On 04/21/2018 11:29 PM, Steve Longerbeam wrote: > > > On 04/12/2018 03:04 AM, Philipp Zabel wrote: >> On Sat, 2018-04-07 at 15:04 +0200, Marek Vasut wrote: >>> In VDIC direct mode, the VDIC applies combing filter during and >>> doubles the framerate, that is, after the first two half-frames >>> are received and the first frame is emitted by the VDIC, every >>> subsequent half-frame is patched into the result and a full frame >>> is produced. The half-frame order in the full frames is as follows >>> 12 32 34 54 etc. >> Is that true? We are only supporting full motion mode (VDI_MOT_SEL=2), >> so I was under the impression that only data from the current field >> makes it into the full frame. The missing lines should be purely >> estimated from the available field using the di_vfilt 4-tap filter. > > Yes, the reference manual states that the VDI_MOT_SEL field > value 2 is "Full motion, only vertical filter is used". Which is > clearly referring to the di_vfilt 4-tap filter. > > There are still some questions about how full motion mode is > supposed to work. For one thing, the di_vfilt only operates on four > consecutive lines of a single field. It makes no sense that the VDIC > can compensate for motion between fields when it is operating > with only one field. > > Marek, where did you get the information that full motion mode > applies some kind of combing filter? By observing how the HW behaves. The input from NXP forum was mostly useless or actually outright wrong. I have to admit it's been a while since I created the patch, but what I saw was basically the hardware producing frames as a combination of halfframe 1-2 2-3 3-4 etc , thus doubling the framerate. Setting the skip normalized the framerate without any loss of image information. > A combing filter would imply > taking previous and/or future fields back into the result, which is > exactly what the other motion modes do, but as I said the reference > manual is clear that full motion mode uses only the (single field) > vertical filter. > > The manual also mentions regarding "real-time mode" which we are > making use of (in which the VDIC FIFO1 receives field F(n-1) directly > from the CSI rather than from DMA): > > "In addition IDMAC read the field from FIFO1 and store in external memory. > Then stored frames are used as F(n) and F(n+1).". > > It is nowhere explicitly stated, but the assumption is that this is IDMAC > channel 13 that stores the CSI field to memory. But many people have > asked Freescale/NXP how this works in the past, and there has never > been a satisfactory answer. And people have reported no success in > getting this channel to work including myself. > > So the approach this driver takes is to use real-time mode to receive > F(n-1) directly from CSI, in concert with full motion mode, so that > the VDIC operates on F(n-1) only. Thus no DMA is necessary. > > Finally I have to say that the other modes are supported in this driver, > but they require DMA transfer of all three fields, and there is no > output device node written to make use of those modes yet. But > from experience, the de-interlaced result is of much better quality > than the real-time/full-motion output. > > > Steve > > -- Best regards, Marek Vasut
