bjacobi/bjkokkos/bjkokkos.kokkos.cxx

92   int                Nblk = end - start, it, m, stride = stride_shared, idx = 0;
164   parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &lsum) { lsum += R[idx] * PetscConj(R[idx]); }, dpi);
183   parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { RP[idx] = R[idx]; });
192   parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &dot) { dot += R[idx] * PetscConj(RP[idx]); }, rhoold);
194   parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
206     parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &dot) { dot += V[idx] * PetscConj(RP[idx]); }, s);
215     parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
221     parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { T[idx] = Diag[idx] * T[idx]; });
225     parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { R[idx] = R[idx] - a * AUQ[idx]; });
227     parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &lsum) { lsum += R[idx] * PetscConj(R[idx]); }, dpi);
240         parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { D[idx] = U[idx] + cf * D[idx]; });
243         parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { D[idx] = Q[idx] + cf * D[idx]; });
246       parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { XX[idx] = XX[idx] + eta * D[idx]; });
284     parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &dot) { dot += R[idx] * PetscConj(RP[idx]); }, rho);
293     parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
300     parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { T[idx] = Diag[idx] * P[idx]; });
311   parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) { XX[idx] = Diag[idx] * XX[idx]; });
322     metad->flops = 2 * (metad->its * (10 * Nblk + 2 * nnz) + 5 * Nblk);
324     metad->flops = 2 * (metad->its * (10 * Nblk + 2 * 50 * Nblk) + 5 * Nblk); // guess
334   int                Nblk = end - start, it, stride = stride_shared, idx = 0; // start in shared mem
390   parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
396   parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &lsum) { lsum += Rr[idx] * PetscConj(Rr[idx]); }, dpi);
416     parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &dot) { dot += Zr[idx] * PetscConj(Rl[idx]); }, beta);
429       parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
437       parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
448     parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &lsum) { lsum += Zr[idx] * PetscConj(Pl[idx]); }, dpi);
459     parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
467     parallel_reduce(Kokkos::TeamVectorRange(team, Nblk), [=](const int idx, PetscScalar &lsum) { lsum += Rr[idx] * PetscConj(Rr[idx]); }, dpi);
501     parallel_for(Kokkos::TeamVectorRange(team, Nblk), [=](int idx) {
518     metad->flops = 2 * (metad->its * (10 * Nblk + 2 * nnz) + 5 * Nblk);
520     metad->flops = 2 * (metad->its * (10 * Nblk + 2 * 50 * Nblk) + 5 * Nblk); // guess