On Fri, 10 May 2019 16:11:54 +0200, Hans Verkuil wrote:
> On 5/10/19 3:52 PM, Michael Tretter wrote:
> > On Fri, 10 May 2019 12:58:43 +0200, Hans Verkuil wrote:
> >> On 5/10/19 12:28 PM, Michael Tretter wrote:
> >>> On Fri, 10 May 2019 10:28:53 +0200, Hans Verkuil wrote:
> >>>> On 5/3/19 2:20 PM, Michael Tretter wrote:
> >>>>> Add a V4L2 mem-to-mem driver for Allegro DVT video IP cores as found in
> >>>>> the EV family of the Xilinx ZynqMP SoC. The Zynq UltraScale+ Device
> >>>>> Technical Reference Manual uses the term VCU (Video Codec Unit) for the
> >>>>> encoder, decoder and system integration block.
> >>>>>
> >>>>> This driver takes care of interacting with the MicroBlaze MCU that
> >>>>> controls the actual IP cores. The IP cores and MCU are integrated in the
> >>>>> FPGA. The xlnx_vcu driver is responsible for configuring the clocks and
> >>>>> providing information about the codec configuration.
> >>>>>
> >>>>> The driver currently only supports the H.264 video encoder.
> >>>>>
> >>>>> Signed-off-by: Michael Tretter <m.tretter@xxxxxxxxxxxxxx>
> >>>>> ---
> >>
> >> <snip>
> >>
> >>>>> +static int allegro_try_fmt_vid_out(struct file *file, void *fh,
> >>>>> + struct v4l2_format *f)
> >>>>> +{
> >>>>> + f->fmt.pix.field = V4L2_FIELD_NONE;
> >>>>> +
> >>>>> + f->fmt.pix.width = clamp_t(__u32, f->fmt.pix.width,
> >>>>> + ALLEGRO_WIDTH_MIN, ALLEGRO_WIDTH_MAX);
> >>>>> + f->fmt.pix.height = clamp_t(__u32, f->fmt.pix.height,
> >>>>> + ALLEGRO_HEIGHT_MIN, ALLEGRO_HEIGHT_MAX);
> >>>>
> >>>> Shouldn't this be rounded up to the macroblock size? Or is the encoder
> >>>> smart enough to do the padding internally?
> >>>
> >>> The driver sends a message with the visible size of the raw frames
> >>> (without macroblock alignment) to the encoder firmware. Therefore, the
> >>> encoder firmware is responsible for handling the padding to macroblock
> >>> size.
> >>
> >> Please add a comment describing this. It is unusual for encoders to be
> >> able to do this so it is good to document this.
> >
> > OK.
> >
> >>
> >>>
> >>> Furthermore, the encoder requires that the stride is 32 byte aligned.
> >>> Therefore, we naturally have a macroblock alignment regarding the
> >>> width, but not regarding the height. This limitation is already
> >>> included in the bytesperline field.
> >>
> >> Ack.
> >>
> >>>
> >>>>
> >>>>> +
> >>>>> + f->fmt.pix.pixelformat = V4L2_PIX_FMT_NV12;
> >>>>> + f->fmt.pix.bytesperline = round_up(f->fmt.pix.width, 32);
> >>>>> + f->fmt.pix.sizeimage =
> >>>>> + f->fmt.pix.bytesperline * f->fmt.pix.height * 3 / 2;
> >>>>> +
> >>>>> + return 0;
> >>>>> +}
> >>>>> +
> >>>>> +static int allegro_s_fmt_vid_out(struct file *file, void *fh,
> >>>>> + struct v4l2_format *f)
> >>>>> +{
> >>>>> + struct allegro_channel *channel = fh_to_channel(fh);
> >>>>> + int err;
> >>>>> +
> >>>>> + err = allegro_try_fmt_vid_out(file, fh, f);
> >>>>> + if (err)
> >>>>> + return err;
> >>>>> +
> >>>>> + channel->width = f->fmt.pix.width;
> >>>>> + channel->height = f->fmt.pix.height;
> >>>>> + channel->stride = f->fmt.pix.bytesperline;
> >>>>> + channel->sizeimage_raw = f->fmt.pix.sizeimage;
> >>>>> +
> >>>>> + channel->colorspace = f->fmt.pix.colorspace;
> >>>>> + channel->ycbcr_enc = f->fmt.pix.ycbcr_enc;
> >>>>> + channel->quantization = f->fmt.pix.quantization;
> >>>>> + channel->xfer_func = f->fmt.pix.xfer_func;
> >>>>> +
> >>>>> + channel->level =
> >>>>> + select_minimum_h264_level(channel->width, channel->height);
> >>>>> + channel->sizeimage_encoded =
> >>>>> + estimate_stream_size(channel->width, channel->height);
> >>>>> +
> >>>>> + return 0;
> >>>>> +}
> >>>>> +
> >>>>> +static int allegro_g_selection(struct file *file, void *priv,
> >>>>> + struct v4l2_selection *s)
> >>>>> +{
> >>>>> + struct v4l2_fh *fh = file->private_data;
> >>>>> + struct allegro_channel *channel = fh_to_channel(fh);
> >>>>> +
> >>>>> + if (!V4L2_TYPE_IS_OUTPUT(s->type))
> >>>>> + return -EINVAL;
> >>>>> +
> >>>>> + switch (s->target) {
> >>>>> + case V4L2_SEL_TGT_CROP:
> >>>>> + case V4L2_SEL_TGT_CROP_DEFAULT:
> >>>>> + case V4L2_SEL_TGT_CROP_BOUNDS:
> >>>>> + s->r.left = 0;
> >>>>> + s->r.top = 0;
> >>>>> + s->r.width = channel->width;
> >>>>> + s->r.height = channel->height;
> >>>>
> >>>> I don't think this is quite right. The CROP target should return the visible
> >>>> width/height (e.g. 1920x1080) whereas the other two targets should return the
> >>>> coded width/height (e.g. 1920x1088 when rounded to the macroblock alignment).
> >>>>
> >>>> Note: if the hardware doesn't require that the raw frame is macroblock aligned,
> >>>> then I need to think a bit more about how the selection handling should be
> >>>> done.
> >>>
> >>> The driver internally calculates the coded width/height in macroblocks
> >>> and cropping and writes it to the SPS. Currently, this isn't exposed to
> >>> userspace, because I don't see a need to tell the userspace about that.
> >>>
> >>> If there is a reason to expose this to userspace, I am fine with
> >>> implementing that.
> >>
> >> There really is no need for the selection API at all. Just drop both
> >> G and S_SELECTION from the driver. Let me know if the compliance test
> >> fails for drivers without selection support, I'll have to fix the test
> >> in that case.
> >
> > The compliance test for VIDIOC_S_FMT fails with the following message
> > if G_SELECTION is not implemented:
> >
> > fail: v4l2-test-formats.cpp(836): sel.r.width != fmt.g_width()
> > test VIDIOC_S_FMT: FAIL
> >
>
> Try this patch:
>
> Signed-off-by: Hans Verkuil <hverkuil-cisco@xxxxxxxxx>
Tested-by: Michael Tretter <m.tretter@xxxxxxxxxxxxxx>
> ---
> diff --git a/utils/v4l2-compliance/v4l2-test-formats.cpp b/utils/v4l2-compliance/v4l2-test-formats.cpp
> index fc497e3c..544ecb5c 100644
> --- a/utils/v4l2-compliance/v4l2-test-formats.cpp
> +++ b/utils/v4l2-compliance/v4l2-test-formats.cpp
> @@ -828,7 +828,11 @@ static int testM2MFormats(struct node *node)
> .type = fmt.g_type(),
> .target = V4L2_SEL_TGT_CROP,
> };
> - node->g_selection(sel);
> + if (node->g_selection(sel) == ENOTTY) {
> + fail_on_test(fmt_cap.g_width() != fmt.g_width());
> + fail_on_test(fmt_cap.g_height() != fmt.g_height());
> + return 0;
> + }
> fail_on_test(sel.r.top || sel.r.left);
> fail_on_test(sel.r.width != fmt.g_width());
> fail_on_test(sel.r.height != fmt.g_height());
> ------------------------------------------------------------
Thanks. First I thought that it is strange to successfully finish the
test early and skip all remaining tests, but all remaining tests are
void anyway if g_selection is not implemented.
Michael
>
> Regards,
>
> Hans
>
