device_math.F90

Go to the documentation of this file.

! Copyright (c) 2021-2024, 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 device_math
  use, intrinsic :: iso_c_binding, only: c_ptr, c_int
  use num_types, only: rp, c_rp
  use utils, only: neko_error
  use comm, only: neko_comm, pe_size, mpi_real_precision
  use mpi_f08, only: mpi_sum, mpi_in_place, mpi_allreduce
 
  ! ========================================================================== !
  ! Device math interfaces
 
  use hip_math
  use cuda_math
  use opencl_math
 
  implicit none
  private
 
  interface device_pwmax
     module procedure device_pwmax_vec2, device_pwmax_vec3, &
          device_pwmax_sca2, device_pwmax_sca3
  interface device_pwmax …
  end interface device_pwmax
 
  interface device_pwmin
     module procedure device_pwmin_vec2, device_pwmin_vec3, &
          device_pwmin_sca2, device_pwmin_sca3
  interface device_pwmin …
  end interface device_pwmin
 
 
  public :: device_copy, device_rzero, device_rone, device_cmult, &
       device_cmult2, device_cadd, device_cadd2, device_cfill, device_add2, &
       device_add3, device_add4, device_add2s1, device_add2s2, &
       device_addsqr2s2, device_add3s2, device_invcol1, device_invcol2, &
       device_col2, device_col3, device_subcol3, device_sub2, device_sub3, &
       device_addcol3, device_addcol4, device_vdot3, device_vlsc3, &
       device_glsc3, device_glsc3_many, device_add2s2_many, device_glsc2, &
       device_glsum, device_masked_copy, device_cfill_mask, &
       device_masked_red_copy, device_vcross, device_absval, &
       device_pwmax, device_pwmin, device_masked_atomic_reduction
 
contains
 
  subroutine device_copy(a_d, b_d, n)
    type(c_ptr) :: a_d, b_d
    integer :: n
#if HAVE_HIP
    call hip_copy(a_d, b_d, n)
#elif HAVE_CUDA
    call cuda_copy(a_d, b_d, n)
#elif HAVE_OPENCL
    call opencl_copy(a_d, b_d, n)
#else
    call neko_error('no device backend configured')
#endif
  subroutine device_copy(a_d, b_d, n) …
  end subroutine device_copy
 
  subroutine device_masked_copy(a_d, b_d, mask_d, n, m)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, m
#if HAVE_HIP
    call hip_masked_copy(a_d, b_d, mask_d, n, m)
#elif HAVE_CUDA
    call cuda_masked_copy(a_d, b_d, mask_d, n, m)
#elif HAVE_OPENCL
    call opencl_masked_copy(a_d, b_d, mask_d, n, m)
#else
    call neko_error('no device backend configured')
#endif
  subroutine device_masked_copy(a_d, b_d, mask_d, n, m) …
  end subroutine device_masked_copy
 
  subroutine device_masked_red_copy(a_d, b_d, mask_d, n, m)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, m
#if HAVE_HIP
    call hip_masked_red_copy(a_d, b_d, mask_d, n, m)
#elif HAVE_CUDA
    call cuda_masked_red_copy(a_d, b_d, mask_d, n, m)
#elif HAVE_OPENCL
    call neko_error('No OpenCL bcknd, masked red copy')
#else
    call neko_error('no device backend configured')
#endif
  subroutine device_masked_red_copy(a_d, b_d, mask_d, n, m) …
  end subroutine device_masked_red_copy
 
  subroutine device_masked_atomic_reduction(a_d, b_d, mask_d, n, m)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, m
#if HAVE_HIP
    call hip_masked_atomic_reduction(a_d, b_d, mask_d, n, m)
#elif HAVE_CUDA
    call cuda_masked_atomic_reduction(a_d, b_d, mask_d, n, m)
#elif HAVE_OPENCL
    call neko_error('No OpenCL bcknd, masked atomic reduction')
#else
    call neko_error('no device backend configured')
#endif
  subroutine device_masked_atomic_reduction(a_d, b_d, mask_d, n, m) …
  end subroutine device_masked_atomic_reduction
 
  subroutine device_cfill_mask(a_d, c, size, mask_d, mask_size)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: size
    type(c_ptr) :: mask_d
    integer :: mask_size
#if HAVE_HIP
    call hip_cfill_mask(a_d, c, size, mask_d, mask_size)
#elif HAVE_CUDA
    call cuda_cfill_mask(a_d, c, size, mask_d, mask_size)
#elif HAVE_OPENCL
    call opencl_cfill_mask(a_d, c, size, mask_d, mask_size)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_cfill_mask(a_d, c, size, mask_d, mask_size) …
  end subroutine device_cfill_mask
 
  subroutine device_rzero(a_d, n)
    type(c_ptr) :: a_d
    integer :: n
#if HAVE_HIP
    call hip_rzero(a_d, n)
#elif HAVE_CUDA
    call cuda_rzero(a_d, n)
#elif HAVE_OPENCL
    call opencl_rzero(a_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_rzero(a_d, n) …
  end subroutine device_rzero
 
  subroutine device_rone(a_d, n)
    type(c_ptr) :: a_d
    integer :: n
    real(kind=rp), parameter :: one = 1.0_rp
#if HAVE_HIP || HAVE_CUDA || HAVE_OPENCL
    call device_cfill(a_d, one, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_rone(a_d, n) …
  end subroutine device_rone
 
  subroutine device_cmult(a_d, c, n)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
#if HAVE_HIP
    call hip_cmult(a_d, c, n)
#elif HAVE_CUDA
    call cuda_cmult(a_d, c, n)
#elif HAVE_OPENCL
    call opencl_cmult(a_d, c, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_cmult(a_d, c, n) …
  end subroutine device_cmult
 
  subroutine device_cmult2(a_d, b_d, c, n)
    type(c_ptr) :: a_d, b_d
    real(kind=rp), intent(in) :: c
    integer :: n
#if HAVE_HIP
    call hip_cmult2(a_d, b_d, c, n)
#elif HAVE_CUDA
    call cuda_cmult2(a_d, b_d, c, n)
#elif HAVE_OPENCL
    call opencl_cmult2(a_d, b_d, c, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_cmult2(a_d, b_d, c, n) …
  end subroutine device_cmult2
 
  subroutine device_cadd(a_d, c, n)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
#if HAVE_HIP
    call hip_cadd(a_d, c, n)
#elif HAVE_CUDA
    call cuda_cadd(a_d, c, n)
#elif HAVE_OPENCL
    call opencl_cadd(a_d, c, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_cadd(a_d, c, n) …
  end subroutine device_cadd
 
  subroutine device_cadd2(a_d, b_d, c, n)
    type(c_ptr) :: a_d
    type(c_ptr) :: b_d
    real(kind=rp), intent(in) :: c
    integer :: n
#if HAVE_HIP
    call hip_cadd2(a_d, b_d, c, n)
#elif HAVE_CUDA
    call cuda_cadd2(a_d, b_d, c, n)
#elif HAVE_OPENCL
    call opencl_cadd2(a_d, b_d, c, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_cadd2(a_d, b_d, c, n) …
  end subroutine device_cadd2
 
  subroutine device_cfill(a_d, c, n)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
#if HAVE_HIP
    call hip_cfill(a_d, c, n)
#elif HAVE_CUDA
    call cuda_cfill(a_d, c, n)
#elif HAVE_OPENCL
    call opencl_cfill(a_d, c, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_cfill(a_d, c, n) …
  end subroutine device_cfill
 
  subroutine device_add2(a_d, b_d, n)
    type(c_ptr) :: a_d, b_d
    integer :: n
#if HAVE_HIP
    call hip_add2(a_d, b_d, n)
#elif HAVE_CUDA
    call cuda_add2(a_d, b_d, n)
#elif HAVE_OPENCL
    call opencl_add2(a_d, b_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_add2(a_d, b_d, n) …
  end subroutine device_add2
 
  subroutine device_add4(a_d, b_d, c_d, d_d, n)
    type(c_ptr) :: a_d, b_d, c_d, d_d
    integer :: n
#if HAVE_HIP
    call hip_add4(a_d, b_d, c_d, d_d, n)
#elif HAVE_CUDA
    call cuda_add4(a_d, b_d, c_d, d_d, n)
#elif HAVE_OPENCL
    call opencl_add4(a_d, b_d, c_d, d_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_add4(a_d, b_d, c_d, d_d, n) …
  end subroutine device_add4
 
  subroutine device_add2s1(a_d, b_d, c1, n)
    type(c_ptr) :: a_d, b_d
    real(kind=rp) :: c1
    integer :: n
#if HAVE_HIP
    call hip_add2s1(a_d, b_d, c1, n)
#elif HAVE_CUDA
    call cuda_add2s1(a_d, b_d, c1, n)
#elif HAVE_OPENCL
    call opencl_add2s1(a_d, b_d, c1, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_add2s1(a_d, b_d, c1, n) …
  end subroutine device_add2s1
 
  subroutine device_add2s2(a_d, b_d, c1, n)
    type(c_ptr) :: a_d, b_d
    real(kind=rp) :: c1
    integer :: n
#if HAVE_HIP
    call hip_add2s2(a_d, b_d, c1, n)
#elif HAVE_CUDA
    call cuda_add2s2(a_d, b_d, c1, n)
#elif HAVE_OPENCL
    call opencl_add2s2(a_d, b_d, c1, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_add2s2(a_d, b_d, c1, n) …
  end subroutine device_add2s2
 
  subroutine device_addsqr2s2(a_d, b_d, c1, n)
    type(c_ptr) :: a_d, b_d
    real(kind=rp) :: c1
    integer :: n
#if HAVE_HIP
    call hip_addsqr2s2(a_d, b_d, c1, n)
#elif HAVE_CUDA
    call cuda_addsqr2s2(a_d, b_d, c1, n)
#elif HAVE_OPENCL
    call opencl_addsqr2s2(a_d, b_d, c1, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_addsqr2s2(a_d, b_d, c1, n) …
  end subroutine device_addsqr2s2
 
  subroutine device_add3(a_d, b_d, c_d, n)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
#if HAVE_HIP
    call hip_add3(a_d, b_d, c_d, n)
#elif HAVE_CUDA
    call cuda_add3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    call opencl_add3(a_d, b_d, c_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_add3(a_d, b_d, c_d, n) …
  end subroutine device_add3
 
  subroutine device_add3s2(a_d, b_d, c_d, c1, c2 , n)
    type(c_ptr) :: a_d, b_d, c_d
    real(kind=rp) :: c1, c2
    integer :: n
#if HAVE_HIP
    call hip_add3s2(a_d, b_d, c_d, c1, c2, n)
#elif HAVE_CUDA
    call cuda_add3s2(a_d, b_d, c_d, c1, c2, n)
#elif HAVE_OPENCL
    call opencl_add3s2(a_d, b_d, c_d, c1, c2, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_add3s2(a_d, b_d, c_d, c1, c2 , n) …
  end subroutine device_add3s2
 
  subroutine device_invcol1(a_d, n)
    type(c_ptr) :: a_d
    integer :: n
#if HAVE_HIP
    call hip_invcol1(a_d, n)
#elif HAVE_CUDA
    call cuda_invcol1(a_d, n)
#elif HAVE_OPENCL
    call opencl_invcol1(a_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_invcol1(a_d, n) …
  end subroutine device_invcol1
 
  subroutine device_invcol2(a_d, b_d, n)
    type(c_ptr) :: a_d, b_d
    integer :: n
#if HAVE_HIP
    call hip_invcol2(a_d, b_d, n)
#elif HAVE_CUDA
    call cuda_invcol2(a_d, b_d, n)
#elif HAVE_OPENCL
    call opencl_invcol2(a_d, b_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_invcol2(a_d, b_d, n) …
  end subroutine device_invcol2
 
  subroutine device_col2(a_d, b_d, n)
    type(c_ptr) :: a_d, b_d
    integer :: n
#if HAVE_HIP
    call hip_col2(a_d, b_d, n)
#elif HAVE_CUDA
    call cuda_col2(a_d, b_d, n)
#elif HAVE_OPENCL
    call opencl_col2(a_d, b_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_col2(a_d, b_d, n) …
  end subroutine device_col2
 
  subroutine device_col3(a_d, b_d, c_d, n)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
#if HAVE_HIP
    call hip_col3(a_d, b_d, c_d, n)
#elif HAVE_CUDA
    call cuda_col3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    call opencl_col3(a_d, b_d, c_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_col3(a_d, b_d, c_d, n) …
  end subroutine device_col3
 
  subroutine device_subcol3(a_d, b_d, c_d, n)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
#if HAVE_HIP
    call hip_subcol3(a_d, b_d, c_d, n)
#elif HAVE_CUDA
    call cuda_subcol3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    call opencl_subcol3(a_d, b_d, c_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_subcol3(a_d, b_d, c_d, n) …
  end subroutine device_subcol3
 
  subroutine device_sub2(a_d, b_d, n)
    type(c_ptr) :: a_d, b_d
    integer :: n
#if HAVE_HIP
    call hip_sub2(a_d, b_d, n)
#elif HAVE_CUDA
    call cuda_sub2(a_d, b_d, n)
#elif HAVE_OPENCL
    call opencl_sub2(a_d, b_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_sub2(a_d, b_d, n) …
  end subroutine device_sub2
 
  subroutine device_sub3(a_d, b_d, c_d, n)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
#if HAVE_HIP
    call hip_sub3(a_d, b_d, c_d, n)
#elif HAVE_CUDA
    call cuda_sub3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    call opencl_sub3(a_d, b_d, c_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_sub3(a_d, b_d, c_d, n) …
  end subroutine device_sub3
 
  subroutine device_addcol3(a_d, b_d, c_d, n)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
#if HAVE_HIP
    call hip_addcol3(a_d, b_d, c_d, n)
#elif HAVE_CUDA
    call cuda_addcol3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    call opencl_addcol3(a_d, b_d, c_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_addcol3(a_d, b_d, c_d, n) …
  end subroutine device_addcol3
 
  subroutine device_addcol4(a_d, b_d, c_d, d_d, n)
    type(c_ptr) :: a_d, b_d, c_d, d_d
    integer :: n
#if HAVE_HIP
    call hip_addcol4(a_d, b_d, c_d, d_d, n)
#elif HAVE_CUDA
    call cuda_addcol4(a_d, b_d, c_d, d_d, n)
#elif HAVE_OPENCL
    call opencl_addcol4(a_d, b_d, c_d, d_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_addcol4(a_d, b_d, c_d, d_d, n) …
  end subroutine device_addcol4
 
  subroutine device_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n)
    type(c_ptr) :: dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d
    integer :: n
#if HAVE_HIP
    call hip_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n)
#elif HAVE_CUDA
    call cuda_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n)
#elif HAVE_OPENCL
    call opencl_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n) …
  end subroutine device_vdot3
 
  subroutine device_vcross(u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, &
       w1_d, w2_d, w3_d, n)
    type(c_ptr) :: u1_d, u2_d, u3_d
    type(c_ptr) :: v1_d, v2_d, v3_d
    type(c_ptr) :: w1_d, w2_d, w3_d
    integer :: n
#if HAVE_HIP
    call hip_vcross(u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, &
         w1_d, w2_d, w3_d, n)
#elif HAVE_CUDA
    call cuda_vcross(u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, &
         w1_d, w2_d, w3_d, n)
#elif HAVE_OPENCL
    call neko_error("no opencl backedn vcross")
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_vcross(u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, & …
  end subroutine device_vcross
 
 
  function device_vlsc3(u_d, v_d, w_d, n) result(res)
    type(c_ptr) :: u_d, v_d, w_d
    integer :: n
    real(kind=rp) :: res
    res = 0.0_rp
#if HAVE_HIP
    res = hip_vlsc3(u_d, v_d, w_d, n)
#elif HAVE_CUDA
    res = cuda_vlsc3(u_d, v_d, w_d, n)
#elif HAVE_OPENCL
    ! Same kernel as glsc3 (currently no device MPI for OpenCL)
    res = opencl_glsc3(u_d, v_d, w_d, n)
#else
    call neko_error('No device backend configured')
#endif
  function device_vlsc3(u_d, v_d, w_d, n) result(res) …
  end function device_vlsc3
 
  function device_glsc3(a_d, b_d, c_d, n) result(res)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n, ierr
    real(kind=rp) :: res
#if HAVE_HIP
    res = hip_glsc3(a_d, b_d, c_d, n)
#elif HAVE_CUDA
    res = cuda_glsc3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    res = opencl_glsc3(a_d, b_d, c_d, n)
#else
    call neko_error('No device backend configured')
#endif
 
#ifndef HAVE_DEVICE_MPI
    if (pe_size .gt. 1) then
       call mpi_allreduce(mpi_in_place, res, 1, &
            mpi_real_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
  function device_glsc3(a_d, b_d, c_d, n) result(res) …
  end function device_glsc3
 
  subroutine device_glsc3_many(h, w_d, v_d_d, mult_d, j, n)
    type(c_ptr), value :: w_d, v_d_d, mult_d
    integer(c_int) :: j, n
    real(c_rp) :: h(j)
    integer :: ierr
#if HAVE_HIP
    call hip_glsc3_many(h, w_d, v_d_d, mult_d, j, n)
#elif HAVE_CUDA
    call cuda_glsc3_many(h, w_d, v_d_d, mult_d, j, n)
#elif HAVE_OPENCL
    call opencl_glsc3_many(h, w_d, v_d_d, mult_d, j, n)
#else
    call neko_error('No device backend configured')
#endif
 
#ifndef HAVE_DEVICE_MPI
    if (pe_size .gt. 1) then
       call mpi_allreduce(mpi_in_place, h, j, &
            mpi_real_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
  subroutine device_glsc3_many(h, w_d, v_d_d, mult_d, j, n) …
  end subroutine device_glsc3_many
 
  subroutine device_add2s2_many(y_d, x_d_d, a_d, j, n)
    type(c_ptr), value :: y_d, x_d_d, a_d
    integer(c_int) :: j, n
#if HAVE_HIP
    call hip_add2s2_many(y_d, x_d_d, a_d, j, n)
#elif HAVE_CUDA
    call cuda_add2s2_many(y_d, x_d_d, a_d, j, n)
#elif HAVE_OPENCL
    call opencl_add2s2_many(y_d, x_d_d, a_d, j, n)
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_add2s2_many(y_d, x_d_d, a_d, j, n) …
  end subroutine device_add2s2_many
 
  function device_glsc2(a_d, b_d, n) result(res)
    type(c_ptr) :: a_d, b_d
    integer :: n, ierr
    real(kind=rp) :: res
#if HAVE_HIP
    res = hip_glsc2(a_d, b_d, n)
#elif HAVE_CUDA
    res = cuda_glsc2(a_d, b_d, n)
#elif HAVE_OPENCL
    res = opencl_glsc2(a_d, b_d, n)
#else
    call neko_error('No device backend configured')
#endif
 
#ifndef HAVE_DEVICE_MPI
    if (pe_size .gt. 1) then
       call mpi_allreduce(mpi_in_place, res, 1, &
            mpi_real_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
  function device_glsc2(a_d, b_d, n) result(res) …
  end function device_glsc2
 
  function device_glsum(a_d, n) result(res)
    type(c_ptr) :: a_d
    integer :: n, ierr
    real(kind=rp) :: res
#if HAVE_HIP
    res = hip_glsum(a_d, n)
#elif HAVE_CUDA
    res = cuda_glsum(a_d, n)
#elif HAVE_OPENCL
    res = opencl_glsum(a_d, n)
#else
    call neko_error('No device backend configured')
#endif
 
#ifndef HAVE_DEVICE_MPI
    if (pe_size .gt. 1) then
       call mpi_allreduce(mpi_in_place, res, 1, &
            mpi_real_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
  function device_glsum(a_d, n) result(res) …
  end function device_glsum
 
  subroutine device_absval(a_d, n)
    integer, intent(in) :: n
    type(c_ptr) :: a_d
#ifdef HAVE_HIP
    call hip_absval(a_d, n)
#elif HAVE_CUDA
    call cuda_absval(a_d, n)
#elif HAVE_OPENCL
    call neko_error('OPENCL is not implemented for device_absval')
#else
    call neko_error('No device backend configured')
#endif
 
  subroutine device_absval(a_d, n) …
  end subroutine device_absval
 
  ! ========================================================================== !
  ! Device point-wise max
 
  subroutine device_pwmax_vec2(a_d, b_d, n)
    type(c_ptr) :: a_d, b_d
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmax_vec2')
#elif HAVE_CUDA
    call cuda_pwmax_vec2(a_d, b_d, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmax_vec2')
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_pwmax_vec2(a_d, b_d, n) …
  end subroutine device_pwmax_vec2
 
  subroutine device_pwmax_vec3(a_d, b_d, c_d, n)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmax_vec3')
#elif HAVE_CUDA
    call cuda_pwmax_vec3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmax_vec3')
#else
    call neko_error('No device backend configured')
#endif
 
  subroutine device_pwmax_vec3(a_d, b_d, c_d, n) …
  end subroutine device_pwmax_vec3
 
  subroutine device_pwmax_sca2(a_d, c, n)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmax_sca2')
#elif HAVE_CUDA
    call cuda_pwmax_sca2(a_d, c, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmax_sca2')
#else
    call neko_error('No device backend configured')
#endif
 
  subroutine device_pwmax_sca2(a_d, c, n) …
  end subroutine device_pwmax_sca2
 
  subroutine device_pwmax_sca3(a_d, b_d, c, n)
    type(c_ptr) :: a_d, b_d
    real(kind=rp), intent(in) :: c
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmax_sca3')
#elif HAVE_CUDA
    call cuda_pwmax_sca3(a_d, b_d, c, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmax_sca3')
#else
    call neko_error('No device backend configured')
#endif
 
  subroutine device_pwmax_sca3(a_d, b_d, c, n) …
  end subroutine device_pwmax_sca3
 
  ! ========================================================================== !
  ! Device point-wise min
 
  subroutine device_pwmin_vec2(a_d, b_d, n)
    type(c_ptr) :: a_d, b_d
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmin_vec2')
#elif HAVE_CUDA
    call cuda_pwmin_vec2(a_d, b_d, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmin_vec2')
#else
    call neko_error('No device backend configured')
#endif
  subroutine device_pwmin_vec2(a_d, b_d, n) …
  end subroutine device_pwmin_vec2
 
  subroutine device_pwmin_vec3(a_d, b_d, c_d, n)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmin_vec3')
#elif HAVE_CUDA
    call cuda_pwmin_vec3(a_d, b_d, c_d, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmin_vec3')
#else
    call neko_error('No device backend configured')
#endif
 
  subroutine device_pwmin_vec3(a_d, b_d, c_d, n) …
  end subroutine device_pwmin_vec3
 
  subroutine device_pwmin_sca2(a_d, c, n)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmin_sca2')
#elif HAVE_CUDA
    call cuda_pwmin_sca2(a_d, c, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmin_sca2')
#else
    call neko_error('No device backend configured')
#endif
 
  subroutine device_pwmin_sca2(a_d, c, n) …
  end subroutine device_pwmin_sca2
 
  subroutine device_pwmin_sca3(a_d, b_d, c, n)
    type(c_ptr) :: a_d, b_d
    real(kind=rp), intent(in) :: c
    integer :: n
 
#if HAVE_HIP
    call neko_error('No HIP backend for device_pwmin_sca3')
#elif HAVE_CUDA
    call cuda_pwmin_sca3(a_d, b_d, c, n)
#elif HAVE_OPENCL
    call neko_error('No OpenCL backend for device_pwmin_sca3')
#else
    call neko_error('No device backend configured')
#endif
 
  subroutine device_pwmin_sca3(a_d, b_d, c, n) …
  end subroutine device_pwmin_sca3
 
end module device_math