How to enable the GPU to run a FORTRAN program?

[eliopoulos]

How can I enable the GPU to run a parallel do loop in a FORTRAN program? I have tried: "!$omp target teams distribute parallel do" but the GPU does not run. Any advice?

[eliopoulos]

I have created a reproducer that produces the same errors with my program.

[Barbara_P_Intel]

Ah! You use the matmul() instrinsic. When you compile with /Qopenmp on Windows or -qopenmp on Linux the matmul() intrinsic is automatically parallelized using OpenMP.

That can be disabled with /Qopt-matmul- on Windows or -no-opt-matmul on LInux. See the Developer Guide.

So far I've just run this on CPU. iGPU is next.

Attached is an update of z.for called z.r1.for. Here's the output I got for CPU:

Q:\06129705>ifx /Qopenmp z.r1.for
Intel(R) Fortran Compiler for applications running on Intel(R) 64, Version 2024.1.0 Build 20240103
Copyright (C) 1985-2023 Intel Corporation. All rights reserved.

Microsoft (R) Incremental Linker Version 14.38.33130.0
Copyright (C) Microsoft Corporation.  All rights reserved.

-out:z.r1.exe
-subsystem:console
-defaultlib:libiomp5md.lib
-nodefaultlib:vcomp.lib
-nodefaultlib:vcompd.lib
z.r1.obj

Q:\06129705>z.r1.exe
 Number of procs is            8
   6000000000000.00        6000000000000.00        6000000000000.00
   6000000000000.00        6000000000000.00        6000000000000.00
   6000000000000.00        6000000000000.00        6000000000000.00

[Barbara_P_Intel]

I have the Intel GPU version running and printing the same answers as the CPU version. It's attached as z.r2.for.

Two MAJOR changes:

• double precision is not supported on Intel(R) Iris(R) Xe Graphics. I made the arrays real.
• It's poor coding practice to use the same variable for the loop index and the end of the loop. This loop construct yields the wrong answer on Intel GPU. 

!   do i=1,i

! Using this construct gives the same answer as running on CPU.

   do j=1,i

When I run it in a oneAPI command window I type:

set LIBOMPTARGET_PLUGIN_PROFILE=T

to get a profile printout. I set that to be sure it runs on Intel GPU. No profile, it didn't offload.

>set LIBOMPTARGET_PLUGIN_PROFILE=T
>ifx /Qopenmp /Qopenmp-targets=spir64 z.r2.for
Intel(R) Fortran Compiler for applications running on Intel(R) 64, Version 2024.1.0 Build 20240103
Copyright (C) 1985-2023 Intel Corporation. All rights reserved.

Microsoft (R) Incremental Linker Version 14.38.33130.0
Copyright (C) Microsoft Corporation.  All rights reserved.

-out:z.r2.exe
-subsystem:console
-defaultlib:libiomp5md.lib
-nodefaultlib:vcomp.lib
-nodefaultlib:vcompd.lib
C:\Users\bperz\AppData\Local\Temp\1932873.obj
C:\Users\bperz\AppData\Local\Temp\19328414.o
-defaultlib:omptarget.lib

>z.r2.exe
 Number of procs is            8
  6.0000001E+12  6.0000001E+12  6.0000001E+12  6.0000001E+12  6.0000001E+12
  6.0000001E+12  6.0000001E+12  6.0000001E+12  6.0000001E+12
======================================================================================================================
LIBOMPTARGET_PLUGIN_PROFILE(LEVEL_ZERO) for OMP DEVICE(0) Intel(R) Iris(R) Xe Graphics, Thread 0
----------------------------------------------------------------------------------------------------------------------
Kernel 0                  : __omp_offloading_80ff4b02_5bc1261f_MAIN___l27
----------------------------------------------------------------------------------------------------------------------
                          : Host Time (msec)                        Device Time (msec)
Name                      :      Total   Average       Min       Max     Total   Average       Min       Max     Count
----------------------------------------------------------------------------------------------------------------------
Compiling                 :     397.03    397.03    397.03    397.03      0.00      0.00      0.00      0.00      1.00
DataAlloc                 :       0.58      0.06      0.00      0.42      0.00      0.00      0.00      0.00      9.00
DataRead (Device to Host) :       0.00      0.00      0.00      0.00      0.00      0.00      0.00      0.00      1.00
DataWrite (Host to Device):       0.00      0.00      0.00      0.00      0.00      0.00      0.00      0.00      1.00
Kernel 0                  :       0.51      0.51      0.51      0.51      0.10      0.10      0.10      0.10      1.00
Linking                   :       0.00      0.00      0.00      0.00      0.00      0.00      0.00      0.00      1.00
OffloadEntriesInit        :       3.88      3.88      3.88      3.88      0.00      0.00      0.00      0.00      1.00
======================================================================================================================

There are some other comments in z.r2.for.

@eliopoulos, let me know how this works for you.

[eliopoulos]

It works. I have the same output. Do better Intel GPU support double precision?

[Barbara_P_Intel]

Yes, the higher end Intel GPUs support double precision.

The Intel(R) Iris(R) Xe Graphics GPUs are designed for gaming, not HPC workloads. Gamers don't need those extra bits for their graphics.

[eliopoulos]

But /Qmkl is not supported because /size-llp64 is not supported by ifx, I guess.

[Barbara_P_Intel]

I don't know how MKL works on Intel(R) Iris(R) Xe Graphics. You might ask on the MKL Forum. Have a simple reproducer handy. Here's the link to the MKL Forum.

ifx definitely supports double precision on CPU and Intel GPUs that support double precision.

[eliopoulos]

Thank you.