CUDA Essentials

This repository is a good starting point to learn about CUDA basics. Details about CUDA specific syntaxes are added as comments. Currently I have covered vector addition and 2D convolution operation.

Content

1. Vector Addition - Baseline, Pinned Memory, Unified Memory
2. Convolution - Baseline, Tiled, cuDNN

Running Instructions

1. Requires CUDA installed in the machine. We use CUDA v11.7
2. To compile the program (program.cu) use

nvcc program.cu -o program

3. To compile a program using cuDNN library

nvcc program.cu -lcudnn -o program

4. To profile program.cu using nvidia profiler

nvprof ./program [arg1] [arg2] [arg3]

Number of arguments are program specific. Go through the code to know more.

Vector Addition Observations:

Tested the code on following cases:

1. 1 block, 1 thread : slight slow down compared to CPU due to lower clock speeds.
2. 1 block, 256 threads : Speedup > 10
3. n blocks, 256 threads : Speedup > 100 (n is selected such that n * 256 = size of vector)

2D Convolution Results

Hardware: Intel Xeon, NVIDIA Tesla P4

Algorithm	Runtime (in ms)	Speedup
Baseline	88.626	-
Tiled	74.520	1.1892
CuDNN Library	35.914	2.4677

License

MIT