[Flang][RT][OpenMP] system_clock behavior with multiple threads

[vzakhari]

Reproducer:

program main
  use omp_lib
  integer, parameter :: n = 1000000
  integer(8) :: t1, t2, rate
  real :: a(n), b(n)
  real(8) :: start, end
  b = 1.0
  call system_clock(t1)
  start = omp_get_wtime()
  call work(a, b, n)
  call system_clock(t2, rate)
  end = omp_get_wtime()
  print *, 'system_clock: ', dble(t2-t1)/dble(rate)
  print *, 'omp_get_wtime: ', end-start
contains
  subroutine work(a, b, n)
    real :: a(:), b(:)
    !$omp parallel do
    do i=1,n
       a(i) = sin(b(i))
    end do
    !$omp end parallel do
  end subroutine work
end program main

Compile with: flang-new -fopenmp clock.f90 -O0
Run with OMP_NUM_THREADS=1 ./a.out and OMP_NUM_THREADS=128 ./a.out

In the multithreaded case, the program prints the wall time multiplied by the number of threads. I guess this might be an acceptable behavior, but many other compilers return the wall time.

I think this might be related to our preference of CLOCK_PROCESS_CPUTIME_ID over other timers:

llvm-project/flang/runtime/time-intrinsic.cpp

Line 64 in e3720bb

#elif defined CLOCK_PROCESS_CPUTIME_ID

Note that gcc, for example, tries to use CLOCK_MONOTONIC and then CLOCK_REALTIME: https://github.com/gcc-mirror/gcc/blob/9693459e030977d6e906ea7eb587ed09ee4fddbd/libgfortran/intrinsics/system_clock.c#L39

@rovka , @Leporacanthicus, you've made changes in this code - would you agree that we should change the order of the ifdefs to match other compilers?

[llvmbot]

@llvm/issue-subscribers-flang-runtime

Author: Slava Zakharin (vzakhari)

Reproducer: ``` program main use omp_lib integer, parameter :: n = 1000000 integer(8) :: t1, t2, rate real :: a(n), b(n) real(8) :: start, end b = 1.0 call system_clock(t1) start = omp_get_wtime() call work(a, b, n) call system_clock(t2, rate) end = omp_get_wtime() print *, 'system_clock: ', dble(t2-t1)/dble(rate) print *, 'omp_get_wtime: ', end-start contains subroutine work(a, b, n) real :: a(:), b(:) !$omp parallel do do i=1,n a(i) = sin(b(i)) end do !$omp end parallel do end subroutine work end program main ```

Compile with: flang-new -fopenmp clock.f90 -O0
Run with OMP_NUM_THREADS=1 ./a.out and OMP_NUM_THREADS=128 ./a.out

In the multithreaded case, the program prints the wall time multiplied by the number of threads. I guess this might be an acceptable behavior, but many other compilers return the wall time.

I think this might be related to our preference of CLOCK_PROCESS_CPUTIME_ID over other timers:

llvm-project/flang/runtime/time-intrinsic.cpp

Line 64 in e3720bb

#elif defined CLOCK_PROCESS_CPUTIME_ID

Note that gcc, for example, tries to use CLOCK_MONOTONIC and then CLOCK_REALTIME: https://github.com/gcc-mirror/gcc/blob/9693459e030977d6e906ea7eb587ed09ee4fddbd/libgfortran/intrinsics/system_clock.c#L39

@rovka , @Leporacanthicus, you've made changes in this code - would you agree that we should change the order of the ifdefs to match other compilers?

[foxtran]

Close to #74202

[vzakhari]

Close to #74202

Thanks! I should have used the search.

[mgjf]

The Fortran standard clearly states that (quote) SYSTEM_CLOCK [...] returns numeric data from a real-time clock. So this is not open to interpretation.

[tblah]

On my system I get only a very small time difference with 48 thread enabled:

 system_clock:   4.45249E-03
 omp_get_wtime: 4.43720817565918E-03

@vzakhari is this fixed for you now?