pyx++

A collection of low-level Python performance libraries that push CPython beyond its defaults. This is not meant to be used in production, this is for learning purposes.

---

Packages

sharedx

True CPU parallelism for Python using CPython sub-interpreters — async-style API, zero-copy shared memory, no pickle overhead. Requires Python 3.12+.

Single Core:      1.2731s
Sharedx (4 workers):  0.1980s   →  6.4x speedup

jitctl

Hints-based developer control over CPython's JIT compiler via decorators (@jit_focus, @jit_ignore, @jit_once). Targets Python 3.13+. Currently in design/planning phase — see jitctl/PLANNING.md.

---

Install

Each package is installed independently.

# sharedx
pip install git+https://github.com/ado11231/pyx-plus-plus.git#subdirectory=sharedx

# or clone and install locally
git clone https://github.com/ado11231/pyx-plus-plus.git
cd pyx-plus-plus/sharedx
pip install -e .

---

Quick start (sharedx)

import numpy as np
from sharedx.pool import InterpreterPool
from sharedx.dispatcher import Dispatcher

# All imports must be inside the function — sub-interpreters start blank
def process(data):
    import numpy as np
    return np.sqrt(np.abs(data * 1.1 + data / 1.5))

chunks = [np.random.randint(0, 255, (256, 256, 3), dtype=np.uint8) for _ in range(100)]

with InterpreterPool(workers=4) as pool:
    dispatcher = Dispatcher(pool)
    futures = [dispatcher.submit(process, (chunk,)) for chunk in chunks]
    results = [f.result() for f in futures]
    dispatcher.shutdown()

The one rule: every function you submit must be self-contained — all imports must live inside the function body. Sub-interpreters start with no imports or variables from your main program.

---

How it works

Your code
    ↓
Dispatcher        creates a Future, submits task to thread pool
    ↓
InterpreterPool   manages N sub-interpreters, each with its own GIL
    ↓
SharedArray       numpy arrays live in shared memory — zero copy
    ↓
Sub-interpreter   runs your function truly in parallel on its own core
    ↓
Future            result comes back, no pickling needed