XinfengZhang
commented
6 months ago suppose it is related with CMC compiler version, how about to use a older version?
Open frantisekz opened 6 months ago
XinfengZhang
commented
6 months ago suppose it is related with CMC compiler version, how about to use a older version?
frantisekz
commented
6 months ago suppose it is related with CMC compiler version, how about to use a older version?
Yeah, it probably is. This is for Fedora packaging of the media-driver, and we no longer have llvm 8.0 in the distribution repositories, so older cm is unfortunately no-go.
I can imagine updating/fixing the media-driver for the newer cm-compiler isn't a priority, but I'd guess it'll be done someday?
(Edit: I did test the previous cm-compiler from cmc_monorepo_80, it didn't have this issue, but failed elsewhere as reported here: #1356 )
XinfengZhang
commented
6 months ago :) why not "DBUILD_KERNELS:BOOL=OFF", the final iHD_video_drv.so should be same, I am expecting no one will change media kernel code, "open source media kernel" just was used to prove that the binary could be compiled.
frantisekz
commented
6 months ago :) why not "DBUILD_KERNELS:BOOL=OFF", the final iHD_video_drv.so should be same, I am expecting no one will change media kernel code, "open source media kernel" just was used to prove that the binary could be compiled.
That is unfortunately prohibited in Fedora packages, it has to be built from the source, or removed, so we're currently shipping the media-driver without kernels :(
intel-mediadev
commented
6 months ago Auto Created VSMGWL-72132 for further analysis.
XinfengZhang
commented
2 months ago hi @frantisekz , sorry for inconvenience firstly. Actually, we are expecting all user will use default build options, the "non free" kernel is actually free to use, it is under MIT license, they are existing in media driver as arrays. the source file is close source just because these reasons : 1. maintenance effort; 2. partial code are assembly code, partial are c for media , because of history; 3. some algorithm should be reserved to protect IP; 4. some algorithm will be placed by HW in future. such as: most of color format conversion / scaling will be handled by HW directly. no need to maintain an shader path. so, our main focus is delivering the full feature media driver. the trending is shader will be replaced by HW. maintaining shader path (actually there are two shader path, full feature and free kernel, considering there are also an build_kernel option) is complex. and suppose no customer will use this free kernel path. because the feature is restricted. is it possible we only provide one iHD driver version in future?
frantisekz
commented
1 month ago @XinfengZhang thanks for the reply, I do greatly appreciate that!
For the non-free option - since I am working/maintaining the media-driver component in Fedora, there are strict rules what can be the part of the package, and having anything close-sourced or potentially targeted by IP protection is a "big no-no".
(In order to get the package in, I am required to even do the source tarball stripping of the "non-free" kernels before uploading the archive to the pool for the build infrastructure. The codecs/multimedia world is a patent minefield unfortunately...)
The current status of the media-driver in the distribution is that we ship it without any kernels, because the build_kernel option doesn't work with cm-compiler from llvm 11 branch.
Being at least able to have "some" form of vaapi support for Intel out of the box in Fedora is a great deal for our users, and a huge improvement compared to the previous situation.
Ideally, we'd be able to at least somehow continue building media-driver (even if restricted in features) without any kernels, and being able to rebuild the free kernels, even if in the future and for just some architectures, would be the "optimal" outcome for Fedora and potentially its downstream RHEL.
If the driver ceased to work without any kernels, and/or an ability to rebuild free kernels, it'd have to be removed from Fedora.
Which component impacted?
Build
Is it regression? Good in old configuration?
No, this issue exist a long time
What happened?
What's the usage scenario when you are seeing the problem?
Others
What impacted?
No response
Debug Information
No response
Do you want to contribute a patch to fix the issue?
None