Adding a quick python module for pretty NVTX markers to be used inside Python CUDA code.
A compile_ppc64le.sh and compile_x86.sh file has been provided with flags to setup the markers for your usecase.
Shown below is the compile_ppc64le.sh file. The x86 file works in a similar manner.
#!/bin/bash
# Please export CUDA_PATH to give the path to your CUDA installation, typically /usr/local/cuda
NVCC=$CUDA_PATH/bin/nvcc
# Things to note:
# This assumes a conda setup but if you don't have conda, just replace this value with your python include folder
# Default installation path would generally be /usr/include/python3.x without conda or virtualenv
# Make sure that the python library folder is in your $LD_LIBRARY_PATH
PYTHON_INC_PATH=$CONDA_PREFIX/include/python3.6m
# Set this up to the folder you'll be running your python code from.
OUT_FOLDER=''
$NVCC profilerModule.c -I$CUDA_PATH/include -L$CUDA_PATH/lib64 -lnvToolsExt -I$PYTHON_INC_PATH -L/usr/lib/powerpc64le-linux-gnu -lpython3.6m --shared -o $OUT_FOLDER/profilerModule.so
The functions in profilerModule.c provide a simple python wrapper for the CUDA C/C++ API for the nvtx functions.
#!/usr/bin/python
from profilerModule import *
print("Calling profilerStart: ")
profilerStart("Start")
print("\nCalling a range push to mark this part: ")
profilerRangePush("MARKER")
print("\nThis portion shall be marked")
profilerRangePop("MARKER")
print("\nCalling profilerStop: ")
profilerStop("Stop")