CUDA Matrix Multiplication Benchmarking

This project implements and benchmarks different approaches to matrix multiplication using CUDA:

• Sequential CPU implementation
• Naive GPU implementation
• GPU implementation with memory coalescing (via thread remapping)
• Tiled GPU implementation using shared memory
• Tiled GPU implementation using shared memory and memory coalescing (via matrix B transposition)

Future work:

• More optimizations planned (vectorized memory access, register tiling, etc.)

The implementations and results are discussed in detail in these blog posts on my personal site:

• Part 1: Naive GPU Implementation, Explanation, and CPU vs naive GPU Benchmarking
• Part 2: Tiled Matrix Multiplication Explained and Implemented, Benchmarking against naive GPU, and Performance Analysis with Nsight Compute

Building the Project

mkdir build && cd build
cmake ..
make

Usage

The executable supports different modes:

Run Full Benchmark Suite

./matmul

This runs benchmarks for all implementations across matrix sizes: 32×32, 256×256, 1024×1024, and 2048×2048.

Profile Specific Implementation

./matmul profile <type> <dim>

• <type>: Implementation type (naive_gpu, coalesced_gpu, tiled_gpu, tiled_coalesced_gpu)
• <dim>: Matrix dimension (creates dim×dim matrices)

Example:

./matmul profile tiled_gpu 1024

NVIDIA Nsight Compute Profiling

For detailed GPU metrics:

ncu --set full -o naive_2048_full.ncu-rep ./matmul profile naive_gpu 2048
ncu --set full -o tiled_2048_full.ncu-rep ./matmul profile tiled_gpu 2048
ncu --set full -o tiled_coalesced_2048_full.ncu-rep ./matmul profile tiled_coalesced_gpu 2048

Implementation Details

The project implements matrix multiplication using different approaches:

• Each thread computes one element of the output matrix (naive_gpu)
• Uses shared memory tiling to improve memory access patterns (tiled_gpu)
• Basic CPU implementation for baseline comparison (sequential_cpu)

Each implementation can be benchmarked and profiled independently to compare performance across different metrics. For now these metrics include GFLOPS and time (ms).