Something weird happend here, read on!
I've expected some performance drop because this algorithm is using the password data itself inside the iteration loop.
That is different to PBKDF2, which I've converted in mode 2100 before and which did not show any performance as expected.
So after I've finished converting this kernel and testing everything works using the unit test, I did some benchmarks to see how much the
performance drop is.
On my 750ti, the speed dropped (minimal) from 981kH/s -> 948kH/s, that's mostly because of the SIMD support i had to drop.
If I'd turn off the SIMD support in the original, the drop would be even less, that us 967kH/s -> 948kH/s which is a bit of a more reasable
comparison in case we just want to rate the drop that is actually caused by the code change itself.
The drop was acceptable for me, so I've decided to check on my GTX1080.Now the weird thing: The performance increased from 6619kH/s to
7134kH/s!!
When I gave it a second thought, it turned out that:
1. The GTX1080 is a scalar GPU so it wont suffer from the drop of the SIMD code as the 750ti did
2. There's a change in how the global data (password) is read into the registers, it reads only that amount of data it actually needs by using
the pw_len information
3. I've added a barrier for CLK_GLOBAL_MEM_FENCE as it turned out to increase the performance in the 750ti
Note that this kernel is now branched into password length < 40 and larger.
There's a large drop on performance where SIMD is really important, for example CPU.
We could workaround this issue by sticking to SIMD inside the length < 40 branch, but I don't know yet how this can be done efficiently.
This is a test commit using buffers large enough to handle both passwords and salts up to length 256.
It requires changes to the kernel code, which is not included in here.
It also requires some of the host code to be modified. Before we're going to modify kernel code to support the larger lengths I want to be
sure of:
1. Host code modification is ok (no overflows or underflows)
2. Passwords and Salts are printed correctly to status, outfile, show, left, etc.
3. Performance does not change (or only very minimal)
This is not a patch that supports actual cracking both passwords and salts up to length 256, but it can not fail anyway.
If if it does, there's no reason to continue to add support for both passwords and salts up to length 256.
1) done + got rid of all u8 datatypes in shared struct
2) cf. previous
3) necessary as this is computed in _init then used in _comp
4) done
5) done
6) done => switch to 16
7) done
