I have a special algorithm where as one of the lasts steps I need to carry out a multiplication of a 3-D array with a 2-D array such that each matrix-slice of the 3-D array is multiplied wich each column of the 2-D array. In other words, if, say A is an N x N x N matrix and B is an N x N matrix, I need to compute a matrix C of size N x N where C(:,i) = A(:,:,i)*B(:,i);.
The naive way to implement this is a loop, i.e.,
C = zeros(N,N);
for i = 1:N
C(:,i) = A(:,:,i)*B(:,i);
end
However, loops aren't the fastest in Matlab and should be avoided. I'm looking for faster ways of doing this. Right now, what I do is to use the fact that (now Mathjax would be great!):
[A1 b1, A2 b2, ..., AN bN] = [A1, A2, ..., AN]*blkdiag(b1,b2,...,bN)
This allows to get rid of the loop, however, we have to create a block-diagonal matrix of size N^2 x N. I'm making it via sparse to be efficient, i.e., like this:
A_long = reshape(A,N,N^2);
b_cell = mat2cell(B,N,ones(1,N)); % convert matrix to cell array of vectors
b_cell{1} = sparse(b_cell{1}); % make first element sparse, this is enough to trigger blkdiag into sparse mode
B_blk = blkdiag(b_cell{:});
C = A_long*B_blk;
According to my benchmarks, this approach is faster than the loop by a factor of around two (for large N), despite the necessary preparations (the multiplication alone is 3 to 4-fold faster than the loop).
Here is a quick benchmark I did, varying the problem size N and measuring the time for the loop and the alternative approach (with and without the preparation steps). For large N the speedup is around 2...2.5.
Still, this looks awfully complicated to me. Is there a simpler or better way to achieve this? This looks like it's a quite generic/standard problem so I could imagine that solutions are around, I just don't know what to search for really.
P.S.: blkdiag(A1,...,AN)*B is an obvious alternative but here the block diagonal is already N^2 x N^2 so I don't think it can be better than what I did.
edit: Thanks to everyone for commenting! I have carried out a new benchmark on a Matlab R2016b. Unfortunately, I do not have both versions on the same computer so we cannot compare the absolute numbers but the relative comparison is still interesting, since it has changed a bit. Here it is:
And here is a zoom on the high-N area:
Couple of observations:
SumRepDot is the solution proposed by Divakar, namely, to use squeeze(sum(bsxfun(#times,A,permute(B,[3,1,2])),2)) which on R2016b simplifies to squeeze(sum(A.*permute(B,[3,1,2]),2)). It is faster than the loop for high N by a factor of around 1.2...1.4.
The loop is still "slow" in a sense that the multiplication with the sparse block diagonal matrix is much faster.
For the latter, the preparation overhead seems to become negligible for high N which makes it overall a factor of 3...4 faster than the loop. This is a nice result.
Related
I have a linear system of about 2000 sparse equations in Matlab. For my final result, I only really need the value of one of the variables: the other values are irrelevant. While there is no real problem in simply solving the equations and extracting the correct variable, I was wondering whether there was a faster way or Matlab command. For example, as soon as the required variable is calculated, the program could in principle stop running.
Is there anyone who knows whether this is at all possible, or if it would just be easier to keep solving the entire system?
Most of the computation time is spent inverting the matrix, if we can find a way to avoid completely inverting the matrix then we may be able to improve the computation time. Lets assume I'm only interested in the solution for the last variable x(N). Using the standard method we compute
x = A\b;
res = x(N);
Assuming A is full rank, we can instead use LU decomposition of the augmented matrix [A b] to get x(N) which looks like this
[~,U] = lu([A b]);
res = U(end,end-1)/U(end,end);
This is essentially performing Gaussian elimination and then solving for x(N) using back-substitution.
We can extend this to find any value of x by swapping the columns of A before LU decomposition,
x_index = 123; % the index of the solution we are interested in
A(:,[x_index,end]) = A(:,[end,x_index]);
[~,U] = lu([A b]);
res = U(end,end)/U(end,end-1);
Bench-marking performance in MATLAB2017a with 10,000 random 200 dimensional systems we get a slight speed-up
Total time direct method : 4.5401s
Total time LU method : 3.9149s
Note that you may experience some precision issues if A isn't well conditioned.
Also, this approach doesn't take advantage of the sparsity of A. In my experiments even with 2000x2000 sparse matrices everything significantly slowed down and the LU method is significantly slower. That said full matrix representation only requires about 30MB which shouldn't be a problem on most computers.
If you have access to theory manuals on NASTRAN, I believe (from memory) there is coverage of partial solutions of linear systems. Also try looking for iterative or tri diagonal solvers for A*x = b. On this page, review the pqr solution answer by Shantachhani. Another reference.
Spefifically, the following two kinds of code can get the same S and V idealy. However, the second one's speed is usually faster than the first one in Matlab. Can someone tell me the reason?
Moreover, which method is more numerically stable?
Thanks.
[~,S,V] = svd(B,'econ');
[Qc,Rc] = qr(B',0);
[U,S,~] = svd(Rc,'econ');
V = Qc*U;
The second method does not have to be faster. For almost squared matrices it can be slower. Consider as example the Golub-Reinsch SVD-algorithm:
Its work depends on the output you want to calculate (only S, Sand V or S,V and U).
If you want to calculate Sand V without performing any preprocessing the required work is 4mn^2+8n^3.
If you perform QR-decomposition before this the needed amount of work is: 2/3n^3+n^2+1/3n-2 for the Housholder transformation. Now if your Matrix was almost squared, i.e m=n, you will have gained not much as R is still m x n. However if m is larger than n you can reduce R to an n x n matrix (called thin QR factorization). Now you want to calculate Uand S which will add 12n^3 for your SVD-algorithm.
So only SVD: 4mn^2+8n^3
SVD with QR: (12+2/3)n^3+n^2+1/3n-2
However most SVD-algorithms should inculde some (R-) bidiagonalizations which will reduce the work to: 2mn^2+11n^3
You can also apply QR, the R-bifactorization and then SVD to make it even faster but it all depends on your matrix dimensions.
Matlab uses for SVD the Lapack libraries. You can look up the exact runtimes here. They're approximately the same as above algorithm.
Hope this helps.
I'm looking to duplication a 784x784 matrix in matlab along a 3rd axis. The following code seems to work:
mat = reshape(repmat(mat, 1,10000),784,784,10000);
Unfortunately, it takes so long to run it's worthless (changing the 10,000s to 1000 makes it take a few minutes, and using 10,000 makes my whole machine freeze up practically). is there a faster way to do this?
For reference, I'm looking to use mvnpdf on 10,000 vectors each of length 784, using the same covariance matrix for each. So my final call looks like
mvnpdf(X,mu,mat)
%size(X) = (10000,784), size(mu) = (10000,784), size(mat) = 784,784,10000
If there's a way to do this that's not repeating the covariance matrix 10,000 times, that'd be helpful too. Thanks!
For replication in more than 2 dimensions, you need to supply the replication counts as an array:
out = repmat(mat,[1,1,10000])
Creating a 784x784 matrix 10,000 times isn't going to take advantage of the vectorization in MATLAB, which is going to be more useful for small arrays. Avoiding a for loop also won't help too much, given the following:
The main speedup you can gain here is by computing the inverse of the covariance matrix once, and then computing the pdf yourself. The inverse of sigma takes O(n^3), and you are needlessly doing that 10,000 times. (Also, the square root determinant can be precomputed.) For reference, the PDF of the multivariate normal distribution is computed as follows:
http://en.wikipedia.org/wiki/Multivariate_normal_distribution#Properties
Better to just compute the inverse once, and then compute z = x - mu for each value, then doing z'Sz for each pdf value, and applying a simple function and a constant. But wait! You can vectorize that, too.
I don't have MATLAB in front of me, but this is basically what you need to do, and it'll run in an instant.
s = inv(sigma);
c = -0.5*log(det(s)) - (k/2)*log(2*pi);
z = x - mu; % 10000 x 784 matrix
ans = exp( c - 0.5 .* dot(z*s, z, 2) ); % 10000 x 1 vector
Sorry if this is obvious but I searched a while and did not find anything (or missed it).
I'm trying to solve linear systems of the form Ax=B with A a 4x4 matrix, and B a 4x1 vector.
I know that for a single system I can use mldivide to obtain x: x=A\B.
However I am trying to solve a great number of systems (possibly > 10000) and I am reluctant to use a for loop because I was told it is notably slower than matrix formulation in many MATLAB problems.
My question is then: is there a way to solve Ax=B using vectorization with A 4x4x N and B a matrix 4x N ?
PS: I do not know if it is important but the B vector is the same for all the systems.
You should use a for loop. There might be a benefit in precomputing a factorization and reusing it, if A stays the same and B changes. But for your problem where A changes and B stays the same, there's no alternative to solving N linear systems.
You shouldn't worry too much about the performance cost of loops either: the MATLAB JIT compiler means that loops can often be just as fast on recent versions of MATLAB.
I don't think you can optimize this further. As explained by #Tom, since A is the one changing, there is no benefit in factoring the various A's beforehand...
Besides the looped solution is pretty fast given the dimensions you mention:
A = rand(4,4,10000);
B = rand(4,1); %# same for all linear systems
tic
X = zeros(4,size(A,3));
for i=1:size(A,3)
X(:,i) = A(:,:,i)\B;
end
toc
Elapsed time is 0.168101 seconds.
Here's the problem:
you're trying to perform a 2D operation (mldivide) on a 3d matrix. No matter how you look at it, you need reference the matrix by index which is where the time penalty kicks in... it's not the for loop which is the problem, but it's how people use them.
If you can structure your problem differently, then perhaps you can find a better option, but right now you have a few options:
1 - mex
2 - parallel processing (write a parfor loop)
3 - CUDA
Here's a rather esoteric solution that takes advantage of MATLAB's peculiar optimizations. Construct an enormous 4k x 4k sparse matrix with your 4x4 blocks down the diagonal. Then solve all simultaneously.
On my machine this gets the same solution up to single precision accuracy as #Amro/Tom's for-loop solution, but faster.
n = size(A,1);
k = size(A,3);
AS = reshape(permute(A,[1 3 2]),n*k,n);
S = sparse( ...
repmat(1:n*k,n,1)', ...
bsxfun(#plus,reshape(repmat(1:n:n*k,n,1),[],1),0:n-1), ...
AS, ...
n*k,n*k);
X = reshape(S\repmat(B,k,1),n,k);
for a random example:
For k = 10000
For loop: 0.122570 seconds.
Giant sparse system: 0.032287 seconds.
If you know that your 4x4 matrices are positive definite then you can use chol on S to improve the accuracy.
This is silly. But so is how slow matlab's for loops still are in 2015, even with JIT. This solution seems to find a sweet spot when k is not too large so everything still fits into memory.
I know this post is years old now, but I'll contribute my two cents anyway. You CAN put all of your A matricies into a bigger block diagonal matrix, where there will be 4x4 blocks on the diagonal of a big matrix. The right hand side will be all of your b vectors stacked on top of each other over and over. Once you set this up, it is represented as a sparse system, and can be efficiently solved with the algorithms mldivide chooses. The blocks are numerically decoupled, so even if there are singular blocks in there, the answers for the nonsingular blocks should be right when you use mldivide. There is a code that took this approach on MATLAB Central:
http://www.mathworks.com/matlabcentral/fileexchange/24260-multiple-same-size-linear-solver
I suggest experimenting to see if the approach is any faster than looping. I suspect it can be more efficient, especially for large numbers of small systems. In particular, if there are nice formulas for the coefficients of A across the N matricies, you can build the full left hand side using MATLAB vector operations (without looping), which could give you additional cost savings. As others have noted, vectorized operations aren't always faster, but they often are in my experience.
I've written this code to perform the 1-d convolution of a 2-d matrix valued function (k is my time index, kend is on the order of 10e3). Is there a faster or cleaner way to do this, perhaps using built in functions?
for k=1:kend
C(:,:,k)=zeros(3);
for l=0:k-1
C(:,:,k)=C(:,:,k)+A(:,:,k-l)*B(:,:,l+1);
end
end
NEW SOLUTION:
This is a newer solution built on the older solution, which solved the previously given formula. The code in the question is actually a modification of that formula, in which the overlap between the two matrices in the third dimension is repeatedly shifted (it's akin to a convolution along the third dimension of the data). The previous solution I gave only computed the result for the last iteration of the code in the question (i.e. k = kend). So, here's a full solution that should be much more efficient than the code in the question for kend on the order of 1000:
kend = size(A,3); %# Get the value for kend
C = zeros(3,3,kend); %# Preallocate the output
Anew = reshape(flipdim(A,3),3,[]); %# Reshape A into a 3-by-3*kend matrix
Bnew = reshape(permute(B,[1 3 2]),[],3); %# Reshape B into a 3*kend-by-3 matrix
for k = 1:kend
C(:,:,k) = Anew(:,3*(kend-k)+1:end)*Bnew(1:3*k,:); %# Index Anew and Bnew so
end %# they overlap in steps
%# of three
Even when using just kend = 100, this solution came out to be about 30 times faster for me than the one in the question and about 4 times faster than a pure for-loop-based solution (which would involve 5 loops!). Note that the discussion below of floating-point accuracy still applies, so it is normal and expected that you will see slight differences between the solutions on the order of the relative floating-point accuracy.
OLD SOLUTION:
Based on this formula you linked to in a comment:
it appears that you actually want to do something different than the code you provided in the question. Assuming A and B are 3-by-3-by-k matrices, the result C should be a 3-by-3 matrix and the formula from your link written out as a set of nested for loops would look like this:
%# Solution #1: for loops
k = size(A,3);
C = zeros(3);
for i = 1:3
for j = 1:3
for r = 1:3
for l = 0:k-1
C(i,j) = C(i,j) + A(i,r,k-l)*B(r,j,l+1);
end
end
end
end
Now, it is possible to perform this operation without any for loops by reshaping and reorganizing A and B appropriately:
%# Solution #2: matrix multiply
Anew = reshape(flipdim(A,3),3,[]); %# Create a 3-by-3*k matrix
Bnew = reshape(permute(B,[1 3 2]),[],3); %# Create a 3*k-by-3 matrix
C = Anew*Bnew; %# Perform a single matrix multiply
You could even rework the code you have in your question to create a solution with a single loop that performs a matrix multiply of your 3-by-3 submatrices:
%# Solution #3: mixed (loop and matrix multiplication)
k = size(A,3);
C = zeros(3);
for l = 0:k-1
C = C + A(:,:,k-l)*B(:,:,l+1);
end
So now the question: Which one of these approaches is faster/cleaner?
Well, "cleaner" is very subjective, and I honestly couldn't tell you which of the above pieces of code makes it any easier to understand what the operation is doing. All the loops and variables in the first solution make it a little hard to track what's going on, but it clearly mirrors the formula. The second solution breaks it all down into a simple matrix operation, but it's difficult to see how it relates to the original formula. The third solution seems like a middle-ground between the two.
So, let's make speed the tie-breaker. If I time the above solutions for a number of values of k, I get these results (in seconds needed to perform 10,000 iterations of the given solution, MATLAB R2010b):
k | loop | matrix multiply | mixed
-----+--------+-----------------+--------
5 | 0.0915 | 0.3242 | 0.1657
10 | 0.1094 | 0.3093 | 0.2981
20 | 0.1674 | 0.3301 | 0.5838
50 | 0.3181 | 0.3737 | 1.3585
100 | 0.5800 | 0.4131 | 2.7311 * The matrix multiply is now fastest
200 | 1.2859 | 0.5538 | 5.9280
Well, it turns out that for smaller values of k (around 50 or less) the for-loop solution actually wins out, showing once again that for loops are not as "evil" as they used to be considered in older versions of MATLAB. Under certain circumstances, they can be more efficient than a clever vectorization. However, when the value of k is larger than around 100, the vectorized matrix-multiply solution starts to win out, scaling much more nicely with increasing k than the for-loop solution does. The mixed for-loop/matrix-multiply solution scales atrociously for reasons that I'm not exactly sure of.
So, if you expect k to be large, I'd go with the vectorized matrix-multiply solution. One thing to keep in mind is that the results you get from each solution (the matrix C) will differ ever so slightly (on the level of the floating-point precision) since the order of additions and multiplications performed for each solution are different, thus leading to a difference in accumulation of rounding errors. In short, the difference between the results for these solutions should be negligible, but you should be aware of it.
Have you looked into Matlab's conv method?
I can't compare it against your provided code, because what you provided gives me a problem with trying to access the zeroth element of A. (When k=1, k-1=0.)
Have you considered using FFTs to convolve? A convolution operation is simply a point-wise multiplication in the frequency domain. You'll have to take some precaution with finite sequences, as you'll end up with circular convolution if you're not careful (but this is trivial to take care of).
Here's a simple example for a 1D case.
>> a=rand(4,1);
>> b=rand(3,1);
>> c=conv(a,b)
c =
0.1167
0.3133
0.4024
0.5023
0.6454
0.3511
The same using FFTs
>> A=fft(a,6);
>> B=fft(b,6);
>> C=real(ifft(A.*B))
C =
0.1167
0.3133
0.4024
0.5023
0.6454
0.3511
A convolution of an M point vector and an N point vector results in an M+N-1 point vector. So, I've padded each of the vectors a and b with zeros before taking the FFT (this is automatically taken care of when I take the 4+3-1=6 point FFT of it).
EDIT
Although the equation that you showed is similar to a circular convolution, it's not exactly it. So you can ditch the FFT approach, and the built-in conv* functions. To answer your question, here's the same operation done without explicit loops:
dim1=3;dim2=dim1;
dim3=10;
a=rand(dim1,dim2,dim3);
b=rand(dim1,dim2,dim3);
mIndx=cellfun(#(x)(1:x),num2cell(1:dim3),'UniformOutput',false);
fun=#(x)sum(reshape(cell2mat(cellfun(#(y,z)a(:,:,y)*b(:,:,z),num2cell(x),num2cell(fliplr(x)),'UniformOutput',false)),[dim1,dim2,max(x)]),3);
c=reshape(cell2mat(cellfun(#(x)fun(x),mIndx,'UniformOutput',false)),[dim1,dim2,dim3]);
mIndx here is a cell, where the ith cell contains a vector 1:i. This is your l index (as others have noted, please don't use l as a variable name).
The next line is an anonymous function that does the convolution operation, making use of the fact that the k index is just the l index flipped around. The operations are carried out on individual cells, and then assembled.
The last line actually performs the operations on the matrices.
The answer is the same as that obtained with the loops. However, you'll find that the looped solution is actually an order of magnitude faster (I averaged 0.007s for my code and 0.0006s for the loop). This is because the loop is pretty straightforward, whereas with this sort of nested construction, there's plenty of function call overheads and repeated reshaping that slow it down.
MATLAB's loops have come a long way since the early days when loops were dreaded. Certainly, vectorized operations are blazing fast; but not everything can be vectorized, and sometimes, loops are more efficient than such convoluted anonymous functions. I could probably shave off a few more tenths here and there by optimizing my construction (or maybe taking a different approach), but I'm not going to do that.
Remember that good code should be readable, as well as efficient and minor optimization at the cost of readability serves no one. Although I wrote the code above, I certainly will not be able to decipher what it does if I revisited it a month later. Your looped code was clear, readable and fast and I would suggest that you stick with it.