tree_amg_smoother.f90

Go to the documentation of this file.

! Copyright (c) 2024-2026, The Neko Authors
! All rights reserved.
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
!
!   * Redistributions of source code must retain the above copyright
!     notice, this list of conditions and the following disclaimer.
!
!   * Redistributions in binary form must reproduce the above
!     copyright notice, this list of conditions and the following
!     disclaimer in the documentation and/or other materials provided
!     with the distribution.
!
!   * Neither the name of the authors nor the names of its
!     contributors may be used to endorse or promote products derived
!     from this software without specific prior written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
! LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
! FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
! COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
! INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
! BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
! LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
! CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
! LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
! ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
module tree_amg_smoother
  use tree_amg, only : tamg_hierarchy_t
  use tree_amg_utils, only : tamg_sample_matrix_val
  use num_types, only : rp
  use math, only : col2, add2, add2s2, glsc2, glsc3, sub2, cmult, &
       cmult2, copy, add3s2
  use device_math, only : device_glsc2, device_glsc3, device_rzero, &
       device_cmult2, device_sub2, device_add2, device_add3s2, &
       device_copy
  use krylov, only : ksp_monitor_t
  use scalar_bc_projector, only : scalar_bc_projector_t
  use gather_scatter, only : gs_t, gs_op_add
  use logger, only : neko_log, log_size
  use profiler, only : profiler_start_region, profiler_end_region
  use device, only: device_map, device_unmap, device_memcpy, host_to_device
  use device_tree_amg_smoother, only : amg_device_cheby_solve_part1, &
       amg_device_cheby_solve_part2
  use neko_config, only : neko_bcknd_device
  use, intrinsic :: iso_c_binding
  implicit none
  private
 
  type, public :: amg_jacobi_t
     real(kind=rp), allocatable :: d(:)
     real(kind=rp), allocatable :: w(:)
     real(kind=rp), allocatable :: r(:)
     real(kind=rp) :: omega
     integer :: lvl
     integer :: n
     integer :: max_iter = 10
     logical :: recompute_diag = .true.
   contains
     procedure, pass(this) :: init => amg_jacobi_init
     procedure, pass(this) :: solve => amg_jacobi_solve
     procedure, pass(this) :: comp_diag => amg_jacobi_diag
     procedure, pass(this) :: free => amg_jacobi_free
  end type amg_jacobi_t
 
  type, public :: amg_cheby_t
     real(kind=rp), allocatable :: d(:)
     type(c_ptr) :: d_d = c_null_ptr
     real(kind=rp), allocatable :: w(:)
     type(c_ptr) :: w_d = c_null_ptr
     real(kind=rp), allocatable :: r(:)
     type(c_ptr) :: r_d = c_null_ptr
     real(kind=rp), allocatable :: ev(:)
     type(c_ptr) :: ev_d = c_null_ptr
     real(kind=rp) :: tha, dlt
     integer :: lvl
     integer :: n
     integer :: power_its = 250
     integer :: power_its_refresh = 20
     logical :: warm_start_eigs = .false.
     logical :: eigs_computed = .false.
     integer :: max_iter = 10
     logical :: recompute_eigs = .true.
   contains
     procedure, pass(this) :: init => amg_cheby_init
     procedure, pass(this) :: solve => amg_cheby_solve
     procedure, pass(this) :: comp_eig => amg_cheby_power
     procedure, pass(this) :: device_solve => amg_device_cheby_solve
     procedure, pass(this) :: device_comp_eig => amg_device_cheby_power
     procedure, pass(this) :: free => amg_cheby_free
  end type amg_cheby_t
 
contains
 
  subroutine amg_cheby_init(this, n, lvl, max_iter)
    class(amg_cheby_t), intent(inout), target :: this
    integer, intent(in) :: n
    integer, intent(in) :: lvl
    integer, intent(in) :: max_iter
 
    allocate(this%d(n))
    allocate(this%w(n))
    allocate(this%r(n))
    if (neko_bcknd_device .eq. 1) then
       call device_map(this%d, this%d_d, n)
       call device_map(this%w, this%w_d, n)
       call device_map(this%r, this%r_d, n)
    end if
    this%n = n
    this%lvl = lvl
    this%max_iter = max_iter
    this%recompute_eigs = .true.
 
    call amg_smoo_monitor(lvl, this)
 
  end subroutine amg_cheby_init
 
  subroutine amg_cheby_free(this)
    class(amg_cheby_t), intent(inout), target :: this
    if (allocated(this%d)) then
       if (neko_bcknd_device .eq. 1 .and. c_associated(this%d_d)) then
          call device_unmap(this%d, this%d_d)
       end if
       deallocate(this%d)
    end if
    if (allocated(this%w)) then
       if (neko_bcknd_device .eq. 1 .and. c_associated(this%w_d)) then
          call device_unmap(this%w, this%w_d)
       end if
       deallocate(this%w)
    end if
    if (allocated(this%r)) then
       if (neko_bcknd_device .eq. 1 .and. c_associated(this%r_d)) then
          call device_unmap(this%r, this%r_d)
       end if
       deallocate(this%r)
    end if
    if (allocated(this%ev)) then
       if (neko_bcknd_device .eq. 1 .and. c_associated(this%ev_d)) then
          call device_unmap(this%ev, this%ev_d)
       end if
       deallocate(this%ev)
    end if
  end subroutine amg_cheby_free
 
 
  subroutine amg_cheby_power(this, amg, n)
    class(amg_cheby_t), intent(inout) :: this
    type(tamg_hierarchy_t), intent(inout) :: amg
    integer, intent(in) :: n
    real(kind=rp) :: lam, b, a, rn
    real(kind=rp), parameter :: boost = 1.1_rp
    real(kind=rp), parameter :: lam_factor = 30.0_rp
    real(kind=rp) :: wtw, dtw, dtd
    integer, allocatable :: fixed_seed(:), saved_seed(:)
    integer :: i, rnd_n, its
    logical :: warm
 
    call profiler_start_region('AMG_cheby_power')
    warm = this%warm_start_eigs .and. this%eigs_computed .and. &
         allocated(this%ev)
    associate(w => this%w, d => this%d, coef => amg%coef, gs_h => amg%gs_h, &
         msh => amg%msh, xh => amg%Xh, bc_projector => amg%bc_projector)
 
      if (warm) then
         its = this%power_its_refresh
         call copy(d, this%ev, n)
      else
         its = this%power_its
 
         ! Save current random seed and set a fixed seed
         call random_seed(size = rnd_n)
         allocate(saved_seed(rnd_n))
         allocate(fixed_seed(rnd_n))
         fixed_seed = 3901
         call random_seed(get = saved_seed)
         call random_seed(put = fixed_seed)
 
         do i = 1, n
            call random_number(rn)
            d(i) = rn + 10.0_rp
         end do
 
         ! Restore saved random seed
         call random_seed(put = saved_seed)
 
         if (this%lvl .eq. 0) then
            call gs_h%op(d, n, gs_op_add)!TODO
            call bc_projector%apply(d, n)
         end if
      end if
      !Power method to get lamba max
      do i = 1, its
         call amg%matvec(w, d, this%lvl)
 
         if (this%lvl .eq. 0) then
            wtw = glsc3(w, coef%mult, w, n)
         else
            wtw = glsc2(w, w, n)
         end if
 
         call cmult2(d, w, 1.0_rp/sqrt(wtw), n)
      end do
 
      call amg%matvec(w, d, this%lvl)
 
      if (this%lvl .eq. 0) then
         dtw = glsc3(d, coef%mult, w, n)
         dtd = glsc3(d, coef%mult, d, n)
      else
         dtw = glsc2(d, w, n)
         dtd = glsc2(d, d, n)
      end if
      lam = dtw / dtd
      b = lam * boost
      a = lam / lam_factor
      this%tha = (b+a)/2.0_rp
      this%dlt = (b-a)/2.0_rp
 
      if (this%warm_start_eigs) then
         if (.not. allocated(this%ev)) then
            allocate(this%ev(this%n))
         end if
         call copy(this%ev, d, n)
      end if
      this%eigs_computed = .true.
 
      this%recompute_eigs = .false.
      call amg_cheby_monitor(this%lvl, lam)
    end associate
    call profiler_end_region('AMG_cheby_power')
  end subroutine amg_cheby_power
 
  subroutine amg_cheby_solve(this, x, f, n, amg, zero_init)
    class(amg_cheby_t), intent(inout) :: this
    integer, intent(in) :: n
    real(kind=rp), dimension(n), intent(inout) :: x
    real(kind=rp), dimension(n), intent(inout) :: f
    class(tamg_hierarchy_t), intent(inout) :: amg
    type(ksp_monitor_t) :: ksp_results
    logical, optional, intent(in) :: zero_init
    integer :: iter, max_iter, i
    real(kind=rp) :: rtr, rnorm
    real(kind=rp) :: rhok, rhokp1, s1, thet, delt, tmp1, tmp2
    logical :: zero_initial_guess
 
    if (this%recompute_eigs) then
       call this%comp_eig(amg, n)
    end if
    if (present(zero_init)) then
       zero_initial_guess = zero_init
    else
       zero_initial_guess = .false.
    end if
    max_iter = this%max_iter
 
    associate(w => this%w, r => this%r, d => this%d, &
         bc_projector => amg%bc_projector)
      call copy(r, f, n)
      if (.not. zero_initial_guess) then
         call amg%matvec(w, x, this%lvl)
         call sub2(r, w, n)
      end if
 
      thet = this%tha
      delt = this%dlt
      s1 = thet / delt
      rhok = 1.0_rp / s1
 
      ! First iteration
      !OCL NORECURRENCE, NOVREC, NOALIAS
      !DIR$ CONCURRENT
      !DIR$ IVDEP
      !GCC$ ivdep
      !$omp parallel do
      do i = 1, n
         d(i) = 1.0_rp/thet * r(i)
         x(i) = x(i) + d(i)
      end do
      !$omp end parallel do
 
      ! Rest of iterations
      do iter = 2, max_iter
         call amg%matvec(w, d, this%lvl)
 
         rhokp1 = 1.0_rp / (2.0_rp * s1 - rhok)
         tmp1 = rhokp1 * rhok
         tmp2 = 2.0_rp * rhokp1 / delt
         rhok = rhokp1
 
         !$omp parallel private(i)
         !OCL NORECURRENCE, NOVREC, NOALIAS
         !DIR$ CONCURRENT
         !DIR$ IVDEP
         !GCC$ ivdep
         !$omp do
         do i = 1, n
            r(i) = r(i) - w(i)
            d(i) = tmp1 * d(i) + tmp2 * r(i)
            x(i) = x(i) + d(i)
         end do
         !$omp end do
         !$omp end parallel
      end do
    end associate
  end subroutine amg_cheby_solve
 
  subroutine amg_device_cheby_power(this, amg, n)
    class(amg_cheby_t), intent(inout) :: this
    type(tamg_hierarchy_t), intent(inout) :: amg
    integer, intent(in) :: n
    real(kind=rp) :: lam, b, a, rn
    real(kind=rp), parameter :: boost = 1.1_rp
    real(kind=rp), parameter :: lam_factor = 30.0_rp
    real(kind=rp) :: wtw, dtw, dtd
    integer, allocatable :: fixed_seed(:), saved_seed(:)
    integer :: i, rnd_n, its
    logical :: warm
 
    call profiler_start_region('AMG_cheby_power')
    warm = this%warm_start_eigs .and. this%eigs_computed .and. &
         c_associated(this%ev_d)
    associate(w => this%w, d => this%d, coef => amg%coef, gs_h => amg%gs_h, &
         msh => amg%msh, xh => amg%Xh, bc_projector => amg%bc_projector)
 
      if (warm) then
         its = this%power_its_refresh
         call device_copy(this%d_d, this%ev_d, n)
      else
         its = this%power_its
 
         ! Save current random seed and set a fixed seed
         call random_seed(size = rnd_n)
         allocate(saved_seed(rnd_n))
         allocate(fixed_seed(rnd_n))
         fixed_seed = 3901
         call random_seed(get = saved_seed)
         call random_seed(put = fixed_seed)
 
         do i = 1, n
            call random_number(rn)
            d(i) = rn + 10.0_rp
         end do
         call device_memcpy(this%d, this%d_d, n, host_to_device, .true.)
 
         ! Restore saved random seed
         call random_seed(put = saved_seed)
 
         if (this%lvl .eq. 0) then
            call gs_h%op(d, n, gs_op_add)!TODO
            call bc_projector%apply(d, n)
         end if
      end if
      do i = 1, its
         call amg%device_matvec(w, d, this%w_d, this%d_d, this%lvl)
 
         if (this%lvl .eq. 0) then
            wtw = device_glsc3(this%w_d, coef%mult_d, this%w_d, n)
         else
            wtw = device_glsc2(this%w_d, this%w_d, n)
         end if
 
         call device_cmult2(this%d_d, this%w_d, 1.0_rp/sqrt(wtw), n)
      end do
 
      call amg%device_matvec(w, d, this%w_d, this%d_d, this%lvl)
 
      if (this%lvl .eq. 0) then
         dtw = device_glsc3(this%d_d, coef%mult_d, this%w_d, n)
         dtd = device_glsc3(this%d_d, coef%mult_d, this%d_d, n)
      else
         dtw = device_glsc2(this%d_d, this%w_d, n)
         dtd = device_glsc2(this%d_d, this%d_d, n)
      end if
      lam = dtw / dtd
      b = lam * boost
      a = lam / lam_factor
      this%tha = (b+a)/2.0_rp
      this%dlt = (b-a)/2.0_rp
 
      if (this%warm_start_eigs) then
         if (.not. c_associated(this%ev_d)) then
            allocate(this%ev(this%n))
            call device_map(this%ev, this%ev_d, this%n)
         end if
         call device_copy(this%ev_d, this%d_d, n)
      end if
      this%eigs_computed = .true.
 
      this%recompute_eigs = .false.
      call amg_cheby_monitor(this%lvl, lam)
    end associate
    call profiler_end_region('AMG_cheby_power')
  end subroutine amg_device_cheby_power
 
  subroutine amg_device_cheby_solve(this, x, f, x_d, f_d, n, amg, zero_init)
    class(amg_cheby_t), intent(inout) :: this
    integer, intent(in) :: n
    real(kind=rp), dimension(n), intent(inout) :: x
    real(kind=rp), dimension(n), intent(inout) :: f
    type(c_ptr) :: x_d
    type(c_ptr) :: f_d
    class(tamg_hierarchy_t), intent(inout) :: amg
    type(ksp_monitor_t) :: ksp_results
    logical, optional, intent(in) :: zero_init
    integer :: iter, max_iter
    real(kind=rp) :: rtr, rnorm
    real(kind=rp) :: rhok, rhokp1, s1, thet, delt, tmp1, tmp2
    logical :: zero_initial_guess
 
    if (this%recompute_eigs) then
       call this%device_comp_eig(amg, n)
    end if
    if (present(zero_init)) then
       zero_initial_guess = zero_init
    else
       zero_initial_guess = .false.
    end if
    max_iter = this%max_iter
 
    associate( w_d => this%w_d, r_d => this%r_d, d_d => this%d_d )
 
      if (.not. zero_initial_guess) then
         call amg%device_matvec(this%w, x, w_d, x_d, this%lvl)
      end if
 
      thet = this%tha
      delt = this%dlt
      s1 = thet / delt
      rhok = 1.0_rp / s1
 
      ! First iteration
      tmp1 = 1.0_rp / thet
      call amg_device_cheby_solve_part1(r_d, f_d, w_d, x_d, d_d, &
           tmp1, n, zero_initial_guess)
      ! Rest of iterations
      do iter = 2, max_iter
         call amg%device_matvec(this%w, this%d, w_d, d_d, this%lvl)
 
         rhokp1 = 1.0_rp / (2.0_rp * s1 - rhok)
         tmp1 = rhokp1 * rhok
         tmp2 = 2.0_rp * rhokp1 / delt
         rhok = rhokp1
 
         call amg_device_cheby_solve_part2(r_d, w_d, d_d, x_d, tmp1, tmp2, n)
 
      end do
    end associate
  end subroutine amg_device_cheby_solve
 
  subroutine amg_jacobi_init(this, n, lvl, max_iter)
    class(amg_jacobi_t), intent(inout), target :: this
    integer, intent(in) :: n
    integer, intent(in) :: lvl
    integer, intent(in) :: max_iter
 
    allocate(this%d(n))
    allocate(this%w(n))
    allocate(this%r(n))
    this%n = n
    this%lvl = lvl
    this%max_iter = max_iter
    this%omega = 0.7_rp
 
  end subroutine amg_jacobi_init
 
  subroutine amg_jacobi_free(this)
    class(amg_jacobi_t), intent(inout), target :: this
    if (allocated(this%d)) then
       deallocate(this%d)
    end if
    if (allocated(this%w)) then
       deallocate(this%w)
    end if
    if (allocated(this%r)) then
       deallocate(this%r)
    end if
  end subroutine amg_jacobi_free
 
  subroutine amg_jacobi_diag(this, amg, n)
    class(amg_jacobi_t), intent(inout) :: this
    type(tamg_hierarchy_t), intent(inout) :: amg
    integer, intent(in) :: n
    real(kind=rp) :: val
    integer :: i
    do i = 1, n
       call tamg_sample_matrix_val(val, amg, this%lvl, i, i)
       this%d(i) = 1.0_rp / val
    end do
    this%recompute_diag = .false.
  end subroutine amg_jacobi_diag
 
  subroutine amg_jacobi_solve(this, x, f, n, amg, niter)
    class(amg_jacobi_t), intent(inout) :: this
    integer, intent(in) :: n
    real(kind=rp), dimension(n), intent(inout) :: x
    real(kind=rp), dimension(n), intent(inout) :: f
    class(tamg_hierarchy_t), intent(inout) :: amg
    type(ksp_monitor_t) :: ksp_results
    integer, optional, intent(in) :: niter
    integer :: iter, max_iter
    real(kind=rp) :: rtr, rnorm
    integer :: i
 
    if (this%recompute_diag) then
       call this%comp_diag(amg, n)
    end if
 
    if (present(niter)) then
       max_iter = niter
    else
       max_iter = this%max_iter
    end if
 
    ! x = x + omega * Dinv( f - Ax )
    associate( w => this%w, r => this%r, d => this%d)
      do iter = 1, max_iter
         w = 0.0_rp
         call amg%matvec(w, x, this%lvl)
         !$omp parallel private(i)
         !$omp do
         do i = 1, n
            r(i) = f(i) - w(i)
         end do
         !$omp end do
         !$omp do
         do i = 1, n
            r(i) = r(i) * d(i)
         end do
         !$omp end do
         !$omp do
         do i = 1, n
            x(i) = x(i) + this%omega * r(i)
         end do
         !$omp end do
         !$omp end parallel
      end do
    end associate
  end subroutine amg_jacobi_solve
 
  subroutine amg_smoo_monitor(lvl, smoo)
    integer, intent(in) :: lvl
    class(amg_cheby_t), intent(in) :: smoo
    character(len=LOG_SIZE) :: log_buf
 
    write(log_buf, '(A8,I2,A28)') '-- level', lvl, '-- init smoother: Chebyshev'
    call neko_log%message(log_buf)
    write(log_buf, '(A22,I6)') 'Iterations:', smoo%max_iter
    call neko_log%message(log_buf)
  end subroutine amg_smoo_monitor
 
  subroutine amg_cheby_monitor(lvl, lam)
    integer, intent(in) :: lvl
    real(kind=rp), intent(in) :: lam
    character(len=LOG_SIZE) :: log_buf
 
    write(log_buf, '(A12,I2,A29,F12.3)') '-- AMG level', lvl, &
         '-- Chebyshev approx. max eig', lam
    call neko_log%message(log_buf)
  end subroutine amg_cheby_monitor
 
end module tree_amg_smoother