TaCuda — Python Bindings

Python package for the native tacuda CUDA library. TA-Lib 0.7.1 baseline coverage (161 functions): indicators, MACDEXT, 19 vector math kernels, candlestick patterns, plus typed core APIs with runtime dtype selection and the zero-copy device API.

Prerequisites

• NVIDIA GPU (Compute Capability >= 6.0)
• CUDA Toolkit 11.x or 12.x
• Built tacuda shared library (tacuda.dll / libtacuda.so / libtacuda.dylib)
• Python >= 3.8, NumPy >= 1.20

Package Scope

• High-level NumPy wrappers for indicator functions
• Precision-aware core indicators (sma, ema, rsi, macd, atr, bbands) via dtype="float16|float32|float64"
• OHLCV convenience container and input validation helpers

Installation

pip install tacuda

Note: The tacuda native library must be built separately and placed where the Python bindings can find it. See "Library Location" below.

Quick Start

import numpy as np
from tacuda import sma, ema, rsi, macd, bbands, OHLCV

# Generate sample data
prices = np.random.randn(100_000).cumsum().astype(np.float32) + 100

# Moving averages
sma20 = sma(prices, 20)
ema12 = ema(prices, 12)

# RSI
rsi14 = rsi(prices, 14)

# MACD (returns macd_line, signal, histogram)
macd_line, signal, hist = macd(prices, 12, 26, 9)

# Bollinger Bands (returns upper, middle, lower)
upper, middle, lower = bbands(prices, 20, 2.0, 2.0)

# OHLCV container
ohlcv = OHLCV.from_columns(open_arr, high_arr, low_arr, close_arr, volume_arr)

MACDEXT and vector math

import tacuda

# Independent MA types per line (0=SMA 1=EMA 2=WMA ... 8=T3)
ext_macd, ext_signal, ext_hist = tacuda.macdext(
    prices, fastPeriod=12, fastType=1, slowPeriod=26, slowType=0,
    signalPeriod=9, signalType=2)

# Element-wise GPU kernels
log_prices = tacuda.ln(prices)
spread = tacuda.sub(high_arr, low_arr)

Zero-copy device pipeline

n = len(prices)
d_in = tacuda.device_alloc(n)
d_out = tacuda.device_alloc(n)
try:
    tacuda.device_upload(prices, d_in, n)   # 1 PCIe transfer in
    tacuda.sma_d(d_in, d_out, n, 20)        # compute on GPU
    tacuda.sin_d(d_out, d_out, n)           # chain kernels, zero copies
    tacuda.device_sync()
    result = tacuda.device_download(d_out, n)  # 1 PCIe transfer out
finally:
    tacuda.device_free(d_in)
    tacuda.device_free(d_out)

Available Indicators (TA-Lib 0.7.1 baseline coverage)

Library Location

The bindings search for the native library in this order:

1. TACUDA_LIBRARY environment variable (full path to shared library)
2. TACUDA_LIBRARY_PATH / TACUDA_LIBRARY_DIR + library name
3. System library path (ctypes.util.find_library)
4. Common build directories relative to the package

Regenerating Metadata

When the C header changes, regenerate from the repository root:

python bindings/generate_bindings.py

Build & Publish (PyPI)

Releases are built and published only by the release-packages workflow (see CONTRIBUTING.md): one platform wheel per native plus a source-only sdist. For a local build, make the sdist before any native is in tacuda/ (setuptools packs package data into it), then put exactly one native there and build only the wheel. Wipe the setuptools cache before each build: build/lib is never pruned, so a second wheel would otherwise also carry the previous build's native:

cd bindings/python
rm -rf build dist ./*.egg-info
python -m build --sdist          # tacuda/ must not contain a native yet
rm -rf build ./*.egg-info
cp /path/to/libtacuda.so tacuda/ # or tacuda.dll for a Windows wheel
python -m build --wheel
python -m twine check dist/*

Changelog

Canonical release history is maintained in CHANGELOG.md.

License

Apache License 2.0 — see LICENSE.

Streaming and version pairing

The 0.8.0 bindings add incremental single-series and portfolio updates for 10 indicator families. Use a matching 0.8.0 native library; an older binary does not export the streaming functions. See the Streaming API guide. The 161-function compatibility baseline is native TA-Lib 0.7.1, not TA-Lib 0.8.1.