[amdgpu] Compute kernels still run when the host process exit?

[amdgpu] Compute kernels still run when the host process exit?

[Smith John]

[-- Attachment #1.1: Type: text/plain, Size: 278 bytes --]

Hi, when I killed an OpenCL host process, the kernels it launched were not
terminated, and still run.

My OpenCL runtime is AMDGPU-PRO 20.20. OS Ubuntu 18.04.5 with  Linux Kernel
5.4.53

I was wondering if it was a bug or the driver did not implement this
"watchdog" mechanism.

[-- Attachment #1.2: Type: text/html, Size: 372 bytes --]

[-- Attachment #2: Type: text/plain, Size: 154 bytes --]

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [amdgpu] Compute kernels still run when the host process exit?

[Alex Deucher]

On Thu, Apr 1, 2021 at 10:08 AM Smith John <ls.cat.kth@gmail.com> wrote:
>
> Hi, when I killed an OpenCL host process, the kernels it launched were not terminated, and still run.
>
> My OpenCL runtime is AMDGPU-PRO 20.20. OS Ubuntu 18.04.5 with  Linux Kernel 5.4.53
>
> I was wondering if it was a bug or the driver did not implement this "watchdog" mechanism.

In general, once you issue work on the GPU it has to run to
completion.  It is not stopped if the application that issued it goes
away.

Alex
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [amdgpu] Compute kernels still run when the host process exit?

[Felix Kuehling]

Am 2021-04-01 um 2:22 p.m. schrieb Alex Deucher:
> On Thu, Apr 1, 2021 at 10:08 AM Smith John <ls.cat.kth@gmail.com> wrote:
>> Hi, when I killed an OpenCL host process, the kernels it launched were not terminated, and still run.
>>
>> My OpenCL runtime is AMDGPU-PRO 20.20. OS Ubuntu 18.04.5 with  Linux Kernel 5.4.53
>>
>> I was wondering if it was a bug or the driver did not implement this "watchdog" mechanism.
> In general, once you issue work on the GPU it has to run to
> completion.  It is not stopped if the application that issued it goes
> away.

If this is using KFD, we destroy user mode queues and free all the
process' memory at process termination. Because we support CWSR, we can
interrupt running shaders. Is this using the legacy OpenCL?

Regards,
  Felix


>
> Alex
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [amdgpu] Compute kernels still run when the host process exit?

[Christian König]

Am 01.04.21 um 20:22 schrieb Alex Deucher:
> On Thu, Apr 1, 2021 at 10:08 AM Smith John <ls.cat.kth@gmail.com> wrote:
>> Hi, when I killed an OpenCL host process, the kernels it launched were not terminated, and still run.
>>
>> My OpenCL runtime is AMDGPU-PRO 20.20. OS Ubuntu 18.04.5 with  Linux Kernel 5.4.53
>>
>> I was wondering if it was a bug or the driver did not implement this "watchdog" mechanism.
> In general, once you issue work on the GPU it has to run to
> completion.  It is not stopped if the application that issued it goes
> away.

Well that is not 100% correct.

When the job has already been pushed to the hardware it indeed runs till 
the end.

But when the job is still in the scheduler and the process which has 
submitted it is killed we also throw away the job.

This obviously only counts for the classic stack and not for the KFD.

Christian.

>
> Alex
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx