panosantoniadis / cuda-exercises-ntua

Lab exercise of Parallel Processing course in NTUA regarding CUDA programming

Parallel Programming for GPUs - Matrix Multiplication

Dense Matrix Multiplication (DMM) is one of the core components in many scientific computations. In this repository, we implement the DMM algorithm for GPUs in CUDA using 4 algorithms, increasing each time the total performance.

Algorithms

• Naive: Simple implementation where each thread just computes one element from the output matrix.
• Coalesced memory acceses of A: Load tiles of the input matrix A in the shared memory.
• Reduced memory accesses: Load tiles of the input matrices A and B in the shared memory.
• Using cuBLAS library

Brief results

All experiments were performed in a NVIDIA Tesla K40c (kepler architecture and compute capability=3.5)

• Total Performance in 2048×2048 matrices

• Choosing the optimal thread block size

• Performance in different problem sizes

Project Structure

• cuda: Source code for DMM.
• common: Helper source code.
• make: Scripts for compiling the source code.
• plots: Plots in order to analyze our results.
• results: Performance of different scenarios.
• report: Final report in Greek.

Contributors:

• Antoniadis Panagiotis
• Bazotis Nikolaos