Avoid block copy operations in scrypt's SMix step to improve performance

[palant]

Function smix() in the scrypt implementation currently does something like that:

• Initialize X
• Copy X to V[0]
• Mix X to Y
• Copy Y to V[1]
• Mix Y to X
• Copy X to V[2]
• ...

If you look closely, the copy operations are unnecessary. Instead, one can do the following:

• Initialize V[0]
• Mix V[0] to V[1]
• Mix V[1] to V[2]
• ...

Only the very last mix operation is slightly problematic because there is no space in V for the result any more and it needs to go into X. This issue can be avoided by allocating X and Y at the end of the V array. That's what I have done now, improves performance by around 10% for me (in both Firefox and Chrome).

[dchest]

Thanks, I'll take a deeper look a bit later.

[palant]

Any progress on that?

[dchest]

Sorry, didn't have time yet, but will definitely do this this week.

[dchest]

Thank you! Merged and published.