Performance of GPU vs CPU

Illustrating Three Approaches to GPU Computing: The Mandelbrot Set

As you may know (a few post earlier), our research lab has equipped by a GPU (Graphics Processing Unit) server, called Tesla k40, and today I did my first experiment to realize the computational power of this parallel hardware versus 5th Generation Intel® Core™ i7 Processor.

The following example, taken from Matlab, shows how fast GPU performance can speed up computing. Using Parallel Computing Toolbox™ this code is then adapted to make use of GPU hardware in three ways:

1. Using the existing algorithm but with GPU data as input 
2. Using arrayfun to perform the algorithm on each element independently 
3. Using the MATLAB/CUDA interface to run some existing CUDA/C++ code 

Setup

The values below specify a highly zoomed part of the Mandelbrot Set in the valley between the main cardioid and the p/q bulb to its left.

A 1000x1000 grid of real parts (X) and imaginary parts (Y) is created between these limits and the Mandelbrot algorithm is iterated at each grid location. For this particular location 500 iterations will be enough to fully render the image.

Below is an implementation of the Mandelbrot Set using standard MATLAB commands running on the CPU.

Using gpuArray

When MATLAB encounters data on the GPU, calculations with that data are performed on the GPU. The class gpuArray provides GPU versions of many functions that you can use to create data arrays, including the linspace, logspace, and meshgrid functions needed here. Similarly, the count array is initialized directly on the GPU using the functionones.

With these changes to the data initialization the calculations will now be performed on the GPU:

Element-wise Operation

Noting that the algorithm is operating equally on every element of the input, we can place the code in a helper function and call it using arrayfun. For GPU array inputs, the function used with arrayfun gets compiled into native GPU code. In this case we placed the loop in pctdemo_processMandelbrotElement.m:

Note that an early abort has been introduced because this function processes only a single element. For most views of the Mandelbrot Set a significant number of elements stop very early and this can save a lot of processing. The for loop has also been replaced by a while loop because they are usually more efficient. This function makes no mention of the GPU and uses no GPU-specific features - it is standard MATLAB code.

Using arrayfun means that instead of many thousands of calls to separate GPU-optimized operations (at least 6 per iteration), we make one call to a parallelized GPU operation that performs the whole calculation. This significantly reduces overhead.

Working with CUDA

In Experiments in MATLAB improved performance is achieved by converting the basic algorithm to a C-Mex function. If you are willing to do some work in C/C++, then you can use Parallel Computing Toolbox to call pre-written CUDA kernels using MATLAB data. You do this with the parallel.gpu.CUDAKernel feature.

A CUDA/C++ implementation of the element processing algorithm has been hand-written in pctdemo_processMandelbrotElement.cu. This must then be manually compiled using nVidia's NVCC compiler to produce the assembly-level pctdemo_processMandelbrotElement.ptx (.ptx stands for "Parallel Thread eXecution language").

The CUDA/C++ code is a little more involved than the MATLAB versions we have seen so far, due to the lack of complex numbers in C++. However, the essence of the algorithm is unchanged:

                         

Summary

This example has shown three ways in which a MATLAB algorithm can be adapted to make use of GPU hardware:

1. Convert the input data to be on the GPU using gpuArray, leaving the algorithm unchanged
2. Use arrayfun on a gpuArray input to perform the algorithm on each element of the input independently
3. Use parallel.gpu.CUDAKernel to run some existing CUDA/C++ code using MATLAB data

Code

 
maxIterations = 500;
gridSize = 1000;
xlim = [-0.748766713922161, -0.748766707771757];
ylim = [ 0.123640844894862,  0.123640851045266];
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
t = tic();
x = linspace( xlim(1), xlim(2), gridSize );
y = linspace( ylim(1), ylim(2), gridSize );
[xGrid,yGrid] = meshgrid( x, y );
z0 = xGrid + 1i*yGrid;
count = ones( size(z0) );
% Calculate
z = z0;
for n = 0:maxIterations
    z = z.*z + z0;
    inside = abs( z )<=2;
    count = count + inside;
end
count = log( count );
% Show
figure
hold on
cpuTime = toc( t );
fig = gcf;
fig.Position = [200 200 600 600];
imagesc( x, y, count );
axis image
colormap( [jet();flipud( jet() );0 0 0] );
title( sprintf( '%1.2fsecs (without GPU)', cpuTime ) );
hold off
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Setup
t = tic();
x = gpuArray.linspace( xlim(1), xlim(2), gridSize );
y = gpuArray.linspace( ylim(1), ylim(2), gridSize );
[xGrid,yGrid] = meshgrid( x, y );
z0 = complex( xGrid, yGrid );
count = ones( size(z0), 'gpuArray' );

% Calculate
z = z0;
for n = 0:maxIterations
    z = z.*z + z0;
    inside = abs( z )<=2;
    count = count + inside;
end
count = log( count );

% Show
figure
fig = gcf;
fig.Position = [200 200 600 600];
hold on
count = gather( count ); % Fetch the data back from the GPU
naiveGPUTime = toc( t );
imagesc( x, y, count )
axis image
colormap( [jet();flipud( jet() );0 0 0] );
title( sprintf( '%1.3fsecs (naive GPU) = %1.1fx faster', ...
    naiveGPUTime, cpuTime/naiveGPUTime ) )
hold off
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% function count = pctdemo_processMandelbrotElement(x0,y0,maxIterations)
% z0 = complex(x0,y0);
% z = z0;
% count = 1;
% while (count <= maxIterations) && (abs(z) <= 2)
%     count = count + 1;
%     z = z*z + z0;
% end
% count = log(count);

% Setup
t = tic();
x = gpuArray.linspace( xlim(1), xlim(2), gridSize );
y = gpuArray.linspace( ylim(1), ylim(2), gridSize );
[xGrid,yGrid] = meshgrid( x, y );

% Calculate
count = arrayfun( @pctdemo_processMandelbrotElement, ...
                  xGrid, yGrid, maxIterations );

% Show
figure
fig = gcf;
fig.Position = [200 200 600 600];
hold on
count = gather( count ); % Fetch the data back from the GPU
gpuArrayfunTime = toc( t );
imagesc( x, y, count )
axis image
colormap( [jet();flipud( jet() );0 0 0] );
title( sprintf( '%1.3fsecs (GPU arrayfun) = %1.1fx faster', ...
    gpuArrayfunTime, cpuTime/gpuArrayfunTime ) );
hold off
% Load the kernel
cudaFilename = 'pctdemo_processMandelbrotElement.cu';
ptxFilename = ['pctdemo_processMandelbrotElement.',parallel.gpu.ptxext];
kernel = parallel.gpu.CUDAKernel( ptxFilename, cudaFilename );

% Setup
t = tic();
x = gpuArray.linspace( xlim(1), xlim(2), gridSize );
y = gpuArray.linspace( ylim(1), ylim(2), gridSize );
[xGrid,yGrid] = meshgrid( x, y );

% Make sure we have sufficient blocks to cover all of the locations
numElements = numel( xGrid );
kernel.ThreadBlockSize = [kernel.MaxThreadsPerBlock,1,1];
kernel.GridSize = [ceil(numElements/kernel.MaxThreadsPerBlock),1];

% Call the kernel
count = zeros( size(xGrid), 'gpuArray' );
count = feval( kernel, count, xGrid, yGrid, maxIterations, numElements );

% Show
figure
fig = gcf;
fig.Position = [200 200 600 600];
hold on
count = gather( count ); % Fetch the data back from the GPU
gpuCUDAKernelTime = toc( t );
imagesc( x, y, count )
axis image
colormap( [jet();flipud( jet() );0 0 0] );
title( sprintf( '%1.3fsecs (GPU CUDAKernel) = %1.1fx faster', ...
    gpuCUDAKernelTime, cpuTime/gpuCUDAKernelTime ) );
hold off