diff mbox series

[v5,2/5] drm/msm/dsi: Adjust pclk rate for compression

Message ID 20230405-add-dsc-support-v5-2-028c10850491@quicinc.com
State New
Headers show
Series Add DSC v1.2 Support for DSI | expand

Commit Message

Jessica Zhang May 23, 2023, 1:08 a.m. UTC
Adjust the pclk rate to divide hdisplay by the compression ratio when DSC
is enabled.

Signed-off-by: Jessica Zhang <quic_jesszhan@quicinc.com>
---
 drivers/gpu/drm/msm/dsi/dsi_host.c | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)

Comments

Marijn Suijten June 8, 2023, 8:36 p.m. UTC | #1
Same title suggestion as earlier: s/adjust/reduce

On 2023-05-22 18:08:56, Jessica Zhang wrote:
> Adjust the pclk rate to divide hdisplay by the compression ratio when DSC
> is enabled.
> 
> Signed-off-by: Jessica Zhang <quic_jesszhan@quicinc.com>
> ---
>  drivers/gpu/drm/msm/dsi/dsi_host.c | 21 ++++++++++++++++++---
>  1 file changed, 18 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c b/drivers/gpu/drm/msm/dsi/dsi_host.c
> index a448931af804..88f370dd2ea1 100644
> --- a/drivers/gpu/drm/msm/dsi/dsi_host.c
> +++ b/drivers/gpu/drm/msm/dsi/dsi_host.c
> @@ -561,7 +561,18 @@ void dsi_link_clk_disable_v2(struct msm_dsi_host *msm_host)
>  	clk_disable_unprepare(msm_host->byte_clk);
>  }
>  
> -static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool is_bonded_dsi)
> +static unsigned long dsi_adjust_compressed_pclk(const struct drm_display_mode *mode,

Nit: adjust_pclk_for_compression

As discussed before we realized that this change is more-or-less a hack,
since downstream calculates pclk quite differently - at least for
command-mode panels.  Do you still intend to land this patch this way,
or go the proper route by introducing the right math from the get-go?
Or is the math at least correct for video-mode panels?

This function requires a documentation comment to explain that all.

> +		const struct drm_dsc_config *dsc)
> +{
> +	int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * drm_dsc_get_bpp_int(dsc),

This sounds like a prime candidate for msm_dsc_get_bytes_per_line(), if
bits_per_component==8 is assumed.  In fact, it then becomes identical
to the following line in dsi_host.c which you added previously:

	hdisplay = DIV_ROUND_UP(msm_dsc_get_bytes_per_line(msm_host->dsc), 3);

If not, what is the difference between these two calculations?  Maybe
they both need to be in a properly-named helper.

> +			dsc->bits_per_component * 3);

As we established in the drm/msm issue [2] there is currently a
confusion whether this /3 (and the /3 in dsi_timing_setup) come from the
ratio between dsi_get_bpp() and dsc->bpp or something else.  Can you
clarify that with constants and comments?

[2]: https://gitlab.freedesktop.org/drm/msm/-/issues/24

> +
> +	return (new_hdisplay + (mode->htotal - mode->hdisplay))
> +			* mode->vtotal * drm_mode_vrefresh(mode);

As clarified in [1] I was not necessarily suggesting to move this math
to a separate helper, but to also use a few more properly-named
intermediate variables to not have multi-line math and self-documenting
code.  These lines could be split to be much more clear.

[1]: https://lore.kernel.org/linux-arm-msm/u4x2vldkzsokfcpbkz3dtwcllbdk4ljcz6kzuaxt5frx6g76o5@uku6abewvye7/

> +}
> +
> +static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode,
> +		const struct drm_dsc_config *dsc, bool is_bonded_dsi)
>  {
>  	unsigned long pclk_rate;
>  
> @@ -576,6 +587,10 @@ static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool
>  	if (is_bonded_dsi)
>  		pclk_rate /= 2;
>  
> +	/* If DSC is enabled, divide hdisplay by compression ratio */
> +	if (dsc)
> +		pclk_rate = dsi_adjust_compressed_pclk(mode, dsc);

The confusion with this comment (and the reason the aforementioned
discussion [2] carried on so long) stems from the fact a division makes
sense for a bit/byte clock, but not for a pixel clock: we still intend
to send the same number of pixels, just spending less bytes on them.  So
as you clarify the /3 above, can you also clarify that here or drop this
comment completely when the function is correctly documented instead?

- Marijn

> +
>  	return pclk_rate;
>  }
>  
> @@ -585,7 +600,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
>  	struct msm_dsi_host *msm_host = to_msm_dsi_host(host);
>  	u8 lanes = msm_host->lanes;
>  	u32 bpp = dsi_get_bpp(msm_host->format);
> -	unsigned long pclk_rate = dsi_get_pclk_rate(mode, is_bonded_dsi);
> +	unsigned long pclk_rate = dsi_get_pclk_rate(mode, msm_host->dsc, is_bonded_dsi);
>  	unsigned long pclk_bpp;
>  
>  	if (lanes == 0) {
> @@ -604,7 +619,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
>  
>  static void dsi_calc_pclk(struct msm_dsi_host *msm_host, bool is_bonded_dsi)
>  {
> -	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, is_bonded_dsi);
> +	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, msm_host->dsc, is_bonded_dsi);
>  	msm_host->byte_clk_rate = dsi_byte_clk_get_rate(&msm_host->base, is_bonded_dsi,
>  							msm_host->mode);
>  
> 
> -- 
> 2.40.1
>
Jessica Zhang June 9, 2023, 12:56 a.m. UTC | #2
On 6/8/2023 1:36 PM, Marijn Suijten wrote:
> Same title suggestion as earlier: s/adjust/reduce

Hi Marijn,

Acked.

> 
> On 2023-05-22 18:08:56, Jessica Zhang wrote:
>> Adjust the pclk rate to divide hdisplay by the compression ratio when DSC
>> is enabled.
>>
>> Signed-off-by: Jessica Zhang <quic_jesszhan@quicinc.com>
>> ---
>>   drivers/gpu/drm/msm/dsi/dsi_host.c | 21 ++++++++++++++++++---
>>   1 file changed, 18 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c b/drivers/gpu/drm/msm/dsi/dsi_host.c
>> index a448931af804..88f370dd2ea1 100644
>> --- a/drivers/gpu/drm/msm/dsi/dsi_host.c
>> +++ b/drivers/gpu/drm/msm/dsi/dsi_host.c
>> @@ -561,7 +561,18 @@ void dsi_link_clk_disable_v2(struct msm_dsi_host *msm_host)
>>   	clk_disable_unprepare(msm_host->byte_clk);
>>   }
>>   
>> -static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool is_bonded_dsi)
>> +static unsigned long dsi_adjust_compressed_pclk(const struct drm_display_mode *mode,
> 
> Nit: adjust_pclk_for_compression

Acked.

> 
> As discussed before we realized that this change is more-or-less a hack,
> since downstream calculates pclk quite differently - at least for
> command-mode panels.  Do you still intend to land this patch this way,
> or go the proper route by introducing the right math from the get-go?
> Or is the math at least correct for video-mode panels?

Sorry but can you please clarify what exactly is incorrect or different 
about this math when compared to downstream? And, if you think that this 
math is incorrect, what exactly has to be changed to make it the "right 
math"?

We've already shown step-by-step [1] not only how the resulting pclk 
from the downstream code matches out upstream calculations, but also how 
the inclusion of porches in the upstream math would make up for the fact 
that upstream has no concept of transfer time [2].

If the lack of transfer time in the upstream math is the issue, I 
believe that that's not within the scope of this series, as transfer 
time is not something specific to DSC.

Moreover, as stated in an earlier revision [3], there is no way to 
validate DSC over DSI for video mode. As far as I know, we do not have a 
way to validate video mode datapath for DSI in general.

[1] https://gitlab.freedesktop.org/drm/msm/-/issues/24#note_1936144
[2] https://gitlab.freedesktop.org/drm/msm/-/issues/24#note_1945792
[3] 
https://patchwork.freedesktop.org/patch/535117/?series=117219&rev=1#comment_970492

> 
> This function requires a documentation comment to explain that all.
> 
>> +		const struct drm_dsc_config *dsc)
>> +{
>> +	int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * drm_dsc_get_bpp_int(dsc),
> 
> This sounds like a prime candidate for msm_dsc_get_bytes_per_line(), if
> bits_per_component==8 is assumed.  In fact, it then becomes identical
> to the following line in dsi_host.c which you added previously:
> 
> 	hdisplay = DIV_ROUND_UP(msm_dsc_get_bytes_per_line(msm_host->dsc), 3);
> 
> If not, what is the difference between these two calculations?  Maybe
> they both need to be in a properly-named helper.

While the math technically matches up (assuming, also, that 
mode->hdisplay == slice_width * slice_count for all cases), there are 
conceptual differences between the pclk and hdisplay calculations.

Just to reiterate what was already said on IRC:

In the pclk calculation, we're multiplying pclk by the compression ratio 
(which would be target_bpp / src_bpp) -- please note that here, we 
calculate src_bpp by doing bpc * 3.

In the hdisplay calculation, we calculate the bytes per line and divide 
by 3 (bytes) to account for the fact that we can only process 3 bytes 
per pclk cycle.

So while I understand why you would want to put this behind a shared 
helper, I think abstracting the pclk and hdisplay math away would 
obfuscate the conceptual difference between the 2 calculations.

> 
>> +			dsc->bits_per_component * 3);
> 
> As we established in the drm/msm issue [2] there is currently a
> confusion whether this /3 (and the /3 in dsi_timing_setup) come from the
> ratio between dsi_get_bpp() and dsc->bpp or something else.  Can you
> clarify that with constants and comments?

Sure, we are planning to add a patch to the end of this series 
documenting the math.

> 
> [2]: https://gitlab.freedesktop.org/drm/msm/-/issues/24
> 
>> +
>> +	return (new_hdisplay + (mode->htotal - mode->hdisplay))
>> +			* mode->vtotal * drm_mode_vrefresh(mode);
> 
> As clarified in [1] I was not necessarily suggesting to move this math
> to a separate helper, but to also use a few more properly-named
> intermediate variables to not have multi-line math and self-documenting
> code.  These lines could be split to be much more clear.

Acked.

> 
> [1]: https://lore.kernel.org/linux-arm-msm/u4x2vldkzsokfcpbkz3dtwcllbdk4ljcz6kzuaxt5frx6g76o5@uku6abewvye7/
> 
>> +}
>> +
>> +static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode,
>> +		const struct drm_dsc_config *dsc, bool is_bonded_dsi)
>>   {
>>   	unsigned long pclk_rate;
>>   
>> @@ -576,6 +587,10 @@ static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool
>>   	if (is_bonded_dsi)
>>   		pclk_rate /= 2;
>>   
>> +	/* If DSC is enabled, divide hdisplay by compression ratio */
>> +	if (dsc)
>> +		pclk_rate = dsi_adjust_compressed_pclk(mode, dsc);
> 
> The confusion with this comment (and the reason the aforementioned
> discussion [2] carried on so long) stems from the fact a division makes
> sense for a bit/byte clock, but not for a pixel clock: we still intend
> to send the same number of pixels, just spending less bytes on them.  So
> as you clarify the /3 above, can you also clarify that here or drop this
> comment completely when the function is correctly documented instead?

Acked.

Thanks,

Jessica Zhang

> 
> - Marijn
> 
>> +
>>   	return pclk_rate;
>>   }
>>   
>> @@ -585,7 +600,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
>>   	struct msm_dsi_host *msm_host = to_msm_dsi_host(host);
>>   	u8 lanes = msm_host->lanes;
>>   	u32 bpp = dsi_get_bpp(msm_host->format);
>> -	unsigned long pclk_rate = dsi_get_pclk_rate(mode, is_bonded_dsi);
>> +	unsigned long pclk_rate = dsi_get_pclk_rate(mode, msm_host->dsc, is_bonded_dsi);
>>   	unsigned long pclk_bpp;
>>   
>>   	if (lanes == 0) {
>> @@ -604,7 +619,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
>>   
>>   static void dsi_calc_pclk(struct msm_dsi_host *msm_host, bool is_bonded_dsi)
>>   {
>> -	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, is_bonded_dsi);
>> +	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, msm_host->dsc, is_bonded_dsi);
>>   	msm_host->byte_clk_rate = dsi_byte_clk_get_rate(&msm_host->base, is_bonded_dsi,
>>   							msm_host->mode);
>>   
>>
>> -- 
>> 2.40.1
>>
Abhinav Kumar June 9, 2023, 1:09 a.m. UTC | #3
On 6/8/2023 5:56 PM, Jessica Zhang wrote:
> 
> 
> On 6/8/2023 1:36 PM, Marijn Suijten wrote:
>> Same title suggestion as earlier: s/adjust/reduce
> 
> Hi Marijn,
> 
> Acked.
> 
>>
>> On 2023-05-22 18:08:56, Jessica Zhang wrote:
>>> Adjust the pclk rate to divide hdisplay by the compression ratio when 
>>> DSC
>>> is enabled.
>>>
>>> Signed-off-by: Jessica Zhang <quic_jesszhan@quicinc.com>
>>> ---
>>>   drivers/gpu/drm/msm/dsi/dsi_host.c | 21 ++++++++++++++++++---
>>>   1 file changed, 18 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c 
>>> b/drivers/gpu/drm/msm/dsi/dsi_host.c
>>> index a448931af804..88f370dd2ea1 100644
>>> --- a/drivers/gpu/drm/msm/dsi/dsi_host.c
>>> +++ b/drivers/gpu/drm/msm/dsi/dsi_host.c
>>> @@ -561,7 +561,18 @@ void dsi_link_clk_disable_v2(struct msm_dsi_host 
>>> *msm_host)
>>>       clk_disable_unprepare(msm_host->byte_clk);
>>>   }
>>> -static unsigned long dsi_get_pclk_rate(const struct drm_display_mode 
>>> *mode, bool is_bonded_dsi)
>>> +static unsigned long dsi_adjust_compressed_pclk(const struct 
>>> drm_display_mode *mode,
>>
>> Nit: adjust_pclk_for_compression
> 
> Acked.
> 
>>
>> As discussed before we realized that this change is more-or-less a hack,
>> since downstream calculates pclk quite differently - at least for
>> command-mode panels.  Do you still intend to land this patch this way,
>> or go the proper route by introducing the right math from the get-go?
>> Or is the math at least correct for video-mode panels?
> 
> Sorry but can you please clarify what exactly is incorrect or different 
> about this math when compared to downstream? And, if you think that this 
> math is incorrect, what exactly has to be changed to make it the "right 
> math"?
> 

Agree with Jessica, just calling the math a hack without explaining why 
you think it is so is not justified especially when a great detail of 
explanation was given on the bug. Sorry but its a bit harsh on the 
developers.

> We've already shown step-by-step [1] not only how the resulting pclk 
> from the downstream code matches out upstream calculations, but also how 
> the inclusion of porches in the upstream math would make up for the fact 
> that upstream has no concept of transfer time [2].
> 
> If the lack of transfer time in the upstream math is the issue, I 
> believe that that's not within the scope of this series, as transfer 
> time is not something specific to DSC.
> 
> Moreover, as stated in an earlier revision [3], there is no way to 
> validate DSC over DSI for video mode. As far as I know, we do not have a 
> way to validate video mode datapath for DSI in general.
> 
> [1] https://gitlab.freedesktop.org/drm/msm/-/issues/24#note_1936144
> [2] https://gitlab.freedesktop.org/drm/msm/-/issues/24#note_1945792
> [3] 
> https://patchwork.freedesktop.org/patch/535117/?series=117219&rev=1#comment_970492
> 
>>
>> This function requires a documentation comment to explain that all.
>>
>>> +        const struct drm_dsc_config *dsc)
>>> +{
>>> +    int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * 
>>> drm_dsc_get_bpp_int(dsc),
>>
>> This sounds like a prime candidate for msm_dsc_get_bytes_per_line(), if
>> bits_per_component==8 is assumed.  In fact, it then becomes identical
>> to the following line in dsi_host.c which you added previously:
>>
>>     hdisplay = DIV_ROUND_UP(msm_dsc_get_bytes_per_line(msm_host->dsc), 
>> 3);
>>
>> If not, what is the difference between these two calculations?  Maybe
>> they both need to be in a properly-named helper.
> 
> While the math technically matches up (assuming, also, that 
> mode->hdisplay == slice_width * slice_count for all cases), there are 
> conceptual differences between the pclk and hdisplay calculations.
> 
> Just to reiterate what was already said on IRC:
> 
> In the pclk calculation, we're multiplying pclk by the compression ratio 
> (which would be target_bpp / src_bpp) -- please note that here, we 
> calculate src_bpp by doing bpc * 3.
> 
> In the hdisplay calculation, we calculate the bytes per line and divide 
> by 3 (bytes) to account for the fact that we can only process 3 bytes 
> per pclk cycle.
> 
> So while I understand why you would want to put this behind a shared 
> helper, I think abstracting the pclk and hdisplay math away would 
> obfuscate the conceptual difference between the 2 calculations.
> 
>>
>>> +            dsc->bits_per_component * 3);
>>
>> As we established in the drm/msm issue [2] there is currently a
>> confusion whether this /3 (and the /3 in dsi_timing_setup) come from the
>> ratio between dsi_get_bpp() and dsc->bpp or something else.  Can you
>> clarify that with constants and comments?
> 
> Sure, we are planning to add a patch to the end of this series 
> documenting the math.
> 
>>
>> [2]: https://gitlab.freedesktop.org/drm/msm/-/issues/24
>>
>>> +
>>> +    return (new_hdisplay + (mode->htotal - mode->hdisplay))
>>> +            * mode->vtotal * drm_mode_vrefresh(mode);
>>
>> As clarified in [1] I was not necessarily suggesting to move this math
>> to a separate helper, but to also use a few more properly-named
>> intermediate variables to not have multi-line math and self-documenting
>> code.  These lines could be split to be much more clear.
> 
> Acked.
> 
>>
>> [1]: 
>> https://lore.kernel.org/linux-arm-msm/u4x2vldkzsokfcpbkz3dtwcllbdk4ljcz6kzuaxt5frx6g76o5@uku6abewvye7/
>>
>>> +}
>>> +
>>> +static unsigned long dsi_get_pclk_rate(const struct drm_display_mode 
>>> *mode,
>>> +        const struct drm_dsc_config *dsc, bool is_bonded_dsi)
>>>   {
>>>       unsigned long pclk_rate;
>>> @@ -576,6 +587,10 @@ static unsigned long dsi_get_pclk_rate(const 
>>> struct drm_display_mode *mode, bool
>>>       if (is_bonded_dsi)
>>>           pclk_rate /= 2;
>>> +    /* If DSC is enabled, divide hdisplay by compression ratio */
>>> +    if (dsc)
>>> +        pclk_rate = dsi_adjust_compressed_pclk(mode, dsc);
>>
>> The confusion with this comment (and the reason the aforementioned
>> discussion [2] carried on so long) stems from the fact a division makes
>> sense for a bit/byte clock, but not for a pixel clock: we still intend
>> to send the same number of pixels, just spending less bytes on them.  So
>> as you clarify the /3 above, can you also clarify that here or drop this
>> comment completely when the function is correctly documented instead?
> 
> Acked.
> 
> Thanks,
> 
> Jessica Zhang
> 
>>
>> - Marijn
>>
>>> +
>>>       return pclk_rate;
>>>   }
>>> @@ -585,7 +600,7 @@ unsigned long dsi_byte_clk_get_rate(struct 
>>> mipi_dsi_host *host, bool is_bonded_d
>>>       struct msm_dsi_host *msm_host = to_msm_dsi_host(host);
>>>       u8 lanes = msm_host->lanes;
>>>       u32 bpp = dsi_get_bpp(msm_host->format);
>>> -    unsigned long pclk_rate = dsi_get_pclk_rate(mode, is_bonded_dsi);
>>> +    unsigned long pclk_rate = dsi_get_pclk_rate(mode, msm_host->dsc, 
>>> is_bonded_dsi);
>>>       unsigned long pclk_bpp;
>>>       if (lanes == 0) {
>>> @@ -604,7 +619,7 @@ unsigned long dsi_byte_clk_get_rate(struct 
>>> mipi_dsi_host *host, bool is_bonded_d
>>>   static void dsi_calc_pclk(struct msm_dsi_host *msm_host, bool 
>>> is_bonded_dsi)
>>>   {
>>> -    msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, 
>>> is_bonded_dsi);
>>> +    msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, 
>>> msm_host->dsc, is_bonded_dsi);
>>>       msm_host->byte_clk_rate = 
>>> dsi_byte_clk_get_rate(&msm_host->base, is_bonded_dsi,
>>>                               msm_host->mode);
>>>
>>> -- 
>>> 2.40.1
>>>
Dmitry Baryshkov June 9, 2023, 4:58 p.m. UTC | #4
On 08/06/2023 23:36, Marijn Suijten wrote:
> Same title suggestion as earlier: s/adjust/reduce
> 
> On 2023-05-22 18:08:56, Jessica Zhang wrote:
>> Adjust the pclk rate to divide hdisplay by the compression ratio when DSC
>> is enabled.
>>
>> Signed-off-by: Jessica Zhang <quic_jesszhan@quicinc.com>
>> ---
>>   drivers/gpu/drm/msm/dsi/dsi_host.c | 21 ++++++++++++++++++---
>>   1 file changed, 18 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c b/drivers/gpu/drm/msm/dsi/dsi_host.c
>> index a448931af804..88f370dd2ea1 100644
>> --- a/drivers/gpu/drm/msm/dsi/dsi_host.c
>> +++ b/drivers/gpu/drm/msm/dsi/dsi_host.c
>> @@ -561,7 +561,18 @@ void dsi_link_clk_disable_v2(struct msm_dsi_host *msm_host)
>>   	clk_disable_unprepare(msm_host->byte_clk);
>>   }
>>   
>> -static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool is_bonded_dsi)
>> +static unsigned long dsi_adjust_compressed_pclk(const struct drm_display_mode *mode,
> 
> Nit: adjust_pclk_for_compression
> 
> As discussed before we realized that this change is more-or-less a hack,
> since downstream calculates pclk quite differently - at least for
> command-mode panels.  Do you still intend to land this patch this way,
> or go the proper route by introducing the right math from the get-go?
> Or is the math at least correct for video-mode panels?

Can we please postpone the cmd-vs-video discussion? Otherwise I will 
reserve myself a right to push a patch dropping CMD mode support until 
somebody comes with a proper way to handle CMD clock calculation.


It is off-topic for the sake of DSC 1.2 support. Yes, all CMD panel 
timings are a kind of a hack and should be improved. No, we can not do 
this as a part of this series. I think everybody agrees that with the 
current way of calculating CMD panel timings, this function does some 
kind of a trick.

> 
> This function requires a documentation comment to explain that all.
> 
>> +		const struct drm_dsc_config *dsc)
>> +{
>> +	int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * drm_dsc_get_bpp_int(dsc),
> 
> This sounds like a prime candidate for msm_dsc_get_bytes_per_line(), if
> bits_per_component==8 is assumed.  In fact, it then becomes identical
> to the following line in dsi_host.c which you added previously:
> 
> 	hdisplay = DIV_ROUND_UP(msm_dsc_get_bytes_per_line(msm_host->dsc), 3);

This would imply a simple /3, but as far as I understand it is not 
correct here.

> 
> If not, what is the difference between these two calculations?  Maybe
> they both need to be in a properly-named helper.
> 
>> +			dsc->bits_per_component * 3);

I hope to see a documentation patch to be posted, telling that this 
scales hdisplay and thus pclk by the factor of compressed_bpp / 
uncompressed_bpp.

This is not how it is usually done, but I would accept a separate 
documentation patch going over the calculation here and in 
dsi_timing_setup (and maybe other unobvious cases, if there is anything 
left).

> 
> As we established in the drm/msm issue [2] there is currently a
> confusion whether this /3 (and the /3 in dsi_timing_setup) come from the
> ratio between dsi_get_bpp() and dsc->bpp or something else.  Can you
> clarify that with constants and comments?
> 
> [2]: https://gitlab.freedesktop.org/drm/msm/-/issues/24
> 
>> +
>> +	return (new_hdisplay + (mode->htotal - mode->hdisplay))
>> +			* mode->vtotal * drm_mode_vrefresh(mode);
> 
> As clarified in [1] I was not necessarily suggesting to move this math
> to a separate helper, but to also use a few more properly-named
> intermediate variables to not have multi-line math and self-documenting
> code.  These lines could be split to be much more clear.

I think it's fine more or less. One pair of parenthesis is unnecessary, 
but that's mostly it. Maybe `new_htotal' variable would make some sense.

Also, please excuse me if this was discussed somewhere. This calculation 
means that only the displayed data is compressed, but porches are not 
touched. Correct?

> 
> [1]: https://lore.kernel.org/linux-arm-msm/u4x2vldkzsokfcpbkz3dtwcllbdk4ljcz6kzuaxt5frx6g76o5@uku6abewvye7/
> 
>> +}
>> +
>> +static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode,
>> +		const struct drm_dsc_config *dsc, bool is_bonded_dsi)
>>   {
>>   	unsigned long pclk_rate;
>>   
>> @@ -576,6 +587,10 @@ static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool
>>   	if (is_bonded_dsi)
>>   		pclk_rate /= 2;
>>   
>> +	/* If DSC is enabled, divide hdisplay by compression ratio */
>> +	if (dsc)
>> +		pclk_rate = dsi_adjust_compressed_pclk(mode, dsc);

Looking for the perfection, I'd also move the pclk adjustment to come 
before the is_bonded_dsi check.

> 
> The confusion with this comment (and the reason the aforementioned
> discussion [2] carried on so long) stems from the fact a division makes
> sense for a bit/byte clock, but not for a pixel clock: we still intend
> to send the same number of pixels, just spending less bytes on them.  So
> as you clarify the /3 above, can you also clarify that here or drop this
> comment completely when the function is correctly documented instead?
> 
> - Marijn
> 
>> +
>>   	return pclk_rate;
>>   }
>>   
>> @@ -585,7 +600,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
>>   	struct msm_dsi_host *msm_host = to_msm_dsi_host(host);
>>   	u8 lanes = msm_host->lanes;
>>   	u32 bpp = dsi_get_bpp(msm_host->format);
>> -	unsigned long pclk_rate = dsi_get_pclk_rate(mode, is_bonded_dsi);
>> +	unsigned long pclk_rate = dsi_get_pclk_rate(mode, msm_host->dsc, is_bonded_dsi);
>>   	unsigned long pclk_bpp;
>>   
>>   	if (lanes == 0) {
>> @@ -604,7 +619,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
>>   
>>   static void dsi_calc_pclk(struct msm_dsi_host *msm_host, bool is_bonded_dsi)
>>   {
>> -	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, is_bonded_dsi);
>> +	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, msm_host->dsc, is_bonded_dsi);
>>   	msm_host->byte_clk_rate = dsi_byte_clk_get_rate(&msm_host->base, is_bonded_dsi,
>>   							msm_host->mode);
>>   
>>
>> -- 
>> 2.40.1
>>
Jessica Zhang June 9, 2023, 5:24 p.m. UTC | #5
On 6/9/2023 9:58 AM, Dmitry Baryshkov wrote:
> On 08/06/2023 23:36, Marijn Suijten wrote:
>> Same title suggestion as earlier: s/adjust/reduce
>>
>> On 2023-05-22 18:08:56, Jessica Zhang wrote:
>>> Adjust the pclk rate to divide hdisplay by the compression ratio when 
>>> DSC
>>> is enabled.
>>>
>>> Signed-off-by: Jessica Zhang <quic_jesszhan@quicinc.com>
>>> ---
>>>   drivers/gpu/drm/msm/dsi/dsi_host.c | 21 ++++++++++++++++++---
>>>   1 file changed, 18 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c 
>>> b/drivers/gpu/drm/msm/dsi/dsi_host.c
>>> index a448931af804..88f370dd2ea1 100644
>>> --- a/drivers/gpu/drm/msm/dsi/dsi_host.c
>>> +++ b/drivers/gpu/drm/msm/dsi/dsi_host.c
>>> @@ -561,7 +561,18 @@ void dsi_link_clk_disable_v2(struct msm_dsi_host 
>>> *msm_host)
>>>       clk_disable_unprepare(msm_host->byte_clk);
>>>   }
>>> -static unsigned long dsi_get_pclk_rate(const struct drm_display_mode 
>>> *mode, bool is_bonded_dsi)
>>> +static unsigned long dsi_adjust_compressed_pclk(const struct 
>>> drm_display_mode *mode,
>>
>> Nit: adjust_pclk_for_compression
>>
>> As discussed before we realized that this change is more-or-less a hack,
>> since downstream calculates pclk quite differently - at least for
>> command-mode panels.  Do you still intend to land this patch this way,
>> or go the proper route by introducing the right math from the get-go?
>> Or is the math at least correct for video-mode panels?
> 
> Can we please postpone the cmd-vs-video discussion? Otherwise I will 
> reserve myself a right to push a patch dropping CMD mode support until 
> somebody comes with a proper way to handle CMD clock calculation.
> 
> 
> It is off-topic for the sake of DSC 1.2 support. Yes, all CMD panel 
> timings are a kind of a hack and should be improved. No, we can not do 
> this as a part of this series. I think everybody agrees that with the 
> current way of calculating CMD panel timings, this function does some 
> kind of a trick.
> 
>>
>> This function requires a documentation comment to explain that all.
>>
>>> +        const struct drm_dsc_config *dsc)
>>> +{
>>> +    int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * 
>>> drm_dsc_get_bpp_int(dsc),
>>
>> This sounds like a prime candidate for msm_dsc_get_bytes_per_line(), if
>> bits_per_component==8 is assumed.  In fact, it then becomes identical
>> to the following line in dsi_host.c which you added previously:
>>
>>     hdisplay = DIV_ROUND_UP(msm_dsc_get_bytes_per_line(msm_host->dsc), 
>> 3);
> 
> This would imply a simple /3, but as far as I understand it is not 
> correct here.
> 
>>
>> If not, what is the difference between these two calculations?  Maybe
>> they both need to be in a properly-named helper.
>>
>>> +            dsc->bits_per_component * 3);
> 
> I hope to see a documentation patch to be posted, telling that this 
> scales hdisplay and thus pclk by the factor of compressed_bpp / 
> uncompressed_bpp.
> 
> This is not how it is usually done, but I would accept a separate 
> documentation patch going over the calculation here and in 
> dsi_timing_setup (and maybe other unobvious cases, if there is anything 
> left).
> 
>>
>> As we established in the drm/msm issue [2] there is currently a
>> confusion whether this /3 (and the /3 in dsi_timing_setup) come from the
>> ratio between dsi_get_bpp() and dsc->bpp or something else.  Can you
>> clarify that with constants and comments?
>>
>> [2]: https://gitlab.freedesktop.org/drm/msm/-/issues/24
>>
>>> +
>>> +    return (new_hdisplay + (mode->htotal - mode->hdisplay))
>>> +            * mode->vtotal * drm_mode_vrefresh(mode);
>>
>> As clarified in [1] I was not necessarily suggesting to move this math
>> to a separate helper, but to also use a few more properly-named
>> intermediate variables to not have multi-line math and self-documenting
>> code.  These lines could be split to be much more clear.
> 
> I think it's fine more or less. One pair of parenthesis is unnecessary, 
> but that's mostly it. Maybe `new_htotal' variable would make some sense.
> 
> Also, please excuse me if this was discussed somewhere. This calculation 
> means that only the displayed data is compressed, but porches are not 
> touched. Correct?

Hi Dmitry,

Correct, we will apply the compression ratio to only the hdisplay but 
keep the porches as is.

> 
>>
>> [1]: 
>> https://lore.kernel.org/linux-arm-msm/u4x2vldkzsokfcpbkz3dtwcllbdk4ljcz6kzuaxt5frx6g76o5@uku6abewvye7/
>>
>>> +}
>>> +
>>> +static unsigned long dsi_get_pclk_rate(const struct drm_display_mode 
>>> *mode,
>>> +        const struct drm_dsc_config *dsc, bool is_bonded_dsi)
>>>   {
>>>       unsigned long pclk_rate;
>>> @@ -576,6 +587,10 @@ static unsigned long dsi_get_pclk_rate(const 
>>> struct drm_display_mode *mode, bool
>>>       if (is_bonded_dsi)
>>>           pclk_rate /= 2;
>>> +    /* If DSC is enabled, divide hdisplay by compression ratio */
>>> +    if (dsc)
>>> +        pclk_rate = dsi_adjust_compressed_pclk(mode, dsc);
> 
> Looking for the perfection, I'd also move the pclk adjustment to come 
> before the is_bonded_dsi check.

Acked.

Thanks,

Jessica Zhang

> 
>>
>> The confusion with this comment (and the reason the aforementioned
>> discussion [2] carried on so long) stems from the fact a division makes
>> sense for a bit/byte clock, but not for a pixel clock: we still intend
>> to send the same number of pixels, just spending less bytes on them.  So
>> as you clarify the /3 above, can you also clarify that here or drop this
>> comment completely when the function is correctly documented instead?
>>
>> - Marijn
>>
>>> +
>>>       return pclk_rate;
>>>   }
>>> @@ -585,7 +600,7 @@ unsigned long dsi_byte_clk_get_rate(struct 
>>> mipi_dsi_host *host, bool is_bonded_d
>>>       struct msm_dsi_host *msm_host = to_msm_dsi_host(host);
>>>       u8 lanes = msm_host->lanes;
>>>       u32 bpp = dsi_get_bpp(msm_host->format);
>>> -    unsigned long pclk_rate = dsi_get_pclk_rate(mode, is_bonded_dsi);
>>> +    unsigned long pclk_rate = dsi_get_pclk_rate(mode, msm_host->dsc, 
>>> is_bonded_dsi);
>>>       unsigned long pclk_bpp;
>>>       if (lanes == 0) {
>>> @@ -604,7 +619,7 @@ unsigned long dsi_byte_clk_get_rate(struct 
>>> mipi_dsi_host *host, bool is_bonded_d
>>>   static void dsi_calc_pclk(struct msm_dsi_host *msm_host, bool 
>>> is_bonded_dsi)
>>>   {
>>> -    msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, 
>>> is_bonded_dsi);
>>> +    msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, 
>>> msm_host->dsc, is_bonded_dsi);
>>>       msm_host->byte_clk_rate = 
>>> dsi_byte_clk_get_rate(&msm_host->base, is_bonded_dsi,
>>>                               msm_host->mode);
>>>
>>> -- 
>>> 2.40.1
>>>
> 
> -- 
> With best wishes
> Dmitry
>
Marijn Suijten June 11, 2023, 9:42 p.m. UTC | #6
On 2023-06-09 19:58:00, Dmitry Baryshkov wrote:
> On 08/06/2023 23:36, Marijn Suijten wrote:
> > Same title suggestion as earlier: s/adjust/reduce
> > 
> > On 2023-05-22 18:08:56, Jessica Zhang wrote:
> >> Adjust the pclk rate to divide hdisplay by the compression ratio when DSC
> >> is enabled.
> >>
> >> Signed-off-by: Jessica Zhang <quic_jesszhan@quicinc.com>
> >> ---
> >>   drivers/gpu/drm/msm/dsi/dsi_host.c | 21 ++++++++++++++++++---
> >>   1 file changed, 18 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c b/drivers/gpu/drm/msm/dsi/dsi_host.c
> >> index a448931af804..88f370dd2ea1 100644
> >> --- a/drivers/gpu/drm/msm/dsi/dsi_host.c
> >> +++ b/drivers/gpu/drm/msm/dsi/dsi_host.c
> >> @@ -561,7 +561,18 @@ void dsi_link_clk_disable_v2(struct msm_dsi_host *msm_host)
> >>   	clk_disable_unprepare(msm_host->byte_clk);
> >>   }
> >>   
> >> -static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool is_bonded_dsi)
> >> +static unsigned long dsi_adjust_compressed_pclk(const struct drm_display_mode *mode,
> > 
> > Nit: adjust_pclk_for_compression
> > 
> > As discussed before we realized that this change is more-or-less a hack,
> > since downstream calculates pclk quite differently - at least for
> > command-mode panels.  Do you still intend to land this patch this way,
> > or go the proper route by introducing the right math from the get-go?
> > Or is the math at least correct for video-mode panels?
> 
> Can we please postpone the cmd-vs-video discussion?

If you had read Jessica's reply (and the discussions thus far) this
patch is intended for CMD-mode:

    Moreover, as stated in an earlier revision [3], there is no way to
    validate DSC over DSI for video mode. As far as I know, we do not
    have a way to validate video mode datapath for DSI in general.

Hence my hopefully-relevant question whether this workaround - to reduce
the hdisplay portion that comprises the full pclk signal - is at the
very least right for video mode?  After all, CMD mode doesn't have
porches so it makes no sense to account for those (except that it
currently pretends to represent the "transfer time").

Furthermore there is *zero* documentation describing this workaround,
not even in v6.

> Otherwise I will reserve myself a right to push a patch dropping CMD
> mode support until somebody comes with a proper way to handle CMD
> clock calculation.

Please do.  That should force me or someone else to submit the right
calculations instead of repeatedly getting stuck in this loop of
complaints and incomprehensible patches with no fix in sight.

> It is off-topic for the sake of DSC 1.2 support. Yes, all CMD panel 
> timings are a kind of a hack and should be improved. No, we can not do 
> this as a part of this series. I think everybody agrees that with the 
> current way of calculating CMD panel timings, this function does some 
> kind of a trick.

I understand that you want to be pragrmatic about this situation, but it
seems wrong to build new math on top of fundamentally broken values.  If
there's one thing I learned while contributing to mainline here, it is
that sometimes things are fundamentally broken and you cannot
ship/contribute a shiny new feature before first fixing the
fundamentals.  What makes this different?

> > This function requires a documentation comment to explain that all.
> > 
> >> +		const struct drm_dsc_config *dsc)
> >> +{
> >> +	int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * drm_dsc_get_bpp_int(dsc),
> > 
> > This sounds like a prime candidate for msm_dsc_get_bytes_per_line(), if
> > bits_per_component==8 is assumed.  In fact, it then becomes identical
> > to the following line in dsi_host.c which you added previously:
> > 
> > 	hdisplay = DIV_ROUND_UP(msm_dsc_get_bytes_per_line(msm_host->dsc), 3);
> 
> This would imply a simple /3, but as far as I understand it is not 
> correct here.

If you read Jessica's comment on that duplicate line in v6, that is
exactly what is happening:

    * hdisplay will be divided by 3 here to account for the fact
    * that DPU sends 3 bytes per pclk cycle to DSI.

That comment acknowledges that it is the same: but why isn't this?  Why
is bits_per_component (number of bits per pixel component _in the source
picture_) suddenly introduced in this pclk calculation while it is not
used anywhere else? 

> > If not, what is the difference between these two calculations?  Maybe
> > they both need to be in a properly-named helper.
> > 
> >> +			dsc->bits_per_component * 3);

How is this different from dsi_get_bpp()?

> I hope to see a documentation patch to be posted, telling that this 
> scales hdisplay and thus pclk by the factor of compressed_bpp / 
> uncompressed_bpp.
> 
> This is not how it is usually done, but I would accept a separate 
> documentation patch going over the calculation here and in 
> dsi_timing_setup (and maybe other unobvious cases, if there is anything 
> left).

I'd love to see additional documentation for existing logic, but the
next revision of this patch could atomically add these comments while
adding the logic (seems to not have been done for v6...)?

> > As we established in the drm/msm issue [2] there is currently a
> > confusion whether this /3 (and the /3 in dsi_timing_setup) come from the
> > ratio between dsi_get_bpp() and dsc->bpp or something else.  Can you
> > clarify that with constants and comments?
> > 
> > [2]: https://gitlab.freedesktop.org/drm/msm/-/issues/24
> > 
> >> +
> >> +	return (new_hdisplay + (mode->htotal - mode->hdisplay))
> >> +			* mode->vtotal * drm_mode_vrefresh(mode);
> > 
> > As clarified in [1] I was not necessarily suggesting to move this math
> > to a separate helper, but to also use a few more properly-named
> > intermediate variables to not have multi-line math and self-documenting
> > code.  These lines could be split to be much more clear.
> 
> I think it's fine more or less. One pair of parenthesis is unnecessary, 
> but that's mostly it. Maybe `new_htotal' variable would make some sense.
> 
> Also, please excuse me if this was discussed somewhere. This calculation 
> means that only the displayed data is compressed, but porches are not 
> touched. Correct?

Porches don't exist in CMDmode (but they do in mainline, because
transfer time is not calculated yet).  For video mode I don't know.

> > [1]: https://lore.kernel.org/linux-arm-msm/u4x2vldkzsokfcpbkz3dtwcllbdk4ljcz6kzuaxt5frx6g76o5@uku6abewvye7/
> > 
> >> +}
> >> +
> >> +static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode,
> >> +		const struct drm_dsc_config *dsc, bool is_bonded_dsi)
> >>   {
> >>   	unsigned long pclk_rate;
> >>   
> >> @@ -576,6 +587,10 @@ static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool
> >>   	if (is_bonded_dsi)
> >>   		pclk_rate /= 2;
> >>   
> >> +	/* If DSC is enabled, divide hdisplay by compression ratio */
> >> +	if (dsc)
> >> +		pclk_rate = dsi_adjust_compressed_pclk(mode, dsc);
> 
> Looking for the perfection, I'd also move the pclk adjustment to come 
> before the is_bonded_dsi check.

Perfection?

- Marijn

> > The confusion with this comment (and the reason the aforementioned
> > discussion [2] carried on so long) stems from the fact a division makes
> > sense for a bit/byte clock, but not for a pixel clock: we still intend
> > to send the same number of pixels, just spending less bytes on them.  So
> > as you clarify the /3 above, can you also clarify that here or drop this
> > comment completely when the function is correctly documented instead?
> > 
> > - Marijn
> > 
> >> +
> >>   	return pclk_rate;
> >>   }
> >>   
> >> @@ -585,7 +600,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
> >>   	struct msm_dsi_host *msm_host = to_msm_dsi_host(host);
> >>   	u8 lanes = msm_host->lanes;
> >>   	u32 bpp = dsi_get_bpp(msm_host->format);
> >> -	unsigned long pclk_rate = dsi_get_pclk_rate(mode, is_bonded_dsi);
> >> +	unsigned long pclk_rate = dsi_get_pclk_rate(mode, msm_host->dsc, is_bonded_dsi);
> >>   	unsigned long pclk_bpp;
> >>   
> >>   	if (lanes == 0) {
> >> @@ -604,7 +619,7 @@ unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
> >>   
> >>   static void dsi_calc_pclk(struct msm_dsi_host *msm_host, bool is_bonded_dsi)
> >>   {
> >> -	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, is_bonded_dsi);
> >> +	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, msm_host->dsc, is_bonded_dsi);
> >>   	msm_host->byte_clk_rate = dsi_byte_clk_get_rate(&msm_host->base, is_bonded_dsi,
> >>   							msm_host->mode);
> >>   
> >>
> >> -- 
> >> 2.40.1
> >>
> 
> -- 
> With best wishes
> Dmitry
>
Marijn Suijten June 11, 2023, 9:59 p.m. UTC | #7
On 2023-06-08 17:56:47, Jessica Zhang wrote:
<snip>
> > As discussed before we realized that this change is more-or-less a hack,
> > since downstream calculates pclk quite differently - at least for
> > command-mode panels.  Do you still intend to land this patch this way,
> > or go the proper route by introducing the right math from the get-go?
> > Or is the math at least correct for video-mode panels?
> 
> Sorry but can you please clarify what exactly is incorrect or different 
> about this math when compared to downstream? And, if you think that this 
> math is incorrect, what exactly has to be changed to make it the "right 
> math"?
> 
> We've already shown step-by-step [1] not only how the resulting pclk 
> from the downstream code matches out upstream calculations, but also how 
> the inclusion of porches in the upstream math would make up for the fact 
> that upstream has no concept of transfer time [2].

But it doesn't match, unless one hardcodes the desired clock (and/or
tranfer_time calculations) in a panel driver and uses that to come up
with artificial porches in the DRM mode.

> If the lack of transfer time in the upstream math is the issue, I 
> believe that that's not within the scope of this series, as transfer 
> time is not something specific to DSC.

Yes, that is the issue, and there is zero documentation in this patch
describing that it is currently a workaround to rescale the hdisplay
portion.

After all, when there are no porches pretending to make up for the lack
of transfer time, this code will be obsolete.

> Moreover, as stated in an earlier revision [3], there is no way to 
> validate DSC over DSI for video mode. As far as I know, we do not have a 
> way to validate video mode datapath for DSI in general.

It was just a question wheter a calculation of this form is correct for
video mode, where porches are used and not - afaik - tranfer time?

> [1] https://gitlab.freedesktop.org/drm/msm/-/issues/24#note_1936144
> [2] https://gitlab.freedesktop.org/drm/msm/-/issues/24#note_1945792
> [3] 
> https://patchwork.freedesktop.org/patch/535117/?series=117219&rev=1#comment_970492
> 
> > 
> > This function requires a documentation comment to explain that all.
> > 
> >> +		const struct drm_dsc_config *dsc)
> >> +{
> >> +	int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * drm_dsc_get_bpp_int(dsc),
> > 
> > This sounds like a prime candidate for msm_dsc_get_bytes_per_line(), if
> > bits_per_component==8 is assumed.  In fact, it then becomes identical
> > to the following line in dsi_host.c which you added previously:
> > 
> > 	hdisplay = DIV_ROUND_UP(msm_dsc_get_bytes_per_line(msm_host->dsc), 3);
> > 
> > If not, what is the difference between these two calculations?  Maybe
> > they both need to be in a properly-named helper.
> 
> While the math technically matches up (assuming, also, that 
> mode->hdisplay == slice_width * slice_count for all cases), there are 
> conceptual differences between the pclk and hdisplay calculations.
> 
> Just to reiterate what was already said on IRC:
> 
> In the pclk calculation, we're multiplying pclk by the compression ratio 
> (which would be target_bpp / src_bpp) -- please note that here, we 
> calculate src_bpp by doing bpc * 3.
> 
> In the hdisplay calculation, we calculate the bytes per line and divide 
> by 3 (bytes) to account for the fact that we can only process 3 bytes 
> per pclk cycle.

Your comment in v6 explains that hdisplay is divided by 3 because "DPU
sends 3 bytes per pclk cycle to DSI".  That inherently describes **a
relation between hdisplay and pclk** so why is the math then different?
After all, pclk is calculated based on the number of bytes (after DSC
compression) that need to be sent from DPU to DSI... and that has
nothing to do with the number of source bytes.

Note that the original hdisplay neither has any notion of bytes: it is
the _number of horizontal pixels_.

> So while I understand why you would want to put this behind a shared 
> helper, I think abstracting the pclk and hdisplay math away would 
> obfuscate the conceptual difference between the 2 calculations.

That difference is still unclear, FWIW.

> >> +			dsc->bits_per_component * 3);

And bits_per_component hasn't been used before yet... It is the number
of bits in the original image, so this could just be dsi_get_bpp() as
used elsewhere?

> > 
> > As we established in the drm/msm issue [2] there is currently a
> > confusion whether this /3 (and the /3 in dsi_timing_setup) come from the
> > ratio between dsi_get_bpp() and dsc->bpp or something else.  Can you
> > clarify that with constants and comments?
> 
> Sure, we are planning to add a patch to the end of this series 
> documenting the math.

Why can't you - at least for the new math introduced right here -
document it right from the get-go?  Having a separate patch to add it
seems extraneous; though extra documentation for existing code is always
welcome.

- Marijn

<snip>
diff mbox series

Patch

diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c b/drivers/gpu/drm/msm/dsi/dsi_host.c
index a448931af804..88f370dd2ea1 100644
--- a/drivers/gpu/drm/msm/dsi/dsi_host.c
+++ b/drivers/gpu/drm/msm/dsi/dsi_host.c
@@ -561,7 +561,18 @@  void dsi_link_clk_disable_v2(struct msm_dsi_host *msm_host)
 	clk_disable_unprepare(msm_host->byte_clk);
 }
 
-static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool is_bonded_dsi)
+static unsigned long dsi_adjust_compressed_pclk(const struct drm_display_mode *mode,
+		const struct drm_dsc_config *dsc)
+{
+	int new_hdisplay = DIV_ROUND_UP(mode->hdisplay * drm_dsc_get_bpp_int(dsc),
+			dsc->bits_per_component * 3);
+
+	return (new_hdisplay + (mode->htotal - mode->hdisplay))
+			* mode->vtotal * drm_mode_vrefresh(mode);
+}
+
+static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode,
+		const struct drm_dsc_config *dsc, bool is_bonded_dsi)
 {
 	unsigned long pclk_rate;
 
@@ -576,6 +587,10 @@  static unsigned long dsi_get_pclk_rate(const struct drm_display_mode *mode, bool
 	if (is_bonded_dsi)
 		pclk_rate /= 2;
 
+	/* If DSC is enabled, divide hdisplay by compression ratio */
+	if (dsc)
+		pclk_rate = dsi_adjust_compressed_pclk(mode, dsc);
+
 	return pclk_rate;
 }
 
@@ -585,7 +600,7 @@  unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
 	struct msm_dsi_host *msm_host = to_msm_dsi_host(host);
 	u8 lanes = msm_host->lanes;
 	u32 bpp = dsi_get_bpp(msm_host->format);
-	unsigned long pclk_rate = dsi_get_pclk_rate(mode, is_bonded_dsi);
+	unsigned long pclk_rate = dsi_get_pclk_rate(mode, msm_host->dsc, is_bonded_dsi);
 	unsigned long pclk_bpp;
 
 	if (lanes == 0) {
@@ -604,7 +619,7 @@  unsigned long dsi_byte_clk_get_rate(struct mipi_dsi_host *host, bool is_bonded_d
 
 static void dsi_calc_pclk(struct msm_dsi_host *msm_host, bool is_bonded_dsi)
 {
-	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, is_bonded_dsi);
+	msm_host->pixel_clk_rate = dsi_get_pclk_rate(msm_host->mode, msm_host->dsc, is_bonded_dsi);
 	msm_host->byte_clk_rate = dsi_byte_clk_get_rate(&msm_host->base, is_bonded_dsi,
 							msm_host->mode);