• State
  • Locked Locked
  • Replies 2 replies
  • Subscribers 98 subscribers
  • Views 833 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

mathlib divsp versus libc __c66x_divf

Alberto Chessa
Mastermind 6670 points
Other Parts Discussed in Thread: MATHLIB

Hello,

Since the default __c66x_divf  implementation supplied with the libc (C6678, little endian) has very poor performance, I tried to substitute it with the mathlib divsp. In general, the results seem to be correct but in some case the results are different. In particular, with the C66 optimized divsp, it seems that sometime the modf function loops forever.

I'm tyring to isolate the problem (the whole algorithm is complex and maybe there are also some problems in the input data). Meanwhile, I wonder if someone have already tried to use the mathlib optimized divsp and if he has found any problems.

 

To substitute the default __c66x_divf  I simple add to my executable a __c66x_divf function that calls the optimized divsp. By the way, the optimized one is very fast, respect to the built-in.

 

 

  • 0
    TI__Guru* 84110 points

    Alberto,

    Although I will not know the answer, to help whoever does try to answer you, would you please supply some more information:

    CCS version
    CGT version
    PC platform OS (WinXP, Linux, etc.)
    mathlib version and/or MCSDK version
    example code snippets that show the cases where you have problems or see difference

    This may belong in one of the other forums, but it seems that a lot of mathlib questions are handled here, so we can see how it goes.

    Regards,
    RandyP

  • 0 in reply to RandyP
    Mastermind 6670 points

     

    Sorry, I forgot::

    OS: WinXP

    CCS: 5.1.0.06

    MCSDK: 2.0.1.12

    MATHLIB: 3.0.0.0

    Compiler: 7.2.4

     

    About the example, we are working to isolate the problem but so far I have nothing of usable to post here.

    Anyway I rerun another test (not target to the divsp, but for our algorithms) and I saw that the divsp produce more imprecise result respect to the "bult-in" one, where "more imprecise" is respect to the expected results produced with VisualC. In general the difference is acceptable but in one case it is too big.

     

    RandyP said:

    Alberto,

    Although I will not know the answer, to help whoever does try to answer you, would you please supply some more information:

    CCS version
    CGT version
    PC platform OS (WinXP, Linux, etc.)
    mathlib version and/or MCSDK version
    example code snippets that show the cases where you have problems or see difference

    This may belong in one of the other forums, but it seems that a lot of mathlib questions are handled here, so we can see how it goes.

    Regards,
    RandyP