karm-patel/Blocked_Matrix_Multiplication

Analysis of blocked matrix multiplication

What this repo contains

In this repo, we have analyzed 96 various versions of the blocked matrix multiplication programs on combinations of following factors.

1. Matrix size: 2048, 8192
2. Outer loop order (variants): ‘ijk’, ‘jik’, ‘jki’, ‘ikj’, ‘kij’, ‘kji’
3. Tile size: 4, 8, 16, 32, 64, 128, 256, 512

Analysis contains various performance counters including but not limited to cache-misses, TLB misses, cpu cycles, branch instructions, etc. We use perforator CLI tool for analysis.

Analysis Summary

How to use this code?

Perforater installation

curl https://zyedidia.github.io/eget.sh | sh ./eget zyedidia/perforator

pip3 install regex pandas

DELETE FOLLOWING DIRECTORIES

out2048/
out8192/
csv/

RUN FOLLOWING PYTHON FILES IN SEQUENCE

1. genereate .out files in out2048/

python3 outgen2048.py

2. generate csv files in csv/

python3 execout2048.py

3. genereate .out files in out2048/

python3 outgen8192.py

4. generate csv files in csv/

python3 execout8192.py

FOLLOWING COMMAND IS BEING EXECUTED

sudo ./perforator -e cpu-cycles,branch-misses,branch-instructions,cache-misses,dtlb-read-accesses,dtlb-read-accesses,dtlb-read-misses,dtlb-write-misses,l1d-read-accesses,l1d-read-misses,l1d-write-accesses,l1i-read-misses,ll-read-accesses,ll-read-misses,ll-write-accesses,ll-write-misses -r multiplication --csv ./out2048/v3_2048_32 > csv/v3_2048_32.csv