Skip to content

Programming / vision / nvidia

Download NVIDIA Nsight Systems

Nvidia profiling tools

nsys

1
sudo apt install nvidia-nsight-systems

TODO: check

1
nsys profile -o output python3 your_vpi_script.py

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
#!/usr/bin/env python3
import vpi
import numpy as np
import time

# Maximum performance on Jetson
import subprocess
subprocess.run("sudo jetson_clocks", shell=True)

img = np.random.randint(0, 255, (1080, 1920, 3), dtype=np.uint8)

start = time.time()
with vpi.Backend.CUDA:
    frame = vpi.asimage(img, vpi.Format.BGR8).convert(vpi.Format.U8)
    corners, scores = frame.harriscorners()

# This causes GPU→CPU transfer (shows in profiler)
corners_cpu = corners.cpu()
print(f"Time: {(time.time()-start)*1000:.2f}ms")

1
nsys profile -t cuda --stats=true python3 test.py
  • [CUDA memcpy HtoD] = CPU→GPU transfers (minimize these)
  • [CUDA memcpy DtoH] = GPU→CPU transfers (minimize these)
  • [CUDA kernel] = VPI operations on GPU