On the use of @sync during benchmarking in the documentation #279
Description
The documentation suggests: you need to ensure the GPU is synchronized at the end of every sample, e.g. by calling synchronize(). However this is generally overkill -- the overhead from @sync can be at the same order of magnitude as the actual cost of the kernel call or even higher which makes the measurement highly inaccurate. I usually end up calling @sync every N kernel calls to mitigate this. Also @benchmark generally gives a good estimate without the sync if you ignore the minimum time and use the median.