nvmetrics

Introduction

nvmetrics is a library designed to facilitate the measurement of GPU metrics using NVIDIA CUPTI (CUDA Profiling Tools Interface). While tools like NCU (NVIDIA Command-Line Profiler) and GUI-based profilers exist, nvmetrics offers a programmatic approach to GPU profiling, enabling seamless integration into custom workflows and applications.

The core of this library is based on this code. The main changes and additions are:

• Some refactoring (without altering the behavior of the code)
• Added linting using pre-commit
• Added CMake to build this code as a library
• Added Python interface using pybind11
• Added C++ and Python demo program to illustrate the use of this library

Features

• Provides a simple C++ interface for measuring GPU metrics.
• Enables programmatic access to NVIDIA CUPTI for GPU profiling.
• Includes a Python interface using pybind11 for easy integration into Python projects.

Installation

Requirements

• NVIDIA GPU with CUDA support
• CUDA Toolkit
• pybind11 (for Python interface)

Build from source

1. Clone this repository:

git clone https://git.astron.nl/RD/recruit/nvmetrics.git

2. Navigate to the repository directory:

cd nvmetrics

3. Build the library:

cmake -S. -Bbuild
make -C build

4. (Optional) Install the library:

make -C build install

Usage

C++

#include <vector>
#include <string>
#include "nvmetrics.h"

using namespace nvmetrics;

int main() {
    std::vector<std::string> metrics = {"metric1", "metric2", "metric3"};
    
    // Start measuring metrics
    measureMetricsStart(metrics);
    
    // Perform GPU operations
    
    // Stop measuring metrics and retrieve results
    std::vector<double> results = measureMetricsStop();
    
    // Process results
    
    return 0;
}

Python

import nvmetrics

# Define metrics to measure
metrics = ["metric1", "metric2", "metric3"]

# Start measuring metrics
nvmetrics.measureMetricsStart(metrics)

# Perform GPU operations

# Stop measuring metrics and retrieve results
results = nvmetrics.measureMetricsStop()

# Process results