Why don't I get "thread_inst_executed"

Question

When I list nvprof's metrics with

nvprof --query-events

I see:

thread_inst_executed: Number of instructions executed by the active threads. For each instruction it increments by number of threads, including predicated-off threads, that execute the instruction. It does not include replays.

I would like to use this metric, so I collect metrics using:

nvprof --csv --metrics thread_inst_executed,inst_executed,inst_executed_global_loads,inst_executed_global_stores,inst_executed_local_loads,inst_executed_local_stores,inst_executed_shared_loads,inst_executed_shared_stores,gld_transactions,gst_transactions,local_load_transactions,local_store_transactions,shared_load_transactions,shared_store_transactions,l2_read_transactions,l2_write_transactions,dram_read_transactions,dram_write_transactions,sysmem_read_transactions,sysmem_write_transactions ./my_program my arguments

The output has every metric I asked for... except thread_inst_executed. Why is it missing? How can I get it?

Answer 1

When I list nvprof's metrics with

nvprof --query-events

That isn't consistent usage (emphasis added).

Using nvprof (or nvvp), events and metrics are not the same thing.

To query events, you would use:

--query-events

To query metrics, you would use:

--query-metrics

To profile, asking for an event measurement, you would use

--events name_of_event,...

To profile, asking for a metric measurement, you would use

--metrics name_of_metric,...

If you do something like this:

--metrics name_of_event,...

or

--events name_of_metric,...

I don't know what the behavior is, but I would not expect it to work properly.