Mismatch on sample with error resilence enabled

err.ivf

Created today with aomenc git head. Almost every frame, if not all, has artifacts when decoded with dav1d.

Mismatch on second frame, maybe post filters

The mismatch does not originate from the post filters. I disabled LR, CDEF and LPF and I still get the mismatch.

added mismatch label

The above sample was encoded after https://aomedia.googlesource.com/aom/+/7a9bb78320f29b7f4da2f81c4db6cc955ac65247 but before https://aomedia.googlesource.com/aom/+/29147cfe56c41a91b4e1eb636edc0e9c1bf84f98, so perhaps it's not a "good" sample.

err2.ivf is encoded after both commits, using the same soruce video and encoding settings. There are still some mismatches, but apparently not as many.

mentioned in issue #71 (closed)

changed milestone to %0.1.0

So, let's scratch all I wrote earlier. I'm displaying invisible frames (because the mismatches don't align well with block boundaries) and am seeing this weird thing where invisible frames (after --skip=158) for 9 decodes correctly, but for 4/2 they do not. However, this is really weird, the show_existing_frame for these do display correctly. [edit later: we do not use show-existing-frame in this sample at all.] It appears that some of the postprocessing we do is not included in the reference buffers? Or something like that. First block in this frame with a mismatch is the arm (by4=8-16,bx4=24-26). I'm focussing on the smaller block under it (8x8) at by4=22-24,bx4=20-22.

MVs read in dav1d:

Post-residual_mv[1:y=2,x=-16,2:y=2,x=-16]: r=47696

MVs read in aom:

Post-residual_mv: 47696, mv=2,-16 and -2,16

So it appears my earlier guess was correct and we did invert signs for the MV somewhere, but I don't know why/how yet.

mentioned in merge request !200 (merged)

closed via merge request !200 (merged)

closed via commit 3eff6189