Re: [PATCH] [media] v4l2-dv-timings: Introduce v4l2_calc_fps()

From: Hans Verkuil
Date: Tue Mar 14 2017 - 03:25:07 EST

On 03/13/2017 08:03 PM, Jose Abreu wrote:
> Hi Hans,
> On 09-03-2017 15:40, Hans Verkuil wrote:
>> On 09/03/17 16:15, Jose Abreu wrote:
>>> Hi Hans,
>>> Thanks for the review!
>>> On 09-03-2017 12:29, Hans Verkuil wrote:
>>>> On 07/03/17 17:48, Jose Abreu wrote:
>>>>> HDMI Receivers receive video modes which, according to
>>>>> CEA specification, can have different frames per second
>>>>> (fps) values.
>>>>> This patch introduces a helper function in the media core
>>>>> which can calculate the expected video mode fps given the
>>>>> pixel clock value and the horizontal/vertical values. HDMI
>>>>> video receiver drivers are expected to use this helper so
>>>>> that they can correctly fill the v4l2_streamparm structure
>>>>> which is requested by vidioc_g_parm callback.
>>>>> We could also use a lookup table for this but it wouldn't
>>>>> correctly handle 60Hz vs 59.94Hz situations as this all
>>>>> depends on the pixel clock value.
>>>>> Signed-off-by: Jose Abreu <joabreu@xxxxxxxxxxxx>
>>>>> Cc: Carlos Palminha <palminha@xxxxxxxxxxxx>
>>>>> Cc: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
>>>>> Cc: Hans Verkuil <hans.verkuil@xxxxxxxxx>
>>>>> Cc: Charles-Antoine Couret <charles-antoine.couret@xxxxxxxxxxxx>
>>>>> Cc: linux-media@xxxxxxxxxxxxxxx
>>>>> Cc: linux-kernel@xxxxxxxxxxxxxxx
>>>>> ---
>>>>> drivers/media/v4l2-core/v4l2-dv-timings.c | 29 +++++++++++++++++++++++++++++
>>>>> include/media/v4l2-dv-timings.h | 8 ++++++++
>>>>> 2 files changed, 37 insertions(+)
>>>>> diff --git a/drivers/media/v4l2-core/v4l2-dv-timings.c b/drivers/media/v4l2-core/v4l2-dv-timings.c
>>>>> index 5c8c49d..19946c6 100644
>>>>> --- a/drivers/media/v4l2-core/v4l2-dv-timings.c
>>>>> +++ b/drivers/media/v4l2-core/v4l2-dv-timings.c
>>>>> @@ -814,3 +814,32 @@ struct v4l2_fract v4l2_calc_aspect_ratio(u8 hor_landscape, u8 vert_portrait)
>>>>> return aspect;
>>>>> }
>>>>> EXPORT_SYMBOL_GPL(v4l2_calc_aspect_ratio);
>>>>> +
>>>>> +struct v4l2_fract v4l2_calc_fps(const struct v4l2_dv_timings *t)
>>>>> +{
>>>>> + const struct v4l2_bt_timings *bt = &t->bt;
>>>>> + struct v4l2_fract fps_fract = { 1, 1 };
>>>>> + unsigned long n, d;
>>>>> + unsigned long mask = GENMASK(BITS_PER_LONG - 1, 0);
>>>> This is wrong since v4l2_fract uses u32, and LONG can be 64 bits.
>>> Yes, its wrong. I will remove the variable and just use fps, 100
>>> instead of mask, mask.
>>>>> + u32 htot, vtot, fps;
>>>>> + u64 pclk;
>>>>> +
>>>>> + if (t->type != V4L2_DV_BT_656_1120)
>>>>> + return fps_fract;
>>>>> +
>>>>> + htot = V4L2_DV_BT_FRAME_WIDTH(bt);
>>>>> + vtot = V4L2_DV_BT_FRAME_HEIGHT(bt);
>>>>> + pclk = bt->pixelclock;
>>>>> + if (bt->interlaced)
>>>>> + htot /= 2;
>>>> This can be dropped. This is the timeperframe, not timeperfield. So for interleaved
>>>> formats the time is that of two fields (aka one frame).
>>> Ok, but then there is something not correct in
>>> v4l2_dv_timings_presets structure field values because I get
>>> wrong results in double clocked modes. I checked the definition
>>> and the modes that are double clocked are defined with half the
>>> clock, i.e., V4L2_DV_BT_CEA_720X480I59_94 is defined with a pixel
>>> clock of 13.5MHz but in CEA spec this mode is defined with pixel
>>> clock of 27MHz.
>> It's defined in the CEA spec as 1440x480 which is the double clocked
>> version of 720x480.
>> The presets are defined without any pixel repeating. In fact, no driver
>> that is in the kernel today supports pixel repeating. Mostly because there was
>> never any need since almost nobody uses resolutions that require this.
>> If you decide to add support for this, then it would not surprise me if
>> some of the core dv-timings support needs to be adjusted.
>> To be honest, I never spent time digging into the pixel repeating details,
>> so I am not an expert on this at all.
>>>>> +
>>>>> + fps = (htot * vtot) > 0 ? div_u64((100 * pclk), (htot * vtot)) : 0;
>>>>> +
>>>>> + rational_best_approximation(fps, 100, mask, mask, &n, &d);
>>>> I think you can just use fps, 100 instead of mask, mask.
>>>> What is returned if fps == 0?
>>> I will add a check for this.
>>>> I don't have a problem as such with this function, but just be aware that the
>>>> pixelclock is never precise: there are HDMI receivers that are unable to report
>>>> the pixelclock with enough precision to even detect if it is 60 vs 59.94 Hz.
>>>> And even for those that can, it is often not reliable.
>>> My initial intention for this function was that it should be used
>>> with v4l2_find_dv_timings_cea861_vic, when possible. That is,
>>> HDMI receivers have access to AVI infoframe contents. Then they
>>> should get the vic, call v4l2_find_dv_timings_cea861_vic, get
>>> timings and then call v4l2_calc_fps to get fps. If no AVI
>>> infoframe is available then it should resort to pixel clock and
>>> H/V measures as last resort.
>> Right, but there are no separate VIC codes for 60 vs 59.94 Hz. Any vertical
>> refresh rate that can be divided by 6 can also support these slightly lower
>> refresh rates. The timings returned by v4l2_find_dv_timings_cea861_vic just
>> report if that is possible, but the pixelclock is set for 24, 30 or 60 fps.
> Right, I was forgetting about this ...
> So:
> 1) Most of HDMI receivers do not have the expected precision in
> measuring pixel clock value;


Newer HDMI receivers tend to have better precision.

However, the 1000/1001 factor is within the error of margin that the HDMI
spec has for the pixelclock, so even if it is 59.94 you still (theoretically)
do not know if that is because it really has that fps or if the source just has
a bad clock.

It's a bit theoretical, in practice you can assume the source really is sending
at 59.94 AFAIK.

> 2) Most (I would guess all of them?) have access to AVI infoframe
> contents;

All will have that.

> 3) The FPS value is generally used by applications to calculate
> expected frame rate and number of frames dropped (right?);

Not really. Most HDMI drivers do not implement g_parm, instead they fill in
the detected pixelclock in QUERY_DV_TIMINGS, leaving it up to the application
to calculate the fps from that.

> 4) The factor in FPS value can be adjusted by 1000/1001;
> From these points I would propose in just using the vic and drop
> the resolution in fps a little bit, do you agree?

The reality is that how to detect the 1000/1001 reduced fps is fuzzy. Part of
the reason for that is that most of the HDMI receivers we have in the kernel
were developed by Cisco/Tandberg (i.e. mostly me) for our video conferencing
systems, and those all run at 60 Hz. So we never had the need to detect 59.94 vs
60 Hz. In addition, some of the older Analog Devices devices didn't have the
resolution to detect the difference.

So I always held off a bit with defining exactly how to do this since I had
no experience with it.

My question to you is: can you reliably detect the difference between 60 and 59.94
Hz and between 24 and 23.976 Hz by just the measured pixelclock?

You need to test this with different sources, not just signal generators. You
probably get a range of pixelclock values for the same framerate for different
sources, since each source has their own clock.

My preference would be to extend query_dv_timings a bit for this:

<brainstorm mode on>
Add a flag V4L2_DV_FL_CAN_DETECT_REDUCED_FPS. If set, then the hw can detect the
difference between regular fps and 1000/1001 fps. Note: this is only valid for
timings of VIC codes with the V4L2_DV_FL_CAN_REDUCE_FPS flag set.

Allow V4L2_DV_FL_REDUCED_FPS to be used for receivers if V4L2_DV_FL_CAN_DETECT_REDUCED_FPS
is set.

For standard VIC codes the pixelclock returned by query_dv_timings is that of the
corresponding VIC timing, not what is measured. This will ensure fixed fps values

g_parm should calculate the fps based on the v4l2_bt_timings struct, looking at the

For those receivers that cannot detect the difference, the fps will be 24/30/60 Hz,
for those that can detect the difference g_parm can check if both V4L2_DV_FL_CAN_DETECT_REDUCED_FPS
and V4L2_DV_FL_REDUCED_FPS are set and reduce the fps by 1000/1001.
<brainstorm mode off>

If your hw can reliably detect the difference, then now is a good time to close
this gap in the DV_TIMINGS API.