OpenCL AES: Optionally use inverse tables for decryption key schedule #5806
Conversation
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
magnumripper
force-pushed
the
opencl-aes-inv-tables
branch
2 times, most recently
from
343bb99 to
9709487
Compare
|
Ready to merge. Like I said in #5800, AMD Vega got 42% boost and nvidia 1080ti got 44% when running cryptosafe-opencl which has no KDF at all and decrypts just a single block of AES-256 per key setup. Good stuff. BTW it also boosts key setup for encryption a tiny bit (~1%) as it no longer copies unneeded tables to local memory. |
|
Using relbench I see that only cryptosafe get this huge boost but I believe/hope that's mainly because test vectors are usually smaller than real world hashes. Cryptosafe is so poorly designed our format needs on-device mask to fully exploit it. Oh and another reason might be that some formats need a cost parameter to actually use AES or only AES (where default cost may be eg. RC4). I used |
magnumripper
force-pushed
the
opencl-aes-inv-tables
branch
from
b4c3e16 to
a5c921c
Compare
|
For completeness I should probably improve the non-table inverse code path similar to the more efficient code we now have in the inverse table code path. Then we can compare apples to apples. |
magnumripper
force-pushed
the
opencl-aes-inv-tables
branch
from
b21e38f to
b8237ab
Compare
This boosts AES_set_decrypt_key() by halving the number of table lookups. It mostly affects formats that decrypt a small amount per key (several formats only decrypt one or two blocks). Closes openwall#5800, see openwall#5613 (comment)
Now only a macro definition is different between tables or not. This revealed that the inverse tables does no good, the improved swap/invert code does. Consequently we disable the use of tables - but leave the option in there.
magnumripper
force-pushed
the
opencl-aes-inv-tables
branch
from
b8237ab to
9d5193c
Compare
As I said in #5800 this revealed it was that code that gave us a 40% boost, not the tables. Let's not merge this yet, I'll try to get to the bottom of it. |
This made little difference except maybe on CPU, at least under MacOS.
This is my current take on #5800, subject to change.