Colliding password protected MS office 97-2003 documents

[magnum]

Good stuff. I just have to ask... Would you care to share any (high or low level or just conceptual) details on how you managed to speed up RC4 on GPU?

I tried these obvious things when I played with 40-bit RC4 BF (even older office) on GPU a while ago:

• Using char or uint for the state array
  
• Putting state array in local memory or not
  
• memcpy IV from a constant array [instead of using swap_state() in a for loop]
  
• Unrolled set_key() for fixed length
  
• Fixed-length decryption, unrolled to 32-bit stores
  

For a given GPU, a good combination of the above is significantly faster than a bad combination of it... but I just can't come up with any more ideas and iirc I was way below your speeds.