SVD speed in CPU and GPU

Question

I\'m testing svd in Matlab R2014a and it seems that there is no CPU vs GPU speedup. I\'m using a GTX 460 car

Answer 1

Generally SVD is a difficult to paralellize routine. You can check here that with a high end Tesla card, the speedup is not very impressive.

You have a GTX460 card - Fermi architecture. The card is optimized for gaming (single precision computations), not HPC (double precision computation). The Single Precision / Double Precision throughput ratio is 12. So the card has 873 GFLOPS SP / 72 GFLOPS DP. Check here.

So if the Md array uses double precision elements, then the computation on it would be rather slow. Also there's a high chance that when calling the CPU routine, all CPU cores will get utilized, reducing the possible gain of running the routine on the GPU. Plus, in the GPU run you pay time for transferring the buffer to the device.

Per Divakar's suggestion, you could use Md = single(Md) to convert your array to single precision and run the benchmark again. You can try and go with a bigger dataset size to see if something changes. I don't expect to much gain for this routine on your GPU.

Update 1:

After you posted the results, I saw that the DP/SP time ratio is 2. On the CPU side this is normal, because you can fit 2 times less double values in SSE registers. However, a ratio of only 2 on the GPU side means that the gpu code does not make best use of the SM cores - because the theoretical ratio is 12. In other words, I would have expected much better SP performance for an optimized code, compared to DP. It seems that this is not the case.

Answer 2

The issue

First of all, I have replicated your issue in Matlab2016b using the following code:

clear all
close all
clc

Nrows = 2500;
Ncols = 2500;

NumTests = 10;

h_A = rand(Nrows, Ncols);
d_A = gpuArray.rand(Nrows, Ncols);

timingCPU = 0;
timingGPU = 0;

for k = 1 : NumTests
    % --- Host
    tic
    [h_U, h_S, h_V] = svd(h_A);
%     h_S = svd(h_A);
    timingCPU = timingCPU + toc;

    % --- Device
    tic
    [d_U, d_S, d_V] = svd(d_A);
%     d_S = svd(d_A);
    timingGPU = timingGPU + toc;
end

fprintf('Timing CPU = %f; Timing GPU = %f\n', timingCPU / NumTests, timingGPU / NumTests);

By the above code, it is possible to either compute the singular values only or compute the full SVD including the singular vectors. It is possible also to compare the different behavior of the CPU and GPU versions of the SVD code.

The timing is reported in the following table (timing in s; Intel Core i7-6700K CPU @ 4.00GHz, 16288 MB, Max threads(8), GTX 960):

              Sing. values only | Full SVD         | Sing. val. only | Full
                                |                  |                 |
Matrix size   CPU      GPU      | CPU       GPU    |                 |
                                |                  |                 |
 200 x  200   0.0021    0.043   |  0.0051    0.024 |   0.098         |  0.15
1000 x 1000   0.0915    0.3     |  0.169     0.458 |   0.5           |  2.3
2500 x 2500   3.35      2.13    |  4.62      3.97  |   2.9           |  23
5000 x 5000   5.2      13.1     | 26.6      73.8   |  16.1           | 161

The first 4 columns refer to a comparison between the CPU and GPU Matlab versions of the svd routine when it is used to calculate the singular values only or the full SVD. As it can be seen, the GPU version can be significantly slower than the GPU one. The motivation has been already pointed out in some answers above: there is an inherent difficulty to parallelize the SVD computation.

Using cuSOLVER?

At this point, the obvious question is: can we get some speedup with cuSOLVER? Indeed, we could use mexFiles to make the cuSOLVER routines run under Matlab. Unfortunately, the situation with cuSOLVER is even worse, as it can be deduced from the last two columns of the above table. Such columns report the timing of the codes at Singular values calculation only with CUDA and Parallel implementation for multiple SVDs using CUDA using cusolverDnSgesvd for the singular values only calculation and full SVD calculation, respectively. As it can be seen, cuSOLVER's cusolverDnSgesvd performs even worser than Matlab, if one takes into account that it deals with single precision, while Matlab with double precision.

The motivation for this behavior is further explained at cusolverDnCgesvd performance vs MKL where Joe Eaton, manager of cuSOLVER library, says

I understand the confusion here. We do provide a decent speedup for LU, QR and LDL^t factorizations, which is what we would like to say for SVD as well. Our purpose with cuSOLVER is to provide dense and sparse direct solvers as part of the CUDA toolkit for the first time; we have to start somewhere. Since CULA is no longer supported, we felt it was urgent to get some functionality into the hands of developers in CUDA 7.0. Since CUDA runs on more that x86 host CPUs these days, cuSOLVER fills a need where there is no MKL. That being said, we can do better with SVD, but it will have to wait for the next CUDA release, priorities and timelines being tight already.

Using other libraries

At this point, other possibilities are using other libraries like

1. CULA;
2. MAGMA;
3. ArrayFire.

CULA is not offered for free, so I have not tried it.

I had some installation issues with MAGMA dependencies, so I have not investigated this point further (disclaimer: I expect that, with some more time, I could be able to solve such issues).

I then finally ended up with using ArrayFire.

Using ArrayFire, I had the following timing for the full SVD computation:

 200 x  200      0.036
1000 x 1000      0.2
2500 x 2500      4.5
5000 x 5000     29

As it can be seen, the timing is slightly higher, but now comparable, to the CPU case.

Here is the ArrayFire code:

#include <arrayfire.h>
#include <cstdio>
#include <cstdlib>
#include <fstream>

using namespace af;

int main(int argc, char *argv[])
{
    const int N = 1000;

    try {

        // --- Select a device and display arrayfire info
        int device = argc > 1 ? atoi(argv[1]) : 0;
        af::setDevice(device);
        af::info();

        array A = randu(N, N, f64);
        af::array U, S, Vt;

        // --- Warning up
        timer time_last = timer::start();
        af::svd(U, S, Vt, A);
        S.eval();
        af::sync();
        double elapsed = timer::stop(time_last);
        printf("elapsed time using start and stop = %g ms \n", 1000.*elapsed);

        time_last = timer::start();
        af::svd(U, S, Vt, A);
        S.eval();
        af::sync();
        elapsed = timer::stop(time_last);
        printf("elapsed time using start and stop = %g ms \n", 1000.*elapsed);

    }
    catch (af::exception& e) {

        fprintf(stderr, "%s\n", e.what());
        throw;
    }

    return 0;
}

Answer 3

As VAndrei has already stated, the SVD is an algorithm which is difficult to parallelize.

Your main problem is the size of your matrix. The performance of the SVD drops rapidly with a growing matrix size. So your main goal should be to reduce the size of the matrix. This can be accomplished using Gaussian normal equations (which is basically a reduction of an overdetermined linear system in the least-squares sense).

This can be done by simply multiplying the transpose onto the matrix:

MhReduced = Mh' * Mh;

This reduces your matrix to the size of cols*cols (if cols is the number of columns of Mh). Then you just call [U,S,V] = svd(MhReduced);

Note: Using this method may yield singular vectors with opposite sign (just important if you're comparing these methods).

If your matix is well-conditioned this should work without problems. However, in case of an ill-conditioned matrix, this method may fail to produce a usable result, whereas applying SVD directly could still yield a usable result due to SVD's robustness.

This should increase your performance immensly, at least with matrices big enough. Another advantage is that you can use much larger matrices. You'll probably won't have to use the GPU at all (since either matrices are so big that copying to GPU costs too much or after reduction the matrix is so small that the speedup of the GPU won't be big enough).

Also note that a large chunk of performance is lost, if you use return values. If you're only interested in the performance of the SVD caluclation, don't take any return values. If you are only interested in the "solution vector", just get V (and access the last column): [~,~, V] = svd(Mh);.

EDIT:

I've looked at your sample code, but I'm not sure what it is, you are calculating. Also I realized that it's rather hard to understand what I did with A'*A, so I will explain in detail.

Given a linear system with A*x=b, A denoting the coefficient matrix with m rows and n cols, x the solution vector and b the constant vector (both with m rows), a solution can be calculated as follows:

• if A is square (m=n): x = A^-1 * b,

• if A is not square (m!=n, m > n):

  A * x = b

  A'* A * x = A' * b

  x = (A' * A)^-1 * A'*b

A" = (A'*A)^-1 * A' is typically called pseudo-inverse. However this calculation does influence the condition number of the matrix negatively. A solution to this problem is using a singular value decomposition (SVD). If USV = svd(A) denotes the results of the SVD, the pseudo-inverse is given by VS"U', with S" is formed by taking the inverse of the non-zero elements of S. So A" = VS"U'.

x = A"*b

However since a SVD is rather costly, especially with large matrices. If matrix A is well-conditioned and very precicse results are not necessarily required (we're talking 1e-13 or 1e-14), the much faster approach by calculating the peseudo-inverse via (A'*A)^-1 * A can be used.

If your case actually is A*x=0, just use a SVD and read the last column vector from V, it is the solution.

If you use the SVD not to solve a linear system but for the results of U and S (as your example suggests), I'm not sure what I've posted will help you.

Sources: 1, 2, 3

Here is some sample code for you to test. Test it with large matrices, you will see that using (A'*A)^-1 * A' is much faster than the alternatives.

clear all

nbRows = 30000;
nbCols = 100;
% Matrix A
A = rand(nbRows,nbCols);

% Vector b
b = rand(nbRows,1);

% A*x=b

% Solve for x, using SVD
% [U,S,V]=svd(A,0);
% x= V*((U'*b)./diag(S))
tic
[U1,S1,V1]=svd(A,0);
x1= V1*((U1'*b)./diag(S1));
toc

tic
[U1,S1,V1]=svd(A,0);
x2 = V1*inv(S1)*U1'*b;
toc

% Solve for x, using manual pseudo-inverse
% A*x=b
% A'*A*x = A'*b
% x = (A'*A)^-1 * A'*b
tic
x3 = inv(A'*A) * A'*b;
toc

% Solve for x, let Matlab decide how (most likely SVD)
tic
x4 = A\b;
toc

Answer 4

I have tried to parallelize SVD on my laptop equipped with GTX 460 for over one months, which was also a part of my undergraduate thesis, I did so many experiments that I later discovered that MATLAB is extremely fast and outperforms my code, by the way, I used one side Jacobi, and I have not yet seen any paper that reveals an algorithm faster than svd of MATLAB. On GPU, the time cost of memory copy can be very high if you are not using an elegant model, I refer you to read more about CUDA. If you need any help, please contact me.