Forum Discussion

Altera_Forum's avatar
Altera_Forum
Icon for Honored Contributor rankHonored Contributor
12 years ago

FP Custom Instructions - exp() won't get any faster

Hi everybody.

I'm working on a design using Nios II/f that heavily relies on floating point math.

In Qsys I addedd Floating Point Hardware and I connected it to the Custom Instruction Master of the Nios II.

Apparently everything works.

My design is a Neural Network, so basically I have a loop of multiplication\addition and a loop of exponential calculus.

The first loop gets faster (about 10x), the second one doesen't change at all.

I tried to make a benchmark code, something like: 

Performance Start
float a = 10;
for(i = 0; i < 1000; i++)
a = exp(a);
Performance End

And run it with and without the following pragmas:

# pragma no_custom_fadds# pragma no_custom_fsubs# pragma no_custom_fmuls# pragma no_custom_fdivs

The performance (about 6Mcycles if I remember correctly) doesen't change.

Hardware divide is enabled both in the Floating Point Hardware module and in the Nios II/f.

My doubt is that since i'm using qsys and not sopc i didn't find the "custom instructions" tab in the Nios II module where to specify the usage of FP Hardware, I assumed that in Qsys when you connect the module the FP hardware is always used unless de-activating pragmas are declared.

Maybe I was wrong here?

These are my includes: 

# include <stdio.h># include "io.h"# include <sys/alt_alarm.h># include <altera_avalon_performance_counter.h># include "system.h"# include <math.h># include <float.h>

Thanks for the help! :)

12 Replies

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    I suspect that exp() is using double, not float. Try expf() instead.

    You'll need to make sure the library code is built with the correct compiler options.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    Yes I tried expf too, the global performance increase but I still don't see any difference with and without pragmas.

    About the compiler options... are you talking about the BSP settings? I'm using the settings from the "hello world" template.

    Any particular option I should check?
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    You need to make sure the float instructions are used when the BSP is built.

    I might be worth looking at the symbol table for the final image - you probably don't want the float match functions in your image.

    (You probably don't want the double ones either, but printf() might pull those in.)

    You might also decide that you only need a limited accuracy exp() function - and write one that is faster but less accurate.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    In the BSP editor I've searched for some reference on the FP custom functions.

    Under Advanced -> Hal -> Make

    fpu_present

    hardware_divide_present

    hardware_fp_cust_inst_divider_present

    hardware_fp_cust_inst_no_divider_present

    hardware_multiplier_present

    I tried to turn these options On and Off but with no effect.

    I'm not quite following you on the symbol table. Where can I find it?

    About the exp() accuracy, since I need to compute a much more complex function (a TanSig) I'll surely build a custom instruction using a LUT.

    Still I'd like to be sure to have implemented the FP Hardware correctly ;)
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    You might need to just make sure that the BSP is all recompiled. Somewhere there is a box for its compiler options (IMHO it should default to -O3, but it doesn't). I think you have to pass the custom instruction numbers for the fp instructions on the compiler command line - so they probably have to go in there.

    Alternatively, find the code for expf() and compile it as part of your program.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    I have a same trouble with divide in Floating Point Hardware. But i don't known why,too. :cry:

    Hardware divide is enabled both in the Floating Point Hardware module and in the Nios II/f. And i sure divide avalible in Floating Point Hardware because i compare total logic elements in 2 cases (turn on/off dividion Floating Point Hardware):

    Turn on: 9724 LEs

    Turn off: 5041 LEs

    i sure Floating Point Hardware is used because it work with multiply float.

    Who is resolved it? Can you help me?
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    1) Check that you are doing 'float' division, not 'double' division (would call divdf3).

    2) Check that the relevant file(s) have been compiled to use the custom instruction, not calling divsf3.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    i didn't find divsf3 in my nios ii application and BSP. Can you show me where it is?

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    Check the generated code (objdump -d) and/or the linker map file (might be generated by default).

    I might have got the function name slightly wrong!
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    12 years ago

    Hi! I resolved it. Input, output type in FP is float but in my project is integer. Thank you very much!