powertechpreview / libfsmc

Library for Fast Small Matrix Computation
BSD 3-Clause "New" or "Revised" License
0 stars 2 forks source link

Create generic method for matrix with size power of two #18

Open pdrocaldeira opened 3 years ago

pdrocaldeira commented 3 years ago

One of our main targets is to find out where our efficiency ends comparing with different BLAS libraries.

For that, we need to increase matrix sizes and measure performance until we hit this turning point.

My suggested approach is to increase matrix size by power of two so we could implement a divide and conquer approach, like this video: https://www.youtube.com/watch?v=auw2Nm6ZOqI

Basically you will consider each matrix as a matrix block and will stop until you hit a baseline: 2x2 size.

After this is done, we need performance results from 2x2 until 1024x1024.

So, in short:

pdrocaldeira commented 3 years ago

From our first tests divide and conquer approach was utterly bad. We are moving to a loop approach.