[v2,3/4] media: staging: rkisp1: rsz: set output format to YUV422 if cap format is YUV444
diff mbox series

Message ID 20200515142952.20163-4-dafna.hirschfeld@collabora.com
State New
Headers show
Series
  • media: staging: rkisp1: various fixes to capture formats
Related show

Commit Message

Dafna Hirschfeld May 15, 2020, 2:29 p.m. UTC
If the capture format is YUV444M then the memory input format
should be YUV422, so the resizer should not change the default
hdiv, vdiv in that case.

Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
---
 drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

Comments

Helen Koike May 20, 2020, 10:23 p.m. UTC | #1
Hi Dafna,

On 5/15/20 11:29 AM, Dafna Hirschfeld wrote:
> If the capture format is YUV444M then the memory input format
> should be YUV422, so the resizer should not change the default
> hdiv, vdiv in that case.

I didn't understand why YUV444M is special, do you mind elaborating a bit
more to help me understand please?

Thanks
Helen

> 
> Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
> ---
>  drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> index 04a29af8cc92..5f9740ddd558 100644
> --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
> +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
>  	 * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
>  	 * streams should be set according to the pixel format in the capture.
>  	 * The resizer always gets the input as YUV422. If the capture format
> -	 * is RGB then the memory input (the resizer output) should be YUV422
> -	 * so we use the hdiv, vdiv of the YUV422 info in this case.
> +	 * is RGB or YUV444 then the memory input (the resizer output) should
> +	 * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
>  	 */
> -	if (v4l2_is_format_yuv(cap->pix.info)) {
> +	if (v4l2_is_format_yuv(cap->pix.info) &&
> +	    cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
>  		src_c.width = cap->pix.info->hdiv;
>  		src_c.height = cap->pix.info->vdiv;
>  	} else {
>
Tomasz Figa May 26, 2020, 11:09 p.m. UTC | #2
On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
<dafna.hirschfeld@collabora.com> wrote:
>
> If the capture format is YUV444M then the memory input format
> should be YUV422, so the resizer should not change the default
> hdiv, vdiv in that case.
>
> Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
> ---
>  drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> index 04a29af8cc92..5f9740ddd558 100644
> --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
> +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
>          * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
>          * streams should be set according to the pixel format in the capture.
>          * The resizer always gets the input as YUV422. If the capture format
> -        * is RGB then the memory input (the resizer output) should be YUV422
> -        * so we use the hdiv, vdiv of the YUV422 info in this case.
> +        * is RGB or YUV444 then the memory input (the resizer output) should
> +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
>          */
> -       if (v4l2_is_format_yuv(cap->pix.info)) {
> +       if (v4l2_is_format_yuv(cap->pix.info) &&
> +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
>                 src_c.width = cap->pix.info->hdiv;
>                 src_c.height = cap->pix.info->vdiv;

As pointed out in another thread, this should have been the original
size divided by the divisor and not just the latter alone.

It seems a bit suspicious to me that we don't need to upscale the
chroma planes here, because it would mean that the MI itself would be
doing some horizontal pixel doubling. The hardware documentation
doesn't really explain this, though.

Have you been able to validate that the setting without upscaling
indeed produces correct output?

Best regards,
Tomasz
Dafna Hirschfeld June 12, 2020, 12:45 p.m. UTC | #3
On 27.05.20 01:09, Tomasz Figa wrote:
> On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
> <dafna.hirschfeld@collabora.com> wrote:
>>
>> If the capture format is YUV444M then the memory input format
>> should be YUV422, so the resizer should not change the default
>> hdiv, vdiv in that case.
>>
>> Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
>> ---
>>   drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
>>   1 file changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>> index 04a29af8cc92..5f9740ddd558 100644
>> --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
>> +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>> @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
>>           * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
>>           * streams should be set according to the pixel format in the capture.
>>           * The resizer always gets the input as YUV422. If the capture format
>> -        * is RGB then the memory input (the resizer output) should be YUV422
>> -        * so we use the hdiv, vdiv of the YUV422 info in this case.
>> +        * is RGB or YUV444 then the memory input (the resizer output) should
>> +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
>>           */
>> -       if (v4l2_is_format_yuv(cap->pix.info)) {
>> +       if (v4l2_is_format_yuv(cap->pix.info) &&
>> +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
>>                  src_c.width = cap->pix.info->hdiv;
>>                  src_c.height = cap->pix.info->vdiv;
> 
> As pointed out in another thread, this should have been the original
> size divided by the divisor and not just the latter alone.
> 
> It seems a bit suspicious to me that we don't need to upscale the
> chroma planes here, because it would mean that the MI itself would be
> doing some horizontal pixel doubling. The hardware documentation
> doesn't really explain this, though.
> 
> Have you been able to validate that the setting without upscaling
> indeed produces correct output?

Hi,
I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
but there is a problem on selfpath: https://imgur.com/a/vQJPqAn

This patch somehow solves the problem on selfpath but ruins the frames on mainpath.

I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
But the value should be set according to the configuration. For some reason the problem arises only
when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
Setting the the SP_INPUT_* according to the configuration seems to solve the problem.

Thanks,
Dafna


> 
> Best regards,
> Tomasz
>
Tomasz Figa June 18, 2020, 5:47 p.m. UTC | #4
On Fri, Jun 12, 2020 at 02:45:00PM +0200, Dafna Hirschfeld wrote:
> 
> 
> On 27.05.20 01:09, Tomasz Figa wrote:
> > On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
> > <dafna.hirschfeld@collabora.com> wrote:
> > > 
> > > If the capture format is YUV444M then the memory input format
> > > should be YUV422, so the resizer should not change the default
> > > hdiv, vdiv in that case.
> > > 
> > > Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
> > > ---
> > >   drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
> > >   1 file changed, 4 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > index 04a29af8cc92..5f9740ddd558 100644
> > > --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
> > >           * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
> > >           * streams should be set according to the pixel format in the capture.
> > >           * The resizer always gets the input as YUV422. If the capture format
> > > -        * is RGB then the memory input (the resizer output) should be YUV422
> > > -        * so we use the hdiv, vdiv of the YUV422 info in this case.
> > > +        * is RGB or YUV444 then the memory input (the resizer output) should
> > > +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
> > >           */
> > > -       if (v4l2_is_format_yuv(cap->pix.info)) {
> > > +       if (v4l2_is_format_yuv(cap->pix.info) &&
> > > +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
> > >                  src_c.width = cap->pix.info->hdiv;
> > >                  src_c.height = cap->pix.info->vdiv;
> > 
> > As pointed out in another thread, this should have been the original
> > size divided by the divisor and not just the latter alone.
> > 
> > It seems a bit suspicious to me that we don't need to upscale the
> > chroma planes here, because it would mean that the MI itself would be
> > doing some horizontal pixel doubling. The hardware documentation
> > doesn't really explain this, though.
> > 
> > Have you been able to validate that the setting without upscaling
> > indeed produces correct output?
> 
> Hi,
> I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
> but there is a problem on selfpath: https://imgur.com/a/vQJPqAn
> 
> This patch somehow solves the problem on selfpath but ruins the frames on mainpath.
> 
> I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
> sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
> But the value should be set according to the configuration. For some reason the problem arises only
> when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
> value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
> Setting the the SP_INPUT_* according to the configuration seems to solve the problem.

Looking at the datasheet, SP seems to have some internal format
conversion capability - one can set SP_OUTPUT_FORMAT and SP_INPUT_FORMAT
to different YUV subsampling modes or even some RGB formats. MP doesn't
have this capability - it can only write whatever it receives.

I think this needs to be reflected in the driver, if it isn't yet.
Depending on the resizer source format, the MP video node must allow
only formats with matching subsampling mode.

Best regards,
Tomasz
Dafna Hirschfeld June 18, 2020, 6 p.m. UTC | #5
On 18.06.20 19:47, Tomasz Figa wrote:
> On Fri, Jun 12, 2020 at 02:45:00PM +0200, Dafna Hirschfeld wrote:
>>
>>
>> On 27.05.20 01:09, Tomasz Figa wrote:
>>> On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
>>> <dafna.hirschfeld@collabora.com> wrote:
>>>>
>>>> If the capture format is YUV444M then the memory input format
>>>> should be YUV422, so the resizer should not change the default
>>>> hdiv, vdiv in that case.
>>>>
>>>> Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
>>>> ---
>>>>    drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
>>>>    1 file changed, 4 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>> index 04a29af8cc92..5f9740ddd558 100644
>>>> --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>> +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>> @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
>>>>            * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
>>>>            * streams should be set according to the pixel format in the capture.
>>>>            * The resizer always gets the input as YUV422. If the capture format
>>>> -        * is RGB then the memory input (the resizer output) should be YUV422
>>>> -        * so we use the hdiv, vdiv of the YUV422 info in this case.
>>>> +        * is RGB or YUV444 then the memory input (the resizer output) should
>>>> +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
>>>>            */
>>>> -       if (v4l2_is_format_yuv(cap->pix.info)) {
>>>> +       if (v4l2_is_format_yuv(cap->pix.info) &&
>>>> +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
>>>>                   src_c.width = cap->pix.info->hdiv;
>>>>                   src_c.height = cap->pix.info->vdiv;
>>>
>>> As pointed out in another thread, this should have been the original
>>> size divided by the divisor and not just the latter alone.
>>>
>>> It seems a bit suspicious to me that we don't need to upscale the
>>> chroma planes here, because it would mean that the MI itself would be
>>> doing some horizontal pixel doubling. The hardware documentation
>>> doesn't really explain this, though.
>>>
>>> Have you been able to validate that the setting without upscaling
>>> indeed produces correct output?
>>
>> Hi,
>> I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
>> but there is a problem on selfpath: https://imgur.com/a/vQJPqAn
>>
>> This patch somehow solves the problem on selfpath but ruins the frames on mainpath.
>>
>> I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
>> sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
>> But the value should be set according to the configuration. For some reason the problem arises only
>> when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
>> value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
>> Setting the the SP_INPUT_* according to the configuration seems to solve the problem.
> 
> Looking at the datasheet, SP seems to have some internal format
> conversion capability - one can set SP_OUTPUT_FORMAT and SP_INPUT_FORMAT
> to different YUV subsampling modes or even some RGB formats. MP doesn't
> have this capability - it can only write whatever it receives.
> 
> I think this needs to be reflected in the driver, if it isn't yet.
> Depending on the resizer source format, the MP video node must allow
> only formats with matching subsampling mode.

Hi,
I work on v3 now that does that, it supports several yuv mbus codes on the source pads of the resizers
and then in the link_validation callback of the capture it checks that the subsampling matches.

It is not clear to me what is the meaning of the input/output capapility of the selfpath except of RGB convertion.
Since yuv subsampling is what the resizer does.

Thanks,
Dafna


> 
> Best regards,
> Tomasz
>
Tomasz Figa June 18, 2020, 6:12 p.m. UTC | #6
On Thu, Jun 18, 2020 at 08:00:37PM +0200, Dafna Hirschfeld wrote:
> 
> 
> On 18.06.20 19:47, Tomasz Figa wrote:
> > On Fri, Jun 12, 2020 at 02:45:00PM +0200, Dafna Hirschfeld wrote:
> > > 
> > > 
> > > On 27.05.20 01:09, Tomasz Figa wrote:
> > > > On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
> > > > <dafna.hirschfeld@collabora.com> wrote:
> > > > > 
> > > > > If the capture format is YUV444M then the memory input format
> > > > > should be YUV422, so the resizer should not change the default
> > > > > hdiv, vdiv in that case.
> > > > > 
> > > > > Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
> > > > > ---
> > > > >    drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
> > > > >    1 file changed, 4 insertions(+), 3 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > index 04a29af8cc92..5f9740ddd558 100644
> > > > > --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
> > > > >            * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
> > > > >            * streams should be set according to the pixel format in the capture.
> > > > >            * The resizer always gets the input as YUV422. If the capture format
> > > > > -        * is RGB then the memory input (the resizer output) should be YUV422
> > > > > -        * so we use the hdiv, vdiv of the YUV422 info in this case.
> > > > > +        * is RGB or YUV444 then the memory input (the resizer output) should
> > > > > +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
> > > > >            */
> > > > > -       if (v4l2_is_format_yuv(cap->pix.info)) {
> > > > > +       if (v4l2_is_format_yuv(cap->pix.info) &&
> > > > > +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
> > > > >                   src_c.width = cap->pix.info->hdiv;
> > > > >                   src_c.height = cap->pix.info->vdiv;
> > > > 
> > > > As pointed out in another thread, this should have been the original
> > > > size divided by the divisor and not just the latter alone.
> > > > 
> > > > It seems a bit suspicious to me that we don't need to upscale the
> > > > chroma planes here, because it would mean that the MI itself would be
> > > > doing some horizontal pixel doubling. The hardware documentation
> > > > doesn't really explain this, though.
> > > > 
> > > > Have you been able to validate that the setting without upscaling
> > > > indeed produces correct output?
> > > 
> > > Hi,
> > > I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
> > > but there is a problem on selfpath: https://imgur.com/a/vQJPqAn
> > > 
> > > This patch somehow solves the problem on selfpath but ruins the frames on mainpath.
> > > 
> > > I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
> > > sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
> > > But the value should be set according to the configuration. For some reason the problem arises only
> > > when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
> > > value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
> > > Setting the the SP_INPUT_* according to the configuration seems to solve the problem.
> > 
> > Looking at the datasheet, SP seems to have some internal format
> > conversion capability - one can set SP_OUTPUT_FORMAT and SP_INPUT_FORMAT
> > to different YUV subsampling modes or even some RGB formats. MP doesn't
> > have this capability - it can only write whatever it receives.
> > 
> > I think this needs to be reflected in the driver, if it isn't yet.
> > Depending on the resizer source format, the MP video node must allow
> > only formats with matching subsampling mode.
> 
> Hi,
> I work on v3 now that does that, it supports several yuv mbus codes on the source pads of the resizers
> and then in the link_validation callback of the capture it checks that the subsampling matches.

Is it enough? I believe the principle is that within the same entity the
state needs to stay consistent, i.e. if the sink pad changes, the source
pad or video node must be updated to expose only correct formats,
including resetting the current format.

> 
> It is not clear to me what is the meaning of the input/output capapility of the selfpath except of RGB convertion.
> Since yuv subsampling is what the resizer does.

Could be some historical legacy, a byproduct of some feature (e.g.
RGB could be bundled with support for other formats) or could be there
on purpose.

It would be worth checking whether there are any quality or
performance/power differences depending on whether the resizer or the SP
is used for resampling. For example, one could be using some
interpolation algorithm, while the other just a simple pixel doubling or
decimation.

Best regards,
Tomasz
Dafna Hirschfeld June 18, 2020, 6:50 p.m. UTC | #7
On 18.06.20 20:12, Tomasz Figa wrote:
> On Thu, Jun 18, 2020 at 08:00:37PM +0200, Dafna Hirschfeld wrote:
>>
>>
>> On 18.06.20 19:47, Tomasz Figa wrote:
>>> On Fri, Jun 12, 2020 at 02:45:00PM +0200, Dafna Hirschfeld wrote:
>>>>
>>>>
>>>> On 27.05.20 01:09, Tomasz Figa wrote:
>>>>> On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
>>>>> <dafna.hirschfeld@collabora.com> wrote:
>>>>>>
>>>>>> If the capture format is YUV444M then the memory input format
>>>>>> should be YUV422, so the resizer should not change the default
>>>>>> hdiv, vdiv in that case.
>>>>>>
>>>>>> Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
>>>>>> ---
>>>>>>     drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
>>>>>>     1 file changed, 4 insertions(+), 3 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>>>> index 04a29af8cc92..5f9740ddd558 100644
>>>>>> --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>>>> +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>>>> @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
>>>>>>             * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
>>>>>>             * streams should be set according to the pixel format in the capture.
>>>>>>             * The resizer always gets the input as YUV422. If the capture format
>>>>>> -        * is RGB then the memory input (the resizer output) should be YUV422
>>>>>> -        * so we use the hdiv, vdiv of the YUV422 info in this case.
>>>>>> +        * is RGB or YUV444 then the memory input (the resizer output) should
>>>>>> +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
>>>>>>             */
>>>>>> -       if (v4l2_is_format_yuv(cap->pix.info)) {
>>>>>> +       if (v4l2_is_format_yuv(cap->pix.info) &&
>>>>>> +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
>>>>>>                    src_c.width = cap->pix.info->hdiv;
>>>>>>                    src_c.height = cap->pix.info->vdiv;
>>>>>
>>>>> As pointed out in another thread, this should have been the original
>>>>> size divided by the divisor and not just the latter alone.
>>>>>
>>>>> It seems a bit suspicious to me that we don't need to upscale the
>>>>> chroma planes here, because it would mean that the MI itself would be
>>>>> doing some horizontal pixel doubling. The hardware documentation
>>>>> doesn't really explain this, though.
>>>>>
>>>>> Have you been able to validate that the setting without upscaling
>>>>> indeed produces correct output?
>>>>
>>>> Hi,
>>>> I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
>>>> but there is a problem on selfpath: https://imgur.com/a/vQJPqAn
>>>>
>>>> This patch somehow solves the problem on selfpath but ruins the frames on mainpath.
>>>>
>>>> I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
>>>> sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
>>>> But the value should be set according to the configuration. For some reason the problem arises only
>>>> when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
>>>> value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
>>>> Setting the the SP_INPUT_* according to the configuration seems to solve the problem.
>>>
>>> Looking at the datasheet, SP seems to have some internal format
>>> conversion capability - one can set SP_OUTPUT_FORMAT and SP_INPUT_FORMAT
>>> to different YUV subsampling modes or even some RGB formats. MP doesn't
>>> have this capability - it can only write whatever it receives.
>>>
>>> I think this needs to be reflected in the driver, if it isn't yet.
>>> Depending on the resizer source format, the MP video node must allow
>>> only formats with matching subsampling mode.
>>
>> Hi,
>> I work on v3 now that does that, it supports several yuv mbus codes on the source pads of the resizers
>> and then in the link_validation callback of the capture it checks that the subsampling matches.
> 
> Is it enough? I believe the principle is that within the same entity the
> state needs to stay consistent, i.e. if the sink pad changes, the source
> pad or video node must be updated to expose only correct formats,
> including resetting the current format.
Yes, but video devices have only a sink pad. Also, the sink pad of
a video device is not associated with an mbus code. The video format configuration
is not related to the media controller API.

> 
>>
>> It is not clear to me what is the meaning of the input/output capapility of the selfpath except of RGB convertion.
>> Since yuv subsampling is what the resizer does.
> 
> Could be some historical legacy, a byproduct of some feature (e.g.
> RGB could be bundled with support for other formats) or could be there
> on purpose.
> 
> It would be worth checking whether there are any quality or
> performance/power differences depending on whether the resizer or the SP
> is used for resampling. For example, one could be using some
> interpolation algorithm, while the other just a simple pixel doubling or
> decimation.

I can experiment with that.

Thanks,
Dafna

> 
> Best regards,
> Tomasz
>
Tomasz Figa June 18, 2020, 7:18 p.m. UTC | #8
On Thu, Jun 18, 2020 at 08:50:23PM +0200, Dafna Hirschfeld wrote:
> 
> 
> On 18.06.20 20:12, Tomasz Figa wrote:
> > On Thu, Jun 18, 2020 at 08:00:37PM +0200, Dafna Hirschfeld wrote:
> > > 
> > > 
> > > On 18.06.20 19:47, Tomasz Figa wrote:
> > > > On Fri, Jun 12, 2020 at 02:45:00PM +0200, Dafna Hirschfeld wrote:
> > > > > 
> > > > > 
> > > > > On 27.05.20 01:09, Tomasz Figa wrote:
> > > > > > On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
> > > > > > <dafna.hirschfeld@collabora.com> wrote:
> > > > > > > 
> > > > > > > If the capture format is YUV444M then the memory input format
> > > > > > > should be YUV422, so the resizer should not change the default
> > > > > > > hdiv, vdiv in that case.
> > > > > > > 
> > > > > > > Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
> > > > > > > ---
> > > > > > >     drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
> > > > > > >     1 file changed, 4 insertions(+), 3 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > > > index 04a29af8cc92..5f9740ddd558 100644
> > > > > > > --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > > > +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > > > @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
> > > > > > >             * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
> > > > > > >             * streams should be set according to the pixel format in the capture.
> > > > > > >             * The resizer always gets the input as YUV422. If the capture format
> > > > > > > -        * is RGB then the memory input (the resizer output) should be YUV422
> > > > > > > -        * so we use the hdiv, vdiv of the YUV422 info in this case.
> > > > > > > +        * is RGB or YUV444 then the memory input (the resizer output) should
> > > > > > > +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
> > > > > > >             */
> > > > > > > -       if (v4l2_is_format_yuv(cap->pix.info)) {
> > > > > > > +       if (v4l2_is_format_yuv(cap->pix.info) &&
> > > > > > > +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
> > > > > > >                    src_c.width = cap->pix.info->hdiv;
> > > > > > >                    src_c.height = cap->pix.info->vdiv;
> > > > > > 
> > > > > > As pointed out in another thread, this should have been the original
> > > > > > size divided by the divisor and not just the latter alone.
> > > > > > 
> > > > > > It seems a bit suspicious to me that we don't need to upscale the
> > > > > > chroma planes here, because it would mean that the MI itself would be
> > > > > > doing some horizontal pixel doubling. The hardware documentation
> > > > > > doesn't really explain this, though.
> > > > > > 
> > > > > > Have you been able to validate that the setting without upscaling
> > > > > > indeed produces correct output?
> > > > > 
> > > > > Hi,
> > > > > I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
> > > > > but there is a problem on selfpath: https://imgur.com/a/vQJPqAn
> > > > > 
> > > > > This patch somehow solves the problem on selfpath but ruins the frames on mainpath.
> > > > > 
> > > > > I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
> > > > > sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
> > > > > But the value should be set according to the configuration. For some reason the problem arises only
> > > > > when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
> > > > > value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
> > > > > Setting the the SP_INPUT_* according to the configuration seems to solve the problem.
> > > > 
> > > > Looking at the datasheet, SP seems to have some internal format
> > > > conversion capability - one can set SP_OUTPUT_FORMAT and SP_INPUT_FORMAT
> > > > to different YUV subsampling modes or even some RGB formats. MP doesn't
> > > > have this capability - it can only write whatever it receives.
> > > > 
> > > > I think this needs to be reflected in the driver, if it isn't yet.
> > > > Depending on the resizer source format, the MP video node must allow
> > > > only formats with matching subsampling mode.
> > > 
> > > Hi,
> > > I work on v3 now that does that, it supports several yuv mbus codes on the source pads of the resizers
> > > and then in the link_validation callback of the capture it checks that the subsampling matches.
> > 
> > Is it enough? I believe the principle is that within the same entity the
> > state needs to stay consistent, i.e. if the sink pad changes, the source
> > pad or video node must be updated to expose only correct formats,
> > including resetting the current format.
> Yes, but video devices have only a sink pad. Also, the sink pad of
> a video device is not associated with an mbus code. The video format configuration
> is not related to the media controller API.

The video device interface and the entity it is linked to belong to the
same driver. The V4L2 video device interface requires that the state is
always valid and the implementation of the video device needs to take
into account its links to external entities.

Best regards,
Tomasz
Dafna Hirschfeld June 19, 2020, 7:30 a.m. UTC | #9
On 18.06.20 21:18, Tomasz Figa wrote:
> On Thu, Jun 18, 2020 at 08:50:23PM +0200, Dafna Hirschfeld wrote:
>>
>>
>> On 18.06.20 20:12, Tomasz Figa wrote:
>>> On Thu, Jun 18, 2020 at 08:00:37PM +0200, Dafna Hirschfeld wrote:
>>>>
>>>>
>>>> On 18.06.20 19:47, Tomasz Figa wrote:
>>>>> On Fri, Jun 12, 2020 at 02:45:00PM +0200, Dafna Hirschfeld wrote:
>>>>>>
>>>>>>
>>>>>> On 27.05.20 01:09, Tomasz Figa wrote:
>>>>>>> On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
>>>>>>> <dafna.hirschfeld@collabora.com> wrote:
>>>>>>>>
>>>>>>>> If the capture format is YUV444M then the memory input format
>>>>>>>> should be YUV422, so the resizer should not change the default
>>>>>>>> hdiv, vdiv in that case.
>>>>>>>>
>>>>>>>> Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
>>>>>>>> ---
>>>>>>>>      drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
>>>>>>>>      1 file changed, 4 insertions(+), 3 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>>>>>> index 04a29af8cc92..5f9740ddd558 100644
>>>>>>>> --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>>>>>> +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
>>>>>>>> @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
>>>>>>>>              * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
>>>>>>>>              * streams should be set according to the pixel format in the capture.
>>>>>>>>              * The resizer always gets the input as YUV422. If the capture format
>>>>>>>> -        * is RGB then the memory input (the resizer output) should be YUV422
>>>>>>>> -        * so we use the hdiv, vdiv of the YUV422 info in this case.
>>>>>>>> +        * is RGB or YUV444 then the memory input (the resizer output) should
>>>>>>>> +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
>>>>>>>>              */
>>>>>>>> -       if (v4l2_is_format_yuv(cap->pix.info)) {
>>>>>>>> +       if (v4l2_is_format_yuv(cap->pix.info) &&
>>>>>>>> +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
>>>>>>>>                     src_c.width = cap->pix.info->hdiv;
>>>>>>>>                     src_c.height = cap->pix.info->vdiv;
>>>>>>>
>>>>>>> As pointed out in another thread, this should have been the original
>>>>>>> size divided by the divisor and not just the latter alone.
>>>>>>>
>>>>>>> It seems a bit suspicious to me that we don't need to upscale the
>>>>>>> chroma planes here, because it would mean that the MI itself would be
>>>>>>> doing some horizontal pixel doubling. The hardware documentation
>>>>>>> doesn't really explain this, though.
>>>>>>>
>>>>>>> Have you been able to validate that the setting without upscaling
>>>>>>> indeed produces correct output?
>>>>>>
>>>>>> Hi,
>>>>>> I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
>>>>>> but there is a problem on selfpath: https://imgur.com/a/vQJPqAn
>>>>>>
>>>>>> This patch somehow solves the problem on selfpath but ruins the frames on mainpath.
>>>>>>
>>>>>> I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
>>>>>> sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
>>>>>> But the value should be set according to the configuration. For some reason the problem arises only
>>>>>> when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
>>>>>> value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
>>>>>> Setting the the SP_INPUT_* according to the configuration seems to solve the problem.
>>>>>
>>>>> Looking at the datasheet, SP seems to have some internal format
>>>>> conversion capability - one can set SP_OUTPUT_FORMAT and SP_INPUT_FORMAT
>>>>> to different YUV subsampling modes or even some RGB formats. MP doesn't
>>>>> have this capability - it can only write whatever it receives.
>>>>>
>>>>> I think this needs to be reflected in the driver, if it isn't yet.
>>>>> Depending on the resizer source format, the MP video node must allow
>>>>> only formats with matching subsampling mode.
>>>>
>>>> Hi,
>>>> I work on v3 now that does that, it supports several yuv mbus codes on the source pads of the resizers
>>>> and then in the link_validation callback of the capture it checks that the subsampling matches.
>>>
>>> Is it enough? I believe the principle is that within the same entity the
>>> state needs to stay consistent, i.e. if the sink pad changes, the source
>>> pad or video node must be updated to expose only correct formats,
>>> including resetting the current format.
>> Yes, but video devices have only a sink pad. Also, the sink pad of
>> a video device is not associated with an mbus code. The video format configuration
>> is not related to the media controller API.
> 
> The video device interface and the entity it is linked to belong to the
> same driver. The V4L2 video device interface requires that the state is
> always valid and the implementation of the video device needs to take
> into account its links to external entities.

But if the pad on the other side of the link is configured after the video device
then the state might become invalid.

Thanks,
Dafna

> 
> Best regards,
> Tomasz
>
Tomasz Figa June 19, 2020, 12:04 p.m. UTC | #10
On Fri, Jun 19, 2020 at 09:30:50AM +0200, Dafna Hirschfeld wrote:
> 
> 
> On 18.06.20 21:18, Tomasz Figa wrote:
> > On Thu, Jun 18, 2020 at 08:50:23PM +0200, Dafna Hirschfeld wrote:
> > > 
> > > 
> > > On 18.06.20 20:12, Tomasz Figa wrote:
> > > > On Thu, Jun 18, 2020 at 08:00:37PM +0200, Dafna Hirschfeld wrote:
> > > > > 
> > > > > 
> > > > > On 18.06.20 19:47, Tomasz Figa wrote:
> > > > > > On Fri, Jun 12, 2020 at 02:45:00PM +0200, Dafna Hirschfeld wrote:
> > > > > > > 
> > > > > > > 
> > > > > > > On 27.05.20 01:09, Tomasz Figa wrote:
> > > > > > > > On Fri, May 15, 2020 at 4:30 PM Dafna Hirschfeld
> > > > > > > > <dafna.hirschfeld@collabora.com> wrote:
> > > > > > > > > 
> > > > > > > > > If the capture format is YUV444M then the memory input format
> > > > > > > > > should be YUV422, so the resizer should not change the default
> > > > > > > > > hdiv, vdiv in that case.
> > > > > > > > > 
> > > > > > > > > Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@collabora.com>
> > > > > > > > > ---
> > > > > > > > >      drivers/staging/media/rkisp1/rkisp1-resizer.c | 7 ++++---
> > > > > > > > >      1 file changed, 4 insertions(+), 3 deletions(-)
> > > > > > > > > 
> > > > > > > > > diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > > > > > index 04a29af8cc92..5f9740ddd558 100644
> > > > > > > > > --- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > > > > > +++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
> > > > > > > > > @@ -394,10 +394,11 @@ static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
> > > > > > > > >              * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
> > > > > > > > >              * streams should be set according to the pixel format in the capture.
> > > > > > > > >              * The resizer always gets the input as YUV422. If the capture format
> > > > > > > > > -        * is RGB then the memory input (the resizer output) should be YUV422
> > > > > > > > > -        * so we use the hdiv, vdiv of the YUV422 info in this case.
> > > > > > > > > +        * is RGB or YUV444 then the memory input (the resizer output) should
> > > > > > > > > +        * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
> > > > > > > > >              */
> > > > > > > > > -       if (v4l2_is_format_yuv(cap->pix.info)) {
> > > > > > > > > +       if (v4l2_is_format_yuv(cap->pix.info) &&
> > > > > > > > > +           cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
> > > > > > > > >                     src_c.width = cap->pix.info->hdiv;
> > > > > > > > >                     src_c.height = cap->pix.info->vdiv;
> > > > > > > > 
> > > > > > > > As pointed out in another thread, this should have been the original
> > > > > > > > size divided by the divisor and not just the latter alone.
> > > > > > > > 
> > > > > > > > It seems a bit suspicious to me that we don't need to upscale the
> > > > > > > > chroma planes here, because it would mean that the MI itself would be
> > > > > > > > doing some horizontal pixel doubling. The hardware documentation
> > > > > > > > doesn't really explain this, though.
> > > > > > > > 
> > > > > > > > Have you been able to validate that the setting without upscaling
> > > > > > > > indeed produces correct output?
> > > > > > > 
> > > > > > > Hi,
> > > > > > > I investigated it again, without this patch, the frames on mainpath for YUV444 look good: https://imgur.com/a/NtazWhY
> > > > > > > but there is a problem on selfpath: https://imgur.com/a/vQJPqAn
> > > > > > > 
> > > > > > > This patch somehow solves the problem on selfpath but ruins the frames on mainpath.
> > > > > > > 
> > > > > > > I think the bug is actually in another place in the code. The function 'rkisp1_sp_config'
> > > > > > > sets a constant value RKISP1_MI_CTRL_SP_INPUT_YUV422 as the input format on the RKISP1_CIF_MI_CTRL register.
> > > > > > > But the value should be set according to the configuration. For some reason the problem arises only
> > > > > > > when trying to capture yuv444. When I capture yuv420 on the selfpath the frame looks good although the
> > > > > > > value RKISP1_MI_CTRL_SP_INPUT_YUV422 is set.
> > > > > > > Setting the the SP_INPUT_* according to the configuration seems to solve the problem.
> > > > > > 
> > > > > > Looking at the datasheet, SP seems to have some internal format
> > > > > > conversion capability - one can set SP_OUTPUT_FORMAT and SP_INPUT_FORMAT
> > > > > > to different YUV subsampling modes or even some RGB formats. MP doesn't
> > > > > > have this capability - it can only write whatever it receives.
> > > > > > 
> > > > > > I think this needs to be reflected in the driver, if it isn't yet.
> > > > > > Depending on the resizer source format, the MP video node must allow
> > > > > > only formats with matching subsampling mode.
> > > > > 
> > > > > Hi,
> > > > > I work on v3 now that does that, it supports several yuv mbus codes on the source pads of the resizers
> > > > > and then in the link_validation callback of the capture it checks that the subsampling matches.
> > > > 
> > > > Is it enough? I believe the principle is that within the same entity the
> > > > state needs to stay consistent, i.e. if the sink pad changes, the source
> > > > pad or video node must be updated to expose only correct formats,
> > > > including resetting the current format.
> > > Yes, but video devices have only a sink pad. Also, the sink pad of
> > > a video device is not associated with an mbus code. The video format configuration
> > > is not related to the media controller API.
> > 
> > The video device interface and the entity it is linked to belong to the
> > same driver. The V4L2 video device interface requires that the state is
> > always valid and the implementation of the video device needs to take
> > into account its links to external entities.
> 
> But if the pad on the other side of the link is configured after the video device
> then the state might become invalid.

The pad on the other side of the link is already outside of the scope.
Basically, the entity that interfaces with the video node must have
state consistent with the video node.

To make sure we're on the same page, let me give some examples below:

Initial state:
ISP source pad: YUV 4:2:2
Resizer sink pad: YUV 4:2:2
Resizer source pad: YUV 4:2:0
Video CAPTURE format: NV12

Scenario 1:
- Userspace changes ISP source pad to RAW.
- Userspace starts streaming and link validation fails.
- Still, the state is consistent between the resizer and the video node.

Scenario 2:
- Userspace changes ISP source pad to RAW.
- Userspace changes Resizer sink pad to RAW.
- Kernel updates Resizer source pad to RAW.
- Kernel updates video CAPTURE format to a supported RAW pixelformat.
- All the state is consistent and streaming can start.
- If the userspace calls ENUM_FMT on video CAPTURE, it only gets RAW
  pixelformats.

Scenario 3:
- Userspace changes Resizer source pad to YUV 4:2:2.
- Kernel updates video CAPTURE to a supported YUV 4:2:2 pixelformat
  (e.g. YUYV).
- All the state is consistent and streaming can start.
- If the userspace calls ENUM_FMT on video CAPTURE, it only gets YUV
  4:2:2 pixelformats.

Scenario 4:
- Userspace calls ENUM_FMT on video CAPTURE and only gets YUV 4:2:0
  pixelformats.
- It's not possible to make the state inconsistent between the resizer
  and the video node by using V4L2 video ioctls.

Am I missing something?

Best regards,
Tomasz

Patch
diff mbox series

diff --git a/drivers/staging/media/rkisp1/rkisp1-resizer.c b/drivers/staging/media/rkisp1/rkisp1-resizer.c
index 04a29af8cc92..5f9740ddd558 100644
--- a/drivers/staging/media/rkisp1/rkisp1-resizer.c
+++ b/drivers/staging/media/rkisp1/rkisp1-resizer.c
@@ -394,10 +394,11 @@  static void rkisp1_rsz_config(struct rkisp1_resizer *rsz,
 	 * (4:2:2 -> 4:2:0 for example). So the width/height of the CbCr
 	 * streams should be set according to the pixel format in the capture.
 	 * The resizer always gets the input as YUV422. If the capture format
-	 * is RGB then the memory input (the resizer output) should be YUV422
-	 * so we use the hdiv, vdiv of the YUV422 info in this case.
+	 * is RGB or YUV444 then the memory input (the resizer output) should
+	 * be YUV422 so we use the hdiv, vdiv of the YUV422 info in this case.
 	 */
-	if (v4l2_is_format_yuv(cap->pix.info)) {
+	if (v4l2_is_format_yuv(cap->pix.info) &&
+	    cap->pix.info->format != V4L2_PIX_FMT_YUV444M) {
 		src_c.width = cap->pix.info->hdiv;
 		src_c.height = cap->pix.info->vdiv;
 	} else {