fasthardware 2.1.0

A zero-configuration hardware acceleration library for Python.

⚡ fasthardware

A zero-configuration hardware acceleration library for Python.
Squeeze every last FPS out of your system with a single function call — covering OS scheduling, CPU threading, GPU inference, and memory management.

---

🚀 Installation

pip install fasthardware

---

🔥 Quick Start

from fasthardware import fasthardware

# Call once at startup — automatically detects and optimizes your hardware
fasthardware.speedup()

# Your code here
...

# Periodically sweep RAM fragmentation (recommended every 60s)
fasthardware.manual_sweep()

---

🧠 What does it do?

fasthardware automatically detects your hardware and applies the most effective low-level optimizations — no manual configuration needed.

speedup(mode="DEFAULT")

Layer	Optimization	Description
OS	Process Priority Elevation	Calls kernel APIs directly to set process to HIGH priority
CPU	C Library Thread Binding	Sets OMP / MKL / OPENBLAS / NUMEXPR threads to max cores
GC	Stutter Prevention	Disables Python GC to eliminate random frame drops
CUDA	cuDNN Benchmark + TF32	Auto-selects optimal kernels; enables TF32 on Ampere+ GPUs
OpenVINO	iGPU LATENCY hint + Model Cache	Minimizes inference latency; caches compiled models to disk
OpenCV	SIMD + OpenCL	Forces AVX/SSE acceleration and OpenCL parallel backend

Only installed libraries are optimized — missing ones are silently skipped.

ULTIMATE mode

fasthardware.speedup(mode="ULTIMATE")

Additionally detects all child processes (multiprocessing workers) and:

• Elevates their OS priority
• Pins each worker to a dedicated CPU core for zero cache contention

---

get_core() — OpenVINO Core Singleton

After calling speedup(), fasthardware holds an optimized OpenVINO Core instance internally.
Instead of creating your own ov.Core(), reuse this to inherit all applied optimizations:

from fasthardware import fasthardware
import openvino as ov

fasthardware.speedup()

# ✅ Reuse the pre-optimized Core — LATENCY hint + model cache already applied
core = fasthardware.get_core()

compiled_model = core.compile_model("model.xml", "GPU")

If you create ov.Core() manually after speedup(), the PERFORMANCE_HINT set by fasthardware will not carry over. Use get_core() to avoid this.

---

manual_sweep()

Goes beyond Python's gc.collect() — performs a full 3-generation GC sweep and calls OS-level memory APIs to return fragmented RAM back to the system.

Platform	Method
Windows	SetProcessWorkingSetSize(-1, -1) via kernel32
Linux	malloc_trim(0) via libc

Recommended: call every 60 seconds in long-running inference loops.

---

🎯 Designed For

• Real-time computer vision pipelines (YOLOv8, OpenVINO, OpenCV)
• High-FPS webcam inference loops
• Edge AI on Intel iGPU / N-series / Core Ultra
• Multiprocessing workloads needing tight CPU affinity control
• Any Python app where consistent, stutter-free performance matters

---

📊 Real-World Results

Tested on Intel N100 (integrated GPU) running YOLOv8n-pose with OpenVINO:

Metric	Without fasthardware	With fasthardware
Peak FPS	baseline	+5 ~ 15 FPS
Frame stutters	frequent	eliminated
Memory creep (long runs)	gradual degradation	suppressed

---

⚙️ Full API Reference

# Initialize — call once at startup
fasthardware.speedup(mode="DEFAULT")   # Standard optimization
fasthardware.speedup(mode="ULTIMATE")  # + child process supercharging

# Get the pre-optimized OpenVINO Core singleton
core = fasthardware.get_core()  # Call after speedup()

# Sweep memory — call periodically
fasthardware.manual_sweep()

---

🛡️ Safety

• Priority set to HIGH (not REALTIME) — system stability preserved
• GC disabled during runtime; call gc.enable() on exit if needed
• All kernel API calls wrapped in try/except — fails silently without permissions
• Tested on Windows 11 and Ubuntu 22.04

---

📄 License

MIT License

---

🤝 Contributing

Issues and PRs are welcome.
If fasthardware helped your project hit a new FPS record, feel free to share it!