[Bibill]

Dear all,

I have written a UDF in which i read a text file and store the data in a float vector to perform some computations on it. Typically, some descriptive statistics computations.

I have tested the code in a C compiler and it works well. However when i use it as a UDF in Fluent, the results differ, probably due to the parallel computation since it only does mathematics and does not access any data or information from Fluent in my UDF.

For instance, different nodes give me different values for a same variable.

Can anyone explain me how to force Fluent to treat the code like i would be treated by a simple C compiler (e.g Xcode) ?

Thanks in advance!

[vinerm]

Any field data is only available at nodes, so, that part must be executed on nodes only, e.g., if you want to access pressure using F_P or C_P, use the code within

#if !RP_HOST
#endif

Do note ! before RP for negation.

File reading can be done by any node. If host does the reading, you have to transfer data to node using

host_to_node_transfer_#(var)

macros where # needs to be replaced by the type of data being transferred. To ensure that each node accesses only the cells and faces belonging to it, use _int loops, such as, begin_c_loop_int or begin_f_loop_int.

[Bibill]

Thank you for your answer!

Then, since i do not access any solver/field data but i only read file and do some operations on the data from the text file i can do everything in HOST ?

A shortened version on my UDF is :


void cumulative_sum(float *cum_vector, float *signal, int size,int choice);


DEFINE_EXECUTE_AT_END(udf)
{

/* Import lift coefficient from txt file */

FILE* file_r;
file_r = fopen("cl.txt", "r");
if(file_r==NULL)
{
#if !RP_NODE
Message("\n");
Message("Error opening file \n");
Message(" \n");
#endif
}

float *signal;
signal = (float*)malloc(N_TIME*sizeof(float));
if ((signal==NULL))
{
#if !RP_NODE
Message("Memory allocation error \n");
#endif
}

int i;
for(i=0;i<N_TIME;i++)
{
fscanf(file_r, "%f",&f1);
signal[i] = f1;
}
fclose(file_r);
/* End import lift coefficient */

/* Computation of cummulative sum of C_l and cummulative sum of C_l^2*/
float *cum_x;
float *cum_squared;
cum_x = (float *)malloc((current_ts+1)*sizeof(float));
cum_squared = (float *)malloc((current_ts+1)*sizeof(float));
cumulative_sum(cum_x,signal,current_ts,0);
cumulative_sum(cum_squared,signal,current_ts,1);


...
}


where the auxilliary function cumulative_sum compute the cumsum of the signal for choice==0 and of the signal^2 for choice==1.

[vinerm]

If you are doing everything on host, then put whole of the code within

#if !RP_NODE
#endif

[Bibill]

Even by doing everything on Host there are differences from the results with a C compiler and Fluent (it is exactly the same code)...

Do you have any common mistakes that i can investigate to fix it ?

[vinerm]

What kind of differences do you observe?

[Bibill]

The purpose of my UDF is to estimate a parameter linked to the lift coefficient.

For it, i do some computations of descriptive statistics during the running time for the lift coefficient. It gives me a series of values. The parameter i want to estimate corresponds to the minimum of this series.

In the end, with the C compiler i obtain an increasing series e.g [13,15,15.5,30,35] and with Fluent i obtain basically the same series but with spurious 0 e.g [13,0,0,30,0].


Physically, a zero value makes no sense... It is thus due to how does Fluent treat the data. Moreover, when i print the values obtained by different nodes of computation for the same time it gives the right value (the one of the C compiler) for one or two nodes and 0 for the others...

[AlexanderZ]

input/output must be executed on host only

use Ansys Fluent Customization manual, it has detail explanation on parallelization process

next time put whole your code and full output from console.
By the way, it is good to define all your variables in the head of code not inside the code