchao Zhang   }
1512facf1bSJunchao Zhang 
1612facf1bSJunchao Zhang   PetscErrorCode Update(const PetscScalar *diag_ptr_h)
1712facf1bSJunchao Zhang   {
1812facf1bSJunchao Zhang     PetscFunctionBegin;
1912facf1bSJunchao Zhang     PetscCheck(diag_dual.view_host().data() == diag_ptr_h, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Host pointer has changed since last call");
2012facf1bSJunchao Zhang     PetscCallCXX(diag_dual.modify_host()); /* mark the host has newer data */
2112facf1bSJunchao Zhang     PetscCallCXX(diag_dual.sync_device());
22*3ba16761SJacob Faibussowitsch     PetscFunctionReturn(PETSC_SUCCESS);
2312facf1bSJunchao Zhang   }
2412facf1bSJunchao Zhang };
2512facf1bSJunchao Zhang 
2612facf1bSJunchao Zhang /* Make 'transpose' a template parameter instead of a function input parameter, so that
2712facf1bSJunchao Zhang  it will be a const in template instantiation and gets optimized out.
2812facf1bSJunchao Zhang */
2912facf1bSJunchao Zhang template <PetscBool transpose>
3012facf1bSJunchao Zhang static PetscErrorCode PCApplyOrTranspose_PBJacobi_Kokkos(PC pc, Vec x, Vec y)
3112facf1bSJunchao Zhang {
3212facf1bSJunchao Zhang   PC_PBJacobi               *jac   = (PC_PBJacobi *)pc->data;
3312facf1bSJunchao Zhang   PC_PBJacobi_Kokkos        *pckok = static_cast<PC_PBJacobi_Kokkos *>(jac->spptr);
3412facf1bSJunchao Zhang   ConstPetscScalarKokkosView xv;
3512facf1bSJunchao Zhang   PetscScalarKokkosView      yv;
3612facf1bSJunchao Zhang   PetscScalarKokkosView      Av = pckok->diag_dual.view_device();
3712facf1bSJunchao Zhang   const PetscInt             bs = jac->bs, mbs = jac->mbs, bs2 = bs * bs;
3812facf1bSJunchao Zhang   const char                *label = transpose ? "PCApplyTranspose_PBJacobi_Kokkos" : "PCApply_PBJacobi_Kokkos";
3912facf1bSJunchao Zhang 
4012facf1bSJunchao Zhang   PetscFunctionBegin;
4112facf1bSJunchao Zhang   PetscCall(PetscLogGpuTimeBegin());
4212facf1bSJunchao Zhang   VecErrorIfNotKokkos(x);
4312facf1bSJunchao Zhang   VecErrorIfNotKokkos(y);
4412facf1bSJunchao Zhang   PetscCall(VecGetKokkosView(x, &xv));
4512facf1bSJunchao Zhang   PetscCall(VecGetKokkosViewWrite(y, &yv));
4612facf1bSJunchao Zhang   PetscCallCXX(Kokkos::parallel_for(
4712facf1bSJunchao Zhang     label, bs * mbs, KOKKOS_LAMBDA(PetscInt row) {
4812facf1bSJunchao Zhang       const PetscScalar *Ap, *xp;
4912facf1bSJunchao Zhang       PetscScalar       *yp;
5012facf1bSJunchao Zhang       PetscInt           i, j, k;
5112facf1bSJunchao Zhang 
5212facf1bSJunchao Zhang       k  = row / bs;                                /* k-th block */
5312facf1bSJunchao Zhang       i  = row % bs;                                /* this thread deals with i-th row of the block */
5412facf1bSJunchao Zhang       Ap = &Av(bs2 * k + i * (transpose ? bs : 1)); /* Ap points to the first entry of i-th row */
5512facf1bSJunchao Zhang       xp = &xv(bs * k);
5612facf1bSJunchao Zhang       yp = &yv(bs * k);
5712facf1bSJunchao Zhang       /* multiply i-th row (column) with x */
5812facf1bSJunchao Zhang       yp[i] = 0.0;
5912facf1bSJunchao Zhang       for (j = 0; j < bs; j++) {
6012facf1bSJunchao Zhang         yp[i] += Ap[0] * xp[j];
6112facf1bSJunchao Zhang         Ap += (transpose ? 1 : bs); /* block is in column major order */
6212facf1bSJunchao Zhang       }
6312facf1bSJunchao Zhang     }));
6412facf1bSJunchao Zhang   PetscCall(VecRestoreKokkosView(x, &xv));
6512facf1bSJunchao Zhang   PetscCall(VecRestoreKokkosViewWrite(y, &yv));
6612facf1bSJunchao Zhang   PetscCall(PetscLogGpuFlops(bs * bs * mbs * 2)); /* FMA on entries in all blocks */
6712facf1bSJunchao Zhang   PetscCall(PetscLogGpuTimeEnd());
68*3ba16761SJacob Faibussowitsch   PetscFunctionReturn(PETSC_SUCCESS);
6912facf1bSJunchao Zhang }
7012facf1bSJunchao Zhang 
7112facf1bSJunchao Zhang static PetscErrorCode PCDestroy_PBJacobi_Kokkos(PC pc)
7212facf1bSJunchao Zhang {
7312facf1bSJunchao Zhang   PC_PBJacobi *jac = (PC_PBJacobi *)pc->data;
7412facf1bSJunchao Zhang 
7512facf1bSJunchao Zhang   PetscFunctionBegin;
7612facf1bSJunchao Zhang   PetscCallCXX(delete static_cast<PC_PBJacobi_Kokkos *>(jac->spptr));
7712facf1bSJunchao Zhang   PetscCall(PCDestroy_PBJacobi(pc));
78*3ba16761SJacob Faibussowitsch   PetscFunctionReturn(PETSC_SUCCESS);
7912facf1bSJunchao Zhang }
8012facf1bSJunchao Zhang 
8112facf1bSJunchao Zhang PETSC_INTERN PetscErrorCode PCSetUp_PBJacobi_Kokkos(PC pc)
8212facf1bSJunchao Zhang {
8312facf1bSJunchao Zhang   PC_PBJacobi *jac = (PC_PBJacobi *)pc->data;
8412facf1bSJunchao Zhang   PetscInt     len;
8512facf1bSJunchao Zhang 
8612facf1bSJunchao Zhang   PetscFunctionBegin;
8712facf1bSJunchao Zhang   PetscCall(PCSetUp_PBJacobi_Host(pc)); /* Compute the inverse on host now. Might worth doing it on device directly */
8812facf1bSJunchao Zhang   len = jac->bs * jac->bs * jac->mbs;
8912facf1bSJunchao Zhang   if (!jac->spptr) {
9012facf1bSJunchao Zhang     PetscCallCXX(jac->spptr = new PC_PBJacobi_Kokkos(len, const_cast<PetscScalar *>(jac->diag)));
9112facf1bSJunchao Zhang   } else {
9212facf1bSJunchao Zhang     PC_PBJacobi_Kokkos *pckok = static_cast<PC_PBJacobi_Kokkos *>(jac->spptr);
9312facf1bSJunchao Zhang     PetscCall(pckok->Update(jac->diag));
9412facf1bSJunchao Zhang   }
9512facf1bSJunchao Zhang   PetscCall(PetscLogCpuToGpu(sizeof(PetscScalar) * len));
9612facf1bSJunchao Zhang 
9712facf1bSJunchao Zhang   pc->ops->apply          = PCApplyOrTranspose_PBJacobi_Kokkos<PETSC_FALSE>;
9812facf1bSJunchao Zhang   pc->ops->applytranspose = PCApplyOrTranspose_PBJacobi_Kokkos<PETSC_TRUE>;
9912facf1bSJunchao Zhang   pc->ops->destroy        = PCDestroy_PBJacobi_Kokkos;
100*3ba16761SJacob Faibussowitsch   PetscFunctionReturn(PETSC_SUCCESS);
10112facf1bSJunchao Zhang }
102