Only NVIDIA Gpus with Turing!
EDIT: Cheapest and smallest Turing card is the NVIDIA T400
1 Like
Indeed, it is not supported.
Architecture: Maxwell
Turing+ Nvidia GPU. Complete list is here: GitHub - NVIDIA/open-gpu-kernel-modules: NVIDIA Linux open GPU kernel module source.
2 Likes
Thank you
more character to fill up the post
No I don’t mean the one I selected,I mean that I searched "game-starter"but found no result about it.
Ah ok,
There are plenty of games in HaikuDepot!
This is truly amazing work. Is the work pushed somewhere to take a look at and test?
2 Likes
AFAIK, currently, as a whole, nowhere, but there are some code parts of it in X512 github. The base open-gpu driver code is also in Nvidia repo, but without any Haiku-specific code. Everything is still considerably experimental and under work.
1 Like
I’ve been working on porting the nvidia-open driver to another OS (managarm) and tried to get the published NVK patches working (hence my interest in this effort), but to no success so far (well, modeset works at least).
As a debug measure, I decided to build and test the NVK patches on Linux (by applying the patches on top of upstream mesa, symlinking nvidia-open source and commenting a line in a generated source), but that leads to crashes for me. I’m running vkcube as a testcase and validation layers complain about the swapchain being null. Has anybody run into that before?
My NVK patches are not tested on Linux yet. You may tried another version adapted for Linux: Making sure you're not a bot!.
1 Like
Thank you for pointing out that MR, I hadn’t seen that yet. Looking at it though, it seems to hardcode some hClasses, unlike your branch; it seems like your branch is further along than the MR, so I’ll probably continue working with that.
I’ve been able to make your branch work on Linux on my RTX 2070, where the issue was that NVA06F_CTRL_CMD_GPFIFO_SCHEDULE returned NV_ERR_INVALID_ARGUMENT. For some reason, the parameter size passed to nvRmApiControl was off, and tracing the nvidia-provided ICD revealed that it expects a value of 2 on my host driver (570.153.02), which made me remember that I have a driver version mismatch compared to the headers. 575.51.02 has added a new field to the params, which is why this failed.
It might make sense to print a warning if the driver version that NVK expects is not what nvidia-drm reports; IIRC there was a control function for matching versions? I might look into that soon.
Thank you for your work and everybody involved, this is amazing.
tl;dr: the NVK stuff seems to work on Linux just fine if you’re less of an idiot than me.
2 Likes
Yes, it is basically older version of my branch, but adapted and tested on Linux.
DRM stuff is not needed at all when running NVK with NVRM. No libdrm, no nvidia-drm.ko are needed.
2 Likes
Fair enough. NVRM allows for version checking without involving DRM at all, so that’s what I’ll do then.