CUDA Profiler Feature Requests Nice to have...

– Save each “Summary Comparison” plot to an image file.

– Pop-up help for obtuse columns headers such as “gst 64b”

– Save table data to XML for flexible formating via XSLT (e.g. docbook)

– Annotate each project and each session with text.

For me it feels like the profiler gets worse and worse the more modern the GPU is, for compute capability 1.1 it is possible to see the number of coalesced and uncoalesced reads but not for more modern cards. For 1.3 cards it is possible to see gst32, gld64 etc but not for 2.0 cards. With my new Fermi card I can only see number of load and store requests. It also seems like the texture cache hits and misses are gone for 2.0 cards…

Why not keep all the counters and let the user choose which ones to use?

Another problem is that my occupancy always is reported as 0 or 1 (it never reports floating point values), I guess that’s a bug.