Lines Matching refs:batch_sz

137 PetscErrorCode LandauKokkosStaticDataSet(DM plex, const PetscInt Nq, const PetscInt Nb, const PetscInt batch_sz, const PetscInt num_grids, PetscInt a_numCells[], PetscInt a_species_offset[], PetscInt a_mat_offset[], PetscReal a_nu_alpha[], PetscReal a_nu_beta[], PetscReal a_invMass[], PetscReal a_lambdas[], PetscReal a_invJ[], PetscReal a_x[], PetscReal a_y[], PetscReal a_z[], PetscReal a_w[], LandauStaticData *SData_d)
246     auto ipfdf_data = new Kokkos::View<PetscReal ***, Kokkos::LayoutLeft>("fdf", batch_sz, dim + 1, IPf_sz);
248     auto ipfdf_data = new Kokkos::View<PetscReal ***, Kokkos::LayoutRight>("fdf", batch_sz, dim + 1, IPf_sz);
381 PetscErrorCode LandauKokkosJacobian(DM plex[], const PetscInt Nq, const PetscInt Nb, const PetscInt batch_sz, const PetscInt num_grids, const PetscInt a_numCells[], PetscReal a_Eq_m[], PetscScalar a_elem_closure[], const PetscScalar a_xarray[], const LandauStaticData *SData_d, const PetscReal shift, const PetscLogEvent events[], const PetscInt a_mat_offset[], const PetscInt a_species_offset[], Mat subJ[], Mat JacP)
395   const PetscInt   coo_sz_batch = SData_d->coo_size / batch_sz;                                                 // capture
481   Kokkos::View<PetscScalar ****, Kokkos::LayoutRight> d_elem_mats("element matrices", batch_sz, num_grids, elem_mat_num_cells_max_grid, elem_mat_size_max); // not used (cpu assembly)
499       closure_sz *= batch_sz;
512     PetscCall(PetscInfo(plex[0], "df & f shared memory size: %d bytes in level, %d num cells total=%d, team size=%d, vector size=%d, #face=%d, Nf_max=%d\n", scr_bytes_fdf, KOKKOS_SHARED_LEVEL, (int)(num_cells_batch * batch_sz), team_size, vector_size, nfaces, (int)Nf_max));
514       "f, df", Kokkos::TeamPolicy<>(num_cells_batch * batch_sz, team_size, vector_size).set_scratch_size(KOKKOS_SHARED_LEVEL, Kokkos::PerTeam(scr_bytes_fdf)), KOKKOS_LAMBDA(const team_member team) {
521           const PetscInt moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, d_mat_offset);
593         const PetscInt moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, d_mat_offset);
727     PetscCall(PetscInfo(plex[0], "Jacobian shared memory size: %zu bytes in level %s (max shared=%zu), num cells total=%d, team size=%d, vector size=%d, #face=%d, Nf_max=%d\n", jac_scr_bytes, jac_shared_level == 0 ? "local" : "global", maximum_shared_mem_size, (int)(num_cells_batch * batch_sz), team_size, vector_size, nfaces, (int)Nf_max));
728     Kokkos::parallel_for("Jacobian", Kokkos::TeamPolicy<Kokkos::LaunchBounds<256, lbound2>>(num_cells_batch * batch_sz, team_size, vector_size).set_scratch_size(jac_shared_level, Kokkos::PerTeam(jac_scr_bytes)), jac_lambda);
738       "Mass", Kokkos::TeamPolicy<>(num_cells_batch * batch_sz, team_size, vector_size), KOKKOS_LAMBDA(const team_member team) {
748           const PetscInt moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, d_mat_offset);
790     for (PetscInt b_id = 0; b_id < batch_sz; b_id++) { // OpenMP (once)
795         PetscInt           moffset = LAND_MOFFSET(b_id, grid, batch_sz, num_grids, a_mat_offset), nloc, nzl, colbuf[1024], row;