I’m using nvvp in CUDA Toolkit 4.2. When I am profiling a GPU kernel, I get negative “Global Store Efficiency” on the kernel displayed in the “Properties” tab, e.g. Global Store Efficiency -36.2%. I expect this number should be within the range of 0% and 100%. Does anyone have similar problems like this? Is it normal?
Like you, I would expect this number to be between 0% and 100%, so this looks like a bug to me. If this happens with CUDA 4.2, I would suggest filing a bug against the profiler, attaching a self-contained repro program that demonstrates the problem. Thank you for your help.