Why is my Matlab for-loop code faster than my vect

问题:

I had always heard that vectorized code runs faster than for loops in MATLAB. However, when I tried vectorizing my MATLAB code it seemed to run slower.

I used tic and toc to measure the times. I changed only the implementation of a single function in my program. My vectorized version ran in 47.228801 seconds and my for-loop version ran in 16.962089 seconds.

Also in my main program I used a large number for N, N = 1000000and DataSet's size is 1 301, and I ran each version several times for different data sets with the same size and N.

Why is the vectorized so much slower and how can I improve the speed further?

The "vectorized" version

function [RNGSet] = RNGAnal(N,DataSet)
%Creates a random number generated set of numbers to check accuracy overall
%   This function will produce random numbers and normalize a new Data set
%   that is derived from an old data set by multiply random numbers and
%   then dividing by N/2
randData = randint(N,length(DataSet));
tempData = repmat(DataSet,N,1);
RNGSet = randData .* tempData;
RNGSet = sum(RNGSet,1) / (N/2); % sum and normalize by the N
end

The "for-loop" version

function [RNGData] = RNGAnsys(N,Data)
%RNGAnsys This function produces statistical RNG data using a for loop
%   This function will produce RNGData that will be used to plot on another
%   plot that possesses the actual data
multData = zeros(N,length(Data));
for i = 1:length(Data)
    photAbs = randint(N,1); % Create N number of random 0's or 1's
    multData(:,i) = Data(i) * photAbs; % multiply each element in the molar data by the random numbers
end

sumData = sum(multData,1); % sum each individual energy level's data point
RNGData = (sumData/(N/2))'; % divide by n, but account for 0.5 average by n/2
end

回答1:

Vectorization

First glance at the for-loop code tells us that since photAbs is a binary array each column of which is scaled according to each element of Data, this binary feature could be used for vectorization. This is abused in the code here -

function RNGData = RNGAnsys_vect1(N,Data)

%// Get the 2D Matrix of random ones and zeros
photAbsAll = randint(N,numel(Data));

%// Take care of multData internally by summing along the columns of the
%// binary 2D matrix and then multiply each element of it with each scalar 
%// taken from Data by performing elementwise multiplication
sumData = Data.*sum(photAbsAll,1);

%// Divide by n, but account for 0.5 average by n/2
RNGData = (sumData./(N/2))'; %//'

return;

After profiling, it appears that the bottleneck is the random binary array creating part. So, using a faster random binary array creator as suggested in this smart solution, the above function could be further optimized like so -

function RNGData = RNGAnsys_vect2(N,Data)

%// Create a random binary array and sum along the columns on the fly to
%// save on any variable space that would be required otherwise. 
%// Also perform the elementwise multiplication as discussed before.
sumData = Data.*sum(rand(N,numel(Data))<0.5,1);

%// Divide by n, but account for 0.5 average by n/2
RNGData = (sumData./(N/2))'; %//'

return;

Using the smart binary random array creator, the original code could be optimized as well, that will be used for a fair benchmarking between optimized for-loop and vectorized codes later on. The optimized for-loop code is listed here -

function RNGData = RNGAnsys_opt1(N,Data)

multData = zeros(N,numel(Data));
for i = 1:numel(Data)

    %// Create N number of random 0's or 1's using a smart approach
    %// Then, multiply each element in the molar data by the random numbers
    multData(:,i) = Data(i) * rand(N,1)<.5; 
end

sumData = sum(multData,1); % sum each individual energy level's data point
RNGData = (sumData/(N/2))'; % divide by n, but account for 0.5 average by n/2
return;

Benchmarking

Benchmarking Code

N = 15000; %// Kept at this value as it going out of memory with higher N's.
           %// Size of dataset is more important anyway as that decides how
           %// well is vectorized code against a for-loop code

DS_arr = [50 100 200 500 800 1500 5000]; %// Dataset sizes
timeall = zeros(2,numel(DS_arr));

for k1 = 1:numel(DS_arr)
    DS = DS_arr(k1);
    Data = rand(1,DS);

    f = @() RNGAnsys_opt1(N,Data);%// Optimized for-loop code
    timeall(1,k1) = timeit(f);
    clear f

    f = @() RNGAnsys_vect2(N,Data);%// Vectorized Code
    timeall(2,k1) = timeit(f);
    clear f
end

%// Display benchmark results
figure,hold on, grid on
plot(DS_arr,timeall(1,:),'-ro')
plot(DS_arr,timeall(2,:),'-kx')
legend('Optimized for-loop code','Vectorized code')
xlabel('Dataset size ->'),ylabel('Time(sec) ->')
avg_speedup = mean(timeall(1,:)./timeall(2,:))
title(['Average Speedup with vectorized code = ' num2str(avg_speedup) 'x'])

Results

Concluding remarks

Based on the experience I had so far with MATLAB, neither for loops nor vectorized techniques are fit for all situations, but everything is situation-specific.

回答2:

Try using the matlab profiler to determine which line or lines of code are using the most amount of time. That way you can find out if the repmat function is what is slowing you down as is being suggested. Let us know what you find, I'm interested!

回答3:

randData = randint(N,length(DataSet));

allocates a 1.2GB array. (4*301*1000000). Implicitly you create up to 4 of these monsters in your program, causing continuous cache-misses.

You for-loop code could nearly run in the processor cache (or it does on the bigger xeons).