How can I further improve the performance of this function?

This function is an adaptation for my needs of nanmean and nanstd taken by NanSuite Jan Gläscher.

function [inputstemp4]=OPTIM3_END_GeneticPattern(inputstemp)
  %#codegen

global inputs;

[m,n]=size(inputs);
nset1=1:100;
nimp1=1:n;

STDinputs=zeros(100,n);
AVinputs=zeros(100,n);
inputstempcopy=inputstemp;
dim = 2;
nans = isnan(inputstemp);
inputstemp(isnan(inputstemp)) = 0; 
count = size(inputstemp,dim) - sum(nans,dim);
i = find(count==0);
count(i) = ones(size(i));
AVinputs(nset1,nimp1) = sum(inputstemp,dim)./count;
AVinputs(i) = i + NaN;
avg= sum(inputstemp,dim)./count;
avg(i)= i + NaN;
inputstempcopy = inputstempcopy - repmat(avg,[1,33,1]);
inputstempcopy(isnan(inputstempcopy)) = 0; 
STDinputs(nset1,nimp1) = sqrt(sum(inputstempcopy.*inputstempcopy,dim)./max(count-1,1));
STDinputs(i) = i + NaN;
%----------

inputstemp4=AVinputs-min(STDinputs.*AVinputs,AVinputs);

end

Encoding mex worsens performance, coder.extrinsic (nanmean, nanstd) -the original Matlab functions - lengthens the time. Encoding mex my substitute nanmeam / nanstd code:

inputstemp=reshape(permute(inputstemp ,[2,1,3]),100.*33,n);

for nset=1:100

    for nimp=1:n
        countpart=0; 
         for npart=1:33
             if isnan(inputstemp(npart+(33.*(nset-1)),nimp))==0
              countpart=countpart+1; 
             end
         end 

       if countpart~=0

            %avcount=coder.nullcopy(zeros(1,countpart));
             avcount=zeros(1,countpart);
             countpart=0;
             for npart=1:33
               if isnan(inputstemp(npart+(33.*(nset-1)),nimp))==0
                 countpart=countpart+1; 
                 avcount(1,countpart)=inputstemp(npart+(33.*(nset-1)),nimp);
               end
             end
        else
             avcount=0;
        end
         if countpart~=0
           [qq,ww]=size(avcount);
           mymean=0;
           mysum=0;
           for rr=1:ww
               mymean=mymean+avcount(1,rr);
           end
           mymean=mymean./ww;
           for rr=1:ww
               mysum=mysum+(avcount(1,rr)-mymean).^2;
           end
           AVinputs(nset,nimp)=mymean;
           STDinputs(nset,nimp)=sqrt(mysum./(ww-1));
         else
          AVinputs(nset,nimp)=0;
           STDinputs(nset,nimp)=0;   
         end

     end
  end

inputstemp4=AVinputs-min(STDinputs.*AVinputs,AVinputs);

...It improves performance compared to the original code but it does not beat the performance of nansuite I adapted. I tried using single precision but does not work.