28069M CLA accuracy problem

evs

Genius 5910 points

I'm have a strange problem.

If I run this on the CPU:

    long j;
   float test=0;
   for (j=0;j<20587;j++) test+=0.002879994;

The result is : 59.29 (correct)

if I run the same code on the CLA the result is 59.24

So how do I fix this?

Thanks!

over 9 years ago

0 Fulano de Tal over 9 years ago

Expert 2090 points

This may not be completely related, but what are your floating point options set to? Namely fp_mode and fp_reassoc. I'm not sure if these floating point options only apply to the CPU or both CPU and CLA. This can be checked in CCS by going to your project properties, then CCS Build>C2000 Compiler>Optimization. To avoid any precission loss, fp_mode should be set to 'strict' and fp_reassoc to 'off'.

0 evs over 9 years ago in reply to Fulano de Tal

Genius 5910 points

Yes, I did set FP_mode to strict and fp_reassoc to off.

0 Fulano de Tal over 9 years ago in reply to evs

Expert 2090 points

Does it matter which compiler version you use? Which are you using? Perhaps it's compiler-related, but the experts can chime in on that.

0 evs over 9 years ago in reply to Fulano de Tal

Genius 5910 points

I tried 6.4.9, 6.4.4 and 6.4.2 same result.

0 Vishal_Coelho over 9 years ago in reply to evs

TI__Mastermind 20850 points

This might be related to https://e2e.ti.com/support/microcontrollers/c2000/f/171/t/487649

Try turning on the rounding mode as i've suggested in the post.

0 Fulano de Tal over 9 years ago in reply to Vishal_Coelho

Expert 2090 points

I tried this and it worked. Keep in mind the CLA status register MSTF can't be written to by the C28x CPU, to set the RNDF32 flag, it has to be done from the CLA.

0 evs over 9 years ago in reply to Vishal_Coelho

Genius 5910 points

Yes this solve the problem:

    __meallow();
   __asm(" MSETFLG RNDF32=1") ;
   __medis();

This is way more impotent then hiding it deep a way, in the reference manual. So why? Is there any reason for not using it?

I got really suspicious when really impotent information is hided away in datasheets?

0 Vishal_Coelho over 9 years ago in reply to evs

TI__Mastermind 20850 points

evs said:
So why? Is there any reason for not using it?

IEEE754 provides for 4 different rounding modes. The CLA and C28x+FPU provide two of thoe modes, truncation (RNDF=0) and round to nearest, break ties to even (RNDF=1). Now you can choose to use whichever mode you wish - on the C28x the boot code usually turns on RNDF but there is no such startup code for the CLA so its up to the user whether they should turn on rounding.

0 Fulano de Tal over 9 years ago in reply to Vishal_Coelho

Expert 2090 points

Vishal, what if you have a target (say F28035) that doesn't have FPU but uses RTS routines (softlib) ... what is the default rounding mode for softlib floating point operations?

0 Vishal_Coelho over 9 years ago in reply to Fulano de Tal

TI__Mastermind 20850 points

Hmm honestly I haven't looked at the standard library routines closely to see what kind of rounding they use.

For instance,
C:\ti\ccsv6\tools\compiler\ti-cgt-c2000_15.12.1.LTS\lib\src\fs_div28.inc

the division operation seems to add a half LSb to the mantissa in there at some point which would always round up (toward + infinity), but i cant say that is true of all math operations.

0 Fulano de Tal over 9 years ago in reply to Vishal_Coelho

Expert 2090 points

Hmm, it looks like double precision (fd_add28.inc) says ROUNDING MODE: ROUND TO NEAREST in the comments and uses ROUND64 macro to round to the nearest.

Yet, all single precision stuff (fs_add28.inc) doesn't say in the comments which mode, but suggests to +inf since we are always adding half LSB and it doesn't use any rounding macro.

To evs's point, this is kind of troubling the fact that the rounding mode for CLA, FPU, and CPU softlib can all be different, and in the case of softlib, two different rounding modes depending upon whether using single or double precision float. At least with FPU and CLA you can set them to match, but in the CPU softlib, we can't really change anything unless we want to recompile RTS lib.

C2000™︎ microcontrollers

C2000 microcontrollers forum

28069M CLA accuracy problem