🚀 Nsight Compute Profiling: Custom PyTorch CUDA Kernel

👨‍💻 Author: Dartayous Hunter

AI Engineer & Visual Effects Technologist
Blending low-level performance with high-level visual storytelling

“I profiled my own PyTorch CUDA extension in Nsight Compute, resolved permission-level access to counters, mapped register usage and warp efficiency, and benchmarked my kernel against native ops — then made it faster.”

🎯 Overview

This project benchmarks and profiles a custom CUDA kernel wrapped as a PyTorch extension, using NVIDIA Nsight Compute for full GPU telemetry access. From kernel build to warp execution analysis, this pipeline unlocks performance visibility and measurable speedups — proving that GPU engineering is as cinematic as it is scientific.

---

🧩 Project Structure

nsight_cuda_profiling/
├── vector_add/                # PyTorch CUDA extension
│   ├── vector_add.cpp
│   ├── vector_add_kernel.cu
│   ├── setup.py
│   └── test_vector_add.py
├── run_vector_add.bat         # Launch script (activates venv)
├── profiling_notes/           # Nsight report interpretation
│   └── warp_efficiency.md
├── infographic_guide.md       # Annotated storyboard flow
└── README.md                  # You’re reading it!

⚙️ Tech Stack

• Python 3.11 (torch311 virtual env)

• PyTorch (CUDA 12.1 enabled)

• Nsight Compute GUI

• C++ / CUDA kernel (.cu + .cpp)

• Bat file launcher for controlled execution

• Registry patch for counter access (PerfCounterAccess)

• Windows OS

📈 Performance Highlights

• ✅ Custom kernel: vector_add(a, b, out, size)

• 🔬 Validated with torch.allclose() in Python

• ⚡ Speedup: 4.4x faster than PyTorch native ops

• 🧠 Nsight analysis: warp execution, register pressure, memory stalls

🧠 Steps to Reproduce

1.) Compile extension:

python setup.py install

2.) Benchmark kernel:

python test_vector_add.py

3.) Launch in Nsight Compute:

• Executable: C:\Windows\System32\cmd.exe

• Arguments: /C run_vector_add.bat --target-processes all

• Output: vector_add_report.ncu-rep

4.) Enable GPU performance counters: Run as Admin:

reg add "HKLM\SOFTWARE\NVIDIA Corporation\Global\PerfCounterAccess" /v Enable /t REG_DWORD /d 1 /f

5.) Reboot and relaunch Nsight.

---

🧼 Repository Hygiene

This project includes a .gitignore configured for Python, CUDA profiling, and modular virtual environments to keep the repository clean and focused:

🐍 Python Environment and Build Artifacts

pycache/ *.py[cod] *.egg *.egg-info/ dist/ build/ *.spec

📓 Jupyter Notebook Checkpoints

.ipynb_checkpoints/ -checkpoint.

📈 Nsight Compute/CUDA Profiler Files

*.ncu-rep *.ncu-proj

⚙️ System and IDE Artifacts

*.exe *.dll *.obj *.log *.tmp .vscode/ *.DS_Store

🧳 Virtual Environments (Global or Local)

Venv/ venv/ ENV/ env/ .venv/

📦 Python Egg Packaging Metadata

vector_add.egg-info/

---

GitHub Push Workflow

# Initialize Git if you haven’t yet
git init

# Stage your project files
git add .

# Commit your changes with a descriptive message
git commit -m "Initial commit — project setup and .gitignore configured"

# Link to your GitHub repository (replace with your repo URL)
git remote add origin https://github.com/your-username/your-repo.git

# Push to main branch
git push -u origin main

📊 Visualization Assets

See infographic_guide.md for a breakdown of the visual storyboard. Includes: Build → Benchmark → Launch → Unlock → Profile

✨ About the Author

Dartayous is a seasoned VFX compositing artist turned AI engineer with a passion for GPU performance and storytelling. This repository reflects a crossover journey: where cinematic precision meets CUDA kernel mastery.