128/256 threads/block

[ancahamuraru]

Update the OpenCL kernel for 128/256 threads/block based on the equivalent CUDA kernel - see commit f2b9db2 from the main Gromacs master branch: gromacs@f2b9db2

Evaluate the performance of the new kernel for AMD and NVIDIA GPUs and decide on the final version or versions of the OpenCL kernel that will be used.

[pszi1ard]

As this will increase register pressure, I suggest trying 128 threads/block too. Additionally, reduction will become tricky without the lane-shuffle ops.

[ancahamuraru]

Thanks for the comment. It's my mistake, I forgot to mention 128 threads/block.
The issue title and description are now updated.