10:25 <mszyprow> memeka: ping

10:34 <memeka> mszyprow pong

10:35 <memeka> Sorry busy with sick kid :(

10:35 <mszyprow> memeka: I've fixed the DRM dmabuf import issue, my fault since Linux v4.6

10:35 <memeka> Saw the drm patch, excited to try tmr

10:35 <mszyprow> memeka: but I noticed that gst uses dmabuf in very inefficient way

10:36 <mszyprow> memeka: from the logs I see it imports the buffers for every single frame

10:37 <mszyprow> memeka: I don't know the gst internals, but one should check it, there is huge space for optimizations there

10:37 <mszyprow> memeka: the buffers should be imported/exported once on pipeline configuration then simply used

10:38 <mszyprow> so basically now we benchmark dmabuf setup/import functions instead of real zero-copy buffer processing

10:38 <memeka> I’ll let the gst ppl know, but they might be aware

10:38 <memeka> I know they’re working on this: https://bugzilla.gnome.org/show_bug.cgi?id=783521

10:39 <mszyprow> those 2 pipelines should work at the same speed:

10:39 <mszyprow> filesrc location=bunny_trailer_1080p.mov ! qtdemux ! h264parse ! v4l2h264dec capture-io-mode=dmabuf ! v4l2video2convert output-io-mode=dmabuf-import capture-io-mode=dmabuf ! fpsdisplaysink video-sink=kmssink sync=false

10:39 <mszyprow> filesrc location=bunny_trailer_1080p.mov ! qtdemux ! h264parse ! v4l2h264dec capture-io-mode=dmabuf ! v4l2video2convert output-io-mode=dmabuf-import capture-io-mode=dmabuf-import ! fpsdisplaysink video-sink=kmssink sync=false

10:39 <mszyprow> but in real test there are significant differences

10:40 <mszyprow> it should really not matter which driver exports the buffer

10:40 <memeka> What should the second one do? Is gscaler importing buf from kms?

10:40 <memeka> Why not use userptr?

10:41 <mszyprow> well, I just gave the example you sent me with the iommu bug issue ;)

10:41 <mszyprow> dmabuf should be preffered over userptr if possible

10:42 <mszyprow> in theory it would be faster when used properly

10:42 <mszyprow> userptr relies on a series of hacks

10:42 <memeka> Ok I’ll try them

10:42 <memeka> Oh I see

10:42 <memeka> I thought the idea was for the sink to allocate buffers

10:42 <mszyprow> but in practice the current code imports/exports the buffers every frame

10:43 <mszyprow> what completely kills the idea of zerocopy

10:43 <memeka> And send them to the previous elements (gsc, MFC)

10:43 <mszyprow> as setting up dmabufs between drivers has non-zero overhead

10:44 <memeka> In any case, with iommu disabled I was getting <10% cpu usage

10:44 <memeka> So I’ll compare to that :)

10:44 <mszyprow> if you know the gst people, please report them my complaints ;)

10:44 <memeka> Yup I will

10:45 <mszyprow> at least from the kernel side now it should work fine

10:45 <memeka> Cool

10:45 <memeka> Thx for looking at that, one more bug fixed!

10:45 <mszyprow> I will focus on the dmabuf import for mfc encoding now

10:47 <memeka> Wow cool

10:47 <memeka> I have to go to my son now sorry

10:48 <mszyprow> memeka: no problem

10:48 <memeka> But thx fir the awesome news