# Memory usage in nested loops
**URL:** https://discourse.julialang.org/t/memory-usage-in-nested-loops/32098
**Category:** New to Julia
**Tags:** multithreading, memory-allocation
**Created:** [December 10, 2019, 1:43pm UTC](https://discourse.julialang.org/t/memory-usage-in-nested-loops/32098 "2019-12-10T13:43:29Z")
**Posts on this page:** 3
**Page:** 1
### Author:  [@M\_Kaj](https://discourse.julialang.org/u/M_Kaj)
#### Post date: [December 10, 2019, 1:43pm UTC](https://discourse.julialang.org/t/memory-usage-in-nested-loops/32098/1 "2019-12-10T13:43:29Z")
Hey everyone,
i would like to have some advise how to perform loops over vector multiplication.
I need to do this for alot of vectors, so i try to use multi threading.
The memory consumption of my function is really high and i do not really understand why.
To be more specific, I have 4 arrays with data (a,b,c,d) with dimension (L, N, N).
I use loops over the indices (N,N) and elementwise multiplication in the direction (L).
For multi threading i creat 2 auxillary arrays (aux1,aux2) for each thread which store elementwise multiplication along L-direction of two element of (a,b,c,d).
Afterwards I take a shifted scalar product of the auxillary arrays.
Increasing N leads to a rapid growth in the memory consumption, which is far larger than the increase in size of the arrays. I would appreciate alot suggestions how to improve here.
The example code looks like this
```julia
function test()
in_p=100
N=12
L=2^14
a=rand(Complex{Float64},L,N,N)
b=rand(Complex{Float64},L,N,N)
c=rand(Complex{Float64},L,N,N)
d=rand(Complex{Float64},L,N,N)
aux=zeros(Complex{Float64},Threads.nthreads() )
aux1=[zeros(Complex{Float64},L) for i=1:Threads.nthreads()]
aux2=[zeros(Complex{Float64},L) for i=1:Threads.nthreads()]
Threads.@threads for i1=1:N
t=Threads.threadid()
for i2=1:N , i3=1:N , i4=1:N
aux1[t] .= a[:,i1,i2] .* b[:,i3,i4]
aux2[t] .= c[:,i2,i3] .* d[:,i4 ,i1]
aux[t] += transpose(@views aux1[t][1:end-in_p]) * (@views aux2[t][1+in_p:end])
end
end
return sum(aux)
end
```
---
### Author:  [@tomaklutfu](https://discourse.julialang.org/u/tomaklutfu)
#### Post date: [December 10, 2019, 2:41pm UTC](https://discourse.julialang.org/t/memory-usage-in-nested-loops/32098/2 "2019-12-10T14:41:16Z")
```julia
aux1[t] .= a[:,i1,i2] .* b[:,i3,i4]
aux2[t] .= c[:,i2,i3] .* d[:,i4 ,i1]
```
causes memory allocation on rhs before broadcasting `*`. It doesn’t seem like but
```julia
aux[t] += transpose(@views aux1[t][1:end-in_p]) * (@views aux2[t][1+in_p:end])
```
also allocates in rhs with `aux[t] + transpose(...)*(...)`
---
### Author:  [@stevengj](https://discourse.julialang.org/u/stevengj)
#### Post date: [December 10, 2019, 3:44pm UTC](https://discourse.julialang.org/t/memory-usage-in-nested-loops/32098/3 "2019-12-10T15:44:17Z")
You can just put `@views` in front of `function test()` in order to use views for slices everywhere in the function. (Creating a view still allocates a small amount of memory.)
In any case, doing 3 array passes just to compute a dot product at the end seems pretty wasteful, especially when you don’t even use all of the products in your `aux1` and `aux2` arrays. It looks like you can get the same result with just:
```nohighlight
s = aux[t]
for i = 1:L-in_p
s += a[i,i1,i2]*b[i,i3,i4] * c[i+in_p,i2,i3]*d[i+in_p,i4,i1]
end
aux[t] = s
```
(You might also want `@simd for ...` on this loop as well as `@inbounds s += ...`, for speed.)