[VTune][CLI] Export GPU Hotspots Results from CLI?

[GlowingScrewdriver]

I've been using VTune to profile kernels on a GPU. I'm able to use the VTune CLI to collect profiling info for GPU Hotspots Analysys from my application, like so:

$ vtune -collect gpu-hotspots -- application args ...

I can subsequently open and view those results in the VTune GUI.

However, I'd like to export those results using the CLI. I am aware of the `vtune -report` subcommand; however, it doesn't seem to be able to handle GPU hotspots analysis:

$ vtune -report gpu-hotspots -r r000gh/ -report-output out.csv -format csv
> vtune: Error: Cannot find report `gpu-hotspots': check report name spelling using -help report.

$ vtune -help report
> Available Reports:
>  
>     affinity                 Display binding of a thread to a range of sockets, physical, and logical cores (affinity).
>     callstacks               Display CPU or wait time for callstacks.
>     exec-query               
>     gprof-cc                 Display CPU or wait time in the gprof-like format.
>     hotspots                 Display detailed view by default or selected groping (e.g. functions) with a column set corresponding to GUI Bottom-Up pane.
>     hw-events                Display hardware events.
>     platform-power-analysis  Display CPU sleep time, wake-up reasons and CPU frequency scaling time.
>     summary                  Display data about overall performance.
>     timeline                 Display data over time.
>     top-down                 Display a call tree for your target application and provide CPU and wait time for each function.
>     vectspots                Display statistics that helps identify code regions for tracing on a HW simulator.

Note: I've prefixed command lines with '$' and output lines with '>'.

How can I export my GPU Hotspots Analysis results to CSV (or some other form, if CSV isn't possible) from the CLI?

Help would be appreciated. TIA!

[yuzhang3_intel]

You can try the command line below to get report.

vtune -report exec-query -rep-knob row-by="/GPUComputeTaskType" -sort-desc "Total Time:Self" -rep-knob column-by="ViewpointGUIandCLIColumns" -r <result_dir>

[Svetlana_K_Intel · Accepted solution]

You can try `vtune -r <result-dir> -R hotspots -group-by=gpu-adapter,gpu-stack,computing-task` command  if your intention is to get the list of kernels with all the HW metrics attributed to them. If you have just one gpu on board and it's not PVC - 'gpu-adapter,gpu-stack' can be omitted.

[GlowingScrewdriver]

Thank you for your replies. I will try these suggestions.

[GlowingScrewdriver]

I have a couple of other questions which I believe are related to those in my initial post. I am unable to export the GPU clock speed to the CSV report, although it is captured during profiling (I am able to see it on the VTune GUI). And, I'd like to get readings of the DRAM frequency. How can I achieve these two things?

Note: Please do let me know if this post belongs in a new thread rather than here.

Thanks

[Svetlana_K_Intel]

@GlowingScrewdriver  you can use this command line to get average GPU frequency per kernel:

vtune -R exec-query -r <result_dir> -rep-knob column-by="/GPUAvgGpuCoreFrequencyMHzMetric" -rep-knob row-by="/GPUAdapter/GPUTileName/GPUComputeTask"

 As for the DRAM frequency, I'm not so sure VTune collects anything like that...

[pratchet]

Hi, this was really helpful information, with gpu-hotspots, how do i get LLC Cache Hit counters? Or in the case of PVC L2 Cache HIt rate. I tried overview, full-compute, global-memory-access but i'm not sure where the metric for Cache hits are in the report or maybe i'm grouping them incorrectly. Any advice would be helpful. Thank you!