Imbedded regression use with GPUs

I am trying to speed up a regression model that I have by using a GPU (I have a Tesla C2075). Does anyone have examples of ways to do this they are willing to share. The regression model (shown below) provides great results, but is very slow.

Thanks for any help you can provide,

Mel

%%Create a matrix of every possible combination of predictors

parfor r=n:n1
    xind{r}=nchoosek(1:h,r);% Establishes the xindex for unique subsamples n-n1

end
parfor g=n:n1

      xind2=xind{g};  % Opens X data sets for 'g' iterations from new sets

      z=xind2(:,1);   % Establishes one column of interation 'g' values

      ind3=size(z);   % Establises the index

      ind3=ind3(:,1); % Calculates the number of rows in interation 'g'

      for Z=1:ind3
          x1=xind2(Z,:);  % Starts the regression by selecting columns of 'X'
          % by interations 1 through Z of choosen values
          x2=X(:,x1);   % Stores selected columns
          p = g + 1; % Number of possible X variables plus 1 intercept
          x = [ones(l,1) x2]; % Add ones for intercept to column

          % Turn of the warning that individual regressions may be singular
          warning off

          beta = inv(x'*x)*x'*y; % Creates hat matrix for the coefficients
          pred = x*beta; % Creates predicted value
          ybar = mean(y)*ones(l,1); % Creates a column of the mean Y
          ssreg = (pred-ybar)'*(pred-ybar); % Calculates the Regression Sum of Squares
          sstot = (y-ybar)'*(y-ybar); % Calculates the Total Sum of Squares
          sse = (pred-y)'*(pred-y); % Calculates thes the Sum of Squares Error
          r2 = ssreg/sstot; % Calculates the R square
          adjr2 = 1 - ((1-r2)*((l-1)/(l-p-1)));% Calculates the Adjusted R squares
          aic = 2*g + h*[log(sse/h)]; % Calculates Akaike's information criterion for each run.

          R(g).size=g;    % Stores the size (number) of predictors  for each
          % regression in cell array R

          R(g).adjr2(Z,1)=adjr2;  % Stores the calculated adjr2 for each
          % regression in cell array R

      end

end

Answer 1

[Edric Ellis]

Whether or not you will get much benefit from using GPUArrays depends quite a bit on where the majority of the time is spent in your program, and the size of your data. For example, MATLAB's \ operator runs well on the GPU, see for example this demo.

Answer 2

[bym]

Pre-allocation of variables in for loop will speed things up. Also, use

beta = x\y;

instead of

beta = inv(x'*x)*x'*y;

[Walter Roberson]

Does that work out the same? Isn't x'*x a correlation matrix?

[Mel]

Thanks, I'll give those a try. Still hoping to get some information on GPU use.