Compare this function:
```Julia
function f2(k)
M = spzeros(2*k,2*k)
for i = 1:k
j1 = (i-1)*2+1
j2 = i*2
M[j1,j1] = rand()
M[j2,j1] = rand()
M[j1,j2] = rand()
M[j2,j2] = rand()
end
return M
end
```
which is much faster. It seems your original code has two performance
issues that are unrelated to sparse matrix memory allocation:
(1) `randn` allocates a new matrix every time
(2) Something about indexing sparse matrices with ranges seems slow (I
don't know why)
If you want to continue to use `randn`, then you can use `randn!`
instead, and preallocate the small matrix outside the loop.
-erik
On Mon, Feb 1, 2016 at 9:42 AM, Kristoffer Carlsson
<[email protected]> wrote:
>> However doing in this way is more cumbersome because you need to have a
>> good estimate of the number of entries
>
> Not true. The difference between pre allocating the arrays and just pushing
> into them is not that large due to how julia arrays work (constant
> ammortized time etc).
>
>
> On Monday, February 1, 2016 at 1:34:12 PM UTC+1, alan souza wrote:
>>
>> You could try to use the triplet form (tree vectors containing the row/col
>> indexes and the value of the entry) and call the function sparse.
>> In this way you can preallocate in advance these three vectors.
>> However doing in this way is more cumbersome because you need to have a
>> good estimate of the number of entries and to explicitly calculate the index
>> for all entries.
>>
>> On Sunday, January 31, 2016 at 8:07:56 PM UTC-2, Gabriel Goh wrote:
>>>
>>> Generating a sparse matrix from scratch seems to be quite memory
>>> intensive. and slow. Say I wish to create a large block diagonal matrix with
>>> 2x2 block entries.
>>>
>>> Doing it naively is quite slow
>>>
>>> function f(k)
>>> M = spzeros(2*k,2*k)
>>> for i = 1:k
>>> D = (i-1)*2 + 1:i*2
>>> M[D,D] = randn(2,2)
>>> end
>>> return M
>>> end
>>>
>>> julia> @time f(10000)
>>> 2.534277 seconds (239.26 k allocations: 3.013 GB, 15.58% gc time)
>>>
>>> Is there a way to speed this up by preallocating the memory somehow?
--
Erik Schnetter <[email protected]>
http://www.perimeterinstitute.ca/personal/eschnetter/
