Fortran calling C: How do I get an efficient vectorised function -

i have call c function fortran, want in vectorised loop. working intel 16.0.3 compilers on linux.

so options are: can try , function inline or can use simd function (i want use openmp simd this, want portable , use openmp).

if call fortran fortran works both ways. passing arguments use linear/ref clause pass reference vector of values rather vector of references , seems work efficiently. in c linear/ref clause not recognised.

i can function nominally vectorise inserting gathers , scatters , performance no better scalar (at least small test function).

if put linear(ref(r,s)) in fortran interface block message

the uval or ref modifiers must not used on dummy argument value attribute.

i can performance using trick of passing value fortran , returning value function return. produces vectorised function, , performance good, unfortunately real function needs return more 1 value.

if try inline c function, won't work. opt-report tells me callsite cannot inlined. true scalar function. cannot c functions inline fortran @ all. using ipo try , this. have wondered if inlining problem fortran passing pointer pointer? code gives right answers seems somehow acceptable.

the test code (passing pointers) essentially...

fortran caller

use, intrinsic :: iso_c_binding …. real*8, allocatable :: r(:),s(:)  . ….  interface  integer simd_c_func(r,s) bind(c, name="simd_c_func")  !$omp declare simd(simd_c_func)  import :: c_double  real(kind=c_double), intent(inout):: r,s end interface  allocate....   !$omp simd  i=1,n   ierror=simd_c_func(r(i),s(i)) enddo 

c callee

#pragma omp declare simd int simd_c_func(double *r, double *s) {  (*r)+=(*s);  return 0; } 

the linear(ref()) pretty new, 1 can hope linear using omp , contiguous can joy.

the message:

the uval or ref modifiers must not used on dummy argument value attribute."

indicates doing value , on stack. thought c arrays in heap? maybe need -heap switch?

c-side:

i suggest add linear #pragma

#pragma omp declare simd linear(r,s) 

have looked @ using #pragma vector aligned? , #pragma ivdep? assuming (*r) reference (in heap)

f90-side

i suggest replace:

real(8), allocatable :: r(:), s(:) 

with

!$dir attributes align:64                           :: r, s real(kind=8), dimension(:), allocatable, contiguous :: r, s 

the real change here 64bit alignment , contiguous. 1 can use switch -align array64byte

if ierror not array want one? or perhaps need use first/last private or reduction clause on ierror? use reduction.

!$omp simd goes to:

!$omp simd reduction(ierror) 

but mention need more 1 value maybe need allocate ierror?

and interested in how interface knows gets return value of zero. seem have have integer function somewhere in interface.

linking-compiling side:

if c produced f90- module use allow see deep enough c code inline or otherwise help. may need -ipo on compiler/link allow f90 understand can c callee.