Roofline-Guided FP32 GEMM (Ubuntu 24.04 · 4 vCPUs)

This artifact builds a reproducible Roofline workflow for FP32 GEMM:

1. Calibrate sustained memory bandwidth (STREAM Triad) and compute throughput (OpenBLAS SGEMM).
2. Run two GEMM kernels (naive baseline and cache-blocked).
3. Collect FLOPs, bytes, OI, GFLOP/s to CSV.
4. Plot a Roofline figure using your measured ceilings.
5. Validate with unit tests and Valgrind memcheck.

---

What you get

• Kernels: gemm_baseline (i–j–k), gemm_blocked (tiled MB×NB×KB)
• Calibration tools: STREAM Triad (C, OpenMP) and an SGEMM peak timer
• Data & plots: CSV logs and a Roofline chart (OI vs GFLOP/s, log–log)
• Tests & hygiene: unit tests, Valgrind memcheck, stable OMP settings

---

System Requirements

• Ubuntu 24.04 (tested on a 4-vCPU, 8 GB RAM VM)
• GCC / OpenMP / OpenBLAS
• Python 3 with Matplotlib + Pandas
• jq (parse JSON), awk (parsing logs)
• Valgrind

Install prerequisites:

sudo apt update
sudo apt install -y build-essential libopenblas-dev valgrind jq \
                    python3 python3-matplotlib python3-pandas

Set up Environment settings

export OMP_NUM_THREADS=4
export OMP_PROC_BIND=close
export OMP_PLACES=cores

Clean build folder

make clean

Clean result folder

make distclean

Memory roof B (GB/s) with STREAM Triad, STREAM_ARRAY_SIZE=50,000,000 by default, NTIMES=20 iterations

make calibrate-mem

Compute roof F (GFLOP/s) with an SGEMM peak timer

make sgemm_peak
make calibrate-comp

Check the combined file

cat results/calibration.json

Build kernels

make baseline
make blocked MB=96 NB=96 KB=256   # you can tune block size

Run and collect the results (This will take a while)

./scripts/collect.sh baseline 1024 7
./scripts/collect.sh blocked  1024 7 # n, Trial
./scripts/collect.sh baseline 2048 7
./scripts/collect.sh blocked  2048 7

Plot the roofline graph (There's unit issue in the paper, should've divide by 1000 in the paper)

make roofline # -> results/roofline.png

Unit tests

make test

Large tests

make perf-large N=2048 TRIALS=5 THREADS=4 # -> results/large_runs.csv

Valgrind tests

make memcheck-test       # unit tests under valgrind
make memcheck-baseline   # n=128, baseline kernel
make memcheck-blocked    # n=128, blocked kernel
make memcheck-sgemm      # sgemm_peak sanity