Autotune measures your device on every run and picks a launch size. The tuning database overrides that measurement where a known value does better, or where autotune cannot see what a particular workload needs.
Every file ending in .hctune under tunings/ is read at startup and all the rows go into one table. There is no single database file, and the split across files is for whoever maintains them rather than for hashcat. The shipped tree carries 322 rows over 12 files plus 329 aliases in Alias.hctune. Your own file sits beside them and is read the same way.
The database is skipped entirely for --help, -I, --hash-info, --keyspace, --left, --show, --version and --identify, because none of them launch a kernel.
There are six columns. The first three determine whether the row applies. The last three specify the settings:
| column | meaning |
|---|---|
| Device Name | the device, or an alias, or * for any |
| Attack Kern | the attack kernel the row applies to |
| Hash Type | the hash-mode, or * |
| Vector Width | the vector width to use |
| Kernel Accel | the -n value |
| Kernel Loops | the -u value |
So a row is a measurement someone already took, recorded for a device and hash-mode combination rather than measured again on every run.
tunings/README.md in the release for the file-format reference