I have a loop in which I use ppval to evaluate a set of values from a piecewise polynomial spline. The interpolation is easily the most time consuming part of the loop and I am looking for a way improve the function's efficiency.
More specifically, I'm using a finite difference scheme to calculate transient temperature distributions in friction welds. To do this I need to recalculate the material properties (as a function of temperature and position) at each time step. The rate limiting factor is the interpolation of these values. I could use an alternate finite difference scheme (less restrictive in the time domain) but would rather stick with what I have if at all possible.
I've included a MWE below:
x=0:.1:10;
y=sin(x);
pp=spline(x,y);
tic
for n=1:10000
x_int=10*rand(1000,1);
y_int=ppval(pp,x_int);
end
toc
plot(x,y,x_int,y_int,'*') % plot for sanity of data
Elapsed time is 1.265442 seconds.
Edit - I should probably mention that I would be more than happy with a simple linear interpolation between values but the interp1 function is slower than ppval
x=0:.1:10;
y=sin(x);
tic
for n=1:10000
x_int=10*rand(1000,1);
y_int=interp1(x,y,x_int,'linear');
end
toc
plot(x,y,x_int,y_int,'*') % plot for sanity of data
Elapsed time is 1.957256 seconds.
This is slow, because you're running into the single most annoying limitation of JIT. It's the cause of many many many oh so many questions in the MATLAB tag here on SO:
MATLAB's JIT accelerator cannot accelerate loops that call non-builtin functions.
Both ppval and interp1 are not built in (check with type ppval or edit interp1). Their implementation is not particularly slow, they just aren't fast when placed in a loop.
Now I have the impression it's getting better in more recent versions of MATLAB, but there are still quite massive differences between "inlined" and "non-inlined" loops. Why their JIT doesn't automate this task by simply recursing into non-builtins, I really have no idea.
Anyway, to fix this, you should copy-paste the essence of what happens in ppval into the loop body:
% Example data
x = 0:.1:10;
y = sin(x);
pp = spline(x,y);
% Your original version
tic
for n = 1:10000
x_int = 10*rand(1000,1);
y_int = ppval(pp, x_int);
end
toc
% "inlined" version
tic
br = pp.breaks.';
cf = pp.coefs;
for n = 1:10000
x_int = 10*rand(1000,1);
[~, inds] = histc(x_int, [-inf; br(2:end-1); +inf]);
x_shf = x_int - br(inds);
zero = ones(size(x_shf));
one = x_shf;
two = one .* x_shf;
three = two .* x_shf;
y_int = sum( [three two one zero] .* cf(inds,:), 2);
end
toc
Profiler:
Results on my crappy machine:
Elapsed time is 2.764317 seconds. % ppval
Elapsed time is 1.695324 seconds. % "inlined" version
The difference is actually less than what I expected, but I think that's mostly due to the sum() -- for this ppval case, I usually only need to evaluate a single site per iteration, which you can do without histc (but with simple vectorized code) and matrix/vector multiplication x*y (BLAS) instead of sum(x.*y) (fast, but not BLAS-fast).
Oh well, a ~60% reduction is not bad :)
It is a bit surprising that interp1 is slower than ppval, but having a quick look at its source code, it seems that it has to check for many special cases and has to loop over all the points since it it cannot be sure if the step-size is constant.
I didn't check the timing, but I guess you can speed up the linear interpolation by a lot if you can guarantee that steps in x of your table are constant, and that the values to be interpolated are stricktly within the given range, so that you do not have to do any checking. In that case, linear interpolation can be converted to a simple lookup problem like so:
%data to be interpolated, on grid with constant step
x = 0:0.5:10;
y = sin(x);
x_int = 0:0.1:9.9;
%make sure it is interpolation, not extrapolation
assert(all(x(1) <= x_int & x_int < x(end)));
% compute mapping, this can be precomputed for constant grid
slope = (length(x) - 1) / (x(end) - x(1));
offset = 1 - slope*x(1);
%map x_int to interval 1..lenght(i)
xmapped = offset + slope * x_int;
ind = floor(xmapped);
frac = xmapped - ind;
%interpolate by taking weighted sum of neighbouring points
y_int = y(ind) .* (1 - frac) + y(ind+1) .* frac;
% make plot to check correctness
plot(x, y, 'o-', x_int, y_int, '.')
Related
I was persuaded some time ago to drop my comfortable matlab programming and start programming in Julia. I have been working for a long with neural networks and I thought that, now with Julia, I could get things done faster by parallelising the calculation of the gradient.
The gradient need not be calculated on the entire dataset in one go; instead one can split the calculation. For instance, by splitting the dataset in parts, we can calculate a partial gradient on each part. The total gradient is then calculated by adding up the partial gradients.
Though, the principle is simple, when I parallelise with Julia I get a performance degradation, i.e. one process is faster then two processes! I am obviously doing something wrong... I have consulted other questions asked in the forum but I could still not piece together an answer. I think my problem lies in that there is a lot of unnecessary data moving going on, but I can't fix it properly.
In order to avoid posting messy neural network code, I am posting below a simpler example that replicates my problem in the setting of linear regression.
The code-block below creates some data for a linear regression problem. The code explains the constants, but X is the matrix containing the data inputs. We randomly create a weight vector w which when multiplied with X creates some targets Y.
######################################
## CREATE LINEAR REGRESSION PROBLEM ##
######################################
# This code implements a simple linear regression problem
MAXITER = 100 # number of iterations for simple gradient descent
N = 10000 # number of data items
D = 50 # dimension of data items
X = randn(N, D) # create random matrix of data, data items appear row-wise
Wtrue = randn(D,1) # create arbitrary weight matrix to generate targets
Y = X*Wtrue # generate targets
The next code-block below defines functions for measuring the fitness of our regression (i.e. the negative log-likelihood) and the gradient of the weight vector w:
####################################
## DEFINE FUNCTIONS ##
####################################
#everywhere begin
#-------------------------------------------------------------------
function negative_loglikelihood(Y,X,W)
#-------------------------------------------------------------------
# number of data items
N = size(X,1)
# accumulate here log-likelihood
ll = 0
for nn=1:N
ll = ll - 0.5*sum((Y[nn,:] - X[nn,:]*W).^2)
end
return ll
end
#-------------------------------------------------------------------
function negative_loglikelihood_grad(Y,X,W, first_index,last_index)
#-------------------------------------------------------------------
# number of data items
N = size(X,1)
# accumulate here gradient contributions by each data item
grad = zeros(similar(W))
for nn=first_index:last_index
grad = grad + X[nn,:]' * (Y[nn,:] - X[nn,:]*W)
end
return grad
end
end
Note that the above functions are on purpose not vectorised! I choose not to vectorise, as the final code (the neural network case) will also not admit any vectorisation (let us not get into more details regarding this).
Finally, the code-block below shows a very simple gradient descent that tries to recover the parameter weight vector w from the given data Y and X:
####################################
## SOLVE LINEAR REGRESSION ##
####################################
# start from random initial solution
W = randn(D,1)
# learning rate, set here to some arbitrary small constant
eta = 0.000001
# the following for-loop implements simple gradient descent
for iter=1:MAXITER
# get gradient
ref_array = Array(RemoteRef, nworkers())
# let each worker process part of matrix X
for index=1:length(workers())
# first index of subset of X that worker should work on
first_index = (index-1)*int(ceil(N/nworkers())) + 1
# last index of subset of X that worker should work on
last_index = min((index)*(int(ceil(N/nworkers()))), N)
ref_array[index] = #spawn negative_loglikelihood_grad(Y,X,W, first_index,last_index)
end
# gather the gradients calculated on parts of matrix X
grad = zeros(similar(W))
for index=1:length(workers())
grad = grad + fetch(ref_array[index])
end
# now that we have the gradient we can update parameters W
W = W + eta*grad;
# report progress, monitor optimisation
#printf("Iter %d neg_loglikel=%.4f\n",iter, negative_loglikelihood(Y,X,W))
end
As is hopefully visible, I tried to parallelise the calculation of the gradient in the easiest possible way here. My strategy is to break the calculation of the gradient in as many parts as available workers. Each worker is required to work only on part of matrix X, which part is specified by first_index and last_index. Hence, each worker should work with X[first_index:last_index,:]. For instance, for 4 workers and N = 10000, the work should be divided as follows:
worker 1 => first_index = 1, last_index = 2500
worker 2 => first_index = 2501, last_index = 5000
worker 3 => first_index = 5001, last_index = 7500
worker 4 => first_index = 7501, last_index = 10000
Unfortunately, this entire code works faster if I have only one worker. If add more workers via addprocs(), the code runs slower. One can aggravate this issue by create more data items, for instance use instead N=20000.
With more data items, the degradation is even more pronounced.
In my particular computing environment with N=20000 and one core, the code runs in ~9 secs. With N=20000 and 4 cores it takes ~18 secs!
I tried many many different things inspired by the questions and answers in this forum but unfortunately to no avail. I realise that the parallelisation is naive and that data movement must be the problem, but I have no idea how to do it properly. It seems that the documentation is also a bit scarce on this issue (as is the nice book by Ivo Balbaert).
I would appreciate your help as I have been stuck for quite some while with this and I really need it for my work. For anyone wanting to run the code, to save you the trouble of copying-pasting you can get the code here.
Thanks for taking the time to read this very lengthy question! Help me turn this into a model answer that anyone new in Julia can then consult!
I would say that GD is not a good candidate for parallelizing it using any of the proposed methods: either SharedArray or DistributedArray, or own implementation of distribution of chunks of data.
The problem does not lay in Julia, but in the GD algorithm.
Consider the code:
Main process:
for iter = 1:iterations #iterations: "the more the better"
δ = _gradient_descent_shared(X, y, θ)
θ = θ - α * (δ/N)
end
The problem is in the above for-loop which is a must. No matter how good _gradient_descent_shared is, the total number of iterations kills the noble concept of the parallelization.
After reading the question and the above suggestion I've started implementing GD using SharedArray. Please note, I'm not an expert in the field of SharedArrays.
The main process parts (simple implementation without regularization):
run_gradient_descent(X::SharedArray, y::SharedArray, θ::SharedArray, α, iterations) = begin
N = length(y)
for iter = 1:iterations
δ = _gradient_descent_shared(X, y, θ)
θ = θ - α * (δ/N)
end
θ
end
_gradient_descent_shared(X::SharedArray, y::SharedArray, θ::SharedArray, op=(+)) = begin
if size(X,1) <= length(procs(X))
return _gradient_descent_serial(X, y, θ)
else
rrefs = map(p -> (#spawnat p _gradient_descent_serial(X, y, θ)), procs(X))
return mapreduce(r -> fetch(r), op, rrefs)
end
end
The code common to all workers:
#= Returns the range of indices of a chunk for every worker on which it can work.
The function splits data examples (N rows into chunks),
not the parts of the particular example (features dimensionality remains intact).=#
#everywhere function _worker_range(S::SharedArray)
idx = indexpids(S)
if idx == 0
return 1:size(S,1), 1:size(S,2)
end
nchunks = length(procs(S))
splits = [round(Int, s) for s in linspace(0,size(S,1),nchunks+1)]
splits[idx]+1:splits[idx+1], 1:size(S,2)
end
#Computations on the chunk of the all data.
#everywhere _gradient_descent_serial(X::SharedArray, y::SharedArray, θ::SharedArray) = begin
prange = _worker_range(X)
pX = sdata(X[prange[1], prange[2]])
py = sdata(y[prange[1],:])
tempδ = pX' * (pX * sdata(θ) .- py)
end
The data loading and training. Let me assume that we have:
features in X::Array of the size (N,D), where N - number of examples, D-dimensionality of the features
labels in y::Array of the size (N,1)
The main code might look like this:
X=[ones(size(X,1)) X] #adding the artificial coordinate
N, D = size(X)
MAXITER = 500
α = 0.01
initialθ = SharedArray(Float64, (D,1))
sX = convert(SharedArray, X)
sy = convert(SharedArray, y)
X = nothing
y = nothing
gc()
finalθ = run_gradient_descent(sX, sy, initialθ, α, MAXITER);
After implementing this and run (on 8-cores of my Intell Clore i7) I got a very slight acceleration over serial GD (1-core) on my training multiclass (19 classes) training data (715 sec for serial GD / 665 sec for shared GD).
If my implementation is correct (please check this out - I'm counting on that) then parallelization of the GD algorithm is not worth of that. Definitely you might get better acceleration using stochastic GD on 1-core.
If you want to reduce the amount of data movement, you should strongly consider using SharedArrays. You could preallocate just one output vector, and pass it as an argument to each worker. Each worker sets a chunk of it, just as you suggested.
The goal of this experiment is to compare speed of Matlab and Julia with a small piece of code below.
First the Matlab code:
>> t = 5000; n = 10000; x = 1:t*n;
>> x = reshape(x, t, n);
>> tic(); y1 = sum(x(:) .* x(:)); toc()
Elapsed time is 0.229563 seconds.
>> y1
y1 =
4.1667e+22
>> tic(); y2 = trace(x * x'); toc()
Elapsed time is 15.332694 seconds.
>> y2
y2 =
4.1667e+22
Versus in Julia
julia> t = 5000; n = 10000; x = 1: t*n;
julia> x = reshape(x, t, n);
julia> tic(); y1 = sum(x[:].* x[:]); toc();
elapsed time: 1.235170533 seconds
julia> y1
-4526945843202100544
julia> tic();y2 = trace(x*x'); toc();
The second one did not finish the job in more than 1 minutes. So what is the matter here with Julia? This piece of code happen to run both slower and overflow in Julia? Is there any problem in my style? I think one reason to convert from Matlab to Julia is the speed, and I used to think that Julia handles big number arithmetic operations by default. Now, it looks like these are not correct. Can someone explain?
There are a couple of things going on here.
Firstly, unlike Matlab, the x is an array of machine integers, not floating point values. This appears to be the main difference in speed, as it is unable to use BLAS routines for linear algebra.
You need to do either
x = 1.0:t*n
or explicitly convert it via
x = float(x)
This is also the reason for a different answer: Julia uses machine arithmetic, which for integers will wrap around on overflow (hence the negative number). You won't have this problem with floating point values.
(Julia does have arbitrary-precision integers, but doesn't promote by default: instead you would need to promote them yourself via big(n))
This will give you some speed up, but you can do better. Julia does require a slightly different programming style than Matlab. See the links that Isaiah provided in the comments above.
In regards to your specific examples
sum(x[:].* x[:])
This is slow as it creates 3 intermediate vectors (2 copies of x[:], though hopefully this will change in future, and their product).
Similarly,
trace(x*x')
computes an intermediate matrix (most of which is not used in the trace).
My suggestion would be to use
dot(vec(x),vec(x))
vec(x) just reshapes x into a vector (so no copying), and dot is the usual sum-product. You could also "roll-your-own":
function test(x)
s = zero(eltype(x)) # this prevents type-instability
for xi in x
s += xi*xi
end
s
end
test(x)
(this needs to be in a function for the JIT compiler to work its magic). It should be reasonably fast, though probably still not as fast as dot, which uses BLAS calls underneath.
I have a scalar function f([x,y],[i,j])= exp(-norm([x,y]-[i,j])^2/sigma^2) which receives two 2-dimensional vectors as input (norm here implements the Euclidean norm). The values of x,i range in 1:w and the values y,j range in 1:h. I want to create a cell array X such that X{x,y} will contain a w x h matrix such that X{x,y}(i,j) = f([x,y],[i,j]). This can obviously be done using 4 nested loops like so:
for x=1:w;
for y=1:h;
X{x,y}=zeros(w,h);
for i=1:w
for j=1:h
X{x,y}(i,j)=f([x,y],[i,j])
end
end
end
end
This is however extremely inefficient. I would very much appreciate an efficient way to create X.
The one way to do this is to remove the 2 innermost loops and replace then with a vectorised version. By the look of your f function this shouldn't be too bad
First we need to construct two matrices containing the 1 to w on every row and 1 to h on every column like so
wMat=repmat(1:w,h,1);
hMat=repmat(1:h,w,1)';
This is going to represent the inner two loops, and the transpose will allow us to get all combinations. Now we can vectorise the calculation (f([x,y],[i,j])= exp(-norm([x,y]-[i,j])^2/sigma^2)):
for x=1:w;
for y=1:h;
temp1=sqrt((x-wMat).^2+(y-hMat).^2);
X{x,y}=exp(temp1/(sigma^2));
end
end
Where we have computed the Euclidean norm for all pairs of nodes in the inner loops at once.
Some discussion and code
The trick here is to perform the norm-calculations with numeric arrays and save the results into a cell array version as late as possible. For performing the norm-calculations you can take help of ndgrid, bsxfun and some permute + reshape to give it the "shape" as needed for the final cell array version. So, here's the vectorized approach to perform these tasks -
%// Create x-y/i-j values to be used for calculation of function values
[xi,yi] = ndgrid(1:w,1:h);
%// Get the norm values
normvals = sqrt(bsxfun(#minus,xi(:),xi(:).').^2 + ...
bsxfun(#minus,yi(:),yi(:).').^2);
%// Get the actual function values
vals = exp(-normvals.^2/sigma^2);
%// Get the values into blocks of a 4D array and then re-arrange to match
%// with the shape of numeric array version of X
blks = reshape(permute(reshape(vals, w*h, h, []), [2 1 3]), h, w, h, w);
arranged_blks = reshape(permute(blks,[2 3 1 4]),w,h,w,h);
%// Finally get the cell array version
X = squeeze(mat2cell(arranged_blks,w,h,ones(1,w),ones(1,h)));
Benchmarking and runtimes
After improving the original loopy code with pre-allocation for X and function-inling f, runtime-benchmarks were performed with it against the proposed vectorized approach with datasizes as w, h = 60 and the runtime results thus obtained were -
----------- With Improved loopy code
Elapsed time is 41.227797 seconds.
----------- With Vectorized code
Elapsed time is 2.116782 seconds.
This suggested a whooping close to 20x speedup with the proposed solution!
For extremely huge datasizes
If you are dealing with huge datasizes, essentially you are not giving enough memory for bsxfun to work with, and bsxfun is known to use up a lot of memory for giving you a performance-efficient vectorized solution. So, for such huge-datasize cases, you can use the following loopy approach to replace normvals calculations that was listed in the earlier bsxfun based solution -
%// Get the norm values
nx = numel(xi);
normvals = zeros(nx,nx);
for ii = 1:nx
normvals(:,ii) = sqrt( (xi(:) - xi(ii)).^2 + (yi(:) - yi(ii)).^2 );
end
It seems to me that when you run through the cycle for x=w, y=h, you are calculating all the values you need at once. So you don't need recalculate them. Once you have this:
for i=1:w
for j=1:h
temp(i,j)=f([x,y],[i,j])
end
end
Then, e.g. X{1,1} is just temp(1,1), X{2,2} is just temp(1:2,1:2), and so on. If you can vectorise the calculation of f (norm here is just the Euclidean norm of that vector?) then it will get even simpler.
I need to calculate f(x)=exp(A*x) repeatedly for a tiny, variable column vector x and a huge, constant matrix A (many rows, few columns). In other words, the x are few, but the A*x are many. My problem dimensions are such that A*x takes about as much runtime as the exp() part.
Apart from Taylor expansion and pre-calculating a range of values exp(y) (assuming known the range y of values of A*x), which I haven't managed to speed up considerably (while maintaining accuracy) with respect to what MATLAB is doing on its own, I am thinking about analytically restating the problem in order to be able to precalculate some values.
For example, I find that exp(A*x)_i = exp(\sum_j A_ij x_j) = \prod_j exp(A_ij x_j) = \prod_j exp(A_ij)^x_j
This would allow me to precalculate exp(A) once, but the required exponentiation in the loop is as costly as the original exp() function call, and the multiplications (\prod) have to be carried out in addition.
Is there any other idea that I could follow, or solutions within MATLAB that I may have missed?
Edit: some more details
A is 26873856 by 81 in size (yes, it's that huge), so x is 81 by 1.
nnz(A) / numel(A) is 0.0012, nnz(A*x) / numel(A*x) is 0.0075. I already use a sparse matrix to represent A, however, exp() of a sparse matrix is not sparse any longer. So in fact, I store x non-sparse and I calculate exp(full(A*x)) which turned out to be as fast/slow as full(exp(A*x)) (I think A*x is non-sparse anyway, since x is non-sparse.) exp(full(A*sparse(x))) is a way to have a sparse A*x, but is slower. Even slower variants are exp(A*sparse(x)) (with doubled memory impact for a non-sparse matrix of type sparse) and full(exp(A*sparse(x)) (which again yields a non-sparse result).
sx = sparse(x);
tic, for i = 1 : 10, exp(full(A*x)); end, toc
tic, for i = 1 : 10, full(exp(A*x)); end, toc
tic, for i = 1 : 10, exp(full(A*sx)); end, toc
tic, for i = 1 : 10, exp(A*sx); end, toc
tic, for i = 1 : 10, full(exp(A*sx)); end, toc
Elapsed time is 1.485935 seconds.
Elapsed time is 1.511304 seconds.
Elapsed time is 2.060104 seconds.
Elapsed time is 3.194711 seconds.
Elapsed time is 4.534749 seconds.
Yes, I do calculate element-wise exp, I update the above equation to reflect that.
One more edit: I tried to be smart, with little success:
tic, for i = 1 : 10, B = exp(A*x); end, toc
tic, for i = 1 : 10, C = 1 + full(spfun(#(x) exp(x) - 1, A * sx)); end, toc
tic, for i = 1 : 10, D = 1 + full(spfun(#(x) exp(x) - 1, A * x)); end, toc
tic, for i = 1 : 10, E = 1 + full(spfun(#(x) exp(x) - 1, sparse(A * x))); end, toc
tic, for i = 1 : 10, F = 1 + spfun(#(x) exp(x) - 1, A * sx); end, toc
tic, for i = 1 : 10, G = 1 + spfun(#(x) exp(x) - 1, A * x); end, toc
tic, for i = 1 : 10, H = 1 + spfun(#(x) exp(x) - 1, sparse(A * x)); end, toc
Elapsed time is 1.490776 seconds.
Elapsed time is 2.031305 seconds.
Elapsed time is 2.743365 seconds.
Elapsed time is 2.818630 seconds.
Elapsed time is 2.176082 seconds.
Elapsed time is 2.779800 seconds.
Elapsed time is 2.900107 seconds.
Computers don't really do exponents. You would think they do, but what they do is high-accuracy polynomial approximations.
References:
http://www.math.vanderbilt.edu/~esaff/texts/13.pdf
http://deepblue.lib.umich.edu/bitstream/handle/2027.42/33109/0000495.pdf
http://www.cs.yale.edu/homes/sachdeva/pubs/fast-algos-via-approx-theory.pdf
The last reference looked quite nice. Perhaps it should have been first.
Since you are working on images, you likely have discrete number of intensity levels (255 typically). This can allow reduced sampling, or lookups, depending on the nature of "A". One way to check this is to do something like the following for a sufficiently representative group of values of "x":
y=Ax
cdfplot(y(:))
If you were able to pre-segment your images into "more interesting" and "not as interesting" - like if you were looking at an x-ray being able to trim out all the "outside the human body" locations and clamp them to zero to pre-sparsify your data, that could reduce your number of unique values. You might consider the previous for each unique "mode" inside the data.
My approaches would include:
look at alternate formulations of exp(x) that are lower accuracy but higher speed
consider table lookups if you have few enough levels of "x"
consider a combination of interpolation and table lookups if you have "slightly too many" levels to do a table lookup
consider a single lookup (or alternate formulation) based on segmented mode. If you know it is a bone and are looking for a vein, then maybe it should get less high-cost data processing applied.
Now I have to ask myself why would you be living in so many iterations of exp(A*x)*x and I think you might be switching back and forth between frequency/wavenumber domain and time/space domain. You also might be dealing with probabilities using exp(x) as a basis, and doing some Bayesian fun. I don't know that exp(x) is a good conjugate prior, so I'm going to go with the fourier material.
Other options:
- consider use of fft, fft2, or fftn given your matrices - they are fast and might do part of what you are looking for.
I am sure there is a forier domain variation on the following:
https://mathoverflow.net/questions/34173/fast-matrix-multiplication
http://www-cc.cs.uni-saarland.de/media/oldmaterial/bc.pdf
http://arxiv.org/PS_cache/math/pdf/0511/0511460v1.pdf
You might be able to mix the lookup with a compute using the woodbury matrix. I would have to think about that some to be sure though. (link) At one point I knew that everything that mattered (CFD, FEA, FFT) were all about the matrix inversion, but I have since forgotten the particular details.
Now, if you are living in MatLab then you might consider using "coder" which converts MatLab code to c-code. No matter how much fun an interpreter may be, a good c-compiler can be a lot faster. The mnemonic (hopefully not too ambitious) that I use is shown here: link starting around 13:49. It is really simple, but it shows the difference between a canonical interpreted language (python) and compiled version of the same (cython/c).
I'm sure that if I had some more specifics, and was requested to, then I could engage more aggressively in a more specifically relevant answer.
You might not have a good way to do it on conventional hardware, buy you might consider something like a GPGPU. CUDA and its peers have massively parallel operations that allow substantial speedup for the cost of a few video cards. You can have thousands of "cores" (overglorified pipelines) doing the work of a few ALU's and if the job is properly parallelizable (as this looks like) then it can get done a LOT faster.
EDIT:
I was thinking about Eureqa. One option that I would consider if I had some "big iron" for development but not production would be to use their Eureqa product to come up with a fast enough, accurate enough approximation.
If you performed a 'quick' singular value decomposition of your "A" matrix, you would find that the dominant performance is governed by 81 eigenvectors. I would look at the eigenvalues and see if there were only a few of those 81 eigenvectors providing the majority of the information. If that was the case, then you can clamp the others to zero, and construct a simple transformation.
Now, if it were me, I would want to get "A" out of the exponent. I'm wondering if you can look at the 81x81 eigenvector matrix and "x" and think a little about linear algebra, and what space you are projecting your vectors into. Is there any way that you can make a function that looks like the following:
f(x) = B2 * exp( B1 * x )
such that the
B1 * x
is much smaller rank than your current
Ax.
I am just learning MATLAB and I find it hard to understand the performance factors of loops vs vectorized functions.
In my previous question: Nested for loops extremely slow in MATLAB (preallocated) I realized that using a vectorized function vs. 4 nested loops made a 7x times difference in running time.
In that example instead of looping through all dimensions of a 4 dimensional array and calculating median for each vector, it was much cleaner and faster to just call median(stack, n) where n meant the working dimension of the median function.
But median is just a very easy example and I was just lucky that it had this dimension parameter implemented.
My question is that how do you write a function yourself which works as efficiently as one which has this dimension range implemented?
For example you have a function my_median_1D which only works on a 1-D vector and returns a number.
How do you write a function my_median_nD which acts like MATLAB's median, by taking an n-dimensional array and a "working dimension" parameter?
Update
I found the code for calculating median in higher dimensions
% In all other cases, use linear indexing to determine exact location
% of medians. Use linear indices to extract medians, then reshape at
% end to appropriate size.
cumSize = cumprod(s);
total = cumSize(end); % Equivalent to NUMEL(x)
numMedians = total / nCompare;
numConseq = cumSize(dim - 1); % Number of consecutive indices
increment = cumSize(dim); % Gap between runs of indices
ixMedians = 1;
y = repmat(x(1),numMedians,1); % Preallocate appropriate type
% Nested FOR loop tracks down medians by their indices.
for seqIndex = 1:increment:total
for consIndex = half*numConseq:(half+1)*numConseq-1
absIndex = seqIndex + consIndex;
y(ixMedians) = x(absIndex);
ixMedians = ixMedians + 1;
end
end
% Average in second value if n is even
if 2*half == nCompare
ixMedians = 1;
for seqIndex = 1:increment:total
for consIndex = (half-1)*numConseq:half*numConseq-1
absIndex = seqIndex + consIndex;
y(ixMedians) = meanof(x(absIndex),y(ixMedians));
ixMedians = ixMedians + 1;
end
end
end
% Check last indices for NaN
ixMedians = 1;
for seqIndex = 1:increment:total
for consIndex = (nCompare-1)*numConseq:nCompare*numConseq-1
absIndex = seqIndex + consIndex;
if isnan(x(absIndex))
y(ixMedians) = NaN;
end
ixMedians = ixMedians + 1;
end
end
Could you explain to me that why is this code so effective compared to the simple nested loops? It has nested loops just like the other function.
I don't understand how could it be 7x times faster and also, that why is it so complicated.
Update 2
I realized that using median was not a good example as it is a complicated function itself requiring sorting of the array or other neat tricks. I re-did the tests with mean instead and the results are even more crazy:
19 seconds vs 0.12 seconds.
It means that the built in way for sum is 160 times faster than the nested loops.
It is really hard for me to understand how can an industry leading language have such an extreme performance difference based on the programming style, but I see the points mentioned in the answers below.
Update 2 (to address your updated question)
MATLAB is optimized to work well with arrays. Once you get used to it, it is actually really nice to just have to type one line and have MATLAB do the full 4D looping stuff itself without having to worry about it. MATLAB is often used for prototyping / one-off calculations, so it makes sense to save time for the person coding, and giving up some of C[++|#]'s flexibility.
This is why MATLAB internally does some loops really well - often by coding them as a compiled function.
The code snippet you give doesn't really contain the relevant line of code which does the main work, namely
% Sort along given dimension
x = sort(x,dim);
In other words, the code you show only needs to access the median values by their correct index in the now-sorted multi-dimensional array x (which doesn't take much time). The actual work accessing all array elements was done by sort, which is a built-in (i.e. compiled and highly optimized) function.
Original answer (about how to built your own fast functions working on arrays)
There are actually quite a few built-ins that take a dimension parameter: min(stack, [], n), max(stack, [], n), mean(stack, n), std(stack, [], n), median(stack,n), sum(stack, n)... together with the fact that other built-in functions like exp(), sin() automatically work on each element of your whole array (i.e. sin(stack) automatically does four nested loops for you if stack is 4D), you can built up a lot of functions that you might need just be relying on the existing built-ins.
If this is not enough for a particular case you should have a look at repmat, bsxfun, arrayfun and accumarray which are very powerful functions for doing things "the MATLAB way". Just search on SO for questions (or rather answers) using one of these, I learned a lot about MATLABs strong points that way.
As an example, say you wanted to implement the p-norm of stack along dimension n, you could write
function result=pnorm(stack, p, n)
result=sum(stack.^p,n)^(1/p);
... where you effectively reuse the "which-dimension-capability" of sum.
Update
As Max points out in the comments, also have a look at the colon operator (:) which is a very powerful tool for selecting elements from an array (or even changing it shape, which is more generally done with reshape).
In general, have a look at the section Array Operations in the help - it contains repmat et al. mentioned above, but also cumsum and some more obscure helper functions which you should use as building blocks.
Vectorization
In addition to whats already been said, you should also understand that vectorization involves parallelization, i.e. performing concurrent operations on data as opposed to sequential execution (think SIMD instructions), and even taking advantage of threads and multiprocessors in some cases...
MEX-files
Now although the "interpreted vs. compiled" point has already been argued, no one mentioned that you can extend MATLAB by writing MEX-files, which are compiled executables written in C, that can be called directly as normal function from inside MATLAB. This allows you to implement performance-critical parts using a lower-level language like C.
Column-major order
Finally, when trying to optimize some code, always remember that MATLAB stores matrices in column-major order. Accessing elements in that order can yield significant improvements compared to other arbitrary orders.
For example, in your previous linked question, you were computing the median of set of stacked images along some dimension. Now the order in which those dimensions are ordered greatly affect the performance. Illustration:
%# sequence of 10 images
fPath = fullfile(matlabroot,'toolbox','images','imdemos');
files = dir( fullfile(fPath,'AT3_1m4_*.tif') );
files = strcat(fPath,{filesep},{files.name}'); %'
I = imread( files{1} );
%# stacked images along the 1st dimension: [numImages H W RGB]
stack1 = zeros([numel(files) size(I) 3], class(I));
for i=1:numel(files)
I = imread( files{i} );
stack1(i,:,:,:) = repmat(I, [1 1 3]); %# grayscale to RGB
end
%# stacked images along the 4th dimension: [H W RGB numImages]
stack4 = permute(stack1, [2 3 4 1]);
%# compute median image from each of these two stacks
tic, m1 = squeeze( median(stack1,1) ); toc
tic, m4 = median(stack4,4); toc
isequal(m1,m4)
The timing difference was huge:
Elapsed time is 0.257551 seconds. %# stack1
Elapsed time is 17.405075 seconds. %# stack4
Could you explain to me that why is this code so effective compared to the simple nested loops? It has nested loops just like the other function.
The problem with nested loops is not the nested loops themselves. It's the operations you perform inside.
Each function call (especially to a non-built-in function) generates a little bit of overhead; more so if the function performs e.g. error checking that takes the same amount of time regardless of input size. Thus, if a function has only a 1 ms overhead, if you call it 1000 times, you will have wasted a second. If you can call it once to perform a vectorized calculation, you pay overhead only once.
Furthermore, the JIT compiler (pdf) can help vectorize simple for-loops, where you, for example, only perform basic arithmetic operations. Thus, the loops with simple calculations in your post are sped up by a lot, while the loops calling median are not.
In this case
M = median(A,dim) returns the median values for elements along the dimension of A specified by scalar dim
But with a general function you can try splitting your array with mat2cell (which can work with n-D arrays and not just matrices) and applying your my_median_1D function through cellfun. Below I will use median as an example to show that you get equivalent results, but instead you can pass it any function defined in an m-file, or an anonymous function defined with the #(args) notation.
>> testarr = [[1 2 3]' [4 5 6]']
testarr =
1 4
2 5
3 6
>> median(testarr,2)
ans =
2.5000
3.5000
4.5000
>> shape = size(testarr)
shape =
3 2
>> cellfun(#median,mat2cell(testarr,repmat(1,1,shape(1)),[shape(2)]))
ans =
2.5000
3.5000
4.5000