Inefficient MatrixVectorMult_DDRM.innerProduct

[audetto]

I have noticed that innerProduct

ejml/main/ejml-ddense/src/org/ejml/dense/row/mult/MatrixVectorMult_DDRM.java

Lines 338 to 344 in 2c9d1dc

	for (int k = 0; k < B.numCols; k++) {
	double sum = 0;
	for (int i = 0; i < B.numRows; i++) {
	sum += a[offsetA + i]*B.data[k + i*cols];
	}
	output += sum*c[offsetC + k];
	}

performs a lot worse (2x minimum, but varies with size, 1000s) than my original naive implementation.

I think I figured out the reason.

The access of the matrix data likely trashes the CPU cache, because it keeps jumping column: B.data[k + i*cols], where i is incremented in the inner loop.

If I swap the loops, I get a back the lost speed.

Before I provide a PR, is there any reason this is does this way?

[lessthanoptimal]

There should be two implementations here. One for small matrices where everything is contained inside the CPU cache (where this would be faster) and another for larger matrices that's designed to reduce cache misses.

Now another question is, why does this function exist. It looks like it was not designed to be part of a public API but to perform operations inside of a block-wise algorithm. Are you using this function? I don't see it getting referenced any place so maybe it's best to delete it.

[audetto]

Hi, I used it because it was available in the autocompletion and it did exactly what I needed.

What is not clear from your answer is in which circumstance this implementation is optimal?
If everything stays in the cache, why is this version better?