sven-ola

@brunorro What a read! That's cool stuff - hopefully those guys'n'gals visit the next station on their journey soon: from TH1520, via JH7110, then (hopefully) X1. At least that mailbox communication sounds similar with the SpacemiT SoC: power management via extra rCPU is also a thing here. Thank you for the link // Sven-Ola

@cal5582 Hav you plugged in your Oculink/Intel Arc combination to a standard PC's M.2 slot? Just to make sure, it's not the power supply for our power hungry graphical friends. If it is not the power supply (5V, 12V) then there is one difference from PC to OpiRV2: the OpiRV2's M.2 slot has only 2 PCIe lanes connected, while a standard PC has M.2/M-key connectors with 4 lanes normally. A graphics card may be used to a high number of lanes, so not only hungry for power but also for lanes. Also: there seems to exist Oculink adapters with an optical link between PC/board and card. In that case, I would expect an extra PCIe bridge to be shown with lspci.

@maxsub no, its hacked. At least from my side. For us, that GPU stuff is closed, patented, tighly licensed and so on. LG // Sven-Ola

That "rvdisplay" string is found in the following files: libVK_IMG.so, libsutu_display.so, and libpvr_dri_support.so. When comparing the OpiRV2 versions of these files with https://gitee.com/spacemit-buildroot/img-gpu-powervr I got 9a640=spacemit vs 9a2d0=rvdisplay. Thus, Xunlong has source code for these (different offsets, rvdisplay is one char longer). So no need to hack Armbian kernel, I presume GPU works if we use *.so from gitee repo. HTH // Sven-Ola

@c0rnelius then the GPU probably works, since without GPU e.g the Bianbu first time wizard (language, kbd, timezone, user) needs 20 minutes while uptime reports load=10. I'm damn sure you have noticed an extraslow GUI

Hi @maxsub, no, thats a normal image. Hacking is the looong filename: Armbian-unofficial_26.02.0-trunk_Orangepirv2_trixie_edge_6.18.8_minimal+Orangepirv2_1.0.0_ubuntu_noble_desktop_gnome_linux6.6.63.img.xz

Continued GPU hacks. I uploaded a medley from Armbian (uboot,kernel) plus OrangePi-Ubuntu-Noble (userspace) from downloaded Xunlong image to my https://privat-in.de site (-> Downloads, grab Armbian+OrangePi.img.xz). Root PW is "orangepi" and it has 3 scripts /boot/boot-6.x.x to switch kernels between 6.6.36-ky, 6.6.99 and 6.18.x (both from Armbian). While the original "ky" kernel has GPU support (instantly visible with the Gnome GUI reacting < 2 seconds), the Armbian kernels have no GPU. Turns out: the kernel DRM driver name needs to be "rvdisplay" instead of "spacemit". This works at least with Armbian-Stable-6.6.99, while Armbian-Edge-6.18.x probably needs further kicking. The Armbian kernels have the following (prelim) change: --- a/drivers/gpu/drm/spacemit/spacemit_drm.c +++ b/drivers/gpu/drm/spacemit/spacemit_drm.c @@ -24,7 +24,7 @@ #include "spacemit_dmmu.h" #include "spacemit_gem.h" -#define DRIVER_NAME "spacemit" +#define DRIVER_NAME "rvdisplay" #define DRIVER_DESC "Spacemit SoCs' DRM Driver" #define DRIVER_MAJOR 1 #define DRIVER_MINOR 0 Note, that Xorg has no GPU, the GPU stuff only works with Wayland (gnome-shell + mutter). Also: no luck with that Bianbu-Linux. It's damn slow Gnome (slower that OrangePi with Software rendering) and none of the kernels brings the GPU to live (even the original Bianbu kernel with a orangepirv2 DTB does not work). @c0rnelius Bananapi R3 and Musepi Pro have working GPU? HTH // Sven-Ola

@maxsub: There are some LLVM changes that are fixed with later Mesa, just cherry pick the top 3 commits from https://codeberg.org/sven-ola/mesa-spacemit-k1/commits/branch/spacemit-k1 HTH // Sven-Ola

@c0rnelius I uploaded /etc/apt and /usr/share/keyrings for you on my Opi3z Nextcloud for you, go to https://opi3-3.privat-in.de/index.php/s/by9XnGDbntaDz35, sha256sum is is 9d2a2380f7791630d1303c07e889b77d1acf0789b180f2a6b9157fa3fdf20685

You can change the kernel config for your needs with "kernel-config" argument, or for example ./compile.sh CPUTHREADS=$(nproc) BOARD=orangepirv2 BRANCH=edge RELEASE=trixie KERNEL_CONFIGURE=yes KERNEL_GIT=shallow BUILD_MINIMAL=yes will start a menuconfig before compiling an image.

Yes, I more or less thought about that orangepi.org Ubuntu from their downloads. At least this Ubuntu image shows some OpenGL capabilities with the preinstalled glmark2. I also tried that "Bianbu" image from Spacemit with my kernel, but that Gnome3/Wayland is slower than a slug probably because in case you want to play with that: I've uploaded that image to https://privat-in.de as a playground, root password is "bianbu" and you may need to resize the root partition to your SD card or so. Also I've overlooked that there is a "bianbu 3.0.1" so this is 3.0

Hello @Malay, et.al, the motivation behind my attempt to include Armbian support for OpiRV2 is to have a better device for my Nextcloud-for-private-persons project. For this, I need a device that can be run at home with a current kernel and headless, but with decent storage. I already have a device (Orange Pi Zero 3, right on the photo), but Orange Pi RV2 (to the left) offers two(!) m.2 slots for NVMEs, some Wifi and Eth as well as decent computing power. Also, both have some extra NOR flash to store LUKS keys and decent pricing. Also that RiscV64 has some appeal, admitted 😉 Without GPU support, you can run Mate, LXDE, and XFCE for a GUI, which should be OK for some management tasks. I doubt that RV2 will ever make a good Youtube player, however that may be archived by replacing the Armbian userspace with that Ubuntu-Noble userspace that you can download from Xunlong. This Ubunut also comes with a RiscV64 port of Chromium which is also needed for Youtube and not avail from Debian, but source code *.dsc seems to be offered in the Bianbu pool so that may work out. Anyhow - that GPU porting attempt was based on the PowerVR addons offered as code drop on git@gitee.com:spacemit-buildroot/mesa3d.git. Our Chinese friends grabbed mesa 22.3.5, added their IMG BXE-2-32 this+that to the sources, cherry-picked their way up to mesa 24.0.1 and throw the result over the fence. At least without inking the code via Windows notepad, so no white-space chaos this time. Now, porting this with the closed binary *.so as heavy baggage is all about stable API. Large internal API change -> end of party. I am no graphics specialist, so take this with a pinch of salt. Debian offers mesa since 24.3 with an additional package: mesa-libgallium, you need this for the GPU desktops (Cinnamon, and probably Gnome / KDE). OK - someone or something needs the API from the Mesa internal Gallium driver suite, this needs to be compiled (a *.a is there). So I started to bring that code drop from 24.0 to 24.3, only to stumble over internal API changes. Concrete: we previously have some numeric constants __DRI_IMAGE_COMPONENTS_this+that 0xabcde in Mesa describing image formats (RGB32, BGRA24, 656-16, and so on). The code drop adds a couple of formats probably specific for PowerVR, but the Mesa project completely removed those constants ("nobody uses this"). At that point I thought: this is pointless, I'll pass... HTH // Sven-Ola

Two addons: I compiled on another notebook and stumbled over a RAM error message from Armbian compile.sh. Notebook has 8GB, so I added the recommended KERNEL_BTF=no to the ./compile.sh args. A dkms status says, the bcmdhd is built and active.

@maxsub: I have not really analyzed Armbian build caching until now. Maybe a previous build for another spacemit board prevents the DKMS from being included? Anyhow, re-installing / updating a kernel image should trigger any active DKMS modules to rebuild for the new kernel on the board, so this looks correct. LG // Sven-Ola

Compiled, flashed, uploaded. I got a running XFCE desktop and a working wifi applet. If you still have different results, I may provide a VPN endpoint and an SSH pubkey to log in to your machine. Here's my upload dir: HTH // Sven-Ola

Hello @Malay! Same for you: cannot reproduce. Can you rebuild with SKIP_ARMBIAN_REPO=yes? This builds a little longer but skips cached stuff. Alternatively, rm -r ./cache from the build dir. To help with this, I will start the following commands on my PC and upload the results later to https://privat-in.de/ (Download section). LG // Sven-Ola

Hello @maxsub, we are talking about git log --oneline -1 which says 613737fab. The post_family_tweaks_bsp__orangepirv2_wifi() from above is exactly what's in there. If you run you should get on the board Can you please elaborate on the "got it"? TIA // Sven-Ola

Hu? In difference to Spacemit 6.6.99, the 6.18 kernel has no bcmdhd included. For that reason I put in an adapted version of bcmdhd in a DKMS, installed via a source deb from my repo on codeberg.org/sven-ola. B/c DKMS needs a compiler on the board, the edge image is larger than current image. Code quality of bcmdhd is questionable so it probably never make it into the official Linux tree, but it works (at least on my board with my image). HTH // Sven-Ola

For me, those 57°C is fine without further cooling. Starting cool on my RV2 is: root@orangepirv2:~# uptime;cat $(find /sys -name temp) 18:55:15 up 0 min, 1 user, load average: 0.90, 0.25, 0.08 27000 28000 Now some work, will probably max out at 95° or so root@orangepirv2:~# uptime;cat $(find /sys -name temp) 19:23:11 up 28 min, 3 users, load average: 7.61, 6.79, 3.86 90000 90000 Stop compiling and wait a bit gives this which is fine IMO root@orangepirv2:~# uptime;cat $(find /sys -name temp) 19:35:07 up 40 min, 3 users, load average: 0.00, 0.64, 1.81 56000 58000

@maxsub yes and no. That 900*.patch never was meant to be in the Armbian repo, I wounder how this made it in the Armbian repo. Some black git magic... Patch experimentally rolls back an upstream change (...that removes the kthread under discussion and by this prevents the main CPU from communicating to the remote CPU). Upstream (the Fedora folks that emitted the SpacemiT code drops) has removed the kthread to prevent high CPU load b/c they use RV2 for their RISCV compile farm or so. The 900*.patch was an experiment by me to understand the impact on HDMI audio / ADMA). Long blabla. This proposed and committed change solves the issue for my pull RQ on the topic: https://github.com/armbian/build/pull/9299/changes/613737fab5aadc2d0a288a10a98aa0a433db4e76

As a followup: we have some help from @c0rnelius who digged out a solution to that kthread=high load issue (see latest commit / patches in my branch on https://github.com/sven-ola/armbian-build/tree/orangepi-rv2). Maybe this helps for the R2S too?

Hi @maxsub we see a high load (== 2.00 on idle) on RV2 and probably other SpacemiT devices. Both on 6.6.99 and 6.18.8 kernels. Caused by an active kthread() that is needed for the communication to the rCPU. Note: rCPU means the extra realtime RISCV-CPU that is powered by the realtime OS we load on boot with CONFIG_EXTRA_FIRMWARE="esos.elf". On the RV2 the high load display has no temperature side effects, but maybe this is not true for R2S. We have a photo of the R2S on the PullRQ (see https://github.com/armbian/build/pull/9299#issuecomment-3811665824 ) so the large cooling cap obviously is necessary? Can you retry with a 6.18.8 kernel with disabled kthread on the R2S? Just rename the last 3 patches in build/patch/kernel/archive/spacemit-6.18 from {016,017,018}*.patch to *.disabled and rebuild. TIA // Sven-Ola

I think I've completed all tasks. Thus the OpiRV2 PR now waits for a team member to press the [merge] button probably. @maxsub I compiled 6.6.99 kernel on my board with 2Gb RAM once. Needed 3 hours and a decent swap file. HTH // Sven-Ola

@maxsub: if you still have compile probs, I just uploaded new images to my site on https://privat-in.de (navigate to downloads). LG // Sven-Ola

@maxsub: Out-of-RAM may be possible if it simply spits out "killed". My machine has 32 Gb RAM and 16 Gb swap, does not encountered the OOM-killer lately. Until now, but exhausting RAM is probably kernel-dev-hobby 😗