Hi,
I’m trying to understand the description of available metrics in NCU.
But I couldn’t understand what non-metric names or not directly evaluable mean in the example below.
(Kernel Profiling Guide :: Nsight Compute Documentation )
## non-metric names -- *not* directly evaluable
sm__inst_executed # counter
smsp__average_warp_latency # ratio
sm__throughput # throughput
Could you clarify what evaluable means and why it is not evaluable?
I’m especially interested in smsp__average_warp_latency and related metrics since I want to know warp stall reasons and how much significant it is.
Plus, I wonder why the average latency is represented as a ratio, not cycles or time.
Thanks in advance.
For most metrics, Nsight Compute uses specific naming conventions (as detailed on the documentation page you linked). Metric names are composed of “base names” and “suffixes” and only valid combinations of these are considered valid metric names (there are exceptions, like the --metrics command line parameter, which also accepts base-only names).
As an example, sm__inst_executed would be a base name and you need to complement it with a valid suffix combination to make it a valid metric name. The most simple suffix in this case would probably be .sum, so the fully-qualified name would then be sm__inst_executed.sum.
You can find the valid suffixes for each metric by running the command line
ncu --query-metrics-mode suffix --metrics <metric base name> [--chip <chip>]
I’m especially interested in smsp__average_warp_latency and related metrics since I want to know warp stall reasons and how much significant it is.
I would recommend you to check the metrics included in the following sections:
SchedulerStats
WarpStateStats
SourceCounters
The first two include metrics detailing how warp stalls affect the scheduling and warp latency, averaged over the runtime of the kernel. The latter contains source-correlated metrics collected using sampling which allow you to see the association of stall reasons with source locations and instructions.