blake kernel optimisation for nvidia cards

[tpruvot]

part of my tasks in the RFP-5

Signed-off-by: Tanguy Pruvot tanguy.pruvot@gmail.com

[jcvernaleo]

@tpruvot really sorry for taking so long before getting to this PR.
So I can't test or benchmark this myself (no nvidia hardware) but I did go through the code pretty well. Looks good. One question (which I'll put as a separate comment so I can attach it to the right line).

[jcvernaleo]

Actually, it appears github will not let me do the line by line comment (diff to large maybe?).
Is there a reason that in the code after /* 14 rounds */ you made each round a single line rather than the original code which had each statement on a line? I found the single line with multiple statements much harder to read.

[tpruvot]

no real reason, its how it was done in other blake kernels

[jolan]

tACK

Mined on testnet with this. 2 blocks on fast AMD hardware before quitting, 2 on Intel, and a couple dozen on Nvidia.

[cjepson]

utACK

OK for merge