nvprof metrics: issue_slot_utilization and stall_other

  1. when you query all possible metrics that can be collected using nvprof, one of them is:
    " issue_slot_utilization: Percentage of issue slots that issued at least one instruction, averaged across all cycles"

what is an issue slot? Is it a hardware component in a scheduler, or is it a synonym for instruction cycle?

2.another metric is:
“stall_other: Percentage of stalls occurring due to miscellaneous reasons”

is there a way to find out what these miscellaneous reasons are? I got 46%, which seems to be really high ( or is it not? ). So it would be great if I can figure out what the other reasons are, and decrease the stalling.

  1. another metric is:
    “stall_exec_dependency - percentage of stalled warps because the input is not ready”
    this is talking about register dependency (i.e. when the input of an instruction is a register, which is the output of a previous instruction ), right?

Thanks

1 Like
  1. it’s approximately a synonym for an instruction cycle, not a hardware component

A slot is an opportunity for a warp scheduler to issue one or 2 instructions.

  1. there is no straightforward tool-based way to figure this out, AFAIK, however if you search around you can find descriptions from Greg at NV on forums like this one, where he has detailed out what those reasons might be for your code.

[url]cuda - What are "Other" Issue Stall Reasons displayed by the Nsight profiler? - Stack Overflow

  1. Yes
1 Like