Bug in SimdResizeMethodNearest resizer

[mikeversteeg]

Calling SimdResizerRunwith SimdResizeMethodNearestgives spill on the right of the resized image. E.g. if I have two YUV420P images of 1920*1080 and copy one to the first quadrant of the other (0, 0, 960, 540), there is spill to the right. See image. This does not happen with SimdResizeMethodAreaor SimdResizeMethodBilinear(did not try others) but (very!) unfortunately they are too slow. SimdLib v4.9.111. Windows 64

[ermig1979]

Hello! I have some questions: 1) Is it a source or result image? 2) Could you give code to illustrate the bug.

[mikeversteeg]

1. The spill is in the target/result image.

2. 
   void *rcy, *rcu, *rcv;
   rcy = SimdResizerInit(sr.Width(), sr.Height(), tr.Width(), tr.Height(), 1, SimdResizeChannelByte, SimdResizeMethodNearest);
   if (rcy) {
   SimdResizerRun(rcy, syuv+sr.Left+sr.Top*sw, sw, tyuv+tr.Left+tr.Top*tw, tw);
   SimdRelease(rcy);
   }
   rcu = SimdResizerInit(sr.Width()/2, sr.Height()/2, tr.Width()/2, tr.Height()/2, 1, SimdResizeChannelByte, SimdResizeMethodNearest);
   if (rcu) {
   SimdResizerRun(rcu, syuv+sw*sh + sr.Left/2+sr.Top*sw/2, sw/2, tyuv+tw*th + tr.Left/2+tr.Top/2*tw/2, tw/2);
   SimdRelease(rcu);
   }
   rcv = SimdResizerInit(sr.Width()/2, sr.Height()/2, tr.Width()/2, tr.Height()/2, 1, SimdResizeChannelByte, SimdResizeMethodNearest);
   if (rcv) {
   SimdResizerRun(rcv, syuv+sw*sh*5/4 + sr.Left/2+sr.Top*sw/2, sw/2, tyuv+tw*th*5/4 + tr.Left/2+tr.Top/2*tw/2, tw/2);
   SimdRelease(rcv);
   }

[mikeversteeg]

While you're looking into this.. Do you think even further speed improvement is possible by resizing an YUV frame in 1 pass instead of 3?

[ermig1979]

I added test ResizeYuv420pSpecialTest() to TestResize.cpp.

./Test -m=s -fi=ResizeYuv420p -w=1920 -h=1080

Source image:

Output image:

Green background value in output image is result of Simd::Fill(yuvDst, 0);

[mikeversteeg]

Note the spill is green so won't show up in this test. Also note it is only with nearest filter, if I replace the filter in my code with e.g. area then the bug is gone.

[ermig1979]

Can this spill is consequence of using of uninitialized memory in the output or input image?

[mikeversteeg]

No. Do you understand this only happens with SimdResizeMethodNearest? Spill is always beyond target line and green so it will not show up in your test. Maybe it is also exclusive to Windows 64, I cannot confirm.

[mikeversteeg]

I tested again all 4 resize methods ((Nearest, Bilinear, Bicubic, Area) and can confirm only Nearest has this bug. If it's not a bug then maybe I misunderstood the documentation and my code is wrong, but it's remarkable the other 3 methods work flawlessly.

[ermig1979]

I made my test maximal close to your example (https://github.com/ermig1979/Simd/blob/master/src/Test/TestResize.cpp#L526), but could't reproduce the bug. Could you see at this test and correct me if I did some wrong?

[mikeversteeg]

I have now moved away from Nearest as I noticed in some case Bilinear looks better and it does not use noticeably more CPU. So for me this is no longer relevant, in fact you may drop support for Nearest all together :). Would you still like me to look into this?

[ermig1979]

I think this very specific bug (if it really exists) so I think is better to close this issue.