The ticket mask decides which nonces the chip returns and I set that to 16.
If you set it to 32, you'd get half as many nonces in the same time, with the same average expected difficulty.
Setting it higher just reduces the amount of data that travels back from the chip to cgminer by intercepting it.
It has no effect on the 'quality' of the data, it just affects the average difficulty of the data by removing the lower difficulty nonces from the returned data, the chip still generates this data, it's just removed from the returned data, i.e. less data returned.
Read here, coz it would appear you may not understand the meaning of difficulty:
https://kano.is/index.php?k=workdiffAll mining is simply based on the fact that if you hash some data, 1 in 2^32 hashes will, on average, start with 32 zero bits.
1 in 2^16 hashes will, on average, start with 16 zero bits.
1 in 2^96 hashes will, on average, start with 96 zero bits.
Asic-boost uses the idea that some of the hash processes can be reused for another hash, meaning the hash process is shorter in this case i.e. sometimes you can do the whole hash process a bit faster.
It has zero effect on the value of the hash.
1. I believe that the ticket mask is no need at all. Is it filt nonce output within BM1397 chip as its internal function? or just the code in cgminer do filter? It makes code more complex and cost CPU time, too. Only threshold setup within BM1397 chip, may help to reduce number of ouput nonce with more quality nonces output. One block headers only produce 0 to 3 nonces in average running.
2. I observed total nonces come from asic chips and total nonces pass 1st test, and total nonces fail to pass 1st test. I found that a lot of nonces which are from asic chips fail to pass 1st test in test_nonce() function under ASIC boost mode. Which should not happen, I believe. And this indicate ASIC boost mode can output bad nonces (or less quality nonce).
3. sha256 algorithm is very sensitive to input message. So those midstates will produce error and cause bad nonces output for sure. I believe, turning off some midstate will help the quality of output nonces, which will increase of expectation of hitting bitcoin.
4. " 1 in 2^32 hashes will, on average, start with 32 zero bits.", a lot of miners have done this amount 2^32 of hashing, but, they won't get a 32 zero hash value. This saying is aganist my observation in years. There is always some area (sub-space of sha256 space) hide more quality nonces which can produce better (or smaller) hash values. I called these nonce as quality nonce. I often observed some miners can resolved a bitcoin under 30 seconds. I think they may have few or some universal or nearest universal nonces. Otherwise, it is against the your logic (or theory). In fact, small home miners need few universal nonces to beat big miners, I believe.