[cfe-dev] OpenMP offloaded target region executed in both host and target-device

[Piyumi Lakshani via cfe-dev]

Hi,

I'm working on a project which requires OpenMP offloading to Nvidia GPUs
using Clang.

System specification

OS - Ubuntu 16.04 LTS
Clang -version 4.00
Processor - Intel(R) Core(TM) i7 -4700MQ CPU
Cuda -version - 9.0
Nvidia GPU - GeForce 740M (sm_capability - 35)

But the problem is I when I execute a sample program to test OpenMP
offloading to Nvidia GPUs, part of the target region tends to run in GPU
and then same target region starts executing in the host.

Please find the sample program  attached herewith, This a small C program
written to multiply 2 matrices.
The reason to claim that target region is being executed in both host and
target-device is due to the abnormal output received from the print
function residing in the target region. (My processor has 4 cores capable
of handling 2 hardware level threads per core.).

Please find the image of the command line output attached herewith.

the program was compiled with -

>  clang -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda
>


I can not figure out whether runtime believes that the GPU execution is not
completing successfully?. So the target region is being executed in the
host again.

Thank you!
-- 
*Piyumi Rameshka*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20180409/0ca2443f/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 2mm.c
Type: text/x-csrc
Size: 1278 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20180409/0ca2443f/attachment.c>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: command-line-output.jpg
Type: image/jpeg
Size: 238437 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20180409/0ca2443f/attachment.jpg>