CODEWITHSUNDEEP
cudanvidiaprofiler
warunapww
warunapww

Reputation: 1006

Understanding nvprof event domains

I'm trying to understand the difference between different event domains used in nvprof.

when I try 

nvprof --devices 0 --query-events

It lists the available events categorized into several domains (domain_a, domain_b, domain_c, domain_d, etc...)

I have read that depending on the domain, it varies whether an event is counted for all the SMs or few of them. But I could not find out events in which domain is counted for all, etc...

It would be great, if someone can explain the domains.

Upvotes: 1

Views: 553

Answers (1)

Swapna
Swapna

Reputation: 101

In nvprof, by default the events are profiled for all instances that can be profiled and the data is extrapolated for all available instances. If all available instances cannot be profiled due to hardware limitation then nvprof gives following warning:

==14882== Warning: The following aggregate event values were extrapolated from limited profile data and may therefore be inaccurate. To see the non-aggregate event values, use "--aggregate-mode off".

If all the available instances can be profiled, then you will not see this message.

You can give --aggregate-mode off option before --events to profile events for all instances of a domain. eg:

./nvprof --aggregate-mode off --events warps_launched ./matrixmul

In future release, information about profiled instances and available instances for a domain will be provided in --query-events option.

Upvotes: 1

Related Questions