easobel 0.1.0

SIMD-accelerated Sobel edge detection. 5-12x faster than OpenCV single-threaded.

easobel

SIMD-accelerated Sobel edge detection. 5-12x faster than OpenCV single-threaded.

import easobel
import numpy as np

image = np.random.rand(1080, 1920).astype(np.float32)

edges = easobel.sobel(image)                        # single-threaded
edges = easobel.sobel_mt(image, num_threads=4)      # multi-threaded

Installation

pip install easobel

Pre-built wheels for Linux x86_64, Linux aarch64, and Windows x86_64. No compiler needed.

API

easobel.sobel(image) -> np.ndarray

Single-threaded Sobel edge detection.

• Input: 2D float32 numpy array (height, width), grayscale
• Output: 2D float32 array, same shape
• Border pixels are zero (Sobel stencil needs 1-pixel border)
• Image must be at least 3x3

easobel.sobel_mt(image, num_threads=None) -> np.ndarray

Multi-threaded Sobel with row-striped parallelism.

• Same input/output as sobel()
• num_threads defaults to os.cpu_count()
• ctypes releases the GIL for true thread parallelism

Algorithm

Computes gradient magnitude using L1 norm: |Gx| + |Gy|.

Sobel kernels:

Gx = [-1  0  1]     Gy = [-1 -2 -1]
     [-2  0  2]          [ 0  0  0]
     [-1  0  1]          [ 1  2  1]

The SIMD kernel processes 4 pixels per iteration using f32x4 vectors with a scalar tail loop for remaining pixels.

How it works

The Sobel kernel is written in Ea, a SIMD kernel compiler. Pre-compiled shared libraries are shipped in the wheel. Python handles validation and memory management via ctypes.

License

Apache-2.0