[NORMATIVE] Reduce context model size for inter mode coding
Reduce the context model size for inter mode coding. Before this
change the codec will condition on if the spatial neighbors newmv
count is 0, 1, or 2 and above to decide the context index. This
change makes it condition only on newmv count 0 or 1 and above.
Hence effecitvely reducing it into a binary flag for context
modeling. The compression performance change is below 0.01%.
BUG=aomedia:1288
Change-Id: Iff11222dc45cce142da5caf9809ebb753af0d29f
diff --git a/av1/common/mvref_common.c b/av1/common/mvref_common.c
index 7432168..ccae30e 100644
--- a/av1/common/mvref_common.c
+++ b/av1/common/mvref_common.c
@@ -922,10 +922,8 @@
case 2:
default:
- if (newmv_count[ref_frame] >= 2)
+ if (newmv_count[ref_frame] >= 1)
mode_context[ref_frame] |= 4;
- else if (newmv_count[ref_frame] == 1)
- mode_context[ref_frame] |= 5;
else
mode_context[ref_frame] |= 6;
diff --git a/test/decode_to_md5.sh b/test/decode_to_md5.sh
index d462aac..ef2711c 100755
--- a/test/decode_to_md5.sh
+++ b/test/decode_to_md5.sh
@@ -57,7 +57,7 @@
decode_to_md5_av1() {
# expected MD5 sum for the last frame.
- local expected_md5="fc7565de847d04dc3485b4858c0ed298"
+ local expected_md5="6ebec9f224ad2ad53910c05261b04ee3"
local file="${AV1_IVF_FILE}"
# TODO(urvang): Check in the encoded file (like libvpx does) to avoid
