Re: [RFC 0/5] arm64: imx8mm: Enable Hantro VPUs

[Nicolas Dufresne <nicolas@ndufresne.ca>]

Le lundi 22 novembre 2021 à 09:25 -0800, Tim Harvey a écrit :
> On Sat, Nov 20, 2021 at 7:36 AM Adam Ford <aford173@gmail.com> wrote:
> > 
> > On Fri, Nov 19, 2021 at 5:37 PM Adam Ford <aford173@gmail.com> wrote:
> > > 
> > > On Fri, Nov 19, 2021 at 10:29 AM Nicolas Dufresne <nicolas@ndufresne.ca> wrote:
> > > > 
> > > > Hi Adam, Tim,
> > > > 
> > > > [...]
> > > > > > > > Nicolas and Adam,
> > > > > > > > 
> > > > > > > > For the H1 patches in this series: I've been able to test the IMX8MM
> > > > > > > > H1 JPEG encode using GStreamer 1.18.5:
> > > > > > > > $ gst-inspect-1.0 | grep -e "v4l2.*enc"
> > > > > > > > video4linux2:  v4l2jpegenc: V4L2 JPEG Encoder
> > > > > > > > $ gst-launch-1.0 videotestsrc ! jpegenc ! rtpjpegpay ! udpsink
> > > > > > >                                   ^ v4l2jpegenc
> > > > > > > 
> > > > > > > This is just a transcript error ?
> > > > > > 
> > > > > > Nicolas,
> > > > > > 
> > > > > > No! Thanks for catching my mistake. I was testing with software encode... ooops!
> > > > > > 
> > > > > > 'gst-launch-1.0 videotestsrc ! v4l2jpegenc ! fakesink' actually hangs
> > > > > > the board so likely a power-domain issue there?
> > > > > 
> > > > > The v4l2-compliance tests fail on the h1 decoder with a hang, but I
> > > > > think we're writing to registers which are not documented in the Mini
> > > > > TRM.  The Mini TRM doesn't explicitly show the JPEG encoding as a
> > > > > feature, but some of the registers state JPEG, but because some of the
> > > > > registers written for the H1 are not documented in the TRM.  If those
> > > > > registers are restricted or not in this SoC, I am concerned that it
> > > > > might be related.  I'll try to run some more tests this weekend to
> > > > > check on the status of the power-domain stuff.
> > > > 
> > > > To verify if the HW support JPEG encoding you can read SWREG63 bit 25. This is
> > > > in the TRM, just not labelled properly. To mimic the decoding side, would be "HW
> > > > synthesis config register X" with the bit labelled SW_ENC_JPEG_PROF (but
> > > > PROF/profile is on or off). If your board hang while reading this, you likely
> > > > didn't get the power bit right.
> > > > 
> > > > IMX8 has an undocumented control block thing that we have been fighting with in
> > > > imx8q,  perhaps that's your issue. Few driver was proposed, we are still pending
> > > > on NXP solution to be submitted (they asked us to wait, still waiting =)).
> > > 
> > > Nicolas,
> > > 
> > > Thanks for the suggestion to read offset FC.  There was an attempt
> > > made by Lucas Stach to develop a VPU blk-ctrl driver to coordinate the
> > > power-domains with the GPC driver. Unfortunately, it does appear to
> > > hang, so it might not be operating correctly.
> > > 
> > > Lucas,
> > > 
> > > Do you have any idea of stuff I can try to see if the power domain is
> > > coming online correctly?
> > > 
> > > [   10.434727] imx-pgc imx-pgc-domain.6: request the vpumix domain to power up
> > > [   10.463647] imx-pgc imx-pgc-domain.6: request the vpumix ADB400 to power up
> > > [   10.517155] imx-pgc imx-pgc-domain.6: genpd vpumix success
> > > [   10.728927] vpu: set fuse bits to enable
> > > [   10.825500] imx8m-blk-ctrl 38330000.blk-ctrl: power vpublk-g1 GPC domain
> > > [   10.878986] imx-pgc imx-pgc-domain.7: request the vpu-g1 domain to power up
> > > [   10.932429] imx-pgc imx-pgc-domain.7: genpd vpu-g1 success
> > > [   10.971988] imx8m-blk-ctrl 38330000.blk-ctrl: genpd vpublk-g1 success
> > > [   11.004726] hantro-vpu 38300000.video-codec: registered
> > > nxp,imx8mm-vpu-dec as /dev/video0
> > > [   11.040760] imx8m-blk-ctrl 38330000.blk-ctrl: power vpublk-g2 GPC domain
> > > [   11.066181] imx-pgc imx-pgc-domain.8: request the vpu-g2 domain to power up
> > > [   11.087887] imx-pgc imx-pgc-domain.8: genpd vpu-g2 success
> > > [   11.113808] imx8m-blk-ctrl 38330000.blk-ctrl: genpd vpublk-g2 success
> > > [   11.139634] hantro-vpu 38310000.video-codec: registered
> > > nxp,imx8mm-vpu-g2-dec as /dev/video1
> > > [   11.156463] imx8m-blk-ctrl 38330000.blk-ctrl: power vpublk-h1 GPC domain
> > > [   11.170817] imx-pgc imx-pgc-domain.9: request the vpu-h1 domain to power up
> > > [   11.232990] imx-pgc imx-pgc-domain.9: genpd vpu-h1 success
> > > [   11.252546] imx8m-blk-ctrl 38330000.blk-ctrl: genpd vpublk-h1 success
> > > [   11.266152] hantro-vpu 38320000.video-codec: Checking vpu->enc_base + 0xfc
> > > 
> > > <hang>
> > > 
> > > adam
> > > 
> > 
> > Nicolas, Tim, and Lucas,
> > 
> > I think I have the hanging resolved in the power domains, and I'll be
> > pushing the fix to the GPCv2.
> > 
> > For the H1 Encoder, I added some debugging code to read the offset
> > 0xfc and print some data based on the findings of that VPU-h1 offset.
> > I basically check the various bits per the TRM to see if they are set
> > and print some splat to indicate whether or not the function is
> > supported.
> > 
> > [    8.861865] hantro-vpu 38320000.video-codec: Checking vpu->enc_base + 0xfc
> > [    8.870594] hantro-vpu 38320000.video-codec: Stabilization supported by HW
> > [    8.889341] hantro-vpu 38320000.video-codec: VP8 encoding supported by HW
> > [    8.899386] hantro-vpu 38320000.video-codec: H.264 encoding supported by HW
> > [    8.918171] hantro-vpu 38320000.video-codec: RGB to YUV conversion
> > supported by HW
> > [    8.934067] hantro-vpu 38320000.video-codec: registered
> > nxp,imx8mm-vpu-h1-enc as /dev/video2
> > 
> > Unfortunately, JPEG is not listed as supported.  :-(
> 
> Adam,
> 
> Well not having JPEG encode support is unfortunate, and unexpected. Do
> we not have hantro support yet for VP8/H264 encode?

There is no mainline support yet. You can derive from RK3288 support using Google ChromeOS method (a v4l2 plugins that simulate in userspace a stateful encoder):

- libv4l plugins / https://chromium.googlesource.com/chromiumos/third_party/libv4lplugins/+/refs/heads/master
- Kernel Driver / https://chromium.googlesource.com/chromiumos/third_party/kernel/+/chromeos-4.4/drivers/media/platform/rockchip-vpu/

> 
> I haven't quite figured out how to build a modern mono-repo gstreamer
> on the ubuntu 20.04 rootfs I'm using so I haven't been able to test
> VPU encode/decode properly. I'll keep working on it when I'm back in
> the office the following week.

Did a quick test to make sure there isn't any ubuntu specific blockers, here's a
dirty script that produce a minimal GStreamer, there was really nothing special
compare to other meson projects. Note that I use --wrap-mode=nofallback to avoid
letting GStreamer complete it's feature-set by downloading the planet. This
already build quite a lot and could likely be made smaller by avoid plugins-good
build-dep call, but then you need to check for v4l2odecs and video4linux devs
(mostly gudev a glib udev binding).

# Install ubuntu
podman run -it --rm ubuntu:20.04
sed -i "s/# deb-src/deb-src/" /etc/apt/sources.list
apt update
apt build-dep gstreamer1.0-plugins-good
apt install git python3-pip flex bison

# Need a newer meson
pip3 install --user meson
export PATH=$PATH:~/.local/bin

# Build GStreamer
git clone https://gitlab.freedesktop.org/gstreamer/gstreamer.git
cd gstreamer
meson setup build --wrap-mode=nofallback
ninja -C build

# Run in-place
./gst-env.py
gst-inspect-1.0 v4l2codecs
gst-inspect 1.0 video4linux2

> 
> Best regards,
> 
> Tim
> 
> > 
> > However, the hanging stops occurring, so I'll be posting a patch to
> > update the GPCv2 code.  I can reduce sone device tree duplication, and
> > the G2 throws some splat, but that will be a separate discussion.
> > 
> > I can also run v4l2-compliance on the H1 node, and it responds without hanging.
> > 
> > root@beacon-imx8mm-kit:~# v4l2-compliance -d2
> > v4l2-compliance SHA: not available
> > , 64 bits, 64-bit time_t
> > 
> > Compliance test for hantro-vpu device /dev/video2:
> > 
> > Driver Info:
> > Driver name      : hantro-vpu
> > Card type        : nxp,imx8mm-vpu-h1-enc
> > Bus info         : platform: hantro-vpu
> > Driver version   : 5.16.0
> > Capabilities     : 0x84204000
> > Video Memory-to-Memory Multiplanar
> > Streaming
> > Extended Pix Format
> > Device Capabilities
> > Device Caps      : 0x04204000
> > Video Memory-to-Memory Multiplanar
> > 
> > < snip>
> > 
> > Total for hantro-vpu device /dev/video2: 46, Succeeded: 46, Failed: 0,
> > Warnings: 0
> > 
> > I'll do an RFCv2 on the Hantro G1 and G2 with the H1 removed based on
> > the updated GPCv2 code I'll be pushing shortly, but at least the
> > system doesn't hang, so I'm fairly confident the power domains are
> > working better now even if we cannot support the JPEG.
> > 
> > adam
> > 
> > > > 
> > > > > > 
> > > > > > > 
> > > > > > > > host=192.168.1.146 port=5000
> > > > > > > > viewed on client@192.168.1.146 via:
> > > > > > > > $ gst-launch-1.0 udpsrc port=5000 ! application/x-rtp,payload=96 !
> > > > > > > > rtpjpegdepay ! jpegdec ! autovideosink
> > > > > > > > 
> > > > > > > > For the G1/G2 patches in the series I don't see any Gstreamer
> > > > > > > > 'v4l2.*dec' elements. Perhaps I need a newer version of Gstreamer.
> > > > > > > 
> > > > > > > Most likely yes, I suggest building gstreamer/ branch "main", GStreamer has now
> > > > > > > a single repository. We are very close to 1.20, which will include stable API
> > > > > > > support of H264, MPEG2 and VP8 decoding.
> > > > > > > 
> > > > > > 
> > > > > > Ok, let me see if I can navigate through the build process and I'll
> > > > > > get back to you.
> > > > > > 
> > > > > > Thanks,
> > > > > > 
> > > > > > Tim
> > > > > > 
> > > > > > > > 
> > > > > > > > I have CSI capture and DSI display currently working on
> > > > > > > > imx8mm-venice-gw73xx-0x that I can play with. The CSI sensor only
> > > > > > > > supports RAW8/RAW10 (and gstreamer currently only supports RAW8) and I
> > > > > > > > can't efficiently convert to something the JPEG encoder likes without
> > > > > > > > bayer2rgbneon (a libneon version).
> > > > > > > > 
> > > > > > > > I see from the IMX8MMRM that the 2D GPU supports scaling etc with a
> > > > > > > > wide range of data formats but I'm not sure how to tap into this as
> > > > > > > > that hardware is managed by the vivante driver. On the IMX6QDL there
> > > > > > > > is a separate IPU block that Philipp Zabel wrote a nice mem2mem
> > > > > > > > csc/scaler driver for but I don't see any equivalent currently for
> > > > > > > > IMX8MM.
> > > > > > > > 
> > > > > > > > Best regards,
> > > > > > > > 
> > > > > > > > Tim
> > > > > > > 
> > > >