global memory load efficiency - profiling

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello , I wanted to ask

If I use in my code constant for values that I use as arguments in the kernels , and profile with :

nvprof --metrics gld_efficiency

I am receiving (for 2 kernels )

Global Memory Load Efficiency       0.00%

If I ommit the constant , I am receiving:

kernel1     37.50%      37.50%      37.50%
kernel2     50.00%      50.00%      50.00%

So, use constant or not? I don’t understand.

Also , what is considered a goos gld efficiency.And a bad gld efficiency shows uncoalesced access?
Or , how can I find with nvprof if I have coalesced access?

Thanks!