3

I'm using nvprof to profile the CUDA code. Can someone help me how to calculate the compute to global memory access ratio using the information obtained through performing

nvprof --analysis-metrics

user3176594
  • 41
  • 2
  • 3
    You would have to define what you mean by "compute". And `--analysis-metrics` is intended to create a file for import into `nvvp`. Is that your intent? You can probably get started just by doing something like `nvprof --metrics flop_count_sp,gld_transactions,gst_transactions ./myapp`. The gld/gst data would need to be multiplied by 128 to get bytes. And then divide the flop count by the bytes, to get the ratio of single precision flops per byte (for this example). I think you may need to be more specific about what you want to calculate. – Robert Crovella Sep 20 '14 at 22:22
  • I've used it and the question in detail is given here [link](http://stackoverflow.com/questions/25956891/using-cuda-profiler-nvprof-for-memory-accesses) – user3176594 Sep 21 '14 at 07:32
  • 1
    We don't need a new question per each improvement. Questions and answers on StackOverflow is not for your own benefit, but for the benefit of the whole community. Keeping the questions clear and ordered is a sted towards this end. Please, delete this one it it is redundant. Thanks. – Vitality Sep 21 '14 at 18:33

0 Answers0