fastdotrs 0.1.1

Fast SIMD-optimized dot products (f32·f32, f16·f32)

fast_dot

Fast SIMD-optimized dot products exposed to Python via PyO3 + maturin.

Features

• dot_f32(a: np.ndarray[float32], b: np.ndarray[float32]) -> float
• dot_f16_f32(a: np.ndarray[float16], b: np.ndarray[float32]) -> float (uses F16C when available)

Both functions accept any Python object that implements the buffer protocol (NumPy arrays, PyTorch tensors via .numpy() if on CPU and contiguous) through the NumPy bridge. For non-NumPy objects, convert to NumPy first.

Build and install

• Build wheel: maturin build --release
• Develop install: maturin develop --release

Usage

import numpy as np import fast_dot as fd

x = np.random.rand(1024).astype(np.float32) y = np.random.rand(1024).astype(np.float32) print(fd.dot_f32(x, y))

x16 = x.astype(np.float16) print(fd.dot_f16_f32(x16, y))

Notes

• Requires x86_64. F16C and AVX2 paths are used when available; otherwise falls back to scalar.
• Arrays must be 1-D and contiguous with matching shapes.