srvm / cupti_profiler

CUPTI GPU Profiler

CUDA Profiling Library

This library provides an API for collecting CUDA profiling metrics and events from within a CUDA application. Programmers specify what metrics and events they want, and start the profiler before calling one or more CUDA kernels. The library sets up the appropriate CUPTI callbacks, calculates the number of kernel passes required, gathers values for the specified metrics and events, and returns them to the programmer on a per-kernel basis.

Example Usage:

  vector<string> event_names {
                              "active_warps",
                              "gst_inst_32bit",
                              "active_cycles"
                             };
  vector<string> metric_names {
                               "flop_count_dp",
                               "flop_count_sp",
                               "inst_executed"
                              };

  cupti_profiler::profiler profiler(event_names, metric_names);

  // Get #passes required to compute all metrics and events
  const int passes = profiler.get_passes();

  profiler.start();
  for(int i=0; i<passes; ++i) {
    call_kernel(data);
  }
  profiler.stop();

  printf("Event Trace\n");
  profiler.print_event_values(std::cout);
  printf("Metric Trace\n");
  profiler.print_metric_values(std::cout);

Output:

Event Trace
_Z6kernelIPfEvT_i: (active_warps,1734) (gst_inst_32bit,100) (active_cycles,423) 
_Z7kernel2IPfEvT_i: (active_warps,865) (gst_inst_32bit,50) (active_cycles,418) 

Metric Trace
_Z6kernelIPfEvT_i: (flop_count_dp,0) (flop_count_sp,100) (inst_executed,52) 
_Z7kernel2IPfEvT_i: (flop_count_dp,0) (flop_count_sp,50) (inst_executed,26)