how can i covert following variables as Indexed variables ?

I have following function. I want to improve the execution time. Matlab profiler is showing following messgae. How can i convert 'input' and 'transformedRefVal1Rt' as indexed variables.

If 'input' is an indexed variable, performance can be improved using logical indexing instead of FIND.

If 'transformedRefVal1Rt' is an indexed variable, performance can be improved using logical indexing instead of FIND.

Here Input is 1x49457 and RtRef is 1x3 Matrix. I am also attaching the data of input,RtRef,numberOfInputData

function transformedRefVal1Rt=inputTransformGPU(input,RtRef,numberOfInputData)
    [M,N]=size(RtRef);
    transformedRefVal1Rt=gpuArray(zeros(numberOfInputData,N));
    
    input(find(input>RtRef(1)))=RtRef(1);
    input(find(input<RtRef(N)))=RtRef(N);
    for j=1:N
         transformedRefVal1Rt(find(input== RtRef(j)),j)=1;       
    end
    for j=1:N-1
        aa=find((RtRef(j)> input ) & (input>RtRef(j+1)));
        transformedRefVal1Rt(aa,j+1)=(RtRef(j)-input(aa))/(RtRef(1)-RtRef(j+1));
        transformedRefVal1Rt(aa,j)=1-transformedRefVal1Rt(aa,j+1);
    end

end

Answer 1

[Rik]

No need for a conversion, they already are indexed variables. Lets give a small example of what this message means:

data=1:5;
L= data<2.5; %generate logical array [true true false false false]
ind=find(L); %generate linear index [1 2]
data(ind) %returns [1 2]
data(L) %returns [1 2]

Because you skip a function call if you directly use the logical array to index you get a performance gain.

In your case you can simply remove the find calls, even without needing to edit any other part of your code.

[Raihan Ul Islam]

Thanks for the solution. However the time is not improved . The new code

The old code

Any way to make this code more faster. any idea will be highly appricated

[Rik]

I don't have access to gpuArray, so I removed that. The code below shows that there is a small advantage to skipping find. Your profiler screenshots are unhelpful, because they have a wildly different number of iterations.

clearvars,clc
S=load('data.mat');

timeit(@() fun1(S.input,S.RtRef,S.numberOfInputData))
timeit(@() fun2(S.input,S.RtRef,S.numberOfInputData))

function transformedRefVal1Rt=fun1(input,RtRef,numberOfInputData)
[M,N]=size(RtRef);
transformedRefVal1Rt=(zeros(numberOfInputData,N));
input(find(input>RtRef(1)))=RtRef(1);
input(find(input<RtRef(N)))=RtRef(N);
for j=1:N
    transformedRefVal1Rt(find(input== RtRef(j)),j)=1;
end
for j=1:N-1
    aa=find((RtRef(j)> input ) & (input>RtRef(j+1)));
    transformedRefVal1Rt(aa,j+1)=(RtRef(j)-input(aa))/(RtRef(1)-RtRef(j+1));
    transformedRefVal1Rt(aa,j)=1-transformedRefVal1Rt(aa,j+1);
end
end
function transformedRefVal1Rt=fun2(input,RtRef,numberOfInputData)
[M,N]=size(RtRef);
transformedRefVal1Rt=(zeros(numberOfInputData,N));
input((input>RtRef(1)))=RtRef(1);
input((input<RtRef(N)))=RtRef(N);
for j=1:N
    transformedRefVal1Rt((input== RtRef(j)),j)=1;
end
for j=1:N-1
    aa=((RtRef(j)> input ) & (input>RtRef(j+1)));
    transformedRefVal1Rt(aa,j+1)=(RtRef(j)-input(aa))/(RtRef(1)-RtRef(j+1));
    transformedRefVal1Rt(aa,j)=1-transformedRefVal1Rt(aa,j+1);
end
end

If you want a larger improvement in performance, you should try to replace the loops by array operations.

[Raihan Ul Islam]

Thanks for your comments. it will be very nice if you could give some example of array operation. The N value will not go above 4.

[Rik]

With such a low loop count I doubt you will see much gain in vectorizing this code. I also don't really see a way to vectorize this function, although I expect it can be done.

In case your question is what I mean with vectorization:

a=1:10;
for n=1:9
    if mod(a(n),3)==0
        a(n)=a(n+1);
    end
end

%vectorized:
b=1:10;
L1=mod(b,3)==0;
L2=[false L1(1:(end-1))];
b(L1)=b(L2);

[Raihan Ul Islam]

thanks a lot for all this help