Definition of sass__inst_executed*

May I know what is the description of sass__inst_executed_global_loads? Is that a thread-based counter for executed global loads? I didn’t find that neither in the following commands nor in the manual page.

$ nv-nsight-cu-cli --query-metrics | grep sass__inst_executed_global_loads
$ nv-nsight-cu-cli --list-metrics | grep sass__inst_executed

That pretty much means that metric name or name fragment doesn’t exist.

I do find a metric smsp__sass_inst_executed_op_global_ld

It has the description:

“# of warp instructions executed: LDG, LD”

Not sure if that is what you are referring to.

You may get better help with profiler questions like this on the nsight compute forum.

I mistakenly posted to this forum. I appreciate if you or other moderators move that to the nsight compute forum.

Please see the picture below.

It is correct that sass__inst_executed_global_loads (and similar metrics used by e.g. the Memory Workload Analysis tables) are not listed with the --query-metrics feature. The reason is that they are generated by a different provider internally. Similarly, device__attribute_* metrics would not be listed their, either. Most metrics however will be part of the --query-metrics output.

As indicated by the name, the metric is based on SW-patches, i.e. the kernel code is instrumented at runtime to count a specific property, in this case the number of global load instructions. The counter is incremented for every individual executed instruction, i.e. per GPU thread. It is used for the L1/TEX Cache table’s “Global Load Instructions” cell.

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.