tree 0d6ce838643a3cdf9cca92dcee6e0ce5e00aaf15
parent 79a6f36f4257bf10fac77d521dc31708fd424158
author Neeraj Gadgil <neeraj.gadgil@ittiam.com> 1660813681 +0530
committer Yunqing Wang <yunqingwang@google.com> 1660948555 +0000

rtc: Reduce blk-level MV pel precision

This CL extends the sf force_half_pel_block to speeds 7, 8
for rtc set with a new algorithm. It reduces MV precision for
relatively-static (e.g. background), low-complex large blocks
using source variance, SAD, bsize and qp.

The speed-up / BD Rate trade-off is:
 ---------------------------------------------------------
|cpu|Resolution|Instr. Count|    BD-Rate Drop (%)         |
|   |          |Reduction(%)|avg. psnr|ovr. psnr|  ssim   |
 ---------------------------------------------------------
| 7 |  HDRES   |  2.256     |  0.2120 |  0.2121 | 0.2066  |
| 8 |  HDRES   |  2.195     |  0.1821 |  0.1830 | 0.1826  |
 ---------------------------------------------------------

Worst and Best case-drops:
 ----------------------------------------------------------
|      |   |                 |        BD Rate Drop (%)     |
|      |cpu|    Clip         |   (-ve: Gain, +ve Loss)     |
|      |   |                 |-----------------------------|
|      |   |                 |avg.psnr |ovr. psnr|  ssim   |
|----------------------------------------------------------|
| Best | 7 |testnoise720p    | -0.3958 | -0.2445 | -0.7787 |
|      | 8 |testnoise720p    | -0.7229 | -0.4860 | -0.7079 |
|----------------------------------------------------------|
|Worst | 7 |vidyo1 (avg.psnr)|  0.8261 |  0.7455 |  0.5534 |
|      | 7 |mj1vc720p (ssim) |  0.6386 |  0.6482 |  0.7535 |
|      | 8 |mj1vc720p        |  0.8805 |  0.8451 |  0.9263 |
 ----------------------------------------------------------

No changes to speed 9, 10 or rtc-derf set

STATS_CHANGED

Change-Id: I7f9ba58274e003b8cfc188cd80099b28f187be67