gpu_lob - CUDA/HIP-Accelerated Python Library

A high-performance CUDA/HIP-accelerated Python library for limit order book processing. This project was made to work on Windows, do not try to build it on WSL or any Linux distribution. The goal of this project was to provide support for heavy backtesting of trading algorithms.

Prerequisites

• NVIDIA GPU with CUDA support
• AMD GPU with HIP support
• CUDA Toolkit (11.0+)
• ROCm (6.4)
• Visual Studio (2019 or 2022) C++ Build Tools
• Python 3.8+
• CMake 3.18+

Build and Install

1. Create and Activate Virtual Environment

python -m venv venv
.\venv\Scripts\Activate.ps1

if you get an error about execution policy:

Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
.\venv\Scripts\Activate.ps1

2. Upgrade pip and Install Build Tools

pip install --upgrade pip
pip install scikit-build-core pybind11 numpy

3. Build for Your Platform

For NVIDIA GPUs (CUDA)

Windows/Linux:

cp CMakeLists.txt.cuda_backup CMakeLists.txt
pip install -e .

For AMD GPUs (HIP) - Windows

Windows:

.\build_windows.ps1

Quick Start (Usage Example)

Once installed, you can use the library from Python:

import gpu_lob
import numpy as np

MAX_PRICE = 1000
bids = np.zeros(MAX_PRICE, dtype=np.int32)
asks = np.zeros(MAX_PRICE, dtype=np.int32)

prices = np.array([100, 105], dtype=np.int32)
qtys = np.array([1000, 500], dtype=np.int32)
sides = np.array([gpu_lob.SIDE_BUY, gpu_lob.SIDE_SELL], dtype=np.int32)

gpu_lob.add_liquidity(bids, asks, prices, qtys, sides)

order_prices = np.array([105], dtype=np.int32)
order_qtys = np.array([200], dtype=np.int32)
order_sides = np.array([gpu_lob.SIDE_BUY], dtype=np.int32)
order_ids = np.array([1], dtype=np.int32)

matches = gpu_lob.match_orders(bids, asks, order_prices, order_qtys, order_sides, order_ids)

print(f"Matched: {matches[0]} units")
print(f"Remaining Ask at $105: {asks[105]}")

4. Install in Editable Mode

If you are modifying the C++ or CUDA code, use:

pip install -e . -v --no-build-isolation

Run Benchmarks

python test_lob.py

Check the generated benchmark_plot.png for performance results.