torchvision decoder. fix artifacts, blury images for frame dimensions not multiple of 8 (#7523)

Summary:
Pull Request resolved: #7523

In MUI we found an artifact for video with resolution 426x426, not equal to 8.
We find that root cause is torchvision decoder, which MUI uses extensively

Tiny black bar on the right in the image

Before the fix
{F949912112}

After the fix
{F949913076}

Full video

https://pxl.cl/2CN1G
https://pxl.cl/2CN1L

Differential Revision: D45035410

fbshipit-source-id: 2d4f6a4812df075fc4677e6b96dc2e0029ea6e92

Author: Vadim Zubov

torchvision/csrc/io/decoder/video_sampler.cpp

@@ -181,6 +181,24 @@ bool VideoSampler::init(const SamplerParameters& params) {
   // set output format
   params_ = params;
 
+  if (params.in.video.format == AV_PIX_FMT_YUV420P) {
+      /* When the video width and height are not multiples of 8,
+       * and there is no size change in the conversion,
+       * a blurry screen will appear on the right side
+       * This problem was discovered in 2012 and
+       * continues to exist in version 4.1.3 in 2019
+       * This problem can be avoided by increasing SWS_ACCURATE_RND
+       * details https://trac.ffmpeg.org/ticket/1582
+       */
+      if ((params.in.video.width & 0x7) || (params.in.video.height & 0x7)) {
+          VLOG(1) << "The width " << params.in.video.width
+                  << " and height " << params.in.video.height
+                  << " the image is not a multiple of 8, "
+                  << "the decoding speed may be reduced";
+          swsFlags_ |= SWS_ACCURATE_RND;
+      }
+  }
+
   scaleContext_ = sws_getContext(
       params.in.video.width,
       params.in.video.height,