Re: MIOpen package LFS files
Hi Xuanteng,
On 2024-05-15 17:54, Xuanteng Huang wrote:
>> On May 15, 2024, at 07:14, Cordell Bloor <cgmb@slerp.xyz> wrote:
> I’ve tried to exclude all assembly codes under src/kernels from both the source package and the build process [23], see the diff in src/CMakeLists.txt.
> As for now, we're not sure if the obtained MIOpen is OK to work.
> Maybe the next step is to enable the built-in tests [24] (disabled by patch [25] now) to validate whether the library works?
That would be great! We will (eventually) absolutely need tests for this
package.
I assume Cory disabled the tests for a reason, so perhaps he can expand
on that.
> I have 6900XT and MI100 GPUs in my accessible local machine, and there seems to have some utils [26] for providing isolated GPU environments in our Salsa.
> @ckk, maybe Christian can provide some suggestions?
Sure. I would probably start with an environment using the podman
backend, so you'd need package rocm-podman-support from our own APT
repository [27].
Once installed, you must perform some initial configuration for your
user. There's a utility that will provide you with instructions for all
necessary steps:
$ rocm-podman-setup -u <username>
This configuration is the only step that needs privileges, and it is
only needed once. Everything below is run as <user>.
Once everything reports [OK], you can proceed to build an image:
# builds rocm/debian:unstable
$ rocm-podman-create -m <local-apt-cache> -r unstable
You can run an interactive container including the GPU with:
$ rocm-podman-run -it <other-args> rocm/debian:unstable
You can run autopkgtests with:
$ autopkgtest ... -- podman+rocm rocm/debian:unstable
To see if everythings works correctly, the following should "just work":
$ autopkgtest -B rocrand -- podman+rocm rocm/debian:unstable
Note that podman currently only works properly on systems with one GPU
installed. I've added support for multi-GPU, but I've run into a HSA
issue that I need to resolve first.
The alternative to podman is the QEMU backend. The isolation is
stronger, but the card will no longer be usable by the host, and it
doesn't work for every GPU architecture.
Feel free to ping me if I can help in any way.
Best,
Christian
> [26]: https://salsa.debian.org/rocm-team/rocm-dev-tools
[27]: https://apt.rocm.debian.net/
Reply to: