[Flang] `read` statement is slower than Gfortran in SPEC CPU 2017

[s-watanabe314]

When measuring 649.fotonik3d_s from SPEC CPU 2017 on the Grace (AArch64) machine, the speed of LLVM was slower than GCC as the number of threads increased. The options specified were -Ofast -mcpu=neoverse-v2.

According to our investigation, one of the reasons seems to be that the read statement, called in the non-parallelized initialization process, is slower than Gfortran.
The read statement is called approximately 54 million times within the nested loop, so its impact is significant.

The difference in processing time of the read statement was verified using the test program test.f90. This test was measured using LLVM 20.1.0 built with the build option -DCMAKE_BUILD_TYPE=Release.
To minimize the impact of file systems like NFS, the input file test.dat was placed in the machine's local directory.
test.dat can be generated with the following command:

python3 -c 'for i in range(54000000): print(i,i,i)' > test.dat

• test.f90

program main
  integer, parameter :: num = 54000000

  integer, dimension(num) :: a,b,c
  integer :: t1,t2, count, countmax
  integer :: ii

  open(unit=9, file='test.dat', status='old', IOSTAT=ios)
  call system_clock(t1, count, countmax)

  do ii=1,num
    read(9,*) a(ii),b(ii),c(ii)
  end do

  call system_clock(t2)
  print*, "init time : ",real(t2 - t1)/count,"sec"

  close(9)
end program main

• Compilation commands

$ flang-new -O3 -ffast-math -mcpu=neoverse-v2 test.f90
$ gfortran -O3 -ffast-math -mcpu=neoverse-v2 test.f90

• Measurement results

	version	time
GCC	14.2.0	22.496 [s]
LLVM (Release build)	20.1.0	42.341 [s]

The read statement appears to be nearly twice as slow in Flang compared to Gfortran.

[klausler]

Please try the patch in #134715 and tell me whether it improves performance in your environment to an acceptable level.

[s-watanabe314]

I re-measured test.f90 in my environment, and the results are as follows:

Gfortran 14.2.0	19.71 [s]
LLVM 20.1.0	43.98 [s]
Patch #134715	32.96 [s]

It seems to be still slower than Gfortran, but it has improved. Thank you.

[llvmbot]

@llvm/issue-subscribers-flang-runtime

Author: None (s-watanabe314)

When measuring 649.fotonik3d_s from SPEC CPU 2017 on the Grace (AArch64) machine, the speed of LLVM was slower than GCC as the number of threads increased. The options specified were `-Ofast -mcpu=neoverse-v2`.

According to our investigation, one of the reasons seems to be that the read statement, called in the non-parallelized initialization process, is slower than Gfortran.
The read statement is called approximately 54 million times within the nested loop, so its impact is significant.

The difference in processing time of the read statement was verified using the test program test.f90. This test was measured using LLVM 20.1.0 built with the build option -DCMAKE_BUILD_TYPE=Release.
To minimize the impact of file systems like NFS, the input file test.dat was placed in the machine's local directory.
test.dat can be generated with the following command:

python3 -c 'for i in range(54000000): print(i,i,i)' > test.dat

• test.f90

program main
  integer, parameter :: num = 54000000

  integer, dimension(num) :: a,b,c
  integer :: t1,t2, count, countmax
  integer :: ii

  open(unit=9, file='test.dat', status='old', IOSTAT=ios)
  call system_clock(t1, count, countmax)

  do ii=1,num
    read(9,*) a(ii),b(ii),c(ii)
  end do

  call system_clock(t2)
  print*, "init time : ",real(t2 - t1)/count,"sec"

  close(9)
end program main

• Compilation commands

$ flang-new -O3 -ffast-math -mcpu=neoverse-v2 test.f90
$ gfortran -O3 -ffast-math -mcpu=neoverse-v2 test.f90

• Measurement results

	version	time
GCC	14.2.0	22.496 [s]
LLVM (Release build)	20.1.0	42.341 [s]

The read statement appears to be nearly twice as slow in Flang compared to Gfortran.