Day 1

The focus of Day 1 is vector addition, with an initial naive approach and an improved version.

Code Descriptions

1️vector_add.cu (Naive Vector Addition)

• Implements element-wise vector addition (C = A + B).
• Uses global memory and assigns one thread per element.

2️vector_add_optim.cu (Improved Vector Addition)

• Uses an improved approach with grid-stride loops for better memory access patterns.
• Kernel optimizations:
  • Uses grid-stride loops to ensure all elements are efficiently computed.
  • Launches the initWith() function on the GPU instead of running it on the CPU.
  • dynamically selects the number of blocks based on the number of streaming multiprocessors (SMs).
• Result: Better GPU utilization, reduced memory latency, and improved performance.

Profiling and Running

To compile and profile the CUDA codes, use:

nvcc -o compiled_code_name source_code.cu
nsys profile --stats=true compiled_code_name

My Blog

https://medium.com/@poornima31298/cuda-an-easy-introduction-for-beginners-5e897b62d1bd