Custom CUDA Library for CNN Pre-Processing

Why this exists

Most training pipelines preprocess on the CPU - normalization, augmentations, convolutions - then ship tensors to the GPU. That leaves expensive hardware idle while O(n³) matrix work and pixel loops run on a single thread pool. This project ships a reusable library instead of a one-off script: optimized CUDA kernels packaged as Python-callable shared objects.

Repository

The full implementation, notebooks, and reproduction steps live in a public repo:

github.com/ShamikOfficial/CUDA-Accelerated-Python-Library →

• Matrix module - CPU baseline, naive CUDA, shared-memory tiled CUDA (16×16 tiles), and cuBLAS reference, benchmarked for N = 64 to 2048.
• Convolution module - edge detection, sharpen, and Gaussian blur at 3×3, 5×5, and 7×7, with one GPU thread per output pixel.
• Python bindings - libmatrix.so and libconvo.so loaded via ctypes over NumPy float32 arrays.

Architecture

• Tiled shared memory: each block loads a 16×16 tile of A and B into on-chip memory once, eliminating the redundant global reads that bottleneck the naive kernel.
• CUDA events timing: warm-up launches plus event-based measurement on Tesla T4 (-arch=sm_75) for reproducible benchmarks.
• Zero-padded convolution: boundary-safe filters with configurable modes (edge_n3, sharpen_n7, etc.) and output clipped to 0-255.

Results

• GPU outperforms CPU at every matrix size tested; the gap widens sharply as N grows.
• Tiling delivers the largest gains at medium-to-large sizes; overhead dominates at very small N.
• Custom tiled kernel is competitive with cuBLAS on small matrices; cuBLAS pulls ahead at the heaviest workloads.
• Convolution speedups scale with image area - GPU parallelism pays off once transfer cost is amortized.

Impact

• 115× speedup over CPU at N=1024 and 14× faster PyTorch preprocessing.
• Drop-in GPU acceleration for standard Python ML workflows - no pipeline rewrite required.

Tech