//@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_TRSM_TEAM_INTERNAL_HPP #define KOKKOSBATCHED_TRSM_TEAM_INTERNAL_HPP /// \author Kyungjoo Kim (kyukim@sandia.gov) #include "KokkosBatched_Util.hpp" #include "KokkosKernels_ExecSpaceUtils.hpp" #include "KokkosBlas1_set_impl.hpp" #include "KokkosBlas1_team_scal_impl.hpp" #include "KokkosBatched_InnerTrsm_Serial_Impl.hpp" #include "KokkosBatched_Gemm_Team_Internal.hpp" namespace KokkosBatched { /// /// Team Internal Impl /// ==================== template struct TeamTrsmInternalLeftLower { template KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const bool use_unit_diag, const int m, const int n, const ScalarType alpha, const ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, /**/ ValueType *KOKKOS_RESTRICT B, const int bs0, const int bs1); }; template <> template KOKKOS_INLINE_FUNCTION int TeamTrsmInternalLeftLower::invoke( const MemberType &member, const bool use_unit_diag, const int m, const int n, const ScalarType alpha, const ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, /**/ ValueType *KOKKOS_RESTRICT B, const int bs0, const int bs1) { const ScalarType one(1.0), zero(0.0); if (alpha == zero) KokkosBlas::Impl::TeamSetInternal::invoke(member, m, n, zero, B, bs0, bs1); else { if (alpha != one) KokkosBlas::Impl::TeamScaleInternal::invoke(member, m, n, alpha, B, bs0, bs1); if (m <= 0 || n <= 0) return 0; for (int p = 0; p < m; ++p) { // Made this non-const in order to WORKAROUND issue #349 int iend = m - p - 1; int jend = n; const ValueType *KOKKOS_RESTRICT a21 = iend ? A + (p + 1) * as0 + p * as1 : NULL; ValueType *KOKKOS_RESTRICT b1t = B + p * bs0, *KOKKOS_RESTRICT B2 = iend ? B + (p + 1) * bs0 : NULL; member.team_barrier(); if (!use_unit_diag) { const ValueType alpha11 = A[p * as0 + p * as1]; Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, jend), [&](const int &j) { b1t[j * bs1] = b1t[j * bs1] / 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; B2[i * bs0 + j * bs1] -= a21[i * as0] * b1t[j * bs1]; }); } } return 0; } template <> template KOKKOS_INLINE_FUNCTION int TeamTrsmInternalLeftLower::invoke( const MemberType &member, const bool use_unit_diag, const int m, const int n, const ScalarType alpha, const ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, /**/ ValueType *KOKKOS_RESTRICT B, const int bs0, const int bs1) { constexpr int mbAlgo = Algo::Trsm::Blocked::mb(); const ScalarType one(1.0), zero(0.0), minus_one(-1.0); // note that parallel range is different ( m*n vs m-1*n); if (alpha == zero) KokkosBlas::Impl::TeamSetInternal::invoke(member, m, n, zero, B, bs0, bs1); else { if (alpha != one) KokkosBlas::Impl::TeamScaleInternal::invoke(member, m, n, alpha, B, bs0, bs1); if (m <= 0 || n <= 0) return 0; /// /// case host: team size is small and blocksize (mb,nb) is large /// /// case GPU: team size is large and blocksize (mb,nb) is small InnerTrsmLeftLowerUnitDiag trsm_u(as0, as1, bs0, bs1); InnerTrsmLeftLowerNonUnitDiag trsm_n(as0, as1, bs0, bs1); auto trsm = [&](const int ib, const int jb, const ValueType *KOKKOS_RESTRICT AA, /**/ ValueType *KOKKOS_RESTRICT BB) { const int mb = mbAlgo; const int tsize = member.team_size(); // Made this non-const in order to WORKAROUND issue #349 int nb = (jb / tsize + jb % tsize > 0); int np = jb % nb; for (int p = 0; p < ib; p += mb) { // Made this non-const in order to WORKAROUND issue #349 int pb = ((p + mb) > ib ? (ib - p) : mb); // trsm update const ValueType *KOKKOS_RESTRICT Ap = AA + p * as0 + p * as1; /**/ ValueType *KOKKOS_RESTRICT Bp = BB + p * bs0; member.team_barrier(); Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, (jb / nb) + (np > 0)), [&](const int jj) { // Made this non-const in order to WORKAROUND issue #349 int j = jj * nb, qb = (j + nb) > jb ? np : nb; if (use_unit_diag) trsm_u.serial_invoke(Ap, pb, qb, Bp + j * bs1); else trsm_n.serial_invoke(Ap, pb, qb, Bp + j * bs1); }); member.team_barrier(); // gemm update TeamGemmInternal::invoke(member, ib - p - pb, jb, pb, minus_one, Ap + pb * as0, as0, as1, Bp, bs0, bs1, one, Bp + pb * bs0, bs0, bs1); } }; const bool is_small = true; //(m*n <= 64*64); if (is_small) { trsm(m, n, A, B); } else { // // some cache blocking may need (not priority yet); // trsm(m, n, A, B); } } return 0; } template struct TeamTrsmInternalLeftUpper { template KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const bool use_unit_diag, const int m, const int n, const ScalarType alpha, const ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, /**/ ValueType *KOKKOS_RESTRICT B, const int bs0, const int bs1); }; template <> template KOKKOS_INLINE_FUNCTION int TeamTrsmInternalLeftUpper::invoke( const MemberType &member, const bool use_unit_diag, const int m, const int n, const ScalarType alpha, const ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, /**/ ValueType *KOKKOS_RESTRICT B, const int bs0, const int bs1) { const ScalarType one(1.0), zero(0.0); // note that parallel range is different ( m*n vs m-1*n); if (alpha == zero) KokkosBlas::Impl::TeamSetInternal::invoke(member, m, n, zero, B, bs0, bs1); else { if (alpha != one) KokkosBlas::Impl::TeamScaleInternal::invoke(member, m, n, alpha, B, bs0, bs1); if (m <= 0 || n <= 0) return 0; ValueType *KOKKOS_RESTRICT B0 = B; for (int p = (m - 1); p >= 0; --p) { // Made this non-const in order to WORKAROUND issue #349 int iend = p; int jend = n; const ValueType *KOKKOS_RESTRICT a01 = A + p * as1; /**/ ValueType *KOKKOS_RESTRICT b1t = B + p * bs0; member.team_barrier(); if (!use_unit_diag) { const ValueType alpha11 = A[p * as0 + p * as1]; Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, jend), [&](const int &j) { b1t[j * bs1] = b1t[j * bs1] / alpha11; }); member.team_barrier(); } Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, iend * jend), [&](const int &ij) { int i, j; if (KokkosKernels::Impl::is_gpu_exec_space_v) { i = ij % iend; j = ij / iend; } else { i = ij / jend; j = ij % jend; } B0[i * bs0 + j * bs1] -= a01[i * as0] * b1t[j * bs1]; }); } } return 0; } template <> template KOKKOS_INLINE_FUNCTION int TeamTrsmInternalLeftUpper::invoke( const MemberType &member, const bool use_unit_diag, const int m, const int n, const ScalarType alpha, const ValueType *KOKKOS_RESTRICT A, const int as0, const int as1, /**/ ValueType *KOKKOS_RESTRICT B, const int bs0, const int bs1) { constexpr int mbAlgo = Algo::Trsm::Blocked::mb(); const ScalarType one(1.0), zero(0.0), minus_one(-1.0); // note that parallel range is different ( m*n vs m-1*n); if (alpha == zero) KokkosBlas::Impl::TeamSetInternal::invoke(member, m, n, zero, B, bs0, bs1); else { if (alpha != one) KokkosBlas::Impl::TeamScaleInternal::invoke(member, m, n, alpha, B, bs0, bs1); if (m <= 0 || n <= 0) return 0; InnerTrsmLeftUpperUnitDiag trsm_u(as0, as1, bs0, bs1); InnerTrsmLeftUpperNonUnitDiag trsm_n(as0, as1, bs0, bs1); auto trsm = [&](const int ib, const int jb, const ValueType *KOKKOS_RESTRICT AA, /**/ ValueType *KOKKOS_RESTRICT BB) { const int mb = mbAlgo; //(ib <=5 ? ib : mbAlgo); const int tsize = member.team_size(); // Made this non-const in order to WORKAROUND issue #349 int nb = (jb / tsize + jb % tsize > 0); int np = jb % nb; for (int pp = 0; pp < ib; pp += mb) { const int ptmp = (ib - pp - mb), p = (ptmp < 0 ? 0 : ptmp), pb = (mb + (ptmp < 0) * ptmp); // trsm update const ValueType *KOKKOS_RESTRICT Ap = AA + p * as0 + p * as1; /**/ ValueType *KOKKOS_RESTRICT Bp = BB + p * bs0; member.team_barrier(); Kokkos::parallel_for(Kokkos::TeamThreadRange(member, 0, (jb / nb) + (np > 0)), [&](const int &jj) { const int j = jj * nb, qb = (j + nb) > jb ? np : nb; if (use_unit_diag) trsm_u.serial_invoke(Ap, pb, qb, Bp + j * bs1); else trsm_n.serial_invoke(Ap, pb, qb, Bp + j * bs1); }); member.team_barrier(); // gemm update TeamGemmInternal::invoke(member, p, jb, pb, minus_one, Ap - p * as0, as0, as1, Bp, bs0, bs1, one, BB, bs0, bs1); } }; const bool is_small = true; //(m*n <= 64*64); if (is_small) { trsm(m, n, A, B); } else { // // some cache blocking may need (not priority yet); // trsm(m, n, A, B); } } return 0; } } // namespace KokkosBatched #endif