I am trying to profile some kernels, using Cuda profiling tools. I am using cuda 3.0, with a C1060 nVidia GPU card (GT 280), so 1.3 capable. I have several questions:
[*]From cudaparof, I do not have access to all counters. Some of the missing counters are, for example gld_request, gld_coalesced and gld_uncoalesced. On the other hand, gld32/64/128b are present. I can use all the counters with the command line method (with a CUDA_PROFILE_CONFIG). Is it normal?
[*]Do you some pointers to explain how are count the gld. Naively, I though gld_request=gld_uncoalesced+gld_coalesced. This is not at all!
Thanks