Performance of non-preallocated array

3 次查看（过去 30 天）
Noam
Noam 2015-4-29
编辑： James Tursa 2015-4-29
I'm trying to understand the reason these two options are so differnet in performance:
>> tic;z=[];for i = 1:1e5;z(end+1)=i;end;toc
Elapsed time is 0.022571 seconds.
>> tic;z=[];for i = 1:1e5;z = [z i];end;toc
Elapsed time is 7.752663 seconds.
In both we create an array of the numbers 1 to 100000 while increasing the size of the array (without pre-allocating the array) but in the second option, the concatenation is taking much more time than the isertion option, although in both cases we have to re-allocate the memory and copy the array to the new place in memory.
Thanks Noam

请先登录，再回答此问题。

采纳的回答

Titus Edelhofer
Titus Edelhofer 2015-4-29
Hi,
two comments:
  • You are doing the timing in the command line. This gives typically worse results than in a function, because MATLAB can do within functions some runtime optimizations (acceleration) that is not possible on command line. If you put the code into a function, the results (for my machine) are 0.02 vs 0.10 seconds, so still quite a difference but not as drastic as in your case.
  • Second: the line "z(end+1)=i" is recognized by the accelerator as "grow in a loop without preallocation", so behind the scenes it will grow the array not simply element by element but will do some sort of preallocation itself. That's better than nothing but of course never as good as explicit preallocation.
Titus
  7 个评论
显示 5更早的评论
John D'Errico
John D'Errico 2015-4-29
In fact, MATLAB CAN preallocate an array to have an initial size that is larger than what the array contains, IF the array is sparse. Thus spalloc can allow you to specify the initial size of a sparse zero array.
S = spalloc(M,N,NZMAX) creates an M-by-N all zero sparse matrix
with room to eventually hold NZMAX nonzeros.
It would be nice if the zeros function had this ability. Add an extra argument at the end, to allow the user to specify the initial allocated space for an array that will be grown.
James Tursa
James Tursa 2015-4-29
编辑：James Tursa 2015-4-29
Side Note: The NZMAX value in a MATLAB variable is a size_t type (in C parlance) in the variable structure, and is actually part of the structure definition and available in all MATLAB variables (but to my knowledge unused in everything except sparse). So it is available and could be used for full variables without any changes to the low level variable structure definition if TMW chose to do so.

请先登录，再进行评论。

更多回答（1 个）

James Tursa
James Tursa 2015-4-29
MATLAB's parser can sometimes pre-allocate an array behind the scenes for you (even though you didn't code it that way) if it recognizes a pattern in the for loop. It appears the parser recognizes the pattern in the first case but not the second case.

类别

在 Help Center 和 File Exchange 中查找有关 Sparse Matrices 的更多信息

标签

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by