./VanitySearch -gpu -gpuId 0 -i 5000.txt
VanitySearch v1.16
Search: 133 addresses (Lookup size 133,[1,1]) [Compressed]
Start Wed Jan  8 07:56:31 2020
Base Key: CAFFCF119BAE13EAA1F0C053F98BBA2D3982E39B40FB9A991A475F0B46BA4751
Number of CPU thread: 1
GPU: GPU #0 GeForce RTX 2080 Ti (68x64 cores) Grid(544x128)
[2196.82 Mkey/s][GPU 2192.28 Mkey/s][Total 2^37.02][Prob 0.0%][50% in 1.46226e+31y][Found 0]  GPUEngine: Launch: an illegal memory access was encountered

this happens with 133 addresses.. same thing after 3sec. but if it helps I can give you the 22 million file.. its about 750mb.. I was just courious about performance with big datasets so I dumped the p2pkh and the p2pk addresses from the chainstate of a full node and that's about 22mil at the moment...
(if you want to do the same I have a tool on my GitHub that can batch convert the pub keys compressed uncompressed and leveldb encoded to there corresponding addresses because they are only in utxo form in the chainstate.. search for pub2addr)
i haven't looked too hard in your code but r u using a hash table instead of a bloom filter.. right?
and at the moment its just research 4 me but im thinking about implementing a vanity searcher on a Xilinx u200 FPGA... MAYBEEEEEEEEEE.. as a bitstream.. and so im looking around at the moment..

./VanitySearch  -l
GPU #0 GeForce RTX 2080 Ti (68x64 cores) (Cap 7.5) (11016.3 MB) (Multiple host threads)
GPU #1 GeForce RTX 2080 Ti (68x64 cores) (Cap 7.5) (11019.4 MB) (Multiple host threads)

yes it states Cap 7.5 so compiling with ccap=75 seems right..