[Mesa-dev] [PATCH 1/2] vl: fix low strength denoise filter

[Christian König]

On 13.02.2012 17:31, Andy Furniss wrote:
> Christian König wrote:
>> A filter strength of zero or one doesn't make any
>> sense.
>
> Low strengths work OK now, but this patch seems to have made denoise 
> stronger. Maybe I am looking at the wrong material, but more than 0.2 
> for sharpen or denoise looks too way much.
>
> Benchmark wise sharpen seems cheap, but denoise at full strength is 
> expensive, and comparing full on before and after the patch on 1080 
> material shows that it's doing more work.
Correct, before anything under 0.4 would have given you a zero or one 
pixel window and 1.0 would have given you a 9 pixels window, now 0.1 
gives you a 3 pixels window (the smallest size that makes sense) and 1.0 
gives you an 11 pixels window. Only odd numbers are allowed as window 
size, so if 11 is way to strong what would you say is a good maximum for 
the window size?

> This is with my card (HD4890) on low (239840 kHz) as for htpc it's 
> quiet, testing a 60 sec stream, roughly -
>
> no options = 16 sec
> full sharpen = 24 sec
> full denoise before patch = 48 sec
> full denoise after patch = 66 sec
Yeah, that was to be expected, for a window size of 3 you need to fetch 
5 pixels, for a windows size of 9 you need 17 and for a window size of 
11 you need 21...

> This is current gits so 2D tiling is on, FWIW vdpau is by far the 
> biggest gainer from this of anything I've tested - in the past I could 
> easily compare with R600_TILING= , but I guess I can't turn it off for 
> textures any more.
That was also to be expected, tilling can dramatically increase the 
memory bandwidth, and since we still copy the image data way to often 
around with VDPAU it's no wonder that enabling this has a major effect 
on that performance. You can still easily disable tilling by just using 
an array_mode of zero in r600_texture_create, we have just removed the 
environment variable because tilling now seems to work stable.

Christian.