Re: [Xen-users] Trying CUDA/OpenCL with VGA Passthrough

[Pasi Kärkkäinen <pasik@xxxxxx>]

On Thu, Mar 03, 2011 at 08:31:08PM +0100, Claudio Tanci wrote:
> Hi everyone,
> 
> First post, already a call for help :)
> 
> I'm trying to use a gtx 580 with Xen 4 VGA Passthrough to test CUDA
> performances, the MB is an Asus P7P55 LX with a i7 870, VT-D is enabled
> in bios and reported as active by Xen.
> I'm using Debian Squeeze as Dom0 and DomU OS.
> 
> Dom0 is configured with iommu=1 and iommu=verbose parameters and
> xen-pciback.hide=(01:00.0)(01:00.1) (but I still see them with lspci in
> dom0, is that normal?)
> 
> I configured DomU .cfg VGA passthrough with
> gfx_passthru=1
> pci=['01:00.0']
> 
> and can see the devices as assignable
> root@cloud:~/nvidia# xm pci-list-assignable-devices
> 0000:01:00.0
> 0000:01:00.1
> (0000:01:00.1 refers to the hdmi audio port)
> 
> and then assigned to the DomU
> root@cloud:~/nvidia# xm pci-list squeeze1
> domain bus  slot func
> 0x0000 0x01 0x00 0x0
> 
> I can see the card in the DomU
> root@squeeze1:~# lspci
> 00:00.0 VGA compatible controller: nVidia Corporation Device 1080 (rev
> a1)
> 
> I installed the video driver, cuda toolkit and Nvidia SDK and everything
> seems fine, the module is loaded and the devices exists (No X as I will
> use the machine only as a test bed for CUDA on virtualized hardware)
> root@squeeze1:~# lsmod
> Module                  Size  Used by
> nvidia              10442361  0
> snd_pcm                60551  0
> snd_timer              15582  1 snd_pcm
> snd                    46446  2 snd_pcm,snd_timer
> soundcore               4598  1 snd
> snd_page_alloc          6249  1 snd_pcm
> i2c_core               15712  1 nvidia
> evdev                   7352  0
> pcspkr                  1699  0
> ext3                  106502  1
> jbd                    37085  1 ext3
> mbcache                 5050  1 ext3
> xen_netfront           16257  0
> xen_blkfront            9435  2
> 
> root@squeeze1:~# ls -lh /dev/nvidia*
> crw-rw-rw- 1 root root 195,   0 Mar  3 16:26 /dev/nvidia0
> crw-rw-rw- 1 root root 195, 255 Mar  3 16:26 /dev/nvidiactl
> 
> but deviceQuery fails accessing the card :(
> ./NVIDIA_GPU_Computing_SDK/C/bin/linux/release/deviceQuery Starting...
> 
>  CUDA Device Query (Runtime API) version (CUDART static linking)
> 
> [ 3276.036642] NVRM: failed to register with the ACPI subsystem!
> [ 3276.047716] NVRM: RmInitAdapter failed! (0x23:0x2f:646)
> [ 3276.047737] NVRM: rm_init_adapter(0) failed
> [ 3276.047839] NVRM: failed to unregister from the ACPI subsystem!
> cudaGetDeviceCount FAILED CUDA Driver and Runtime version may be
> mismatched.
> 
> I'm using cuda toolkit 3.2.16, sdk_3.2.16 and I tried both dev driver
> 260.19.26 and 270.29
> The kernel in Dom0 is the 2.6.32-5-xen-amd64 as packaged in Debian, Xen
> package version is 4.0.1-2.
> 
> Any idea?
> 

Most probably you need to do custom patching for your graphics hardware.
I think it's the vBar=pBar patch that you need to modify/add.

-- Pasi


_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users