Today I’d like to introduce guest blogger Sarah Wait Zaranek who works for the MATLAB Marketing team. Sarah previously has written about speeding up code from a customer to get acceptable performance. She will be discussing how to use GPUs to accelerate your MATLAB applications.
- GPU Background
- Learning About Our GPU
- A Simple Example using Overloaded Functions
- Speedup For Our Simple Example
- Understanding and Limiting Your Data Transfer Overhead
- Solving the Wave Equation
- Code Changes to Run Algorithm on GPU
- Code Changes to Transfer Data to the GPU and Back Again
- Comparing CPU and GPU Execution Speeds
- Other Ways to Use the GPU with MATLAB
- Your Thoughts?
In this post, we first will introduce the basics of using the GPU with MATLAB and then move onto solving a 2nd-order wave equation using this GPU functionality. This blog post is inspired by a recent MATLAB Digest article on GPU Computing that I coauthored with one of our developers, Jill Reese. Since the original demo was made, the GPU functions available in MATLAB have grown. If you compare the code below to the code in the paper- they are slightly different, reflecting these new capabilites. Additionally, small changes were made to enable easier explanation of the code in this blog format. The GPU functionality shown in this post requires the Parallel Computing Toolbox.
Originally used to accelerate graphics rendering, GPUs are now increasingly applied to scientific calculations. Unlike a traditional CPU, which includes no more than a handful of cores, a GPU has a massively parallel array of integer and floating-point processors, as well as dedicated, high-speed memory. A typical GPU comprises hundreds of these smaller processors. These processors can be used to greatly speed-up particular types of applications.
A good rule of thumb is that your problem may be a good fit for the GPU if it is:
- Massively parallel: The computations can be broken down into hundreds or thousands of independent units of work. You will see the best performance when all of the cores are kept busy, exploiting the inherent parallel nature of the GPU. Seemingly simple, vectorized MATLAB calculations on arrays with hundreds of thousands of elements often can fit into this category.
- Computationally intensive: The time spent on computation significantly exceeds the time spent on transferring data to and from GPU memory. Because a GPU is attached to the host CPU via the PCI Express bus, the memory access is slower than with a traditional CPU. This means that your overall computational speedup is limited by the amount of data transfer that occurs in your algorithm.
Applications that do not satisfy these criteria might actually run more slowly on a GPU than on a CPU.
With that background, we can now start working with the GPU in MATLAB. Let's start first by querying our GPU to see just what we are working with:
ans = parallel.gpu.CUDADevice handle Package: parallel.gpu Properties: Name: 'Tesla C2050 / C2070' Index: 1 ComputeCapability: '2.0' SupportsDouble: 1 DriverVersion: 4 MaxThreadsPerBlock: 1024 MaxShmemPerBlock: 49152 MaxThreadBlockSize: [1024 1024 64] MaxGridSize: [65535 65535] SIMDWidth: 32 TotalMemory: 3.1820e+009 FreeMemory: 2.6005e+009 MultiprocessorCount: 14 ClockRateKHz: 1147000 ComputeMode: 'Default' GPUOverlapsTransfers: 1 KernelExecutionTimeout: 1 CanMapHostMemory: 1 DeviceSupported: 1 DeviceSelected: 1
We are running on a Tesla C2050. Currently, Parallel Computing Toolbox supports NVDIA GPUs with Compute Capability 1.3 or greater. Here is a link to see if your card is supported.
Over 100 operations (e.g. fft, ifft, eig) are now available as built-in MATLAB functions that can be executed directly on the GPU by providing an input argument of the type GPUArray. These GPU-enabled functions are overloaded—in other words, they operate differently depending on the data type of the arguments passed to them. Here is a list of all the overloaded functions.
Let's create a GPUArray and perform a fft using the GPU. However, let's first do this on the CPU so that we can see the difference in code and performance
A1 = rand(3000,3000); tic; B1 = fft(A1); time1 = toc;
To perform the same operation on the GPU, we first use gpuArray to transfer data from the MATLAB workspace to device memory. Then we can run fft, which is one of the overloaded functions on that data:
A2 = gpuArray(A1); tic; B2 = fft(A2); time2 = toc;
The second case is executed on the GPU rather than the CPU since its input (a GPUArray) is held on the GPU. The result, B2, is stored on the GPU. However, it is still visible in the MATLAB workspace. By running class(B2), we can see that it is also a GPUArray.
ans = parallel.gpu.GPUArray
To bring the data back to the CPU, we run gather.
B2 = gather(B2); class(B2)
ans = double
B2 is now on the CPU and has a class of double.
In this simple example, we can calculate the speedup of performing the fft on our data.
speedUp = time1/time2; disp(speedUp)
It looks like our fft is running ~3.5x faster on the GPU. This is pretty good, especially since fft is multi-threaded in core MATLAB. However, to perform a realistic comparison, we should really include the time spent transferring the vector to and from the GPU. If we do that - we find that our acceleration is greatly reduced. Let's see what happens if we include the time to do our data transfer.
tic; A3 = gpuArray(A1); B3 = fft(A3); B3 = gather(B3); time3 = toc; speedUp = time1/time3; disp(speedUp)
Data transfer overhead can become so significant that it degrades the application's overall performance, especially if you repeatedly exchange data between the CPU and GPU to execute relatively few computationally intensive operations. However not all hope is lost! By limiting the data transfer between the GPU and the CPU, we can still acheive speedup.
Instead of creating the data on the CPU and transferring it to the GPU - we can just create on the GPU directly. There are several constructors available as well as constructor-like functions such as meshgrid. Let's see how that effects our time. To be accurate, we should also retime the original serial code including the time it takes to generate the random matrix on the CPU.
tic; A4 = rand(3000,3000); B4 = fft(A4); time4 = toc; tic; A5 = parallel.gpu.GPUArray.rand(3000,3000); B5 = fft(A5); B5 = gather(B5); time5 = toc; speedUp = time4/time5; disp(speedUp);
This is better, although we still do see the influence of gathering the data back from the GPU. In this simple demo, the effect is exaggerated because we are running a single, already fast operation on the GPU. It is more efficient to perform several operations on the data while it is on the GPU, bringing the data back to the CPU only when required.
To put the above example into context, let's use the same GPU functionality on a more "real-world" problem. To do this, we are going to solve a second-order wave equation:
The algorithm we use to solve the wave equation combines a spectral method in space and a second-order central finite difference method in time. Specifically, we apply the Chebyshev spectral method, which uses Chebyshev polynomials as the basis functions. Our example is largely based on an example in Trefethen's book: "Spectral Methods in MATLAB"
When accelerating our alogrithm, we focus on speeding up the code within the main time stepping while-loop. The operations in that part of the code (e.g. fft and ifft, matrix multiplication) are all overloaded functions that work with the GPU. As a result, we do not need to change the algorithm in any way to execute it on a GPU. We get to simply transfer the data to the GPU and back to the CPU when finished.
function [vv,vvold] = stepsolution(vv,vvold,ii,N,dt,W1T,W2T,W3T,W4T,... WuxxT1,WuxxT2,WuyyT1,WuyyT2) V = [vv(ii,:) vv(ii,N:-1:2)]; U = real(fft(V.')).'; W1test = (U.*W1T).'; W2test = (U.*W2T).'; W1 = (real(ifft(W1test))).'; W2 = (real(ifft(W2test))).'; % Calculating 2nd derivative in x uxx(ii,ii) = W2(:,ii).* WuxxT1 - W1(:,ii).*WuxxT2; uxx([1,N+1],[1,N+1]) = 0; V = [vv(:,ii); vv((N:-1:2),ii)]; U = real(fft(V)); W1 = real(ifft(U.*W3T)); W2 = real(ifft(U.*W4T)); % Calculating 2nd derivative in y uyy(ii,ii) = W2(ii,:).* WuyyT1 - W1(ii,:).*WuyyT2; uyy([1,N+1],[1,N+1]) = 0; % Computing new value using 2nd order central finite difference in % time vvnew = 2*vv - vvold + dt*dt*(uxx+uyy); vvold = vv; vv = vvnew; end
The variables dt, x, index1, and index2 are calculated on the CPU and transferred over to the GPU using gpuArray.
N = 256; index1 = 1i*[0:N-1 0 1-N:-1]; index1 = gpuArray(index1);
For the rest of the variables, we create them directly on the GPU using the overloaded versions of the transpose operator ('), repmat ,and meshgrid.
W1T = repmat(index1,N-1,1);
When, we are done with all our calculations on the GPU, we use gather once to bring data back from the GPU. Note, that we don't have to transfer our data back to the CPU between time steps. This all comes together to look like this:
function vvg = WaveGPU(N,Nsteps) %% Solving 2nd Order Wave Equation Using Spectral Methods % This example solves a 2nd order wave equation: utt = uxx + uyy, with u = % 0 on the boundaries. It uses a 2nd order central finite difference in % time and a Chebysehv spectral method in space (using FFT). % % The code has been modified from an example in Spectral Methods in MATLAB % by Trefethen, Lloyd N. % Points in X and Y x = cos(pi*(0:N)/N); % using Chebyshev points % Send x to the GPU x = gpuArray(x); y = x'; % Calculating time step dt = 6/N^2; % Setting up grid [xx,yy] = meshgrid(x,y); % Calculate initial values vv = exp(-40*((xx-.4).^2 + yy.^2)); vvold = vv; ii = 2:N; index1 = 1i*[0:N-1 0 1-N:-1]; index2 = -[0:N 1-N:-1].^2; % Sending data to the GPU dt = gpuArray(dt); index1 = gpuArray(index1); index2 = gpuArray(index2); % Creating weights used for spectral differentiation W1T = repmat(index1,N-1,1); W2T = repmat(index2,N-1,1); W3T = repmat(index1.',1,N-1); W4T = repmat(index2.',1,N-1); WuxxT1 = repmat((1./(1-x(ii).^2)),N-1,1); WuxxT2 = repmat(x(ii)./(1-x(ii).^2).^(3/2),N-1,1); WuyyT1 = repmat(1./(1-y(ii).^2),1,N-1); WuyyT2 = repmat(y(ii)./(1-y(ii).^2).^(3/2),1,N-1); % Start time-stepping n = 0; while n < Nsteps [vv,vvold] = stepsolution(vv,vvold,ii,N,dt,W1T,W2T,W3T,W4T,... WuxxT1,WuxxT2,WuyyT1,WuyyT2); n = n + 1; end % Gather vvg back from GPU memory when done vvg = gather(vv);
We ran a benchmark in which we measured the amount of time it took to execute 50 time steps for grid sizes of 64, 128, 512, 1024, and 2048 on an Intel Xeon Processor X5650 and then using an NVIDIA Tesla C2050 GPU.
For a grid size of 2048, the algorithm shows a 7.5x decrease in compute time from more than a minute on the CPU to less than 10 seconds on the GPU. The log scale plot shows that the CPU is actually faster for small grid sizes. As the technology evolves and matures, however, GPU solutions are increasingly able to handle smaller problems, a trend that we expect to continue.
To accelerate an algorithm with multiple element-wise operations on a GPU, you can use arrayfun, which applies a function to each element of an GPUarray. Additionally if you have your own CUDA code, you can use the CUDAKernel interface to integrate this code with MATLAB.
Ben Tordoff's previous guest blog post Mandelbrots Sets on the GPU shows a great example using both of these capabilities.
Have you experimented with our new GPU functionality? Let us know what you think or if you have any questions by leaving a comment for this post here.
Get the MATLAB code
Published with MATLAB® 7.13
14 CommentsOldest to Newest
Very useful… if you have an Nvidia GPU… Any plans on supporting ATI?
We have chosen to support CUDA and not Open CL due to the fact that CUDA currently has the only ecosystem with all of the libraries necessary for technical computing. This may change as open CL libraries advance and evolve.
A quick search indicates your examples require special toolboxes like Parallel Computing. Is this correct?
If so, what toolboxes are required?
Indeed you need Parallel Computing Toolbox to access GPU functionality. See the last sentence in the Overview section.
There is also the open-source GPUMat project that can be used to try out GPUs.
The file utilities at
let you switch dynamically between CPU and GPU when loading data from a MAT-file through VM using a memmapfile object and supports the PCT, Jacket and GPUMat.
I have been wanting to use the GPUarray for calculating something similar to gevlike and the other similar loglikelihood functions. This function involves doing the same calculation on every single element in a long vector, and then adding all the results. I have had a case where i had to calculate this so many times that it took me a week to complete. I believe GPUarray would speed it up immensely. Unfortunately my gpu does not fulfill the requirements (it is nvidia, but just a little too old).
I know it is disappointing that we don’t support your card. Development didn’t take the decision lightly. However, they have chosen to require compute capability 1.3 because it was very important for our GPU solution to be able to fully support doubles, guarantee IEEE compliance for double-precision calculations and to provide cross-platform support.
I have a question. If I want to have element-by-element multiplication of each row of, say, 4096 x 4096 matrix with a 4096 row vector, what is the best way to do this on GPU?
On CPU this can be done using a for loop but there is no for loop on GPUs. Do you think making a 4096 x 4096 matrix out of the vector which each row being that same vector and then using matrix.*vecmatrix is a good idea?
Where can i find, how to write my own MATLAB code and run it on gpu (using CUDA)?
If I am understanding what you want to do correctly – I think repmat or bsxfun would the appropriate way to do it on the CPU. When moving to the GPU, repmat is currently supported – so I would do the following:
A = rand(4096,4096); % Using rand as a proxy for your data
Avector = rand(4096,1); % Using rand as a proxy for your data
% Calculation on CPU
AV = repmat(Avector,1,4096);
B = A.*AV;
% Caclulation on GPU
A = gpuArray(A); % Transfering to GPU
Avector = gpuArray(Avector); % Transfering to GPU
AV = repmat(Avector,1,4096);
B = A.*AV;
Hope this helps.
I am not sure specifically what you are wanting to do. If you want to do something besides using GPUArrays and overloaded functions, you can use the arrayfun option. To execute your MATLAB function on the GPU, you can use arrayfun with a function handle to the MATLAB function. To learn more about this – look in the documenation here:
You can also just read through all the general GPU capabilities in the documenation here:
I’m wandering why cpu randn is faster than gpu? It would be much more helpful for simulations to have much faster randn on GPU.
>> tic; parallel.gpu.GPUArray.randn(1e7, 1); toc
Elapsed time is 0.252773 seconds.
>> tic; randn(1e7, 1); toc
Elapsed time is 0.152477 seconds.
The numbers don’t seem to be as drastic one my machine, especially after running the commands once to warm everything up.
The GPU uses the combined multiplicative recursive generator to create uniform random values, and uses inversion for creating normal values. This is not the default stream in a client MATLAB session on the CPU. There it might take longer than the default random number stream in MATLAB. You can change the stream, if you wish.
gpu_stream = parallel.gpu.RandStream(‘CombRecursive’,'Seed’,seed);
A quote from one of our developers about this choice below:
It is rare that the speed of the random number generator has a significant effect on the overall application performance, but it does happen. In most cases, the computations that are done with the random numbers outweigh the cost of their generation. We should note that users rarely worry about the quality of the default random number generator in MATLAB because they trust that it yield “good” random numbers. We wanted that to remain true for the default random number generator the GPU, even if that meant that the random number generation was more computationally intensive.
We chose mrg32k3a (see http://www.mathworks.com/help/techdoc/math/brn4ixh.html#brvku_2 ) as the default random number generator on the GPU as it meets the following criteria:
- It is a very well-tested algorithm
- It has full support for parallel random number generation
- Its statistical properties (i.e. lack of correlation) are known in the parallel setting. Thus, it provides “good” random numbers on a single GPU, on a cluster of GPUs, or on a cluster of CPUs.
- It has a large number of parallel independent substreams, each of which has length 2^76, long enough to make it extremely unlikely that any customer will ever see any effects from exhausting the random number generator.
- It supports both the creation of random matrices and unrestricted usage within arrayfun to generate random scalars(*).
- It supports reproducibility. That is, one can reproduce previous results through retrieving, storing and later re-instating the random number generator state.
- It is supported on the CPU in MATLAB. Furthermore, the random numbers produced on the GPU match exactly those produced by that random number generator when used on the CPU in MATLAB.
Using mrg32k3a with the Inversion method to generate normally distributed random numbers allows us to meet all of the criteria listed above. This makes it a very good choice as default random number generator for the GPU, one that we are very comfortable recommending to our customers.
I really enjoy using PCT’s CUDAKernel functionality to obtain good performance improvement with the GPU in a MATLAB environment. It’s a great tool and dollar-for-dollar very hard to beat.
It seems to me, however, that this lower-level CUDA feature is hardly ever mentioned in your posts. I think that’s too bad.
My question is: are there any plans to provide the mex-like interface for PCT gpuArray objects such as currently exists for Jacket via their SDK? This would give MATLAB PCT users both the improved performance (by fully compiling past the *.ptx stage) as well as greater flexibility/extensibility (by allowing programmers access to native pointers to data on the GPU and interfacing those with other CUDA libraries or CUDA-accelerated codes – other than just individual kernels).