Cryptonight variant 2 support

[SChernykh]

• Added new per-thread parameter "unroll_factor" which can be set to 1, 2, 4, 8, 16, 32 or 64, default is 8.
• CNv2 OpenCL code is in a separate kernel because it uses 1KB more local memory which can hurt other variants performance
• Fixed a bug with "comp_mode"=1 no matter what was set in config.json

Sample thread setting for Radeon RX 560 which gave me the best performance:

"threads": [
    {
        "index": 0,
        "intensity": 1024,
        "worksize": 32,
        "strided_index": 2,
        "mem_chunk": 2,
        "unroll_factor": 16,
        "comp_mode": false,
        "affine_to_cpu": false
    }
],

Performance was the same as in my previous GPU tests.

[SChernykh]

It turned out to be easier than I thought - it took only a few hours to port my OpenCL code. I only tested it on Windows with my RX 560. Command-line parameter setting was not tested, I don't know xmrig well enough to test that.

[xmrig]

What about strided_index? We can't use global variant in OclCache::load() because it's breaks ability to switch variant in runtime, so new code should not depend of -DSTRIDED_INDEX= if it not supported.

About command-line parameter don't worry, I will fix it.
Thank you.

[SChernykh]

Ah, I see. I need to think how to rewrite it properly then.

[SChernykh]

Wow, I've added strided index support to V2 and got better performance with strided_index = 2 and mem_chunk = 2 (64 bytes), but strided_index = 1 kills V2 performance because 16 bytes is too small granularity for V2. I'll test it some more and then submit.

[Bathmat]

@SChernykh Attempting to run CNv2 with dual threads and strided_index=1 results in very poor performance and/or fails. On my RX470/480 GPUS, setting mem_chunk=4 gave slightly worse performance, and mem_chunk=8 was very poor/failed.

Therefore, I would recommend strided_index=2, mem_chunk=2 as the default config.

Additionally, if the default worksize setting is 8, then I would recommend setting the default unroll_factor to 4. Seems to help slightly with performance (1% maybe?)

[xmrig]

@SChernykh Merged, thank you, about default options/better autoconfig new PR welcome.