Slow Integer Matrix Multiplication

[SolidTux]

Currently matrix multiplications are a lot slower for integer matrices than for float matrices, because for floats either BLAS or matrixmultiply is used.

E.g. here for a 1000x1000 matrix the integer version takes more than 5 times as long as the float version and this get worse when the matrix size increases.

Is there anything planned for fast integer matrix multiplications? Or any way I can help to improve this situation?

[bluss]

With f64 you can represent integers up to 1 << 53, so there's your fastest integer matrix multiply. 😉

I don't know if there are any plans. I'm avoiding it since I don't have time for it and the utility is low. It seems like there are a lot more parameters that go into integer MM and the SIMD support is way less developed.

What would the design be with regards to overflow? What's the design for the scalar types, the accumulator and/or the resulting type? Various existing integer MMs have different combinations for these, and this operation seems to show up in specialist libraries.

Someone made an experiment of plugging in integers into matrixmultiply over in bluss/matrixmultiply#28. Maybe one could check in on how the smart people in the thread are doing w.r.t the problem. It would probably be good to bring a design that is already working somewhere else, and bring it to Rust. You don't have to port it - could be a library you call?

[SolidTux]

The matrixmultiply version would be great. I did a short test and used the test branch with ndarray which decreased the time significantly.