Implement a simple algorithm that transposes a no-symmetric matrix of size NxN.

• The algorithm takes the dimension of the matrix as input (use a power of 2).
• For example, ./transpose 10 transposes a 2^10 x 2^10 matrix
• Measure the Effective bandwidth of your implementation by using -00 –O1 –O2 –O3 options. Analyze the cache behavior.

To reduce recording errors and OS noise, multiple executions were performed so that the total exec. time was $~1.5s$.

In this assignment we study how the system behaves with different matrix transpose implementations. Both a simple (sequential, with nested loops) and block implementation are visualized to better understand cache misses depending on how we access the data.

TL:DR simple method workds better when $N \le 10^6$, while block method when $N > 10^6$; prefetching may improve performances when the matrix is big (D1 miss rate, most of the misses are due to write miss).

More info in report.pdf.

Effective Bandwidth	Execution time
L1 miss rate	D1 miss rate
LL miss rate	D references

├── launcher.sh                 # [SLOW] Automatic experiment launcher, do not use in this repo as here there's no plot fn
├── log.txt                     # Param for `launcher.sh`
├── main.pdf                    # Project report
├── Makefile                    
├── plot                        # Images used in the report
│   ├── _bandwidth.png
│   ├── D1 miss rate.png
│   ├── double_bandwidth.png
│   ├── double_time.png
│   ├── D references.png
│   ├── float_bandwidth.png
│   ├── float_time.png
│   ├── L1 miss rate.png
│   ├── LL miss rate.png
│   └── _time.png
├── README.md
├── src                         # My library code
│   ├── matrix.cc
│   ├── matrix.h
│   ├── utils.cc
│   └── utils.h
└── transpose.cc                # Main file