How to debug cuda kernels in python, using vscode (linux)

Question

I use cupy to call cuda kernels, but I don't know how to debug cuda code, here is my wrapper file:

wrapper.py

import math
from pathlib import Path

import cupy as cp
import numpy as np

with open(Path(__file__).parents[1] / 'cuda' / 'lcx_projector_kernels.cu', 'r', encoding='utf-8') as f:
    lines = f.read()
    compute_systemG_kernel = cp.RawKernel(lines, 'compute_systemG_kernel')


def compute_systemG_lm(xstart,
                   xend,
                   img_origin,
                   voxsize,
                   sysG,
                   nLORs,
                   img_dim,
                   tofbin_width,
                   sigma_tof,
                   tofcenter_offset,
                   nsigmas,
                   tofbin,
                   threadsperblock):

        compute_systemG_kernel(
            (math.ceil(nLORs / threadsperblock), ), (threadsperblock, ),
            (xstart.ravel(), xend.ravel(),
            cp.asarray(img_origin), cp.asarray(voxsize), sysG,
            np.int64(nLORs), cp.asarray(img_dim),
            np.float32(tofbin_width), cp.asarray(sigma_tof).ravel(),
            cp.asarray(tofcenter_offset).ravel(), np.float32(nsigmas),
            tofbin)
    )

There are lots of ways to debug cuda code with lib file on Internet, but cupy do not use lib file instead of delivering parameters to cuda kernels. Does anybody know how to debug cuda in my situation?

I have try in this way, but breakpoing only stop in python, not in cuda.

launch.json

{
  "version": "0.2.0",
  "configurations": [
    {
            "name": "train.py",
            "type": "debugpy",
            "request": "launch",
            "program": "/home/fanghaodu/code/LMPDnet/train.py",
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceFolder}"
            }
    },

    {
        "name": "CUDA GDB",
        "type": "cuda-gdb",
        "request": "launch",
        "program": "/home/fanghaodu/.conda/envs/cu128/bin/python", // which python
        "args": ["${file}"],
        "debuggerPath": "/opt/apps/cuda-12.8/bin/cuda-gdb", // which cuda-gdb
    },
    
  ],

    "compounds": [
    {
        "name": "Python and CUDA",
        "configurations": ["train.py", "CUDA GDB"]
    }
],
    
}

I've heard that cuda-gdb can debug cuda file but it needs to build lib file first and should use pybind11, I don't want a troublesome approach. Are there any other convenient ways?

halfer · Accepted Answer · 2025-10-05 22:14:50Z

4

Add debug flags when creating RawKernel:

compute_systemG_kernel = cp.RawKernel(
    lines, "compute_systemG_kernel",
    options=("-G", "--generate-line-info")
)

Launch with:
```
cuda-gdb --args python train.py
```

edited Oct 5 at 22:14

halfer

20.2k20 gold badges110 silver badges207 bronze badges

answered Sep 25 at 13:33

Odilkhon

1563 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

How to debug cuda kernels in python, using vscode (linux)

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related