RandomX 1.0.4 support for all GPUs on Windows

Contains portable OpenCL code for all GPUs and optimized code for AMD Vega GPUs (Windows only).

Update v1.0.4-3: you can use --dataset_host option to avoid allocating 2 GB dataset in GPU memory, but hashrate will be a bit lower because accessing system memory from GPU is slower.

Update v1.0.4-4: fixed invalid hashes on Vega FE with intensity > 2048

Update v1.0.4-5: fixed invalid hashes on AMD GPUs (portable code)

Update v1.0.4-6: fixed OpenCL compilation error in optimized Vega code

Recommended command line for all GPUs:

RandomX_OpenCL.exe --mine --validate --portable --workers 8 --bfactor 5 --fast_fp [--dataset_host] --intensity N

Choose N depending on available GPU memory: 1984 or less for 8 GB GPUs, 1536 or less for 6 GB GPUs, 768 or less for 4 GB GPUs.

Recommended command line for testing AMD Vega:

RandomX_OpenCL.exe --mine --validate [--dataset_host] --intensity 1984

NOTE: 0.01% - 0.012% of failed hashes is normal and expected. The code for FDIV/FSQRT doesn't always produce correct results, but it's much faster than "always correct" version.

To test power usage properly, turn off CPU validation:

RandomX_OpenCL.exe --mine [--dataset_host] --intensity 1984
or
RandomX_OpenCL.exe --mine --portable --workers 8 --bfactor 5 --fast_fp [--dataset_host] --intensity N

The benchmark tests only 1 GPU at a time. If you have more than 1 GPU, you can select GPU to test by adding --device_id N to the command line where N=0,1,2,...