Quark and S3 are now a bit faster (+1 %)
x11 get +0.6 % (+20kH/s on a 750ti, +30kH on a 960)
80 bytes implementation to do/test ... (skein/skein2)
but keep my previous version for older devices...
Look like a stream synch problem, not related to cuda 7 headers or cudart
The threadfence() added doesnt changes performances, and could also
be related to the random cpu validation errors... so keep it for all.
Note: the 80-bytes variant used in skein2 doesn't seems affected.
Implementation based on klausT work.. a bit different
This code must be placed in a common .cu file,
cuda.cpp is not compiled with nvcc and doesnt allow cuda code...
based on klaus commits, will increase a bit speed of most algos
PS: main increase is due to the register count tuning in Makefile
and for skein512 on linux, its the ROTL64
but almost no changes on X11 : 2648MH/s vs 2630 before
Tanguy Pruvot