English | 简体中文
Run
git clone https://github.com/jackfsuia/Vecparser.gitThen install the requirements, run
pip install slyTo vectorize your for-loops from Matlab and CVX, write your loop (please first read the Notice) to the loop_eiditor.m, then run
python vecparser.pyThat's all! The results will be printed in loop_eiditor.m too, please refresh it.
To vectorize the Matlab loops, copy the loops you want to vectorize to loop_eiditor.m, like this one:
% loop_eiditor.m
for n1=1:N1
for n2=1:N2
for n3=1:N3
for n4=1:N4
if n1~=n2*n3 && n3>n4^3
x(n1,n2,n3,n4)= (y(n1,n3)+z(n4))*h(n2,n3,n1); % note: size(z) has to be "N4 1", not "1 N4".
q(n4,n3,n2,n1)= -h(n2,n3,n1)+((y(n1,n3)+z(n4))*h(n2,n3,n1))^2;
end
end
end
end
endthen run
python vecparser.pyThe result will be appended to loop_eiditor.m as
% loop_eiditor.m
for n1=1:N1
for n2=1:N2
for n3=1:N3
for n4=1:N4
if n1~=n2*n3 && n3>n4^3
x(n1,n2,n3,n4)= (y(n1,n3)+z(n4))*h(n2,n3,n1); % note: size(z) has to be "N4 1", not "1 N4".
q(n4,n3,n2,n1)= -h(n2,n3,n1)+((y(n1,n3)+z(n4))*h(n2,n3,n1))^2;
end
end
end
end
end
%-------------------------vectorized by Vecparser as-----------------------
cached_condition_for_this=(permute(repmat(repmat((1:N1)',1,N2,N3)~=permute(repmat(repmat((1:N2)',1,N3).*permute(repmat((1:N3)',1,N2),[2,1]),1,1,N1),[3,1,2]),1,1,1,N4),[2,3,1,4])&permute(repmat(permute(repmat((1:N3)',1,N4),[2,1])>repmat((1:N4)'.^3,1,N3),1,1,N2,N1),[3,2,4,1]));
x=permute(permute((cached_condition_for_this),[4,1,2,3]).*permute((permute(repmat((permute(repmat(y,1,1,N4),[1,3,2])+permute(repmat(z,1,N1,N3),[2,1,3])),1,1,1,N2),[4,3,1,2]).*repmat(h,1,1,1,N4)),[4,1,2,3])+permute((1-permute((cached_condition_for_this),[3,4,1,2])),[2,3,4,1]).*permute(x,[4,2,3,1]),[4,2,3,1]);
q=permute(permute((cached_condition_for_this),[4,1,2,3]).*permute((repmat(-h,1,1,1,N4)+permute(permute((permute(repmat((permute(repmat(y,1,1,N4),[1,3,2])+permute(repmat(z,1,N1,N3),[2,1,3])),1,1,1,N2),[4,3,1,2]).*repmat(h,1,1,1,N4)),[3,4,1,2]).^2,[3,4,1,2])),[4,1,2,3])+permute((1-permute((cached_condition_for_this),[3,4,1,2])),[2,3,4,1]).*permute(q,[1,3,2,4]),[1,3,2,4]);
%-----Please clear this file each time before you write a new loop on------Now copy the results to your matlab to replace the loops, and try them out.
Does this help you save some run time? Give us a ⭐
I ran this performance test on my old computer: Intel(R) Xeon(R) CPU E5-2660 v2 @ 2.20GHz, RAM 16G. Here is what I got:
I also observed that when the loop of iterations is too big, vectorization of it will cause my computer to crash due to memory shortage, therefore it ran slower than traditional loops in those extreme cases. It will be meaningful to see the trade off provided limited RAM, and how it'll perform when GPU come into play.
Now it supports nested 'if' and non-if blocks anywhere in loop, but don't support 'else', so please change it to 'if' instead. Support all the native vectorized operators: +, -, *, /, and so on. Support self-defined vectorized dimension-invariant function without modifications to code. Support CVX style convex inequality like '>=','<=','=='. It might have bugs, for being a experimental project.
- Support reduce operators like
sum,norm,*(matrix multiplication). - Explore its use on other languages (e.g., python)
Vecparser is licensed under the Apache 2.0 License found in the LICENSE file in the root directory of this repository.
If this work is helpful, please kindly cite as:
@article{Vecparser,
title={Vecparser: a parser that vectorizes your nested for-loops (in MATLAB, CVX) as much as possible.},
author={Yannan Luo},
year={2024},
url={https://github.com/jackfsuia/Vecparser}
}This repo uses the matlablexer from pymatlabparser. This repo's pymatlabparser folder is entirely copied from there with nearly zero modifiacations. Thanks for their wonderful works.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent