device_math.F90

Go to the documentation of this file.

! Copyright (c) 2021-2026, 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 device_math
  use, intrinsic :: iso_c_binding, only : c_ptr, c_int
  use num_types, only : rp, xp, c_rp, c_xp
  use utils, only : neko_error
  use comm, only : neko_comm, pe_size, mpi_real_precision, mpi_extra_precision
  use mpi_f08, only : mpi_sum, mpi_min, mpi_max, mpi_in_place, mpi_allreduce
  use device, only : glb_cmd_queue
  ! ========================================================================== !
  ! Device math interfaces
 
  use hip_math
  use cuda_math
  use opencl_math
  use metal_math
 
  implicit none
  private
 
  interface device_cadd
     module procedure device_radd, device_iadd
  end interface device_cadd
 
  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_add4s3, device_add5s4, &
       device_invcol1, device_invcol2, device_col2, device_col3, &
       device_subcol3, device_sub2, device_sub3, device_addcol3, &
       device_addcol4, device_addcol3s2, device_vdot3, device_vlsc3, &
       device_glsc3, device_glsc3_many, device_add2s2_many, device_glsc2, &
       device_glsum, device_masked_copy_0, device_cfill_mask, &
       device_masked_gather_copy_aligned, device_face_masked_gather_copy_0, &
       device_masked_scatter_copy_aligned, device_masked_copy_aligned, &
       device_vcross, device_absval, device_masked_atomic_reduction_0, &
       device_masked_gather_copy_0, device_masked_scatter_copy_0, &
       device_invcol3, device_cdiv, device_cdiv2, device_glsubnorm, &
       device_pwmax2, device_pwmax3, device_cpwmax2, device_cpwmax3, &
       device_pwmin2, device_pwmin3, device_cpwmin2, device_cpwmin3, &
       device_glmax, device_glmin, device_glamax, device_cwrap, &
       device_sqrt_inplace, device_power
 
contains
 
  subroutine device_copy(a_d, b_d, n, strm)
    type(c_ptr) :: a_d, b_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_copy(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    call cuda_copy(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_copy(a_d, b_d, n, strm_)
#elif HAVE_METAL
    call metal_copy(a_d, b_d, n, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_copy
 
  subroutine device_masked_copy_0(a_d, b_d, mask_d, n, n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_masked_copy_0(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_masked_copy_0(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_masked_copy_0(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_METAL
    call metal_masked_copy_0(a_d, b_d, mask_d, n, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_masked_copy_0
 
  subroutine device_masked_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_masked_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_masked_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_masked_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_masked_copy_aligned
 
  subroutine device_masked_gather_copy_0(a_d, b_d, mask_d, n, n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_masked_gather_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_masked_gather_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_masked_gather_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_METAL
    call metal_masked_gather_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_masked_gather_copy_0
 
  subroutine device_face_masked_gather_copy_0(a_d, b_d, mask_d, facet_d, n1, &
       n2, lx, ly, lz, n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d, facet_d
    integer :: n1, n2, lx, ly, lz, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_face_masked_gather_copy(a_d, b_d, mask_d, facet_d, n1, n2, lx, &
         ly, lz, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_face_masked_gather_copy(a_d, b_d, mask_d, facet_d, n1, n2, lx, &
         ly, lz, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_face_masked_gather_copy(a_d, b_d, mask_d, facet_d, n1, n2, &
         lx, ly, lz, n_mask, strm_)
#elif HAVE_METAL
    call metal_face_masked_gather_copy(a_d, b_d, mask_d, facet_d, n1, n2, &
         lx, ly, lz, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_face_masked_gather_copy_0
 
  ! In this case, the mask comes from a mask_t type
  subroutine device_masked_gather_copy_aligned(a_d, b_d, mask_d, n, &
       n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_masked_gather_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_masked_gather_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_masked_gather_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_METAL
    call metal_masked_gather_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_masked_gather_copy_aligned
 
  subroutine device_masked_scatter_copy_0(a_d, b_d, mask_d, n, n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_masked_scatter_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_masked_scatter_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_masked_scatter_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_METAL
    call metal_masked_scatter_copy(a_d, b_d, mask_d, n, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_masked_scatter_copy_0
 
  ! In this case, the mask comes from a mask_t type
  subroutine device_masked_scatter_copy_aligned(a_d, b_d, mask_d, n, &
       n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_masked_scatter_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_masked_scatter_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_masked_scatter_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_METAL
    call metal_masked_scatter_copy_aligned(a_d, b_d, mask_d, n, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_masked_scatter_copy_aligned
 
  subroutine device_masked_atomic_reduction_0(a_d, b_d, mask_d, n, n_mask, strm)
    type(c_ptr) :: a_d, b_d, mask_d
    integer :: n, n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_masked_atomic_reduction(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_masked_atomic_reduction(a_d, b_d, mask_d, n, n_mask, strm_)
#elif HAVE_OPENCL
    call neko_error('No OpenCL bcknd, masked atomic reduction')
#elif HAVE_METAL
    call metal_masked_atomic_reduction(a_d, b_d, mask_d, n, n_mask, strm_)
#else
    call neko_error('no device backend configured')
#endif
  end subroutine device_masked_atomic_reduction_0
 
  subroutine device_cfill_mask(a_d, c, n, mask_d, n_mask, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr) :: mask_d
    integer :: n_mask
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. n_mask .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cfill_mask(a_d, c, n, mask_d, n_mask, strm_)
#elif HAVE_CUDA
    call cuda_cfill_mask(a_d, c, n, mask_d, n_mask, strm_)
#elif HAVE_OPENCL
    call opencl_cfill_mask(a_d, c, n, mask_d, n_mask, strm_)
#elif HAVE_METAL
    call metal_cfill_mask(a_d, c, n, mask_d, n_mask, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cfill_mask
 
  subroutine device_rzero(a_d, n, strm)
    type(c_ptr) :: a_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_rzero(a_d, n, strm_)
#elif HAVE_CUDA
    call cuda_rzero(a_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_rzero(a_d, n, strm_)
#elif HAVE_METAL
    call metal_rzero(a_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_rzero
 
  subroutine device_rone(a_d, n, strm)
    type(c_ptr) :: a_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
    real(kind=rp), parameter :: one = 1.0_rp
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP || HAVE_CUDA || HAVE_OPENCL || HAVE_METAL
    call device_cfill(a_d, one, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_rone
 
  subroutine device_cmult(a_d, c, n, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cmult(a_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_cmult(a_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_cmult(a_d, c, n, strm_)
#elif HAVE_METAL
    call metal_cmult(a_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cmult
 
  subroutine device_cmult2(a_d, b_d, c, n, strm)
    type(c_ptr) :: a_d, b_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cmult2(a_d, b_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_cmult2(a_d, b_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_cmult2(a_d, b_d, c, n, strm_)
#elif HAVE_METAL
    call metal_cmult2(a_d, b_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cmult2
 
  subroutine device_cdiv(a_d, c, n, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cdiv(a_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_cdiv(a_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_cdiv(a_d, c, n, strm_)
#elif HAVE_METAL
    call metal_cdiv(a_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cdiv
 
  subroutine device_cdiv2(a_d, b_d, c, n, strm)
    type(c_ptr) :: a_d, b_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cdiv2(a_d, b_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_cdiv2(a_d, b_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_cdiv2(a_d, b_d, c, n, strm_)
#elif HAVE_METAL
    call metal_cdiv2(a_d, b_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cdiv2
 
  subroutine device_radd(a_d, c, n, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_radd(a_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_radd(a_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_radd(a_d, c, n, strm_)
#elif HAVE_METAL
    call metal_radd(a_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_radd
 
  subroutine device_cadd2(a_d, b_d, c, n, strm)
    type(c_ptr) :: a_d
    type(c_ptr) :: b_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cadd2(a_d, b_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_cadd2(a_d, b_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_cadd2(a_d, b_d, c, n, strm_)
#elif HAVE_METAL
    call metal_cadd2(a_d, b_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cadd2
 
  subroutine device_cwrap(a_d, min_val, max_val, n, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: min_val, max_val
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1 .or. max_val .le. min_val) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cwrap(a_d, min_val, max_val, n, strm_)
#elif HAVE_CUDA
    call cuda_cwrap(a_d, min_val, max_val, n, strm_)
#elif HAVE_OPENCL
    call opencl_cwrap(a_d, min_val, max_val, n, strm_)
#elif HAVE_METAL
    call metal_cwrap(a_d, min_val, max_val, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cwrap
 
  subroutine device_sqrt_inplace(a_d, n, strm)
    type(c_ptr) :: a_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_sqrt_inplace(a_d, n, strm_)
#elif HAVE_CUDA
    call cuda_sqrt_inplace(a_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_sqrt_inplace(a_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_sqrt_inplace
 
  subroutine device_power(ap_d, a_d, p, n, strm)
    type(c_ptr) :: ap_d, a_d
    real(kind=rp), intent(in) :: p
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_power(ap_d, a_d, p, n, strm_)
#elif HAVE_CUDA
    call cuda_power(ap_d, a_d, p, n, strm_)
#elif HAVE_OPENCL
    call opencl_power(ap_d, a_d, p, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_power
 
  subroutine device_cfill(a_d, c, n, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional ::strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_cfill(a_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_cfill(a_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_cfill(a_d, c, n, strm_)
#elif HAVE_METAL
    call metal_cfill(a_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_cfill
 
  subroutine device_add2(a_d, b_d, n, strm)
    type(c_ptr) :: a_d, b_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    call cuda_add2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_add2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    call metal_add2(a_d, b_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add2
 
  subroutine device_add4(a_d, b_d, c_d, d_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d, d_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add4(a_d, b_d, c_d, d_d, n, strm_)
#elif HAVE_CUDA
    call cuda_add4(a_d, b_d, c_d, d_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_add4(a_d, b_d, c_d, d_d, n, strm_)
#elif HAVE_METAL
    call metal_add4(a_d, b_d, c_d, d_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add4
 
  subroutine device_add2s1(a_d, b_d, c1, n, strm)
    type(c_ptr) :: a_d, b_d
    real(kind=rp) :: c1
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add2s1(a_d, b_d, c1, n, strm_)
#elif HAVE_CUDA
    call cuda_add2s1(a_d, b_d, c1, n, strm_)
#elif HAVE_OPENCL
    call opencl_add2s1(a_d, b_d, c1, n, strm_)
#elif HAVE_METAL
    call metal_add2s1(a_d, b_d, c1, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add2s1
 
  subroutine device_add2s2(a_d, b_d, c1, n, strm)
    type(c_ptr) :: a_d, b_d
    real(kind=rp) :: c1
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add2s2(a_d, b_d, c1, n, strm_)
#elif HAVE_CUDA
    call cuda_add2s2(a_d, b_d, c1, n, strm_)
#elif HAVE_OPENCL
    call opencl_add2s2(a_d, b_d, c1, n, strm_)
#elif HAVE_METAL
    call metal_add2s2(a_d, b_d, c1, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add2s2
 
  subroutine device_addsqr2s2(a_d, b_d, c1, n, strm)
    type(c_ptr) :: a_d, b_d
    real(kind=rp) :: c1
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_addsqr2s2(a_d, b_d, c1, n, strm_)
#elif HAVE_CUDA
    call cuda_addsqr2s2(a_d, b_d, c1, n, strm_)
#elif HAVE_OPENCL
    call opencl_addsqr2s2(a_d, b_d, c1, n, strm_)
#elif HAVE_METAL
    call metal_addsqr2s2(a_d, b_d, c1, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_addsqr2s2
 
  subroutine device_add3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_add3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_add3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_add3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add3
 
  subroutine device_add3s2(a_d, b_d, c_d, c1, c2 , n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    real(kind=rp) :: c1, c2
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add3s2(a_d, b_d, c_d, c1, c2, n, strm_)
#elif HAVE_CUDA
    call cuda_add3s2(a_d, b_d, c_d, c1, c2, n, strm_)
#elif HAVE_OPENCL
    call opencl_add3s2(a_d, b_d, c_d, c1, c2, n, strm_)
#elif HAVE_METAL
    call metal_add3s2(a_d, b_d, c_d, c1, c2, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add3s2
 
  subroutine device_add4s3(a_d, b_d, c_d, d_d, c1, c2 , c3, n, strm)
    type(c_ptr) :: a_d, b_d, c_d, d_d
    real(kind=rp) :: c1, c2, c3
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add4s3(a_d, b_d, c_d, d_d, c1, c2, c3, n, strm_)
#elif HAVE_CUDA
    call cuda_add4s3(a_d, b_d, c_d, d_d, c1, c2, c3, n, strm_)
#elif HAVE_OPENCL
    call opencl_add4s3(a_d, b_d, c_d, d_d, c1, c2, c3, n, strm_)
#elif HAVE_METAL
    call metal_add4s3(a_d, b_d, c_d, d_d, c1, c2, c3, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add4s3
 
  subroutine device_add5s4(a_d, b_d, c_d, d_d, e_d, c1, c2 , c3, c4, n, strm)
    type(c_ptr) :: a_d, b_d, c_d, d_d, e_d
    real(kind=rp) :: c1, c2, c3, c4
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add5s4(a_d, b_d, c_d, d_d, e_d, c1, c2, c3, c4, n, strm_)
#elif HAVE_CUDA
    call cuda_add5s4(a_d, b_d, c_d, d_d, e_d, c1, c2, c3, c4, n, strm_)
#elif HAVE_OPENCL
    call opencl_add5s4(a_d, b_d, c_d, d_d, e_d, c1, c2, c3, c4, n, strm_)
#elif HAVE_METAL
    call metal_add5s4(a_d, b_d, c_d, d_d, e_d, c1, c2, c3, c4, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add5s4
 
  subroutine device_invcol1(a_d, n, strm)
    type(c_ptr) :: a_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_invcol1(a_d, n, strm_)
#elif HAVE_CUDA
    call cuda_invcol1(a_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_invcol1(a_d, n, strm_)
#elif HAVE_METAL
    call metal_invcol1(a_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_invcol1
 
  subroutine device_invcol2(a_d, b_d, n, strm)
    type(c_ptr) :: a_d, b_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_invcol2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    call cuda_invcol2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_invcol2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    call metal_invcol2(a_d, b_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_invcol2
 
  subroutine device_invcol3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#ifdef HAVE_HIP
    call hip_invcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_invcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_invcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_invcol3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_invcol3
 
  subroutine device_col2(a_d, b_d, n, strm)
    type(c_ptr) :: a_d, b_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
    if (n .lt. 1) return
#if HAVE_HIP
    call hip_col2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    call cuda_col2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_col2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    call metal_col2(a_d, b_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_col2
 
  subroutine device_col3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_col3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_col3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_col3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_col3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_col3
 
  subroutine device_subcol3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_subcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_subcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_subcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_subcol3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_subcol3
 
  subroutine device_sub2(a_d, b_d, n, strm)
    type(c_ptr) :: a_d, b_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_sub2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    call cuda_sub2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_sub2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    call metal_sub2(a_d, b_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_sub2
 
  subroutine device_sub3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_sub3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_sub3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_sub3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_sub3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_sub3
 
  subroutine device_addcol3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_addcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_addcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_addcol3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_addcol3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_addcol3
 
  subroutine device_addcol4(a_d, b_d, c_d, d_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d, d_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_addcol4(a_d, b_d, c_d, d_d, n, strm_)
#elif HAVE_CUDA
    call cuda_addcol4(a_d, b_d, c_d, d_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_addcol4(a_d, b_d, c_d, d_d, n, strm_)
#elif HAVE_METAL
    call metal_addcol4(a_d, b_d, c_d, d_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_addcol4
 
  subroutine device_addcol3s2(a_d, b_d, c_d, s, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    real(kind=rp) :: s
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_addcol3s2(a_d, b_d, c_d, s, n, strm_)
#elif HAVE_CUDA
    call cuda_addcol3s2(a_d, b_d, c_d, s, n, strm_)
#elif HAVE_OPENCL
    call opencl_addcol3s2(a_d, b_d, c_d, s, n, strm_)
#elif HAVE_METAL
    call metal_addcol3s2(a_d, b_d, c_d, s, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_addcol3s2
 
  subroutine device_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n, strm)
    type(c_ptr) :: dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n, strm_)
#elif HAVE_CUDA
    call cuda_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n, strm_)
#elif HAVE_METAL
    call metal_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  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, strm)
    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
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#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, strm_)
#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, strm_)
#elif HAVE_OPENCL
    call opencl_vcross(u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, &
         w1_d, w2_d, w3_d, n, strm_)
#elif HAVE_METAL
    call metal_vcross(u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, &
         w1_d, w2_d, w3_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_vcross
 
 
  function device_vlsc3(u_d, v_d, w_d, n, strm) result(res)
    type(c_ptr) :: u_d, v_d, w_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
    real(kind=rp) :: res
 
    res = 0.0_rp
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    res = hip_vlsc3(u_d, v_d, w_d, n, strm_)
#elif HAVE_CUDA
    res = cuda_vlsc3(u_d, v_d, w_d, n, strm_)
#elif HAVE_OPENCL
    ! Same kernel as glsc3 (currently no device MPI for OpenCL)
    res = opencl_glsc3(u_d, v_d, w_d, n, strm_)
#elif HAVE_METAL
    ! Same kernel as glsc3 (currently no device MPI for OpenCL)
    res = metal_glsc3(u_d, v_d, w_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end function device_vlsc3
 
  function device_glsc3(a_d, b_d, c_d, n, strm) result(res)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n, ierr
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
    real(kind=rp) :: res
    real(kind=xp) :: res_xp
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
    res_xp = 0.0_xp
#if HAVE_HIP
    res_xp = hip_glsc3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    res_xp = cuda_glsc3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    res_xp = opencl_glsc3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    res = metal_glsc3(a_d, b_d, c_d, n, strm_)
    res_xp = real(res, kind=xp)
#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_xp, 1, &
            mpi_extra_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
    res = real(res_xp, kind=rp)
  end function device_glsc3
 
  subroutine device_glsc3_many(h, w_d, v_d_d, mult_d, j, n, strm)
    type(c_ptr), value :: w_d, v_d_d, mult_d
    integer(c_int) :: j, n
    real(c_rp) :: h(j)
    real(c_xp) :: h_xp(j)
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
    integer :: ierr
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    h_xp = 0.0_c_xp
    call hip_glsc3_many(h_xp, w_d, v_d_d, mult_d, j, n, strm_)
#elif HAVE_CUDA
    h_xp = 0.0_c_xp
    call cuda_glsc3_many(h_xp, w_d, v_d_d, mult_d, j, n, strm_)
#elif HAVE_OPENCL
    h_xp = 0.0_c_xp
    call opencl_glsc3_many(h_xp, w_d, v_d_d, mult_d, j, n, strm_)
#elif HAVE_METAL
    call metal_glsc3_many(h, w_d, v_d_d, mult_d, j, n, strm_)
    h_xp = real(h, kind=c_xp)
#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_xp, j, &
            mpi_extra_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
    h = real(h_xp, kind=c_rp)
  end subroutine device_glsc3_many
 
  subroutine device_add2s2_many(y_d, x_d_d, a_d, j, n, strm)
    type(c_ptr), value :: y_d, x_d_d, a_d
    integer(c_int) :: j, n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_add2s2_many(y_d, x_d_d, a_d, j, n, strm_)
#elif HAVE_CUDA
    call cuda_add2s2_many(y_d, x_d_d, a_d, j, n, strm_)
#elif HAVE_OPENCL
    call opencl_add2s2_many(y_d, x_d_d, a_d, j, n, strm_)
#elif HAVE_METAL
    call metal_add2s2_many(y_d, x_d_d, a_d, j, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_add2s2_many
 
  function device_glsc2(a_d, b_d, n, strm) result(res)
    type(c_ptr) :: a_d, b_d
    integer :: n, ierr
    real(kind=rp) :: res
    real(kind=xp) :: res_xp
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
    res_xp = 0.0_xp
#if HAVE_HIP
    res_xp = hip_glsc2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    res_xp = cuda_glsc2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    res_xp = opencl_glsc2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    res = metal_glsc2(a_d, b_d, n, strm_)
    res_xp = real(res, kind=xp)
#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_xp, 1, &
            mpi_extra_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
    res = real(res_xp, kind=rp)
  end function device_glsc2
 
  function device_glsubnorm(a_d, b_d, n, strm) result(res)
    type(c_ptr), intent(in) :: a_d, b_d
    integer, intent(in) :: n
    integer :: ierr
    real(kind=rp) :: res
    real(kind=xp) :: res_xp
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
    res_xp = 0.0_xp
#if HAVE_HIP
    res_xp = hip_glsubnorm2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    res_xp = cuda_glsubnorm2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    res_xp = opencl_glsubnorm2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    res = metal_glsubnorm2(a_d, b_d, n, strm_)
    res_xp = real(res, kind=xp)
#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_xp, 1, &
            mpi_extra_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
 
    res = real(sqrt(res_xp), kind=rp)
  end function device_glsubnorm
 
  function device_glsum(a_d, n, strm) result(res)
    type(c_ptr) :: a_d
    integer :: n, ierr
    real(kind=rp) :: res
    real(kind=xp) :: res_xp
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
    res_xp = 0.0_xp
#if HAVE_HIP
    res_xp = hip_glsum(a_d, n, strm_)
#elif HAVE_CUDA
    res_xp = cuda_glsum(a_d, n, strm_)
#elif HAVE_OPENCL
    res_xp = opencl_glsum(a_d, n, strm_)
#elif HAVE_METAL
    res = metal_glsum(a_d, n, strm_)
    res_xp = real(res, kind=xp)
#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_xp, 1, &
            mpi_extra_precision, mpi_sum, neko_comm, ierr)
    end if
#endif
    res = real(res_xp, kind=rp)
  end function device_glsum
 
  function device_glmax(a_d, n, strm) result(res)
    type(c_ptr) :: a_d
    integer :: n, ierr
    real(kind=rp) :: res, ninf
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) then
       res = -huge(0.0_rp)
       return
    end if
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
    ninf = -huge(0.0_rp)
#if HAVE_HIP
    res = hip_glmax(a_d, ninf, n, strm_)
#elif HAVE_CUDA
    res = cuda_glmax(a_d, ninf, n, strm_)
#elif HAVE_OPENCL
    res = opencl_glmax(a_d, n, strm_)
#elif HAVE_METAL
    res = metal_glmax(a_d, ninf, n, strm_)
#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_max, neko_comm, ierr)
    end if
#endif
  end function device_glmax
 
  function device_glmin(a_d, n, strm) result(res)
    type(c_ptr) :: a_d
    integer :: n, ierr
    real(kind=rp) :: res, pinf
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) then
       res = huge(0.0_rp)
       return
    end if
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
    pinf = huge(0.0_rp)
#if HAVE_HIP
    res = hip_glmin(a_d, pinf, n, strm_)
#elif HAVE_CUDA
    res = cuda_glmin(a_d, pinf, n, strm_)
#elif HAVE_OPENCL
    res = opencl_glmin(a_d, n, strm_)
#elif HAVE_METAL
    res = metal_glmin(a_d, pinf, n, strm_)
#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_min, neko_comm, ierr)
    end if
#endif
  end function device_glmin
 
  function device_glamax(a_d, n, strm) result(res)
    type(c_ptr) :: a_d
    integer :: n, ierr
    real(kind=rp) :: res
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) then
       res = 0.0_rp
       return
    end if
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    res = hip_glamax(a_d, n, strm_)
#elif HAVE_CUDA
    res = cuda_glamax(a_d, n, strm_)
#elif HAVE_OPENCL
    res = opencl_glamax(a_d, n, strm_)
#elif HAVE_METAL
    res = metal_glamax(a_d, n, strm_)
#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_max, neko_comm, ierr)
    end if
#endif
  end function device_glamax
 
  subroutine device_absval(a_d, n, strm)
    integer, intent(in) :: n
    type(c_ptr) :: a_d
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#ifdef HAVE_HIP
    call hip_absval(a_d, n, strm_)
#elif HAVE_CUDA
    call cuda_absval(a_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_absval(a_d, n, strm_)
#elif HAVE_METAL
    call metal_absval(a_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
 
  end subroutine device_absval
 
  ! ========================================================================== !
  ! Device point-wise max
 
  subroutine device_pwmax2(a_d, b_d, n, strm)
    type(c_ptr) :: a_d, b_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmax_vec2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmax_vec2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmax_vec2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    call metal_pwmax_vec2(a_d, b_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_pwmax2
 
  subroutine device_pwmax3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmax_vec3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmax_vec3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmax_vec3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_pwmax_vec3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
 
  end subroutine device_pwmax3
 
  subroutine device_cpwmax2(a_d, c, n, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmax_sca2(a_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmax_sca2(a_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmax_sca2(a_d, c, n, strm_)
#elif HAVE_METAL
    call metal_pwmax_sca2(a_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
 
  end subroutine device_cpwmax2
 
  subroutine device_cpwmax3(a_d, b_d, c, n, strm)
    type(c_ptr) :: a_d, b_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmax_sca3(a_d, b_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmax_sca3(a_d, b_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmax_sca3(a_d, b_d, c, n, strm_)
#elif HAVE_METAL
    call metal_pwmax_sca3(a_d, b_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
 
  end subroutine device_cpwmax3
 
  ! ========================================================================== !
  ! Device point-wise min
 
  subroutine device_pwmin2(a_d, b_d, n, strm)
    type(c_ptr) :: a_d, b_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmin_vec2(a_d, b_d, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmin_vec2(a_d, b_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmin_vec2(a_d, b_d, n, strm_)
#elif HAVE_METAL
    call metal_pwmin_vec2(a_d, b_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_pwmin2
 
  subroutine device_pwmin3(a_d, b_d, c_d, n, strm)
    type(c_ptr) :: a_d, b_d, c_d
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmin_vec3(a_d, b_d, c_d, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmin_vec3(a_d, b_d, c_d, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmin_vec3(a_d, b_d, c_d, n, strm_)
#elif HAVE_METAL
    call metal_pwmin_vec3(a_d, b_d, c_d, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
 
  end subroutine device_pwmin3
 
  subroutine device_cpwmin2(a_d, c, n, strm)
    type(c_ptr) :: a_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmin_sca2(a_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmin_sca2(a_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmin_sca2(a_d, c, n, strm_)
#elif HAVE_METAL
    call metal_pwmin_sca2(a_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
 
  end subroutine device_cpwmin2
 
  subroutine device_cpwmin3(a_d, b_d, c, n, strm)
    type(c_ptr) :: a_d, b_d
    real(kind=rp), intent(in) :: c
    integer :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
 
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_pwmin_sca3(a_d, b_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_pwmin_sca3(a_d, b_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_pwmin_sca3(a_d, b_d, c, n, strm_)
#elif HAVE_METAL
    call metal_pwmin_sca3(a_d, b_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
 
  end subroutine device_cpwmin3
 
  ! ========================================================================== !
  ! Integer operations
 
  subroutine device_iadd(a_d, c, n, strm)
    type(c_ptr), intent(inout) :: a_d
    integer, intent(in) :: c
    integer, intent(in) :: n
    type(c_ptr), optional :: strm
    type(c_ptr) :: strm_
    if (n .lt. 1) return
 
    if (present(strm)) then
       strm_ = strm
    else
       strm_ = glb_cmd_queue
    end if
 
#if HAVE_HIP
    call hip_iadd(a_d, c, n, strm_)
#elif HAVE_CUDA
    call cuda_iadd(a_d, c, n, strm_)
#elif HAVE_OPENCL
    call opencl_iadd(a_d, c, n, strm_)
#elif HAVE_METAL
    call metal_iadd(a_d, c, n, strm_)
#else
    call neko_error('No device backend configured')
#endif
  end subroutine device_iadd
 
 
end module device_math