//@HEADER // ************************************************************************ // // Kokkos v. 4.0 // Copyright (2022) National Technology & Engineering // Solutions of Sandia, LLC (NTESS). // // Under the terms of Contract DE-NA0003525 with NTESS, // the U.S. Government retains certain rights in this software. // // Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions. // See https://kokkos.org/LICENSE for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //@HEADER #ifndef KOKKOSBATCHED_LU_TEAM_INTERNAL_HPP #define KOKKOSBATCHED_LU_TEAM_INTERNAL_HPP /// \author Kyungjoo Kim (kyukim@sandia.gov) #include "KokkosBatched_Util.hpp" #include "KokkosBatched_Vector.hpp" #include "KokkosBatched_InnerLU_Serial_Impl.hpp" #include "KokkosBatched_InnerTrsm_Serial_Impl.hpp" #include "KokkosBatched_Trsm_Team_Internal.hpp" #include "KokkosBatched_Gemm_Team_Internal.hpp" namespace KokkosBatched { /// /// Team Internal Impl /// ================== template struct TeamLU_Internal { template KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const int m, const int n, ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, const typename MagnitudeScalarType::type tiny); }; template <> template KOKKOS_INLINE_FUNCTION int TeamLU_Internal::invoke( const MemberType &member, const int m, const int n, ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, const typename MagnitudeScalarType::type tiny) { const int k = (m < n ? m : n); if (k <= 0) return 0; const auto abs_tiny = tiny > 0 ? tiny : -tiny; const auto minus_abs_tiny = -abs_tiny; for (int p = 0; p < k; ++p) { // Made this non-const in order to WORKAROUND issue #349 int iend = m - p - 1; int jend = n - p - 1; const ValueType *KOKKOS_RESTRICT a12t = A + (p)*as0 + (p + 1) * as1; ValueType *KOKKOS_RESTRICT a21 = A + (p + 1) * as0 + (p)*as1, *KOKKOS_RESTRICT A22 = A + (p + 1) * as0 + (p + 1) * as1; if (tiny != 0) { if (member.team_rank() == 0) { ValueType &alpha11_reference = A[p * as0 + p * as1]; const auto alpha11_real = Kokkos::ArithTraits::real(alpha11_reference); alpha11_reference += minus_abs_tiny * ValueType(alpha11_real < 0); alpha11_reference += abs_tiny * ValueType(alpha11_real >= 0); } } member.team_barrier(); const ValueType alpha11 = A[p * as0 + p * as1]; Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, iend), [&](const int &i) { // a21[i*as0] *= inv_alpha11; a21[i * as0] /= alpha11; }); member.team_barrier(); Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, iend * jend), [&](const int &ij) { // assume layout right for batched computation const int i = ij / jend, j = ij % jend; A22[i * as0 + j * as1] -= a21[i * as0] * a12t[j * as1]; }); } return 0; } template <> template KOKKOS_INLINE_FUNCTION int TeamLU_Internal::invoke( const MemberType &member, const int m, const int n, ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, const typename MagnitudeScalarType::type /*tiny*/) { constexpr int mbAlgo = Algo::LU::Blocked::mb(); const int k = (m < n ? m : n); if (k <= 0) return 0; const typename MagnitudeScalarType::type one(1.0), minus_one(-1.0); InnerLU lu(as0, as1); InnerTrsmLeftLowerUnitDiag trsm_llu(as0, as1, as0, as1); InnerTrsmLeftLowerNonUnitDiag trsm_run(as1, as0, as1, as0); auto lu_factorize = [&](const int ib, const int jb, ValueType *KOKKOS_RESTRICT AA) { const int tsize = member.team_size(); // Made this non-const in order to WORKAROUND issue #349 int mb = mbAlgo; int nb = ((jb - mb) + (ib - mb)) > 0 ? ((jb - mb) + (ib - mb)) / tsize + (((jb - mb) + (ib - mb)) % tsize > 0) : 1; const int kb = ib < jb ? ib : jb; for (int p = 0; p < kb; p += mb) { const int pb = (p + mb) > kb ? (kb - p) : mb; // diagonal block ValueType *KOKKOS_RESTRICT Ap = AA + p * as0 + p * as1; // lu on a block member.team_barrier(); if (member.team_rank() == 0) lu.serial_invoke(pb, Ap); member.team_barrier(); // Made this non-const in order to WORKAROUND issue #349 int m_abr = ib - p - mb, n_abr = jb - p - mb, mp_abr = m_abr % nb, np_abr = n_abr % nb, mq_abr = (m_abr / nb) + (mp_abr > 0), nq_abr = (n_abr / nb) + (np_abr > 0); // trsm update Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, mq_abr + nq_abr), [&](const int &ij) { if (ij < nq_abr) { const int j = (ij)*nb, qb = (j + nb) > n_abr ? np_abr : nb; trsm_llu.serial_invoke(Ap, pb, qb, Ap + (j + mb) * as1); } else { const int i = (ij - nq_abr) * nb, qb = (i + nb) > m_abr ? mp_abr : nb; trsm_run.serial_invoke(Ap, pb, qb, Ap + (i + mb) * as0); } }); member.team_barrier(); // gemm update TeamGemmInternal::invoke(member, m_abr, n_abr, pb, minus_one, Ap + mb * as0, as0, as1, Ap + mb * as1, as0, as1, one, Ap + mb * as0 + mb * as1, as0, as1); } }; const bool is_small = true; //(m*n <= 64*64); if (is_small) { lu_factorize(m, n, A); } else { // // some cache blocking may need (not priority yet); // lu_factorize(m, n, A); } return 0; } } // namespace KokkosBatched #endif