GPU Coder cannot parallelize loop

I have a for-loop that I am trying to parallelize with GPU Coder, which looks like this

% n_out is of type uint64
% input_array is of type single array
function out = my_func(n_out, input_array) %# codegen
    coder.gpu.kernelfun;
    out = zeros(1, n_out, 'single');
    for i = 1:n_out % loop I want to parallelize
        temp = 0.0;
        %%
        % code that changes temp depending on input_array(i). There are no reads from or writes to
        % variable 'out' here
        %%
        out(i) = temp; % GPU Coder says this is a loop carried dependency?
    end
end

When I run GPU Coder, it does not create a kernel and the build report states:

"Unable to parallelize loop because of loop carried dependencies. Check the use of variable 'out' in function 'my_func'".

1) Why is the assignment

out(i) = temp;

a "loop carried dependency"?

2) How do I remove such a "loop carried dependency"?

EDIT: removed syntax error in for loop index declaration

[Walter Roberson]

I would be curious about what would happen if you wrote into a temporary array, and eventually copied the temporary array to the output variable?

I also wonder whether there are cases where out(i) is not assigned to, leading to a dependancy on the initialization of zeros()

[Chao Luo]

Hi Jeffrey,

Thanks for posting the question. There is a syntax error at line 4,

for i:n_out

I guess you mean

for i = 1:n_out    

After fixing it, I am able to see the loop get parallelized when n_out type is a double scalar.

Answer 1

[Jeffrey]

The syntax error was a copying issue of mine. I've edited the question to reflect my code.

Also, thanks! It was the 'n_out is a type double scalar' that did it. I was using 'n_out' as 'uint64'. forcing a 'double' caused GPU Coder to parallelize the loop.

Do you know why parallelization cares about whether 'n_out' is double or integer?

[Chao Luo]

It is a limitation of the analysis. When a uint64 is used as array index, it is casted into int32. The cast would prevent the analysis to parallelize the loop. double is a special case because it is the default type commonly used as array index, so it is automatic replaced with int32 type so no cast is needed. So, if you have to use integer type, you can use int32 as index type.