Product of three matrices fast

General Usage Performance
1

I want to multiply three matrices together as efficiently as possible.

We all know that if you have tall and skinny or fat and short matrices, the order of operations when multiplying three matrices matter a lot for efficiency.

Example:

n1=5000; n2=10; n3=5000; n4=10
A=randn(n1,n2); B=randn(n2,n3); C=randn(n3,n4)
@btime A*(B*C);

218.357 μs (3 allocations: 391.58 KiB)

@btime (A*B)*C;

105.064 ms (4 allocations: 191.12 MiB)

Is there a way in julia to automatically select the best multiplication order?

I think I know how to create a flop-count based decision. Ideally it would also take the practical performance on the specific computer it is running into account.

2
2 Likes
3

Awesome!

4

Though, It seems it was a while since it was updated so it might not work out of the box with newest julia.

5

The alternative to not specify the order and just A*B*C or equivalently(?) *(A,B,C) already seems to do some smart things which may be sufficient for me. Is this documented somewhere?

6

It does not do anything smart.