Modulo not callable from target regions for int64

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

Compiler: nvfortran 26.3-0 (NVHPC 26.3), Linux x86_64

Summary

The Fortran modulo intrinsic applied to integer(int64) operands fails to compile when used inside
an OpenMP target or OpenACC compute region. The compiler emits:

  NVFORTRAN-S-1058-Call to Compiler runtime function not supported - pgf90_i8modulov_i8

The same call compiles and runs fine on the host, and mod(int64,int64) works inside the target
region — only modulo is affected.

Reproducer (OpenMP)

  program mre_omp_modulo_int64
      use iso_fortran_env, only: int64
      implicit none
      integer(int64) :: a = 10_int64, b = 3_int64, r = 0_int64
      !$omp target map(tofrom: r) map(to: a, b)
      r = modulo(a, b)
      !$omp end target
      print *, 'r = ', r
  end program

  $ nvfortran -mp=gpu -gpu=mem:separate mre_omp.F90
  NVFORTRAN-S-1058-Call to Compiler runtime function not supported - pgf90_i8modulov_i8

Reproducer (OpenACC) — same body, replacing the directives with:

  !$acc serial copy(r) copyin(a, b)
  r = modulo(a, b)
  !$acc end serial

  $ nvfortran -acc=gpu -gpu=mem:separate mre_acc.F90
  NVFORTRAN-S-1058-Call to Compiler runtime function not supported - pgf90_i8modulov_i8

Trigger boundaries

The bug requires all three of: the modulo intrinsic (not mod), int64 operands, and a target/compute
region. int32 is correctly lowered to a device-supported routine; the int64 variant apparently
expects a host-only pgf90_i8modulov_i8 runtime helper that has no device implementation.

Expected

modulo(int64, int64) should be supported on the device, like mod(int64, int64) and modulo(int32,
int32) already are.

Workaround

Open-code it in terms of mod:

  r = mod(a, b)
  if (r /= 0_int64 .and. (r < 0_int64) .neqv. (b < 0_int64)) r = r + b

See here for the full code: fortran-bugs/modulo_int64_target at main · JorgeG94/fortran-bugs · GitHub

2

Hi JorgeG94,

Not all Fortran intrinsics are supported on the device and depends on what’s available on the hardware and with CUDA.

I’m not sure if this particular feature is something that we can support of not, but I’ve added a request, TPR #38537, and asked engineering to see what they can do.

-Mat

3

thanks Mat!