[gigony]

Need to rebase once #141 is merged.

• Instrumented important methods in cuCIM with NVTX so can see performance bottlenecks easily with NSight systems.
• Tracing can be enabled through config file or environment variable or through API and less than 1% performance overheads in normal execution.

Enabling Tracing

Through .cucim.json file

{
        "profiler" : { "trace": true }
}

Through Environment variable

CUCIM_TRACE=1 python

Through API

from cucim import CuImage

CuImage.profiler(trace=True)
#or
CuImage.profiler().trace(True)

CuImage.profiler().config
# {'trace': True}
CuImage.profiler().trace()
# True
CuImage.is_trace_enabled # this is simpler method.
# True

Profiling with NVIDIA Nsight Systems

nsys profile -f true -t cuda,nvtx,osrt -s cpu -x true --trace-fork-before-exec true -o my_profile `which python` benchmark.py
# can add `--stats true`

Then, execute nsight-sys to open the profile results (my_profile.qdrep).

closes #137

[jakirkham]

PR ( #141 ) is in. So this is ready for rebasing

[gigony]

PR ( #141 ) is in. So this is ready for rebasing

@jakirkham
Please help merge this when you have time.

[jakirkham]

Should we raise an issue to track this?

[gigony]

Thanks @jakirkham!

Created #168 and updated the comment in nvtx3.cmake.

[jakirkham]

Had a couple questions above. Otherwise LGTM

[ajschmidt8]

CI is passing. Merging.

[jakirkham]

Thanks AJ! 😄