I am looking for measuring the occupancy and register count for my cuda program now. My code was written by fortran and using “pgfortran” for compiling.
I try to use the command “-ta=nvidia -Minfo=accel” but there is no information printed on the screen, so I am wondering is there any other method for the measurement.
Please help on my problem, thanks a lot.
The compiler feedback messages (-Minfo=accel) will give you the occupancy when using the PGI Accelerator Model. However, since you are not getting any messages, this indicates that you don’t have any PGI ACC directives in your code.
For CUDA Fortran, you can determine the Occupancy by first using the “-Mcuda=ptxinfo” flag to get the register, shared memory, and constant memory usage. This information combined with your block and thread dimensions, can then be used in NVIDIA’s Occupancy Calculator Spreadsheet (http://developer.download.nvidia.com/compute/cuda/CUDA_Occupancy_calculator.xls).
Hope this helps,