CUpti_Error 14: cupti_error_not_compatible with cupti to profile metric: flop_count_sp

I got the 14 error when I am trying to profile my GPU P100 metric: flop_count_sp. I’m following the CUPTI document to enable the 3 events needed for flop_count_sp:

352321813 : flop_count_sp_
352321814 : flop_count_sp_
352321815 : flop_count_sp_

I got this 14 error when I use the API cuptiEventGroupSetEnable or cuptiEventGroupEnable. When I am using nvprof query-metrics, I saw flop_count_sp. This means my P100 GPU supports flop_count_sp. However, the 3 events needed did not show up when I use nvprof query-events. Does this mean my P100 does not support the enabling of these 3 events? I could use nvprof --metrics --metrics flop_count_sp Appliaction to get the flop result I need.

In this case, how should I code using c++ to profile flop_count_sp?

Hi,

CUPTI supports profiling the metric flop_count_sp, refer to the table for all supported metrics on GP100 gpu (compute capability 6.0) CUPTI :: CUPTI Documentation. Please note that “flop_count_sp” is a metric, not an event. Are you using the metric APIs to profile it? CUPTI sample callback_metric can be referenced for the API usage.

Thanks for your reply. I think I got this issue resolved by activating the disable CONCURRENT_KERNEL CUPTI activity. Can you let me know how these two conflict with each other? The CONCURRENT_KERNEL activity and flop_count_sp metric.