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

[Nicolas Dufresne]

Le mardi 16 novembre 2021 à 15:23 -0800, Tim Harvey a écrit :
> On Tue, Nov 9, 2021 at 7:57 AM Nicolas Dufresne <nicolas@xxxxxxxxxxxx> wrote:
> > 
> > Le lundi 08 novembre 2021 à 10:33 -0600, Adam Ford a écrit :
> > > On Mon, Nov 8, 2021 at 7:59 AM Nicolas Dufresne <nicolas@xxxxxxxxxxxx> wrote:
> > > > 
> > > > Hi Adam,
> > > > 
> > > > thanks for you work, I'll try and reply  about the GStreamer questions below, if
> > > > you have further question feel free to ask.
> > > > 
> > > > Le samedi 06 novembre 2021 à 13:37 -0500, Adam Ford a écrit :
> > > > > The i.MX8M has two Hantro video decoders, called G1 and G2 which appear
> > > > > to be related to the video decoders used on the i.MX8MQ, but because of
> > > > > how the Mini handles the power domains, the VPU driver does not need to
> > > > > handle all the functions, so a new compatible flag is required.
> > > > > 
> > > > > This is an RFC because I don't have functional video on my system yet,
> > > > > and I'm hoping there might be people who do and can help test this.
> > > > > I have only tested this far enough to see they enumerate and appear
> > > > > as /dev/videoX and /dev/mediaX devices.
> > > > 
> > > > I will check the patchset, but you need in the mini-variant to disable the G1
> > > > post processor, because this block was fused out. We didn't make it optional
> > > 
> > > Thanks for being willing to review this.
> > > 
> > > > from the start as according to the V1 of the TRM it was there, but that error
> > > > was corrected in V3.
> > > 
> > > Thanks for the clarification.  It wasn't obvious to me, because in
> > > some instances the PP looked like it was there and sometimes not
> > > there.  I'll remove the postproc stuff.
> > > 
> > > > 
> > > > > 
> > > > > I am also curious to know if/what gstreamer plugins are necessary.  In
> > > > > NXP's custom kernel, there are IMX-specific plugins, and I was hoping there
> > > > > would be more generic plugins that I can use to test.  I am hoping some
> > > > > of the linux-media experts might chime in on how to best test.
> > > > 
> > > > I will recommend using GStreamer 1.19.3 or main branch (GStreamer is now a
> > > > single git repo). You will then be able to test Hantro G1 decoding of MPEG2,
> > > > H264 and VP8. Remember that the related plugin depends on libgudev (a glib
> > > > binding of udev).
> > > 
> > > Thanks for the tip.
> > > 
> > > > 
> > > > For the encoder, I believe only JPEG maybe supported, since this is all there is
> > > > mainline for RK3288 (and perhaps other RK). But this will need testing and
> > > > debugging as the G1 version is slightly newer on NXP SoC.
> > > 
> > > For what it's worth the G1 seems to repond cleanly to the inquiries
> > > from v42-compliance.
> > > The G2 throws some splat when I run v4l2-compliance, but I am still
> > > investigating that.
> > > 
> > > [  405.456979] ------------[ cut here ]------------
> > > [  405.464173] WARNING: CPU: 0 PID: 563 at mm/page_alloc.c:5344
> > > __alloc_pages+0x5a4/0xbe0
> > > [  405.472104] Modules linked in: 8021q garp mrp stp llc caam_jr
> > > caamhash_desc caamalg_desc crypto_engine rng_core authenc libdes
> > > imx7_media_csi(C) crct10dif_ce imx_media_common(C)
> > > snd_soc_fsl_asoc_card imx7_mipi_csis(C) snd_soc_imx_audmux
> > > snd_soc_simple_card_utils fsl_imx8_ddr_perf imx8m_ddrc brcmfmac
> > > brcmutil hantro_vpu(C) v4l2_h264 v4l2_mem2mem videobuf2_vmalloc
> > > videobuf2_dma_contig videobuf2_memops cfg80211 ov5640 videobuf2_v4l2
> > > v4l2_fwnode v4l2_async videobuf2_common videodev etnaviv gpu_sched
> > > hci_uart mc btqca btbcm snd_soc_wm8962 at24 spi_imx rtc_pcf85363
> > > rtc_snvs clk_bd718x7 spi_bitbang snvs_pwrkey snd_soc_fsl_sai
> > > imx_pcm_dma caam error bluetooth imx8mm_thermal ecdh_generic
> > > imx_cpufreq_dt ecc rfkill fuse drm ipv6
> > > [  405.535835] CPU: 0 PID: 563 Comm: v4l2-compliance Tainted: G      D
> > >  C        5.15.0-next-20211105-00010-g4bb8e8a25d3c-dirty #28
> > > [  405.547401] Hardware name: Beacon EmbeddedWorks i.MX8M Mini
> > > Development Kit (DT)
> > > [  405.554797] pstate: 20000005 (nzCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> > > [  405.561762] pc : __alloc_pages+0x5a4/0xbe0
> > > [  405.565861] lr : __dma_direct_alloc_pages+0x17c/0x1e0
> > > [  405.570917] sp : ffff800012443810
> > > [  405.574232] x29: ffff800012443810 x28: 0000000000000000 x27: ffff000005288220
> > > [  405.581375] x26: 0000000000000034 x25: 0000000000000000 x24: ffff000000259010
> > > [  405.588517] x23: ffff80001011ab7c x22: ffff000000259010 x21: 00000000ffffffff
> > > [  405.595659] x20: 0000000000000cc1 x19: 0000000000000000 x18: 0000000000000000
> > > [  405.602803] x17: 0000000000000000 x16: 0000000000000000 x15: 0000000000000000
> > > [  405.609947] x14: 0000000000000001 x13: 0000000000000000 x12: 0000000000000000
> > > [  405.617090] x11: ffff80001241d000 x10: ffff00000528833a x9 : ffff00000528832a
> > > [  405.624232] x8 : 0000000000000000 x7 : 0000000000000000 x6 : 0000000000000cc0
> > > [  405.631378] x5 : 00000000bfffffff x4 : ffff000009e30dc0 x3 : 0000000000000000
> > > [  405.638520] x2 : 0000000000000000 x1 : 0000000000000001 x0 : 0000000000000cc1
> > > [  405.645666] Call trace:
> > > [  405.648113]  __alloc_pages+0x5a4/0xbe0
> > > [  405.651862]  __dma_direct_alloc_pages+0x17c/0x1e0
> > > [  405.656569]  dma_direct_alloc+0x70/0x310
> > > [  405.660494]  dma_alloc_attrs+0x7c/0xe4
> > > [  405.664246]  hantro_hevc_get_ref_buf+0x15c/0x184 [hantro_vpu]
> > > [  405.670021]  hantro_g2_hevc_dec_run+0x3b8/0x1910 [hantro_vpu]
> > > [  405.675791]  device_run+0xac/0x110 [hantro_vpu]
> > > [  405.680345]  v4l2_m2m_try_run+0xac/0x1b0 [v4l2_mem2mem]
> > > [  405.685598]  v4l2_m2m_ioctl_streamon+0x84/0xa0 [v4l2_mem2mem]
> > > [  405.691366]  v4l_streamon+0x28/0x34 [videodev]
> > > [  405.695877]  __video_do_ioctl+0x178/0x3dc [videodev]
> > > [  405.700897]  video_usercopy+0x368/0x6dc [videodev]
> > > [  405.705745]  video_ioctl2+0x1c/0x30 [videodev]
> > > [  405.710246]  v4l2_ioctl+0x44/0x64 [videodev]
> > > [  405.714574]  __arm64_sys_ioctl+0xac/0xf0
> > > [  405.718502]  invoke_syscall+0x48/0x114
> > > [  405.722258]  el0_svc_common.constprop.0+0xd4/0xfc
> > > [  405.726969]  do_el0_svc+0x2c/0x94
> > > [  405.730286]  el0_svc+0x28/0x80
> > > [  405.733348]  el0t_64_sync_handler+0xa8/0x130
> > > [  405.737619]  el0t_64_sync+0x1a0/0x1a4
> > > [  405.741287] ---[ end trace 270ed4a899803006 ]---
> > > 
> > > The H1 encoder seems to hang the system when I run v4l2-compliance on
> > > it when H1 is set up as I submitted the patch.  I tried dropping all
> > > the encoder formats except the JPEG format, and it doesn't hang any
> > > more, but it also doesn't really do anything.
> > > The datasheet only references VPU_H1 as supporting VP8 and H.264, so I
> > > am not sure JPEG is even supported.
> > 
> > If JPEG is not supported, then there is nothing left for mainline in this
> > regard. The kernel control interface and encoding flow needs to be designed and
> > specified for encoders like VP8 and H264. Some prototypes and prior-art exist
> > though, but nothing ever got formalized in the form of a specification.
> > 
> > > 
> > > The log from v4l2-compliance on the H1 with everything except the JPEG
> > > removed looks like:
> > > 
> > > root@beacon-imx8mm-kit:~# v4l2-compliance -d2
> > > v4l2-compliance SHA: not available
> > > , 64 bits, 64-bit time_t
> > > 
> > > Segmentation fault
> > > root@beacon-imx8mm-kit:~#
> > > Message from syslogd@  at Thu Jan  1 00:05:07 1970 ...
> > > : Internal error: Oops: 96000004 [#2] SMP
> > > 
> > > Message from syslogd@  at Thu Jan  1 00:05:07 1970 ...
> > > : Code: 52800001 aa1403e0 d2801802 95c31ab9 (b9400aa1)
> > > 
> > > I want to install Gstreamer, but I don't have functioning DSI video,
> > > so I am not entirely sure how I will go about testing the decoders
> > > except by using fakesink
> > 
> > We too don't have an mainline DSI to test the CODECs on recent NXP SoC. For
> > decoders we use fluster, a tool that runs publicly available conformance test.
> > It will simply decode to disk and compare a checksum of the decoded image
> > against the compliant checksum (produced by the reference decoders). For you
> > interested, it uses the new videocodectestsink, which is specialized for
> > producing or calculating conformance image/checksum.
> > 
> > https://github.com/fluendo/fluster
> > 
> > We have added support for GStreamer stateless decoders already.
> > 
> > > 
> > > If the G1 ends up working with some of the newer Gstreamer stuff, I
> > > might just submit a formal patch to just enable the G1 for now.
> > 
> > This looks like a good idea indeed.
> > 
> > > 
> > > adam
> > > > 
> > > > > 
> > > > > Lastly, I didn't update any device tree binding YAML files, because
> > > > > I know there have been some discussions about the power domains on the
> > > > > imx8mq, and I wasn't sure if the imx8mm should get a separate YAML file
> > > > > or if the existing one for te imx8mq should just be modified.
> > > > > 
> > > > > This will likely require the following series in order to apply correctly:
> > > > > https://patchwork.kernel.org/project/linux-arm-kernel/list/?series=576407
> > > > > 
> > > > > Adam Ford (5):
> > > > >   media: hantro: Add support for i.MX8M Mini
> > > > >   arm64: dts: imx8mm:  Enable VPU-G1 and VPU-G2
> > > > >   media: hantro: Rename ROCKCHIP_VPU_ENC_FMT to HANTRO_VPU_ENC_FMT
> > > > >   media: hantro: Add H1 encoder support on i.MX8M Mini
> > > > >   arm64: dts: imx8mm:  Enable Hantro H1 Encoder
> > > > > 
> > > > >  arch/arm64/boot/dts/freescale/imx8mm.dtsi     |  61 ++++++++
> > > > >  drivers/staging/media/hantro/hantro_drv.c     |   3 +
> > > > >  drivers/staging/media/hantro/hantro_hw.h      |  19 ++-
> > > > >  drivers/staging/media/hantro/imx8m_vpu_hw.c   | 143 ++++++++++++++++++
> > > > >  .../staging/media/hantro/rockchip_vpu_hw.c    |  26 ++--
> > > > >  5 files changed, 231 insertions(+), 21 deletions(-)
> > > > > 
> > > > 
> > 
> 
> 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 ?

> 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.

> 
> 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