CUPTI

Setup

Install some dependencies

sudo apt install libnuma-dev libboost-all-dev

Install CUDA and CUDNN.

...

Modify env.sh to point to the right libraries.

Build the profiling library (prof.so).

make

Run on a CUDA application

Make sure your CUDA application is not statically-linked, which is the default when you are building your own CUDA code.

This will record data by appending to an output.cprof file, so usually remove that file first. ./env.sh sets up the LD_PRELOAD environment and invokes your app.

rm -f output.cprof
./env.sh <your app>

Do something with the result:

cprof2<something>.py

Other info

env.sh sets LD_PRELOAD to load the profiling library and its dependences.