Borrar filtros
Borrar filtros

Why is this loop faster than a vectorised version? Could the vectorised version be made faster than the loop?

2 visualizaciones (últimos 30 días)
I'm trying to improve performance in a code that uses a loop. I've written a vectorised version matching the functionality, while avoiding costly transposes. However, I've found that the loop version invariably runs ~25% more quickly. Is there any way to further improve the performance of the vectorised version so that it surpasses the loop?
Of course, this is a tiny sub-function of a much larger, more complex program, but it is called tens of thousands of times in a single run, and is a bottleneck in the run time.
I do have the parallel computing toolbox, so could look into using parfor loops, but these don't always save time, and I was surprised that the vectorised version doesn't perform better!
% Input vectors
x1 = rand(1, 960);
x2 = rand(1, 960);
%% Looped version
tic;
Y1 = 1.75:0.25:39;
Y2 = (10 .^ (Y1 / 21.366) - 1 ) / 0.004368;
EoutLoop = zeros(length(x2), length(Y1));
for i=1:length(Y1)
p1 = 4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1).*ones(1, length(x2));
p2 = 4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1) - 0.35.*(4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1)./4.*1000./24.673.*(0.004368.*1000 + 1)).*(x2 - 51);
p3 = p1.*(x1 >= Y2(i)) + p2 .* (x1 < Y2(i));
g1 = (x1 - Y2(i))./Y2(i);
g1 = abs(g1);
EiLoop = (1 + p3.*min(g1,4)).*exp(-1.*p3.*min(g1,4)).* 10.^(x2./10);
EoutLoop((1:length(x2)), i) = EiLoop(1:length(x2));
end
if (size(EoutLoop,1) > 1)
EoutLoop = sum(EoutLoop);
end
EoutLoop = 10 .* log(EoutLoop) ./ log(10);
% end timer
toc;
%% Vectorised version
% transpose input vector for vectorised version
x2 = x2.';
x1 = x1.';
% start timer
tic;
Y1 = 1.75:0.25:39;
Y2 = (10 .^ (Y1 / 21.366) - 1 ) / 0.004368;
EoutVec = zeros(length(x2), length(Y1));
p1 = 4.*Y2./24.673.*(0.004368.*Y2 + 1).*ones(length(x2), length(Y2));
p2 = 4.*Y2./24.673.*(0.004368.*Y2 + 1) - 0.35.*(4.*Y2./24.673.*(0.004368.*Y2 + 1)./4.*1000./24.673.*(0.004368.*1000 + 1)).*repmat((x2 - 51), 1, length(Y2));
p3 = p1.*(x1 >= repmat(Y2, 1, size(x1, 2))) + p2 .* (x1 < repmat(Y2, 1, size(x1, 2)));
g1 = ((x1 - repmat(Y2, 1, size(x1, 2)))./repmat(Y2, 1, size(x1, 2)));
g1 = abs(g1);
EVec = (1 + p3.*min(g1,4)).*exp(-1.*p3.*min(g1,4)).*repmat(10.^(x2./10), 1, length(Y2));
EoutVec((1:length(x2)), :) = EVec(1:length(x2), :);
if (size(EoutVec,1) > 1)
EoutVec = sum(EoutVec);
end
EoutVec = 10.*log(EoutVec)./log(10);
% end timer
toc;
  3 comentarios
Alexander
Alexander el 22 de Nov. de 2023
I agree. On my old Win7 machine (R2021b) the result is
Loop: Elapsed time is 0.316945 seconds.
Vectorised: Elapsed time is 0.062135 seconds.

Iniciar sesión para comentar.

Respuesta aceptada

Dyuman Joshi
Dyuman Joshi el 22 de Nov. de 2023
Editada: Dyuman Joshi el 22 de Nov. de 2023
Ideally, timeit should be used over tic-toc to get a more accurate idea of run times of the codes. tic-toc is generally used for portions of code.
"Use the timeit function for a rigorous measurement of function execution time. Use tic and toc to estimate time for smaller portions of code that are not complete functions." Reference - Measure the Performance of Your Code
While using tic-toc to measure the time of the code, you can either
> Run the same code multiple times via a for loop and average the data - "Sometimes programs run too fast for tic and toc to provide useful data. If your code is faster than 1/10 second, consider measuring it running in a loop, and then average to find the time for a single run." (Reference - https://in.mathworks.com/help/matlab/ref/tic.html#bswc7ww-3)
or
> Take a large(r) dataset.
I have chosen the latter option below -
% Input vectors
%% Large(r) dataset
x1 = rand(1, 100000);
x2 = rand(1, 100000);
%% Looped version
Y1 = 1.75:0.25:39;
Y2 = (10 .^ (Y1 / 21.366) - 1 ) / 0.004368;
EoutLoop = zeros(length(x2), length(Y1));
tic;
for i=1:length(Y1)
p1 = 4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1).*ones(1, length(x2));
p2 = 4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1) - 0.35.*(4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1)./4.*1000./24.673.*(0.004368.*1000 + 1)).*(x2 - 51);
p3 = p1.*(x1 >= Y2(i)) + p2 .* (x1 < Y2(i));
g1 = (x1 - Y2(i))./Y2(i);
g1 = abs(g1);
EiLoop = (1 + p3.*min(g1,4)).*exp(-1.*p3.*min(g1,4)).* 10.^(x2./10);
EoutLoop((1:length(x2)), i) = EiLoop(1:length(x2));
end
if (size(EoutLoop,1) > 1)
EoutLoop = sum(EoutLoop);
end
EoutLoop = 10 .* log(EoutLoop) ./ log(10);
% end timer
toc;
Elapsed time is 1.298153 seconds.
%% Vectorised version
% transpose input vector for vectorised version
x2 = x2.';
x1 = x1.';
% start timer
Y1 = 1.75:0.25:39;
Y2 = (10 .^ (Y1 / 21.366) - 1 ) / 0.004368;
EoutVec = zeros(length(x2), length(Y1));
tic;
p1 = 4.*Y2./24.673.*(0.004368.*Y2 + 1).*ones(length(x2), length(Y2));
p2 = 4.*Y2./24.673.*(0.004368.*Y2 + 1) - 0.35.*(4.*Y2./24.673.*(0.004368.*Y2 + 1)./4.*1000./24.673.*(0.004368.*1000 + 1)).*repmat((x2 - 51), 1, length(Y2));
p3 = p1.*(x1 >= repmat(Y2, 1, size(x1, 2))) + p2 .* (x1 < repmat(Y2, 1, size(x1, 2)));
g1 = ((x1 - repmat(Y2, 1, size(x1, 2)))./repmat(Y2, 1, size(x1, 2)));
g1 = abs(g1);
EVec = (1 + p3.*min(g1,4)).*exp(-1.*p3.*min(g1,4)).*repmat(10.^(x2./10), 1, length(Y2));
EoutVec((1:length(x2)), :) = EVec(1:length(x2), :);
if (size(EoutVec,1) > 1)
EoutVec = sum(EoutVec);
end
EoutVec = 10.*log(EoutVec)./log(10);
% end timer
toc;
Elapsed time is 0.621410 seconds.
You can see that the time taken by the vectorized approach is less than half of the time taken by the for loop approach.
  4 comentarios
Michael
Michael el 22 de Nov. de 2023
Thanks, I re-wrote the test code as you suggested, and for 10000 runs the vector version took about 30% of the time as the loop (the input vectors are fixed length for the application).
% Input vectors
x1 = rand(1, 960);
x2 = rand(1, 960);
x1T = x1.';
x2T = x2.';
totalRuns = 10000;
%% Looped version
loopTime = 0;
for runs = 1:totalRuns
% start timer
tic
Y1 = 1.75:0.25:39;
Y2 = (10 .^ (Y1 / 21.366) - 1 ) / 0.004368;
EoutLoop = zeros(length(x2), length(Y1));
for i=1:length(Y1)
p1 = 4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1).*ones(1, length(x2));
p2 = 4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1) - 0.35.*(4.*Y2(i)./24.673.*(0.004368.*Y2(i) + 1)./4.*1000./24.673.*(0.004368.*1000 + 1)).*(x2 - 51);
p3 = p1.*(x1 >= Y2(i)) + p2 .* (x1 < Y2(i));
g1 = (x1 - Y2(i))./Y2(i);
g1 = abs(g1);
EiLoop = (1 + p3.*min(g1,4)).*exp(-1.*p3.*min(g1,4)).* 10.^(x2./10);
EoutLoop((1:length(x2)), i) = EiLoop(1:length(x2));
end
if (size(EoutLoop,1) > 1)
EoutLoop = sum(EoutLoop);
end
EoutLoop = 10 .* log(EoutLoop) ./ log(10);
% append time
loopTime = loopTime + toc;
end
loopTime = loopTime/runs;
disp(num2str(loopTime))
%% Vectorised version
vecTime = 0;
for runs = 1:totalRuns
% start timer
tic
Y1 = 1.75:0.25:39;
Y2 = (10 .^ (Y1 / 21.366) - 1 ) / 0.004368;
EoutVec = zeros(length(x2T), length(Y1));
p1 = 4.*Y2./24.673.*(0.004368.*Y2 + 1).*ones(length(x2T), length(Y2));
p2 = 4.*Y2./24.673.*(0.004368.*Y2 + 1) - 0.35.*(4.*Y2./24.673.*(0.004368.*Y2 + 1)./4.*1000./24.673.*(0.004368.*1000 + 1)).*repmat((x2T - 51), 1, length(Y2));
p3 = p1.*(x1T >= repmat(Y2, 1, size(x1T, 2))) + p2 .* (x1T < repmat(Y2, 1, size(x1T, 2)));
g1 = ((x1T - repmat(Y2, 1, size(x1T, 2)))./repmat(Y2, 1, size(x1T, 2)));
g1 = abs(g1);
EVec = (1 + p3.*min(g1,4)).*exp(-1.*p3.*min(g1,4)).*repmat(10.^(x2T./10), 1, length(Y2));
EoutVec((1:length(x2T)), :) = EVec(1:length(x2T), :);
if (size(EoutVec,1) > 1)
EoutVec = sum(EoutVec);
end
EoutVec = 10.*log(EoutVec)./log(10);
vecTime = vecTime + toc;
end
vecTime = vecTime/runs;
disp(num2str(vecTime))
Thanks for the info. I did use Profiler on the full programme, which is how I identified the bottleneck for further testing and optimisation!
Dyuman Joshi
Dyuman Joshi el 22 de Nov. de 2023
You are welcome!
It's good to know that you are utilizing the Profiler, it is an extremely helpful tool!

Iniciar sesión para comentar.

Más respuestas (0)

Categorías

Más información sobre Logical en Help Center y File Exchange.

Productos


Versión

R2023b

Community Treasure Hunt

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

Start Hunting!

Translated by