Superblock qp sweep experiment given fixed rdmult
This commit is to sweep multiple sb qindexes for a
given rdmult(lambda) to optimize the rdcost. The
code changes are controled by a new encoder argument
"sb_qp_sweep", and is disabled by default now.
A unit test is added to ensure end-to-end encoding
and decoding run well.
Borg testing results are provided as below, with
roughly ~4x slower speed. The bdrate gain indicated
some headrooms for better lambda to qp mappings.
avg_psnr ovr_psnr ssim vmaf
speed=1 lowres -1.031 -0.986 -1.231 -1.029
midres2 -1.288 -1.217 -1.424 -1.217
ugc480p -0.888 -0.864 -1.15 -0.968
speed=2 lowres -1.049 -1.005 -1.327 -1.399
midres2 -1.274 -1.207 -1.446 -1.253
ugc480p -0.95 -0.908 -1.172 -0.851
speed=3 lowres -1.238 -1.193 -1.482 -1.254
midres2 -1.48 -1.401 -1.562 -1.219
ugc480p -1.026 -0.959 -1.315 -1.063
speed=4 lowres -1.535 -1.479 -1.812 -1.928
midres2 -1.628 -1.584 -1.905 -1.662
ugc480p -1.193 -1.079 -1.585 -1.269
speed=5 lowres -1.95 -1.874 -2.158 -2.003
midres2 -2.18 -2.098 -2.434 -1.856
ugc480p -1.414 -1.304 -1.842 -1.469
Change-Id: I3dae74c7ac0d557fac498cb7f0cc9ae0d23be29e
diff --git a/av1/arg_defs.c b/av1/arg_defs.c
index acda4f8..abfd4b3 100644
--- a/av1/arg_defs.c
+++ b/av1/arg_defs.c
@@ -679,5 +679,9 @@
"key frame (-1 to 5). When set to -1 (default), it does not have any "
"effect. The actual maximum pyramid height will be the minimum of this "
"value and the value of gf_max_pyr_height."),
+ .sb_qp_sweep =
+ ARG_DEF(NULL, "sb-qp-sweep", 1,
+ "When set to 1, enable the superblock level qp sweep for a "
+ "given lambda to minimize the rdcost."),
#endif // CONFIG_AV1_ENCODER
};
