Profiling Numba CUDA with nsight sys

GHK · July 5, 2022, 12:55pm

Tried profiling this example:

gmarkall/numba-kernel-progress/blob/main/kernel_progress.py

from numba import cuda
from time import sleep
import numpy as np
from progress.bar import Bar


# Timing parameters - adjust if things are too slow / fast for your liking and
# / or setup.

# Number of iterations in the kernel
MAX_VAL = 1000000

# How long the kernel sleeps each iteration
KERNEL_SLEEP = 1000

# How long the host sleeps before checking the progress of the kernel
HOST_SLEEP = 0.001


# A kernel that increments a counter. Simply looping and incrementing may be

This file has been truncated. show original

(which Mr. Markall suggested as an answer to my previous question)

The example runs, nvidia-smi shows GPU activity, but profiling doesn’t show the GPU activity at all only much CPU activity.
Any ideas?

I run on a 7.5 compute capability GPU with this command:

nsys profile -w true -t cuda,nvtx,osrt,cudnn,cublas -s none -o nsight_report -f true -x true python kernel_progress.py

gmarkall · July 5, 2022, 2:04pm

nsys is NSIght Systems, the whole-system profiler. You need to use NSight Compute (ncu) for profiling kernels. See: Nsight Compute CLI :: Nsight Compute Documentation.