__int128 to support more than 64 CPUs

[louisje]

128-bit integer can hold more bits to cpu mask than 64-bit, if GNUC is available.

[Wikiechiu]

Hi Louisje,

May I know

1. what type of CPU to pull together
2. how to put them in use
3. what type of crypto coin can I effectively mine with this?

[louisje]

Still only one CPU in use, just extending 64 core at most to 128 core.

1. It can associate with any type of CPU,
2. without any specific configuration, just use it as usual.
3. And this change is for all algorithm. It just extends the CPU affinity.

The best way to exam this change, is to have a CPU which is more than 64-core.
The CPU I'm using is "Intel(R) Xeon(R) CPU E5-2699 v4", a 88-core CPU.
Without this improvement, I found that it is at most 64 core simultaneously in use, because the "unsigned long" cpu set mask can at most 64-bit wide.

__int128 is 128-bit wide, and ideally as cpu mask, it can control 128-core at most. But __int128 is supported only by GCC, so I put them in #ifdef __GNUC__ closure with original unsigned long in the else part as the fallback. This is working for me, all 88 core are working simultaneously.

cpuminer -o "stratum+tcp://pool.minexmr.com:4444" -u ${WALLET_XMR} -p x -a cryptonight -D

Above is the command line I'm using. I've not try with other crypto function yet, but obviously this change can apply to them as well.

[kyleboddy]

I was able to apply this patch and get 256 threads working with a Xeon Phi x202 but cpuminer is not optimized for the architecture and it is very slow. Only lukminer right now is optimized to use the Phi architecture.

At any rate, thanks for the patch!

[philtimmes]

Considering the fact that I have an 8192 thread system, this patch was helpful. Thanks.