1. I didn't change the get_whole_scratchpad so if it has a bug (looks like it has) it's better to contact the main branch developer as well.
2. m_job.difficulty shouldn't be zero. If it is - something's wrong with a pool. I added a return from the worker thread as it doesn't make any sense to search for nonce having wrong data.
updated github
mbk, can you add m_scratchpad.clear(); at the beginning of the get_whole_scratchpad function. if the miner gets to a point where it wants to download the whole scratchpad again, it just appends instead of starts over.
Also, on my nvidia card it crashes because difficulty is zero at the start and it divides by that. I just put put a check for if zero, set to something else and it doesn't happen again.