From: Ronald B. <rb...@ro...> - 2002-12-27 15:53:47
|
Hi, On Thu, 2002-12-26 at 19:02, Yue Shi Lai wrote: > And even worse, the PSNR drops abruptly to the left of the peak, if you > choose a too small MBCV. That means even if we choose to control MBCV > "intelligently", it must be extremely reliable, or we will face pretty > bad picture quality again. What I'm interested in (something you might have tested), does the optimal MBVCR setting depend on the quality, or is it the same for all qualities? I couldn't directly see this in the graphs, but I'm not sure at what quality/ies you did your measurements (since quality affects the TCV_{NET,DATA}_* reg values too). > So after all, I might have to partially retract my too optimistic > estimate. Yes, we can gain about 1.2 dB of capture quality by constantly > setting the MBCV lower than 0xff, but this means we have to make sure > the value is valid for virtually all inputs. 1,2 isn't that much, but it's worth it, if we're *sure* it's the right value. Also, as stated by Andrew, we need to be sure it doesn't cause timeouts (due to the data transfer limit of the Buz) for capturing at lower qualities than it does now, but I don't think that'll be the case. I'd be interested in seeing whether the 'optimal' MBVCR value differs per-quality, and in whether it's stable across different images. If not, I'd rather just make it 0xff and be 'safe'. For now, I'll change CVS (ZORAN_VIDEODEV_2 branch) to set this reg value to 0xf0, as a quickfix. When we know more, we can add a proper fix. Thanks for all the effort so far! Ronald -- Ronald Bultje <rb...@ro...> Linux Video/Multimedia developer |