Float calculating problems with 548x

[Poloni]

Hi there!

I'm with a problem using LTIB 20080811, kernel 2.6.25 with compiler 4.2-125. I wrote a program that make simple calculations.

The source code follows below:

#include <stdio.h>

#include <math.h>

static float misc_calculations();

int main(int argc, const char * argv[])

{

  float reference = misc_calculations();

  printf("Reference: %f\n", reference);

  while (1)

  {

    float value = misc_calculations();

    if (value != reference)

      printf("Value: %f\n", value);

  }

}

static float misc_calculations()

{

  const float float_a = .2;

  const float float_b = .152;

  float result = 0.0;

  result = float_a – float_b;

  return result;

}

When I execute two instances, the behavior is very strange.

Result:

Value: 2.411886                                                                 
Value: 156705821244071112293643497228883460096.000000                           
Value: 15.301361                                                                
Value: -174085336471250674867240960.000000                                      
Value: 1.250146                                                                 
Value: 2.771876                                                                 
Value: -2.302972                                                                
Value: 2.567792                                                                 
Value: -0.519974                                                                
Value: -1.523481                                                                
Value: 0.132017                                                                 
Value: 0.572767                                                                 
Value: -0.011594                                                                
Value: 2.511886                                                                 
Value: -1.523481                                                                
Value: 7.476519                                                                 
Value: -174085336471250674867240960.000000                                      
Value: 557824153357922598912.000000                                             
Value: 156705821244071112293643497228883460096.000000                           
Value: 1.511886                                                                 
Value: -2.579387                                                                
Value: 0.998819

This problem occurs when I run two process. Only one process, it's executed OK, but when you execute a top program, the errors also appear.

In kernel, I checked the definition CONFIG_FPU. I created the definition, but the result is still the same. For me, the problem is concurrency between process and the kernel is not saving the context.

I used another options as -msoft-float, -mcfv4e and -mcpu=5485.

Anyone ever had this problem?

Regards,

Leandro dos Santos

[kwong]

Hi, I think I've just run into the same problem.  I'm using the a slightly modified 2.6.25 kernel from Freescale on a mcf5472 processor.  If I load down the processor with other tasks, my application that performs the floating point math returns inconsistent results.  If I trim down the number of processes to the bare minimum, my application starts running more reliably.

I looked inside the source code for the kernel, and it looks like the "resume" block within the arch/m68k/coldfire/entry.S file is missing all of the floating point context switching code that existed in the 2.6.10 kernel.

I'm going to try re-adding that code today to see if it makes a difference.  Can anyone else provide any feedback about this in the meantime?

Thanks

[Poloni · Accepted solution]

Hi kwong!

Check the new version of LTIB (2010). In this distribution (kernel 2.6.29) this problem is solved.

I merged the kernel 2.6.29 and solved the problem in the kernel 2.6.25. If you want, I can send my patch.

Regards,

Leandro