host_data use_device with enter data/exit data on 15.3

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

I’m using 15.3 and the following code snippet

!$acc data copy(thermal_energy_updated), copy(planetary_thermal_energy), copy(surface_thermal_energy), &
!$acc& copy(thermal_energy_current)

  time = start_time
  call time_profiling_ini()
  call getTime(t_start_output_cycle)
  call getTime(t_start_output_cycle_w_io)
  do while (.true.)
   if (modulo(time + 0.001d0, output_timestep) < 0.01d0) then
   call write_data(thermal_energy_current, "thermal_energy", time)
   end if
   call run_physics(thermal_energy_current, surface_thermal_energy, planetary_thermal_energy)
   if (use_c) then



!$acc host_data use_device(thermal_energy_current,thermal_energy_updated)
   call diffuse_c(thermal_energy_updated, thermal_energy_current, nx, ny, nz)
!$acc end host_data



   else
   call diffuse (thermal_energy_updated, thermal_energy_current)
   end if
   thermal_energy_temp => thermal_energy_updated
   thermal_energy_updated => thermal_energy_current
   thermal_energy_current => thermal_energy_temp
   time = time + timestep
  end do
!$acc end data

compiles fine, but the same one with enter/exit data does not:

!$acc enter data copyin(thermal_energy_updated), copyin(planetary_thermal_energy), copyin(surface_thermal_energy), &
!$acc& copyin(thermal_energy_current)

  time = start_time
  call time_profiling_ini()
  call getTime(t_start_output_cycle)
  call getTime(t_start_output_cycle_w_io)
  do while (.true.)
   if (modulo(time + 0.001d0, output_timestep) < 0.01d0) then
   call write_data(thermal_energy_current, "thermal_energy", time)
   end if
   call run_physics(thermal_energy_current, surface_thermal_energy, planetary_thermal_energy)
   if (use_c) then
!$acc host_data use_device(thermal_energy_current,thermal_energy_updated)
   call diffuse_c(thermal_energy_updated, thermal_energy_current, nx, ny, nz)
!$acc end host_data
   else
   call diffuse (thermal_energy_updated, thermal_energy_current)
   end if
   thermal_energy_temp => thermal_energy_updated
   thermal_energy_updated => thermal_energy_current
   thermal_energy_current => thermal_energy_temp
   time = time + timestep
   if (time > end_time) then
!$acc exit data copyout(thermal_energy_updated), copyout(planetary_thermal_energy), copyout(surface_thermal_energy), &
!$acc& copyout(thermal_energy_current)
   return
   end if
  end do
!$acc exit data copyout(thermal_energy_updated), copyout(planetary_thermal_energy), copyout(surface_thermal_energy), &
!$acc& copyout(thermal_energy_current)

Compiler output:

pgf90 -Mcuda=cc3x -fast -ta=nvidia,cc3x -Minfo=accel,inline,ipa -Mneginfo -DGPU -c simple_weather.f90 -o simple_weather.o
PGF90-S-0528-Argument number 1 to diffuse_c: device attribute mismatch (simple_weather.f90: 97)
PGF90-S-0528-Argument number 2 to diffuse_c: device attribute mismatch (simple_weather.f90: 97)
  0 inform,   0 warnings,   2 severes, 0 fatal for simulate

Am I guessing correctly that this was fixed somewhere in between 15.3 and 15.7? At least I’ve seen that working on 15.7 with pgc++ at the Lugano Hackathon.

2

There were a couple of fixes to host_data that went into PGI 15.4. This case likely falls into one of those.

3

I was just able to test on Linux PGI 15.7 - the compiler error is still there. I’ve checked again how it worked in our hackathon C++ project: LATfield2_openACC/LATfield2_PlanFFT_ACC.hpp at master · daverio/LATfield2_openACC · GitHub. Then I saw the problem: It requires an “$acc data present” region around the host_data region. A bit clunky I have to say - ideally I think it should work with unstructured data regions as well, no?

4

I’ve been trying to recreate small reproducers for your problem and the hackathon issue and having trouble on both accounts. Can you generate a small reproducer?