I get a somewhat odd error with @. MWE:
julia> y = Array{Float64}(undef, 2, 2); @. y =rand(2,2) + ones(2,2)
ERROR: MethodError: Cannot `convert` an object of type Array{Float64,2} to an object of type Float64
Where is the attempted conversion? Moving @. to after the assignment operator works fine, as does pre-allocating the two matrices.
@. affects every function call, which includes ones and rand in your example. The code is equivalent to:
y .= rand.(2, 2) .+ ones.(2, 2)
which roughly means: “for each index in y, do y[I] = rand(2,2) + ones(2, 2)” , which isn’t what you want.
In this case, you may want to just add the . to the functions you actually want to broadcast, as in:
y .= rand(2, 2) .+ ones(2, 2)
Incidentally, when trying to debug code with macros, @macroexpand (or @macroexpand1) is handy.
It probably would have allowed you to find the problem in this case:
julia> @macroexpand @. y = rand(2,2) + ones(2,2)
:(y .= (+).(rand.(2, 2), ones.(2, 2)))
You can also suppress broadcasting with $, as in
@. y = $rand(2,2) + $ones(2,2)
Note that you don’t need to construct an array of ones here. You can just use .+ 1 as in
@. y = $rand(2,2) + 1
In fact, you don’t need to construct the rand(2,2) array either. Just do:
@. y = rand() + 1
and the rand.() call will call rand() separately for each element of y.
Won’t the last one be slower than the previous? As I understood it, we have optimizations for generating multiple random numbers at once.
Allocating a 2x2 temporary array just to discard it is also slow; it’s a tradeoff. In general, I would err on the side of fewer temporary arrays and more readable code.