CUDA-Accelerated Python Library

GPU-accelerated matrix multiplication and image convolution benchmarks — CPU vs CUDA with reproducible artifacts and a full project report.

---

Problem

Numerical Python workloads on CPU become bottlenecks for large matrix operations and image convolution. Understanding where custom CUDA kernels deliver meaningful speedups requires rigorous, reproducible benchmarking.

Solution

Custom C/CUDA implementations for image convolution and matrix multiplication, benchmarked against CPU baselines with Jupyter-driven analysis, compiled artifacts, and documented results.

Key Results

Workload	Implementations Compared
Matrix multiplication	CPU, naive CUDA, optimized CUDA, cuBLAS
Image convolution	CPU (gcc), CUDA (nvcc), Python bridge (.so)
Artifacts	Reproducible binaries, output images, timing plots
Report	Full methodology in docs/report/CUDA_Project_Report_Final.pdf

---

Project Goals

• Benchmark CPU vs CUDA performance for core numerical workloads.
• Provide executable C/CUDA implementations for image convolution.
• Preserve reproducible benchmarking artifacts and visual outputs.
• Document methodology and results in a final project report.

Repository Layout

CUDA-Accelerated-Python-Library/
├── native_image_convolution_sources/
│   ├── image_convolution_cpu.c
│   ├── image_convolution_cuda.cu
│   └── image_convolution_cuda_python_bridge.cu
├── benchmarks/
│   └── image_convolution/
│       ├── case_01/
│       │   ├── src/
│       │   ├── notebooks/
│       │   ├── data/input/
│       │   ├── results/final/
│       │   ├── results/intermediate/
│       │   └── artifacts/
│       ├── case_02/
│       └── case_03/
├── notebooks/
│   └── matrix_multiplication/
│       └── matrix_multiplication_benchmark.ipynb
├── docs/
│   └── report/
│       ├── CUDA_Project_Report_Final.pdf
│       └── CUDA_Project_Report_Final.docx
├── README.md
└── LICENSE

Workload Modules

Matrix Multiplication

• Notebook: notebooks/matrix_multiplication/matrix_multiplication_benchmark.ipynb
• Covers CPU, naive CUDA, optimized CUDA, and cuBLAS comparisons.
• Includes timing, plotting, and performance analysis.

Image Convolution

Each benchmark case (case_01, case_02, case_03) is organized as:

• native_image_convolution_sources/ (repository root) - canonical source files used by notebooks:
  • image_convolution_cpu.c
  • image_convolution_cuda.cu
  • image_convolution_cuda_python_bridge.cu
• src/ - local case source snapshots
• notebooks/ - image_convolution_benchmark.ipynb
• data/input/ - input.pgm, image.jpg
• results/final/ - final output images (out_*)
• results/intermediate/ - temporary benchmark outputs (tmp_*)
• artifacts/bin and artifacts/lib - compiled executables and shared libraries

Build and Run

From any convolution case directory, for example benchmarks/image_convolution/case_01:

CPU Build

gcc ../../../native_image_convolution_sources/image_convolution_cpu.c -O2 -o artifacts/bin/image_convolution_cpu

CUDA Build

nvcc -O2 -arch=sm_75 ../../../native_image_convolution_sources/image_convolution_cuda.cu -o artifacts/bin/image_convolution_cuda

Python-Interop Shared Library

nvcc -Xcompiler -fPIC -shared ../../../native_image_convolution_sources/image_convolution_cuda_python_bridge.cu -o artifacts/lib/libimage_convolution.so

Sample Execution

./artifacts/bin/image_convolution_cpu data/input/input.pgm results/final/out_edge_n5_512.pgm edge_n5 512
./artifacts/bin/image_convolution_cuda data/input/input.pgm results/final/out_edge_n5_cuda_512.pgm edge_n5 512

Environment

• NVIDIA GPU with CUDA support
• CUDA toolkit (nvcc)
• GCC
• Python 3 + Jupyter for notebook workflows

The benchmark configuration in this project is aligned with Tesla T4-compatible compilation (sm_75), as used in the report workflow.

Report

Project report files are available in docs/report/:

• CUDA_Project_Report_Final.pdf
• CUDA_Project_Report_Final.docx

License

This project is licensed under the MIT License. See LICENSE for details.