[PATCH] drm/omapdrm: dispc: Refuse x-decimation above 4 for all but 8-bit formats

[Jyri Sarha]

On 02/08/17 13:25, Laurent Pinchart wrote:
> Hi Jyri,
> 
> Thank you for the patch.
> 
> On Tuesday 07 Feb 2017 16:41:20 Jyri Sarha wrote:
>> Let's disable all scaling that requires horizontal decimation with
>> higher factor than 4, until we have better estimates of what we can
>> and can not do. However, 1 byte per pixel color format appear to work
>> Ok with all decimation factors.
>>
>> When decimating horizontally by more that 4 the dss is not able to
>> fetch the data in burst mode. When this happens it is hard to tell if
>> there enough bandwidth. Despite what theory says this appears to be
>> true also for 16-bit color formats.
>>
>> Signed-off-by: Jyri Sarha <jsarha at ti.com>
>> ---
>>  drivers/gpu/drm/omapdrm/dss/dispc.c | 19 +++++++++++++++++++
>>  1 file changed, 19 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/omapdrm/dss/dispc.c
>> b/drivers/gpu/drm/omapdrm/dss/dispc.c index 5554b72..61daef6 100644
>> --- a/drivers/gpu/drm/omapdrm/dss/dispc.c
>> +++ b/drivers/gpu/drm/omapdrm/dss/dispc.c
>> @@ -2506,6 +2506,25 @@ static int dispc_ovl_calc_scaling_44xx(unsigned long
>> pclk, unsigned long lclk, return -EINVAL;
>>  	}
>>
>> +	if (*decim_x > 4 && color_mode_to_bpp(color_mode) > 8) {
>> +		/*
>> +		  Let's disable all scaling that requires horizontal
>> +		  decimation with higher factor than 4, until we have
>> +		  better estimates of what we can and can not
>> +		  do. However, 1 byte per pixel color format appear to
>> +		  work Ok with all decimation factors.
>> +
>> +		  When decimating horizontally by more that 4 the dss
>> +		  is not able to fetch the data in burst mode. When
>> +		  this happens it is hard to tell if there enough
>> +		  bandwidth. Despite what theory says this appears to
>> +		  be true also for 16-bit color formats.
>> +		*/
>> +		DSSERR("Not enough bandwidth (x-decimation factor %d > 4)",
>> +			*decim_x);
>> +		return -EINVAL;
> 
> This needs to be validated during the atomic check phase to avoid failures at 
> commit time that are much harder to handle properly.
> 

Sure, but I can not fix everything in one step. This is a first simple
quantitative test that is supposed to tell if scaling can be done
without synclost errors. Next step is to make dispc to not look to HW
for current clock settings, and then finally make all relevant check
callbacks (or some flag in the config callbacks) in dispc.

Anyway, there will be a v2 of this patch that tests color_mode !=
OMAP_DSS_COLOR_NV12 instead of color_mode_to_bpp(color_mode) > 8. The
NV12 appears to be the only color format that works reliably with
current implementation.

Best regards,
Jyri

>> +	}
>> +
>>  	*core_clk = dispc.feat->calc_core_clk(pclk, in_width, in_height,
>>  				out_width, out_height, mem_to_mem);
>>  	return 0;
>