Multiple bugs when Compound prediction is generalised (allowing any combination of references & sign bias)

What is the expected behavior? What do you see instead?
Currently the encoder is restricted to a predefined combination of reference 
frames for compound prediction with specific sign bias. But when the encoder is 
generalised to generate streams allowing any combinations of reference frames & 
sign bias (by adding appropriate code/combinations/thresholds in RD mode 
decision loop, and appropriate entropy coding changes), the encoder/decoder may 
crash due to the following bugs 1) Encoder crashes  due to wrong logic of 
writing compound prediction references in to bitstream in write_ref_frames 
function. 2) Decoder crashing due to wrong way of reference mode coding in 
write_compressed_header function in encoder. 3) Encoder crashes  due to wrong 
update of compound prediction reference counts in update_stats function. For 
the compund prediction, the ref encoder code does not have the feature to 
increment counts in case of fixed reference frame being other than alt ref 
frame. 4) Encoder crashes  due to invalid compound prediction mode combinations 
where respective ref frames are not available/vaild.   

What version are you using? On what operating system?
Version 2.3.0. Windows OS. 

Can you reproduce using the vpxdec or vpxenc tools? What command line are
you using?
You may reproduce these encoder/decoder crashes when you modify encoder to make 
compound prediction completely generalised and generate streams for multiple 
values of reference frame sign bias. 

Please provide any additional information below.
All the mentioned issues can be fixed by modifying the encoder according to 
spec/decoder.

Original issue reported on code.google.com by ynsai...@gmail.com on 17 Mar 2015 at 10:34

matanbs / webm

Multiple bugs when Compound prediction is generalised (allowing any combination of references & sign bias) #977