[PATCH v2 0/7] Host1x IOMMU support + VIC support

[Daniel Vetter]

On Tue, Jan 17, 2017 at 06:22:36PM +0200, Mikko Perttunen wrote:
> On 12/14/2016 04:11 PM, Daniel Vetter wrote:
> > On Wed, Dec 14, 2016 at 03:32:16PM +0200, Mikko Perttunen wrote:
> > > On 14.12.2016 15:05, Daniel Vetter wrote:
> > > > On Wed, Dec 14, 2016 at 02:41:28PM +0200, Mikko Perttunen wrote:
> > > > > On 14.12.2016 14:30, Daniel Vetter wrote:
> > > > > > On Wed, Dec 14, 2016 at 01:16:10PM +0200, Mikko Perttunen wrote:
> > > > > > > This series adds IOMMU support to Host1x and TegraDRM
> > > > > > > and adds support for the VIC (Video Image Compositor)
> > > > > > > host1x client. The series is available as a git repository at
> > > > > > > git://github.com/cyndis/linux.git; branch vic-2.
> > > > > > > 
> > > > > > > A userspace test case for VIC can be found at
> > > > > > > https://github.com/cyndis/drm/tree/work/tegra.
> > > > > > > The testcase is in tests/tegra and is called submit_vic.
> > > > > > > The testcase/TRM include full headers and documentation
> > > > > > > to program the unit. The unit by itself, however, does not
> > > > > > > readily map to existing userspace library interfaces, so
> > > > > > > implementations for those are not provided.
> > > > > > 
> > > > > > Afaik libva has an entire pile of post-processing support. Pretty sure
> > > > > > other video transcode libraries have similar interfaces, so should all be
> > > > > > possible to implement this.
> > > > > 
> > > > > We don't have any actual video transcoding support though, so unless it's
> > > > > possible to just implement a part of libva and defer the rest to some CPU
> > > > > implementation, I don't see how this is useful. I suppose I could implement
> > > > > a GStreamer plugin for colorspace conversion or resizing, since those are
> > > > > very modular.
> > > > 
> > > > Hm, I guess the question then is, how did that get enabled?
> > > 
> > > What is "that"? I'm not exactly sure.
> > > 
> > > Our architecture is such that there's the VIC that handles colorspace
> > > conversion, rescaling, blitting and can do some 2d postprocessing effects as
> > > well.
> > > 
> > > Then there's the separate NVDEC that is a video bitstream decoder. There's
> > > no support for that at the moment. I am working on the IP side of that.
> > > 
> > > The video processing pipeline is then such that NVDEC is fed the bitstream;
> > > NVDEC outputs a YUV picture in a specific format; VIC takes that YUV picture
> > > and converts/rescales it into the desired format. Or if we are encoding
> > > video, VIC takes your RGB image, converts it into a format that NVENC
> > > understands, and so on.
> > > 
> > > So with just VIC support, I could implement some simple 2D things. I don't
> > > know if anyone would want to specifically use the VIC for those since
> > > applications already have fast CPU algorithms. For the video pipeline using
> > > VIC is nice since these units can synchronize work without CPU involvement
> > > and when you're already using NVDEC or NVENC it's barely any extra effort to
> > > involve VIC as well. It can also be useful in power usage sensitive
> > > situations, but we aren't really fit for those situations with the upstream
> > > kernel anyway :)
> > 
> > Ah I thought the nvdec was already enabled, since for i915 that's how we
> > went about things (we have a pretty much exactly matching split in the
> > various video related engines). But if that's not there yet then no
> > worries, all fine.
> > 
> > Since you do seem to plan to enable everything anyway, might be worth it
> > to go directly with something like libva or libvdpau or whatever the cool
> > thing is. libva is my recommendation since it works on non-X11 too afaik,
> > but I have 0 clue. And might be worth it to check out whether you can't do
> > a super-basic libva driver that only does the post processing stuff. With
> > libva you can import/export images, so it might be possible even ... And
> > directly doing the full video engine support instead of a one-off in
> > gstreamer sounds more sensible to me.
> > -Daniel
> > 
> 
> It took a while, but I now have a libva backend to go with this kernel
> driver: https://github.com/cyndis/vaapi-tegra-driver.
> 
> It is far from complete, but the libva putsurface testcase runs and its
> output looks correct when compared to the intel backend. Would this be
> suitable or should something more be implemented before merging the kernel
> driver?

Seems all reasonable to move ahead with this. Thanks for doing it.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch