problem with inlining type bound procedures

renato_v_ — Thu, 17 Sep 2015 08:24:41 GMT

I'm trying to vectorize this loop using !$OMP SIMD instruction:

!$OMP SIMD private(drx, dry, drz, dist, delta_i_sum_Wab) reduction(+:i_sum_Wab)

loop_j2_count: do kk=1,npart_cell

            call add_particle%distance(particle_in_cell(kk), container_size,2.*block%global%h, dist, drx, dry, drz)

            call particle_interaction(add_particle, particle_in_cell(kk),visc_dt,dist, drx,dry,drz,&
                                           block%global%cs0,block%global%p_backGround,block%global%rho0,block%global%const_ker,block%global%viscos_val, block%global%csi,.false., &
                                           delta_i_ax, delta_i_ay, delta_i_az, delta_i_ar, delta_i_deltax, delta_i_deltay, delta_i_deltaz, delta_i_div_pos, delta_i_sum_Wab, delta_i_n_neibou, shifting_enabled)
#ifdef SUMWAB
                i_sum_Wab=i_sum_Wab+delta_i_sum_Wab
#endif

enddo loop_j2_count
!$OMP END SIMD

the problem seems to be the type-bound procedure distance which is defined as follows:

pure subroutine distance(vec1,vec2,container_size, cell_size, dist, drx, dry, drz)

implicit none

class(Type_particle_or), intent(in):: vec1
class(Type_particle_or), intent(in):: vec2
real(R8P), intent(in) :: container_size(3)
real(R_P), intent(in) :: cell_size
real(R8P), intent(out) :: dist
real(R8P), intent(out) :: drx, dry, drz
logical :: plot

drx=vec1%xp-vec2%xp
dry=vec1%yp-vec2%yp
drz=vec1%zp-vec2%zp

plot=.false.
call periodicityCorrection(drx,container_size(1),cell_size, plot)
call periodicityCorrection(dry,container_size(2),cell_size, plot)
call periodicityCorrection(drz,container_size(3),cell_size, plot)

dist=sqrt(drx*drx+dry*dry+drz*drz)

end subroutine distance

this distance subroutine is bound to the following type:

type, public:: Type_particle_or

real(R8P)::                            xp = 0._R8P !< x coords.
real(R8P)::                            yp = 0._R8P !< y coords.
real(R8P)::                            zp = 0._R8P !< z coords.
real(R_P)::                            up = 0._R_P !< v_x
real(R_P)::                            vp = 0._R_P !< v_y
real(R_P)::                            wp = 0._R_P !< v_z
real(R_P)::                            pm = 0._R_P !< mass
real(R_P)::                            rhop = 0._R_P !< density
real(R_P)::                            p = 0._R_P !< pressure
real(R_P)::                            hp = 0._R_P !< smoothing length

integer(I_P) ::                        i=0_I_P

contains

procedure :: assign_particle => assign_particle_or
procedure :: modify_size_particle_or
procedure :: plot_particle
procedure, public :: distance

end type Type_particle_or

Looking at the optimization report file produced by the compiler with flags -qopt-report=5 -qopt-report-phase=all it seems that the compiler is not able to inline type bound procedures and this creates problems when I try to vectorize the code.

thank you very much for your help

Is there anyone who can help

renato_v_ — Fri, 18 Sep 2015 10:18:39 GMT

Is there anyone who can help me with this?

thank you very much

Renato

I'll try to find someone to

Kevin_D_Intel — Fri, 18 Sep 2015 15:09:40 GMT

I'll try to find someone to help. A few questions:

- Which compiler are you using?
- I don't know if its needed, but I'll ask in advance in case someone asks me, is it possible to isolate this into at least a compile-able reproducer?
- Could you provide us wit the optimization report file you mentioned so we can see those details?

One initial thought offered,

Kevin_D_Intel — Fri, 18 Sep 2015 16:00:49 GMT

One initial thought offered, assuming the reason was not the vectorizer refusing to vectorize the loop, was to try !DIR$ FORCEINLINE RECURSIVE

topic I'll try to find someone to in Intel® Fortran Compiler

problem with inlining type bound procedures

Is there anyone who can help

I'll try to find someone to

One initial thought offered,