Question

我有一个MATLAB脚本运行两个相同功能的副本，我想在两个不同的内核上并行运行它们。

这是我正在调用的功能

function [outputVector] = ParallelPitchShift(inputVector, windowSize, hopSize, step)

%% Parameters

% Window size
winSize = windowSize;
% Space between windows
hop = hopSize;
% Pitch scaling factor
alpha = 2^(step/12);

% Intermediate constants
hopOut = round(alpha*hop);

% Hanning window for overlap-add
wn = hann(winSize*2+1);
wn = wn(2:2:end);

%% Source file

x = inputVector;

% Rotate if needed
if size(x,1) < size(x,2)

    x = transpose(x);

end

x = [zeros(hop*3,1) ; x];

%% Initialization

% Create a frame matrix for the current input
[y,numberFramesInput] = createFrames(x,hop,winSize);

% Create a frame matrix to receive processed frames
numberFramesOutput = numberFramesInput;
outputy = zeros(numberFramesOutput,winSize);

% Initialize cumulative phase
phaseCumulative = 0;

% Initialize previous frame phase
previousPhase = 0;

for index=1:numberFramesInput

%% Analysis

    % Get current frame to be processed
    currentFrame = y(index,:);

    % Window the frame
    currentFrameWindowed = currentFrame .* wn' / sqrt(((winSize/hop)/2));

    % Get the FFT
    currentFrameWindowedFFT = fft(currentFrameWindowed);

    % Get the magnitude
    magFrame = abs((currentFrameWindowedFFT));

    % Get the angle
    phaseFrame = angle(currentFrameWindowedFFT);

%% Processing    

    % Get the phase difference
    deltaPhi = phaseFrame - previousPhase;

    previousPhase = phaseFrame;

    % Remove the expected phase difference
    deltaPhiPrime = deltaPhi - hop * 2*pi*(0:(winSize-1))/winSize;

    % Map to -pi/pi range
    deltaPhiPrimeMod = mod(deltaPhiPrime+pi, 2*pi) - pi;

    % Get the true frequency
    trueFreq = 2*pi*(0:(winSize-1))/winSize + deltaPhiPrimeMod/hop;

    % Get the final phase
    phaseCumulative = phaseCumulative + hopOut * trueFreq;    

    % Remove the 60 Hz noise. This is not done for now but could be
    % achieved by setting some bins to zero.

%% Synthesis    

    % Get the magnitude
    outputMag = magFrame;

    % Produce output frame
    outputFrame = real(ifft(outputMag .* exp(j*phaseCumulative)));

    % Save frame that has been processed
    outputy(index,:) = outputFrame .* wn' / sqrt(((winSize/hopOut)/2));

end

%% Finalize

% Overlap add in a vector
outputTimeStretched = fusionFrames(outputy,hopOut);

% Resample with linearinterpolation
outputTime = interp1((0:(length(outputTimeStretched)-1)),outputTimeStretched,(0:alpha:(length(outputTimeStretched)-1)),'linear');

% Return the result
outputVector = outputTime;

return

这就是我要调用函数

clear

x = wavread('x2.wav');

y1 = ParallelPitchShift(x,1024,256,7);

y2 = ParallelPitchShift(x,1024,256,12);

output = x(1:417000)' + y1(1:417000) + y2(1:417000);

sound(x,44100)

pause

sound(output,44100)

可以这样做吗？请让我知道谢谢！

Answer 1

我不明白你所说的相同功能的两个副本是什么意思。你是指同一个函数的两个不同的输入参数，还是两个函数完全不同？

我假设它是前者（因为那是你写的）。在这种情况下，请尝试

somethingVector=[something, something2];
parfor i=1:2
    function(something(i),otherArgs)
end

请记住，在使用parfor循环之前，您应该像这样打开一个工作池：

matlabpool('open',2)

Answer 2

您想要的是并行计算工具箱中的单程序多数据或spmd功能。来自文档：

spmd语句允许您定义同时运行的代码块多个实验室。

要使用它，首先必须定义一个matlabpool。

matlabpool 3

这会在名为labindex的工作区中自动创建一个“隐藏”变量，该变量指定当前正在进行工作的实验室。

labindex

ans =

     1

当我们使用 spmd 时，我们会在 matlabpool 打开的所有实验室中并行执行我们的代码。

spmd 
labindex
end
Lab 1: 

  ans =

       1

Lab 2: 

  ans =

       2

Lab 3: 

  ans =

       3

我们可以使用 labindex 为同一个函数指定不同的输入并并行运行所有内容。如果我们有两个不同的矩阵并且想要并行找到特征值，我们可以做

M = {rand(10) rand(20)}; %Put matrices into cell for easy access
spmd
  if labindex < 3  %We only have two matrices, so only need two labs
     E = eig(M(labindex));
  end
end
E
E =

   Lab 1: class = double, size = [10   1]
   Lab 2: class = double, size = [20   1]
   Lab 3: No data

Answer 3

我测试了并行工具箱并最终没有专门购买它，因为你无法控制在哪个核心上运行的功能。因此，如果启动多个worker（使用matlabpool()）并尝试将函数处理发送到特定的worker / pool，它将无法工作。但是，使用parfor会将工作负载均匀分布在核心上，以便优化处理时间。

对于我想要做的事情，解决方案是让不同的matlab会话运行。希望CPU将下一个会话分配给另一个CPU

MATLAB Parallel Toolbox用于在两个不同的内核上运行两个独立的功能

3 个答案: