Use 8x8 block size for screen content in the first pass
Screen content videos often contain details, for example characters,
and sharp changes in a small neighborhood.
With smaller block sizes, the motion estimation allows the first pass
to get a better estimation of coded errors, thus helping the encoding
decisions in the second pass.
With this change, for screen content test set, on speed 1, 150 frame,
we can see -1.5% gain on psnr, -0.4% ssim, -1.3% vmaf.
Other test sets still use 16x16 block sizes.
STATS_CHANGED
Change-Id: I3d51320b56adc8ae8bf342a1af7053770a5df198
diff --git a/av1/encoder/firstpass.c b/av1/encoder/firstpass.c
index 8a827ea..db098e1 100644
--- a/av1/encoder/firstpass.c
+++ b/av1/encoder/firstpass.c
@@ -1146,7 +1146,8 @@
av1_set_screen_content_options(cpi, features);
}
// Unit size for the first pass encoding.
- const BLOCK_SIZE fp_block_size = BLOCK_16X16;
+ const BLOCK_SIZE fp_block_size =
+ cpi->is_screen_content_type ? BLOCK_8X8 : BLOCK_16X16;
// Number of rows in the unit size.
// Note mi_params->mb_rows and mi_params->mb_cols are in the unit of 16x16.
const int unit_rows = get_unit_rows(fp_block_size, mi_params->mb_rows);
