Sorry, wrong fork requested for pull, please ignore
~~Replaced shifts by 16 and 24 with byte permutations, and shifts by 32 with high low swap. Significant speed gains for pre-compute-5.2 devices, since the slow pre-5.2-rotr64 function is only used for the 63-bit rotate.
One half of a K520 (GK104 chip, 1536 CUDA cores @ 800MHz) went from ~230 MH/s to 335 MH/s (+45.6%).
980 Ti (GM200 chip, 2816 CUDA cores @ 1500 MHz core) went from ~1630 MH/s to 1835MH/s (+12.5%).
Performance gains generally above 40% for 6xx and 7xx cards, with about 10% improvement on Maxwell cards. Successfully mined two blocks.~~
Sorry, wrong fork requested for pull, please ignore
~~Replaced shifts by 16 and 24 with byte permutations, and shifts by 32 with high low swap. Significant speed gains for pre-compute-5.2 devices, since the slow pre-5.2-rotr64 function is only used for the 63-bit rotate.
One half of a K520 (GK104 chip, 1536 CUDA cores @ 800MHz) went from ~230 MH/s to 335 MH/s (+45.6%). 980 Ti (GM200 chip, 2816 CUDA cores @ 1500 MHz core) went from ~1630 MH/s to 1835MH/s (+12.5%).
Performance gains generally above 40% for 6xx and 7xx cards, with about 10% improvement on Maxwell cards. Successfully mined two blocks.~~