A SIMD-optimized implementation of the im2col operation for 2D images (e.g., grayscale). This repository provides C++ source code targeting AVX2 (x86_64) and NEON (Armv8) architectures, as well as a reference (non-SIMD) implementation. You can integrate the resulting library into your own C++ project or access it from Python through a simple Ctypes wrapper.
Core utility developed for Mosaic-SR — used in production to accelerate patch extraction for the super-resolution pipeline on Arm CPU targets without GPU support.
The im2col (“image to column”) operation is commonly used in convolutional neural networks (CNNs) to transform a 2D image (or feature map) into a set of column vectors, facilitating efficient matrix-multiplication-based convolutions.
By taking advantage of SIMD intrinsics, we can significantly speed up the im2col computation on CPUs that support vector operations. This repository includes:
- AVX2 implementation for modern x86_64 processors
- NEON implementation for Armv8 processors
- Reference scalar implementation (no intrinsics) for portability or as a fallback
The repository also contains a Python wrapper and a Jupyter Notebook that can be used as an example on how to use the library in Python.
-
Clone this repository:
git clone https://github.com/Henvezz95/im2col_2D.git cd im2col_2D -
Configure and build with CMake:
mkdir build && cd build cmake -DCMAKE_BUILD_TYPE=Release .. cmake --build .
CMake will detect your CPU architecture and automatically compile the corresponding SIMD implementation (AVX2, NEON, or reference).
-
After building, you should find a shared library (e.g., libim2col.so on Linux or im2col.dll on Windows) in your build folder.
im2col_2D/
├── CMakeLists.txt # Main CMake build script
├── src/
│ ├── im2col_AVX2.cpp # AVX2 implementation
│ ├── im2col_NEON.cpp # NEON implementation
│ └── im2col_ref.cpp # Reference (fallback) implementation
├── python/
│ └── im2col.py # Python ctypes wrapper
├── Test Notebook.ipynb # Jupyter Notebook example- Architecture-aware dispatch: each implementation is tuned to its register width — AVX2 (256-bit, 8 floats) uses 2 cases; NEON (128-bit, 4 floats) uses 3, with Case 2 manually unrolled for
Note: The current implementation supports 2D (grayscale) arrays only. Multichannel (RGB) support via channel-stride is planned for a future release. common kernel sizes (5–8 columns)
- Safe boundary handling: the Python wrapper over-allocates the
output tensor by
mem_tailelements to prevent out-of-bounds reads at image edges, trimming before return - Zero-copy interface: input and output tensors are passed directly via Ctypes pointers — no intermediate copies
