error: OpenMP array shaping operation is not allowed here

[jinz2014]

I tried amdclang to compile the program array_shaping.1.c, and got the error message

array_shaping.1.c:28:39: error: OpenMP array shaping operation is not allowed here
28 | #pragma omp target update from( (([nx][ny+2])a)[0:nx][1],
| ^
array_shaping.1.c:29:39: error: OpenMP array shaping operation is not allowed here
29 | (([nx][ny+2])a)[0:nx][ny] )
| ^

Can you suggest which compilers and options for the program ?

[milfeld]

In godbolt.org I checked nvc++ 25.7 and gcc 15.2 and they appear not to have implemented this feature yet, they didn't understand the shaping syntax.

nvc++ -fopenmp -mp=gpu *.cpp
v25.7  NOT compiled
"<source>", line 27: error: invalid text in pragma
    #pragma omp target update from( (([Nx][Ny+2])a)[0:Nx][1],   (([Nx][Ny+2])a)[0:Nx][ny] )
                                    ^

I have replicated the problem. The x86-64 clang++ compilers 14->21 complained that it is "not allowed here" for the example, which is modeled after the use case: updating discontiguous elements for a target region. I have created a ticket in the examples-internal repo to discuss this issue.

$ clang++ -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda snip.cpp
snip.cpp:29:37: error: OpenMP array shaping operation is not allowed here
   29 |     #pragma omp target update from( (([Nx][Ny+2])a)[0:Nx][1],   (([Nx][Ny+2])a)[0:Nx][Ny] )
      |                                     ^

I have been using this test code as a quick test:

#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
#include <string.h>

#define Nx 4
#define Ny 4

#pragma begin target declare
void change_ghost(int *a, int nx, int ny){
  for(int i=0; i<nx; i++){
    a[i*(ny+2)+1]  += 2;
    a[i*(ny+2)+ny] += 2;
  }
}
#pragma end target declare

int main(){
  int *a;
  a = (int*)malloc(sizeof(int)*Nx*(Ny+2));
  memset(a, 0,     sizeof(int)*Nx*(Ny+2));
  #pragma omp target data map( a[0:Nx*(Ny+2)] )
  {
    #pragma omp target
    change_ghost(a,Nx,Ny);

//  #pragma omp target update from( a[0:Nx*(Ny+2)] )
//  #pragma omp target update from( ([Nx][Ny+2])a )
    #pragma omp target update from( (([Nx][Ny+2])a)[0:Nx][1],   (([Nx][Ny+2])a)[0:Nx][Ny] )
//  #pragma omp target update from( (([Nx][Ny+2])a)[0:Nx][1:1], (([Nx][Ny+2])a)[0:Nx][Ny:1] )
    
    printf(" a[1], a[Ny] =  %d %d\n", a[1],a[Ny]);
  }
  printf(" a[1], a[Ny] =  %d %d\n", a[1],a[Ny]);
  return 0;
}