I have not heard it explicitly from the developers, but it seems obvious to
me that implementing slicing as a view into the original array, is hard to
do right, while a copy is simple. They probably wanted to do the simple
thing first, and then solve the harder problem when things settle down.
Regarding the need for annotation. There is a good reason for this and its
that the type should be kept stable during calculations.
`a = 0` will set a to be an Int with value 0. Later it is updated with
Float64 values and therefore the performance loss.
You can solve this in different variants
a
Excellent, thanks. Replacing `@inbounds a = dot(W[:,n], state)` with the
following gave a ~5x improvement:
a::Float64 = 0
for k in 1:N
@inbounds a += W[k, n] * state[k]
end
The ::Float64 annotation was necessary or it was really slow. Why does
Here are some things you might try:
(1) Don’t use a global const. Pass N as an argument to all functions.
(2) Use in-place multiplication via functions like A_mul_B! in lines like
`@inbounds a = dot(W[:,n], state)`. In Julia 0.3, W[:,n] creates a copy of that
column of W on every iteration.
(3
