How to run this custom-built function on GPU?

Question

O.Hubert 2018-2-26

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/384892-how-to-run-this-custom-built-function-on-gpu

编辑： O.Hubert 2018-2-26

I have a fairly simple code to run multiple times and I was wondering how to use GPU in Matlab for this.

Typically, what the code does is the following:

Draw a random matrix
Have some basic action with this random matrix and other inputs of the function. In this case, a simple matrix multiplication
Check the sign of some cells
Keep the draw if the draw respects signs set a priori; Draw again if not
Repeat this for 500 or 1000 times

Since each repetition is independent of the previous one, I can consider parallelizing.

My question is: *with custom-built functions, how can I use the GPU for 500 or 1000 parallel simulations? In particular, what is the syntax I have to respect for this aim?*

Notes: I have an NVIDIA GeForce GTX 1060 6GB (supported) and MATLAB R2016a with parallel computing toolbox.

I have put here a Minimum Working Example of what I want to do.

    % Minimum Working Example MATLAB - GPU
    A=[0.0027 -0.0001 0.0004 ; -0.0001 0.009 0.0004 ; - 0.00023 -0.0003 0.005]; % variance-covariance matrix
    cholA=chol(A,'lower');
    Nrep=zeros(3);
    Nrep(1,1)=500; % number of repetition for the function ; Nrep is the same size as A because arrayfun only supports same-size matrices.
    gpuArray(cholA); % could have created A as gpuArray directly, but this is an example
    gpuArray(Nrep);
    [output_fct_results,output_fct_draws]=arrayfun(@sign_test,cholA,Nrep,'UniformOutput',false) % launch the function onto the GPU.

Option 1:

    function [output_fct_results,output_fct_draws]=sign_test(cholA,Nrep) % option 1
    % This function tests whether the following signs are respected:
    % test_mat(1,1)=+;test_mat(1,2)=+;
    % test_mat(2,1)=+;test_mat(2,2)=-;
    Nreptemp=Nrep(1,1); % Trick to pass inputs with different sizes
    clear Nrep;
    Nrep=Nreptemp;
    % Initialize output matrices
    output_fct_results=zeros(3,3,Nrep);
    output_fct_draws=zeros(1,Nrep);
    parfor repet=1:Nrep % range of the loop
        fprintf('Repetition %d: \n',repet);
        respect=0;
        draws=0;
        while respect<1 % run as long as the conditions are not respected
            % Draw a normal random matrix
            A=normrnd(0,1,3,3);
            [Q R]=qr(A); % Compute qr decomposition of random matrix A
            for ii=1:3;
                if R(ii,ii)<0
                    Q(:,ii)=-Q(:,ii); % normalize the sign ; makes the QR decomposition unique (Arias et al 2014)
                end
            end
            % Initialize the test matrix
            test_mat=zeros(3);
            % Compute the test matrix
            for j=1:3
                test_mat(:,j)=cholA*Q(:,j);
            end
            % Initialize the test for the signs of the test_matrix
            test_sign=zeros(1,4);
            % Test each relevant cell
            if test_mat(1,1)>=0;
                test_sign(1)=1;
            end
            if test_mat(1,2)>=0;
                test_sign(2)=1;
            end
            if test_mat(2,1)>=0;
                test_sign(3)=1;
            end
            if test_mat(2,2)<=0;
                test_sign(4)=1;
            end
            if sum(test_sign==1)==size(test_mat,2) % If all signs are respected, then exit the loop
                respect=1;
            end
            draws=draws+1;
        end
        output_fct_results(:,:,repet)=Q;
        output_fct_draws(1,repet)=draws;
    end
    end

Option 2:

        function [output_fct_results,output_fct_draws]=sign_test(cholA,Nrep) % option 2
    % This function tests whether the following signs are respected:
    % test_mat(1,1)=+;test_mat(1,2)=+;
    % test_mat(2,1)=+;test_mat(2,2)=-;
    Nreptemp=Nrep(1,1); % Trick to pass inputs with different sizes
    clear Nrep;
    Nrep=Nreptemp;
    % Initialize output matrices
    output_fct_results=zeros(3,3,Nrep);
    output_fct_draws=zeros(1,Nrep);
    draws=0;
    accepted=0;
    while accepted<=Nrep % range of the loop
        % Draw a normal random matrix
        A=normrnd(0,1,3,3);
        [Q R]=qr(A); % Compute qr decomposition of random matrix A
        for ii=1:3;
            if R(ii,ii)<0
                Q(:,ii)=-Q(:,ii); % normalize the sign ; makes the QR decomposition unique (Arias et al 2014)
            end
        end
        % Initialize the test matrix
        test_mat=zeros(3);
        % Compute the test matrix
        for j=1:3
            test_mat(:,j)=cholA*Q(:,j);
        end
        % Initialize the test for the signs of the test_matrix
        test_sign=zeros(1,4);
        % Test each relevant cell
        if test_mat(1,1)>=0;
            test_sign(1)=1;
        end
        if test_mat(1,2)>=0;
            test_sign(2)=1;
        end
        if test_mat(2,1)>=0;
            test_sign(3)=1;
        end
        if test_mat(2,2)<=0;
            test_sign(4)=1;
        end
        if sum(test_sign==1)==size(test_mat,2) % If all signs are respected, then move on to the next iteration
            accepted=accepted+1;
            output_fct_results(:,:,accepted)=Q;
        end
        draws=draws+1;
    end
    output_fct_draws(1,1)=draws;
    end