Invocations to __syncthreads() and barrier() in calcMergeEneGra{.cu/.cl}

[L30nardoSV]

Hi,

By comparing the CUDA and OpenCL versions of gpu_calc_energrad() (develop branch), I found the following invocations to __syncthreads() do not have their barrier() counterparts:

• https://github.com/ccsb-scripps/AutoDock-GPU/blob/develop/cuda/calcMergeEneGra.cu#L414
• https://github.com/ccsb-scripps/AutoDock-GPU/blob/develop/cuda/calcMergeEneGra.cu#L696
• https://github.com/ccsb-scripps/AutoDock-GPU/blob/develop/cuda/calcMergeEneGra.cu#L816

It seems to me these __syncthreads() invocations are not required, but I might be missing something. Also, running few tests after removing such __syncthreads() results in virtually no performance benefits.

Could you please check and comment on that?

[atillack]

@L30nardoSV Those synchronization points (along with thread fences - which i did remove after thorough testing) where put in there during the later stages of bug hunting and since they don't hurt performance I didn't mind leaving them in.

So yes, you are probably correct but I might also be missing something ;-)

Edit: Also, I don't think a corresponding barrier is needed on the OpenCL side ...