On 2019-11-25 4:06 p.m., Timothy Pearson wrote:
----- Original Message -----
From: "Felix Kuehling" <felix.kuehling@xxxxxxx>
To: "Timothy Pearson" <tpearson@xxxxxxxxxxxxxxxxxxxxx>, "amd-gfx" <amd-gfx@xxxxxxxxxxxxxxxxxxxxx>
Sent: Monday, November 25, 2019 11:07:31 AM
Subject: Re: [PATCH 1/1] amdgpu: Enable KFD on POWER systems
Hi Timothy,
Thank you for the patch and for confirming that it works. We did some
experimental work on Power8 a few years ago. I see that Talos II is Power9.
At the time we were working on Power8 we had to add some #ifdef
CONFIG_ACPI guards around some ACPI-specific code in KFD. Do you know to
what extent ACPI is available and working on the Power architecture?
Another problem we ran into with Power, is the physical address map.
System memory can be a physical addresses outside the range accessible
by the GPU. Vega has 44-bit physical addressing. Older Polaris GPUs only
have 40-bits. Did you run into any such problems? Do you need an IOMMU
to make system memory accessible to the GPU?
Regards,
Felix
Yes, we are POWER9. It looks like the ACPI guards are no longer required; as you have surmised, POWER does not use ACPI (the equivalent is OPAL, which is a different interface entirely). What were the APCI calls used for? There may be OPAL equivalents that could be added in to replace them and provide similar functionality.
There are some ACPI calls (e.g. acpi_get_table) in kfd_crat.c for
getting a CRAT table from ACPI. This is only useful for AMD APUs, which
are x86_64. We don't need this for discrete GPUs because on non-APU
systems there is no CRAT table and we build our own. If you can compile
the code without problems on Power and with CONFIG_ACPI not defined,
then I guess this is no longer an issue.
Kernel 5.4 enables a > 32-bit and <=64-bit bypass mode for POWER. This is one reason we came back and revisited the KFD/ROCm functionality on POWER; as it turns out, after fixing up the userspace tools KFD is indeed functional on POWER with 5.4-rc8 and above. My understanding is that the POWER IOMMU is used as a lightweight translation layer between the 64-bit host and the 40/44-bit GPU.
I'm working on getting a Debian PPA set up for POWER to make the userspace tools easier to obtain for testing, but progress is slow due to lack of Debian source packages. Probably the easiest way to replicate / test this with HIP is to use the AOMP repository with my modifications; pull requests are already in place on Github for most of the userspace tooling updates.
Thank you!
Thanks,
Felix
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
