ifx and ifort comparison for do concurrent

Fortran10 — Tue, 06 Aug 2024 06:01:10 GMT

below is my do concurrent code:

Question 1) Why the code does not run in parallel with ifx? Apparently, I do not see any error or warning !

program dc_test implicit none !==== integer(4),parameter::Nx=300,Ny=300 real(8) :: x=0.03d0,y=0.03d0,t_final real(8), parameter :: one=1.0,two=2.0,four=4.0,half=0.5 integer(4) ::steps,i,j,ip,im,jp,jm,t_start,t_end,rate real(8):: dt=1.0d-4,tau=0.0003d0,ep=0.01d0,ka=1.8d0,seed=5.0d0,ppo,t1,t2,th,m real(8):: del=0.02d0,an=6.0d0,al=0.9d0,ga=10.0d0,te=1.0d0,t0=0.2d0,pix=four*atan(one) real(8),dimension(Nx,Ny) :: pp,tt,lpp,ltt,ppx,ppy,eps,deps !==== pp = 0.0 tt = 0.0 do i = 1, Nx do j = 1, Ny if ((i-Nx/two)*(i-Nx/two)+(j-Ny/two)*(j-Ny/two)<seed)pp(i,j)=one end do end do !==== call system_clock (count=t_start, count_rate=rate) do steps = 1,1000 do concurrent (integer::j=1:Ny,i=1:Nx) default (none) & local ( ip,im,jp,jm,th ) & shared ( x,y,lpp,ltt,pp,tt,ppx,ppy,eps,deps,t0,an,del,ep ) jp = j + 1 jm = j - 1 ip = i + 1 im = i - 1 if ( im == 0 ) im = Nx if ( ip == ( Nx + 1) ) ip = 1 if ( jm == 0 ) jm = Ny if ( jp == ( Ny + 1) ) jp = 1 lpp(i,j) = (pp(ip,j)+pp(im,j)+pp(i,jm)+pp(i,jp)-four*pp(i,j))/(x*y) ltt(i,j) = (tt(ip,j)+tt(im,j)+tt(i,jm)+tt(i,jp)-four*tt(i,j))/(x*y) ppx(i,j) = (pp(ip,j) - pp(im,j))/x ppy(i,j) = (pp(i,jp) - pp(i,jm))/y th = atan2( ppy(i,j),ppx(i,j) ) eps(i,j) = ep*(one+del*cos(an*(th-t0))) deps(i,j) = -ep*an*del*sin(an*(th-t0)) end do do concurrent (integer::j=1:Ny,i=1:Nx) default (none) & local( i,j,ip,im,jp,jm,ppo,t1,t2,m) & shared(x,y,pp,tt,eps,deps,ppx,ppy,lpp,ltt,al,pix,te,ga,dt,tau,ka) jp = j + 1 jm = j - 1 ip = i + 1 im = i - 1 if ( im == 0 ) im = Nx if ( ip == ( Nx + 1) ) ip = 1 if ( jm == 0 ) jm = Ny if ( jp == ( Ny + 1) ) jp = 1 ppo = pp(i,j) t1 = ( eps(i,jp)*deps(i,jp)*ppx(i,jp) - eps(i,jm)*deps(i,jm)*ppx(i,jm) ) / y t2 = -( eps(ip,j)*deps(ip,j)*ppy(ip,j) - eps(im,j)*deps(im,j)*ppy(im,j) ) / x m = al/pix*atan(ga*(te-tt(i,j))) pp(i,j) = pp(i,j)+(dt/tau)*(t1+t2+eps(i,j)**2*lpp(i,j) ) & + ppo*(one-ppo)*(ppo-half+m) tt(i,j) = tt(i,j)+dt*ltt(i,j)+ka*(pp(i,j)-ppo) end do end do call system_clock (count=t_end) t_final = real(max(t_end-t_start,1_8))/real(rate) print*, t_final end program dc_test

ifx test

No parallelization to be seen with the output time

>ifx main_dc.f90 /Qopenmp /F500000000 Intel(R) Fortran Compiler for applications running on Intel(R) 64, Version 2024.2.0 Build 20240602 Copyright (C) 1985-2024 Intel Corporation. All rights reserved. Microsoft (R) Incremental Linker Version 14.32.31332.0 Copyright (C) Microsoft Corporation. All rights reserved. -out:main_dc.exe -subsystem:console -stack:500000000 -defaultlib:libiomp5md.lib -nodefaultlib:vcomp.lib -nodefaultlib:vcompd.lib main_dc.obj >main_dc 4.00899982452393

ifort test

parallelization can be confirmed with the output time

>ifort main_dc.f90 /Qopenmp /F500000000 Intel(R) Fortran Intel(R) 64 Compiler Classic for applications running on Intel(R) 64, Version 2021.13.0 Build 20240602_000000 Copyright (C) 1985-2024 Intel Corporation. All rights reserved. ifort: remark #10448: Intel(R) Fortran Compiler Classic (ifort) is now deprecated and will be discontinued late 2024. Intel recommends that customers transition now to using the LLVM-based Intel(R) Fortran Compiler (ifx) for continued Windows* and Linux* support, new language support, new language features, and optimizations. Use '/Qdiag-disable:10448' to disable this message. Microsoft (R) Incremental Linker Version 14.32.31332.0 Copyright (C) Microsoft Corporation. All rights reserved. -out:main_dc.exe -subsystem:console -stack:500000000 -defaultlib:libiomp5md.lib -nodefaultlib:vcomp.lib -nodefaultlib:vcompd.lib main_dc.obj >main_dc 0.282000005245209

Question 2) According to the intel documentation, rules for variable-name in a locality-spec: "variable-name can not be the same as index-name of the same do concurrent statement." So the code is expected to show error or warning because in the concurrent statement

do concurrent (integer::j=1:Ny,i=1:Nx) default (none) & local( i,j,ip,im,jp,jm,ppo,t1,t2,m) ...

i and j are index-name, I think.

Question 3) The option /Qopt-report-phase:openmp does not work with ifx but with ifort. So how to get the report to check if ifx has successfully parallelized the code?

Re: ifx and ifort comparison for do concurrent

Andrew_Smith — Tue, 06 Aug 2024 22:06:26 GMT

The amount of work inside the do concurrent loop is probably too small compared the overhead of parallel running so it actually runs slower. You did not show comparative times for serial and parallel to prove your assumption

Re: ifx and ifort comparison for do concurrent

Fortran10 — Wed, 07 Aug 2024 01:05:18 GMT

The output shows the time for both tests. It is the last line in the output.

Here I it write again

ifort time = 0.282000005245209

ifx time = 4.00899982452393

Re: ifx and ifort comparison for do concurrent

Andrew_Smith — Wed, 07 Aug 2024 14:18:25 GMT

I used task manager and confirmed that without using the /Qopenmp flag, DO CONCURRENT does not produce parallel code.

I had to up the stack size when using /Qopenmp, no idea why as the thread data should be small as only scalar variable are declared as LOCAL

These are my results:

	Serial	Parallel (/Qopenmp)
IFX	2.707999945	4.796
IFORT	2.092999935	0.81

Task manager confirmed that the IFX parallel version did not run in parallel but the IFORT version did.

Re: ifx and ifort comparison for do concurrent

Devorah_H_Intel — Wed, 07 Aug 2024 17:34:43 GMT

For Question # 3
See the Intel Fortran Compiler Porting Guide

As well as other useful information on ifx.

topic Re: ifx and ifort comparison for do concurrent in Intel® Fortran Compiler

ifx and ifort comparison for do concurrent

Re: ifx and ifort comparison for do concurrent

Re: ifx and ifort comparison for do concurrent

Re: ifx and ifort comparison for do concurrent

Re: ifx and ifort comparison for do concurrent