fastcv

fastcv is a C++ CUDA rewrite with Pytorch bindings of the image filters in the OpenCV library.

How to run

1. Clone the repo

https://github.com/JINO-ROHIT/fastcv

2. Move to fastcv

cd fastcv

3. Build the package

pip install -e . --no-build-isolation

You’ll need:

1. CUDA toolkit installed
2. PyTorch with CUDA support
3. A working C++ compiler (MSVC on Windows, gcc/clang on Linux/Mac)

Benchmarks

Tested on: RTX 4060 Ti

Kernel	Image Size	OpenCV (CPU)	fastcv (CUDA)	OpenCV CUDA (soon)	Speedup (×)
RGB2GRAY	4096×4096	3.1841 ms	0.2737 ms	-	11.63
BOX BLUR	4096×4096	44.459 ms	10.012 ms	-	4.44
SOBEL	4096×4096	384.83 ms	0.3223 ms	-	1194.0
EROSION	4096×4096	2.9758 ms	0.5827 ms	-	5.11
DILATION	4096×4096	2.7539 ms	0.5856 ms	-	4.70