ECE 587 Spring 2026

GEMM Project 2 - CPU Optimizations


Report Due: 03/08 (Sun.), by the end of the day (Chicago time)
Late submissions will NOT be graded
I. Objective

In Project 1, we established a single-core baseline for General Matrix Multiplication (GEMM). In this project, we take the next step to leverage OpenMP to utilize the computation and communication resources available from multiple CPU cores, and to explore how memory tiling can further improve performance by optimizing communication patterns.

We rely on the same project setup Project 1 for VM and VS Code. Please refer to Sections II and III in Project 1 instructions for details.

II. CPU Optimizations

Our updated C++ code for this project is available here on Github gemm_tiled.cpp. This is the same Github repository as Project 1 and you may use git to clone it, or download and upload the C++ file. The code has the same structure as Project 1, but with some additions to support OpenMP and tiling. Your job is to add OpenMP, tune tile size, and explain what you observe.

First of all, similar to Project 1, please update the code to include your CWID. Without making any further change, build and run 'gemm_tiled' to confirm that the code works with your CWID in the output. 

ubuntu@ubuntu24:~$ g++ -O3 -march=native -fopenmp -std=c++17 gemm_tiled.cpp -o gemm_tiled

ubuntu@ubuntu24:~$ OMP_NUM_THREADS=4 ./gemm_tiled 1024 128
CWID A12345678 1771013167 N = 1024, M = 128, 4 MiB per matrix
OMP_NUM_THREADS = 4
CWID A12345678 1771013167 gemm_ikj : min 89.279 ms, avg 90.262 ms, max 91.864 ms
CWID A12345678 1771013168 gemm_tiled : min 88.625 ms, avg 90.005 ms, max 92.619 ms
CWID A12345678 1771013168 sums: min 10144.978, max 10145.179, passed
Compared to the 'gemm' program from Project 1, the 'gemm_tiled' program takes one additional argument for tile size M. On the other hand, we will set 'OMP_NUM_THREADS=4' throughout the project to use the 4 cores available on our VM. You may experiment with additional cores if you have access to a more powerful computer to have a better understanding of how performance scales with cores.

Here is what you need to do in this project. Based on these two GEMM implementations 'gemm_ikj()' and 'gemm_tiled()' in the C++ code, you should implement 'gemm_ikj_omp()' and 'gemm_tiled_omp()' to add OpenMP parallelizations, respectively. Don't forget to modify 'main()' to call these new functions to report their performance. Run the program with matrix sizes of 1024, 2048, 4096, 8192, and tile sizes of 16, 32, 64, 128. Collect data from the output. Nevertheless, feel free to structure your code and experiments as you see fit, e.g. to introduce loops in 'main()' to automate runs for different N and M values, but make sure to validate your results and collect data as needed.

III. Project Deliverables

Submit a project report in .doc/.docs or .pdf format to Canvas before the deadline for a total of 15 points. Your project report should include the following.