The core problem was the cuda hefty Thread per block set to high
but took me several hours to find that...
btw... +25% in heavy 12500 with 256 threads per block... vs 128 & 512
if max reg count is set to 80...
Was maybe my fault, but the benchmark mode was
always recomputing from nonce 0.
Also fix blake if -d 1 is used (one thread but second gpu)
stats: do not use thread id as key, prefer gpu id...
There was a different behavior on linux and visual studio
That was making it hard to link functions correctly
That remove some ifdef / extern "C" requirements
note about x86 releases, x86 nvml.dll is not installed on Windows x64!
Displayed data is the average of the last 50 scans in the 5 last minutes
Also move cuda common functions in a new file (cuda.cu)
Signed-off-by: Tanguy Pruvot <tanguy.pruvot@gmail.com>