Extract execution time from the report file *.qdstrm or *.qdrep

Development Tools Nsight Systems Profiling x86 Linux Targets
#1

I profiled my application with nsys and got the .qdstrm report file. By following the instruction on “Using QdstrmImporter” I converted to “qdrep” format and opened it “NIVIDA Nsight Systems 2018.3.1”. Everything works smoothly and it is a wonderful tool to visualize application process.

However, my problem is that I want to access the report file (report1.qdstrm or report1.qdrep) to extract average execution time of some functions. I cannot not find such option in the graphical tools in “NIVIDA Nsight Systems 2018.3.1”. I cannot find any method to export the report file to text format neither. I find a similar thing called NVTXT, but only found the command to convert the NVTXT to this binary format.

It looks like that the data file is encoded using Google Protobuf. Could you please suggest any method to parse the report file programmatically?

#2

Unfortunately we do not currently have a way to programmatically extract information from the .qdrep file yet. We are currently working on a design for extracting data, looking to have this in a version of the tool shipping in Q1 or so of next year, but nothing right now.

Depending on what you want to use the info for, you may be able to approximate by using the function table (the sampling data at the bottom of the screen) and the residency of the function on the call stack. You can filter the timeline and that will also filter what shows in that function table. So you can compare residencies in different parts of your analysis run.

Sorry we don’t have a API for you to get the data right now.

#3

Hi hwilper, thank you so much for your clarification and suggestions! I look forward to see this new feature next year.

#4

Hi hwilper, i’ve came across this post now (december 2020), and I need such feature.
I have some long profiling result inside qdrep file and I want to programmically analysis it using some programming, to check some stats (min / max / avg / std etc.).

Any chance this feature exists already in some API ?