In my last article, I posted a function for calculating one partition of a larger matrix. THe function looked like this
void partial(k, i, g, M, f){
for (m=0; m < n; m++){
j = m * k;
g[i] = g[i] + M[i][j] * f[i];
}
}This is actually wrong. Lets look where I messed up. It was all the way back in the equation.
Continue readingIf we want to distribute the mathematical processing of a matrix multiplication we need an algorithm that can be split and performed in parallel. Looking at the algorithm I alluded to in the last article, I think I can see how to do that. Here’s my thinking.
Continue readingI peridocially fall off the wagon and get drawn back into playing Minecraft. I’ve decided that, in order to make this time not wasted, I need to do something constructive with this urge. Last time I played Minecraft, I found the MultiMC launcher would no longer work. Being a fan of C++ and open source projects, I was not happy with this state. A friend suggested I try the PrismLauncher fork of the code base.
Prism does not seem to have a native Debian based build available, although I admit I did not look very hard. I don’t want to install flatpacks or other binary management software just for one app. So, I figured I would build from sources.
Many of the big scientific computing problems can be solved using matrix mathematics. One of my favorite problems to tackle is implementation of vector times matrix = vector. This has utility in many places, one of which is inference in neural networks.
Since an ARM processor has some support for vector and matrix mathematics, I started wondering what were the size limitations we would hit when trying to solve big problems. Here are some notes: