vector-ta 0.1.3

High-performance technical analysis indicators with optional SIMD/CUDA and language bindings.

vector-ta

VectorTA is a Rust crate of technical analysis indicators focused on speed and predictable allocations, with optional SIMD/CUDA acceleration and optional Python/WASM bindings.

Full documentation (indicator list, API reference, and guides): https://vectoralpha.dev/projects/ta

The CUDA bindings are predominantly only worth using if used in a VRAM-resident workflow. For example, It is possible to achieve a benchmark timing of 6.08 ms for 250 million calculated ALMA indicator data points on an RTX 4090, whereas the CPU (AMD 9950X) AVX-512, AVX2, and scalar timings are approximately 140.61 ms, 188.64 ms, and 386.20 ms, respectively.

The Tauri backtest optimization demo application using this library can achieve 58300 backtests for a double ALMA crossover strategy over 200k data points in only 98.31 milliseconds on the same hardware (RTX 4090 + AMD 9950X).

Rust usage

Example: ADX over HLC slices

use vector_ta::indicators::adx::{adx, AdxInput, AdxParams};

fn compute_adx(
    high: &[f64],
    low: &[f64],
    close: &[f64],
) -> Result<Vec<f64>, Box<dyn std::error::Error>> {
    let input = AdxInput::from_slices(high, low, close, AdxParams { period: Some(14) });
    Ok(adx(&input)?.values)
}

Features

• cuda: GPU acceleration using prebuilt PTX (compute_89) shipped in the crate.
• cuda-build-ptx: You can compile PTX from kernels/cuda/** using nvcc.
• nightly-avx: Runtime-selected AVX2/AVX512 kernels on x86_64 (nightly required).
• python: PyO3 bindings (build from source via maturin).
• wasm: wasm-bindgen bindings (build from source via wasm-pack).

Python (optional)

Build + install into a virtualenv:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip maturin numpy
maturin develop --release --features python

WASM (optional)

Build with wasm-pack:

rustup target add wasm32-unknown-unknown
wasm-pack build --target nodejs --release --features wasm

CUDA (optional)

Enable:

[dependencies]
vector-ta = { version = "0.1.3", features = ["cuda"] }

Notes:

• To force-disable CUDA probing/usage (tests/CI): set CUDA_FORCE_SKIP=1.
• To override where prebuilt PTX is sourced from, set VECTOR_TA_PREBUILT_PTX_DIR (see docs link above).

Nsight Compute profiling (CUDA)

Profile one indicator benchmark with Nsight Compute and export report + summary artifacts:

pwsh -NoProfile -File scripts/profile_cuda_indicator_ncu.ps1 -Indicator rsi -Scenario batch -NcuSet full

Useful flags:

• -ListOnly: list available benchmark IDs for a given indicator.
• -BenchId <id>: profile an exact Criterion benchmark ID instead of auto-picking one.
• -KernelRegex <regex>: override kernel filter (default: .*<indicator>.*).

If you see ERR_NVGPUCTRPERM, enable GPU performance counters in NVIDIA Control Panel (Developer -> Manage GPU Performance Counters) or run with sufficient privileges, then rerun.

License

Apache-2.0 (see LICENSE).