cudaJourney

cudaJourney is a personal CUDA learning repository focused on building GPU intuition through small, self-contained programs.

The codebase currently covers:

• vector operations such as elementwise add, reductions, prefix sums, histograms, and merge-based routines
• matrix operations such as GEMM, 3D stencil kernels, sparse matrix-vector multiplication, and graph traversal experiments
• image processing kernels such as grayscale conversion, blur, and convolution-based edge detection

Most of the work is based on the Oak Ridge National Laboratory CUDA training material, with additional experiments and profiling notes added along the way.

Repository Layout

Path	Focus
vector_operations/	foundational CUDA kernels for scan, reduction, histogram, and related patterns. Results: vector_operations/RESULTS_SUMMARY.md
matrix_operations/	matrix multiplication, stencil, sparse operations, and graph-oriented kernels. Results: matrix_operations/RESULTS_SUMMARY.md
ImageProcessing/	image transforms using CUDA and OpenCV-backed host I/O. Results: ImageProcessing/RESULTS_SUMMARY.md

Requirements

• NVIDIA GPU with a working CUDA runtime
• CUDA Toolkit with nvcc
• OpenCV for the image-processing programs

Build and Run

There is no single build system yet. Each experiment is compiled directly with nvcc.

Example:

nvcc -O2 -o matrix_operations/1_matrix_multiplications matrix_operations/1_matrix_multiplications.cu
./matrix_operations/1_matrix_multiplications 3 1 555 777 999 0

For image-processing programs, link OpenCV in the compile step. One common pattern is:

nvcc -O2 ImageProcessing/1_grayscale_conversion.cu -o ImageProcessing/grayscale_conversion $(pkg-config --cflags --libs opencv4)
./ImageProcessing/grayscale_conversion DEVICE BASIC

The source files print their expected CLI arguments when invoked incorrectly, so the fastest way to explore a kernel is to compile it and run it once with no parameters.

Notes

• Some folders include screenshots or generated images that document kernel results and Nsight observations.
• Performance numbers in the checked-in result markdown files are machine-specific and were collected on the author’s local setup.

Why This Repo Exists

This repository is intentionally incremental. The goal is not a framework or library; it is a documented CUDA learning path with runnable experiments, saved result snapshots, and a clear record of optimization attempts.