(Configuration Scheme 2/3) Installing the GPU Driver

You need to install the GPU driver each time the server is restarted if you use configuration scheme 2/3.

1. Obtain VAGPU-25.03.01.01-RC13-A15.tgz according to Software Environment, upload it to the ~/dependency/ directory, and decompress it to obtain the kernel-space GPU driver.

   1 2	cd ~/dependency/ tar -zxvf VAGPU-25.03.01.01-RC13-A15.tgz

2. Install the PCIe driver for the GPU.

   1 2	cd ~/dependency/VAGPU-25.03.01.01-RC13-A15/openEuler-6.6.0+/ko_fw insmod va_pci.ko

3. Copy the firmware in the driver package to the /lib/firmware/ directory of the OS.

   1	cp rgx* /lib/firmware/

4. Install the GPU driver.

   The GPU driver starts a kworker process for each GPU node. A single DaoCloud DC1000 card has four nodes. To improve the performance of kworker processes, you are advised to use the kworkerCores parameter to bind kworker processes to CPU cores. Each value of the kworkerCores parameter indicates a core bound to the kworker process of the corresponding GPU node.

   When binding GPU driver processes to CPU cores, ensure that the CPU cores bound to the kworker processes and GPU rendering nodes belong to the same CPU chip. For details about how to query the CPU chip to which a GPU rendering node belongs, see Determining the GPU Topology.

   The following core binding methods are for reference only. You can make adjustments based on actual circumstances.

   Configuration scheme 2 (Kunpeng 920 7260 processor + 4 x DaoCloud DC1000)

   1	insmod va_gfx.ko kworkerCores=0,0,1,1,32,32,33,33,64,64,65,65,96,96,97,97

   Configuration scheme 3 (Kunpeng 920 7280Z processor + 8 x DaoCloud DC1000)

   1	insmod va_gfx.ko kworkerCores=80,80,81,81,82,82,83,83,0,0,1,1,2,2,3,3,240,240,241,241,242,242,243,243,160,160,161,161,162,162,163,163

5. Wait until the script execution is complete and check the kernel logs.

   1	dmesg | grep VAGPU | grep version

   In the command output, if the kernel-mode driver version and GPU firmware version are the same (content in bold), the GPU driver is installed.

   PVR_K:  28823: Meta firmware version: 1.18@6276027 build: release branch:  commit: 67e785a8 tag: VAGPU-25.03.01.01-RC13-A15
   ...

To change the driver version, you need to uninstall the drivers and install the drivers of another version.

1. Delete all containers to release the drivers.
2. Uninstall the drivers in sequence.

   1 2	rmmod va_gfx rmmod va_pci