How can I increase the accuracy of floating point?

Hi,

Try using IQMath library.

http://processors.wiki.ti.com/index.php/IQmath_Library_for_C28x

Regards,

Gautam

   it is afrad that IQmath is not working in CLA.

   Because CLA doesn't support 64bit calculation and more than two parameters for function.

  so, I'm finding out the way to increase the accuracy of floating point.

Regards

 - Bang Hyun -

Hi,

Go through this doc: http://www.ti.com/lit/sw/sprc993/sprc993.pdf

It says:

"If your project uses IQmath, then include it in CLAShared.h"

Regards,

Gautam

 I've already done it, but it didn't work.

 if we include IQmath in CLAShared.h,  we can use only IQ data type , not function in CLA.

 Regards

- Bang Hyun -

Ohh... I see! So using only the IQ data type with traditional operands don't help?

Regards,

Gautam

in order to do it like what you saying, we need 64bit data type about multiplication.

and it is impossible to divide integer data type with integer in CLA.

I solved the mutiplication problem by using two 32bit integer. but division is impossible.

IQ data type is just 32bit 'long' type. nothing to do it without function.

Regards

-Bang Hyun-

I hope a TI buddy of mine sheds some light here!

Regards,

Gautam

really I hope so.

plz inform of this to them!

Regards.

-Bang Hyun-

Bang,

You can divide by right shifting.  For example, say you need to divide x by 10.

y = x / 10

Instead, mulitply by 51 then right shift by 9:

y = x * 51 >> 9;

You'll have to trade off accuracy and bit depth.

Regards,

Bill

Bang,

When you say CLA operations have errors what do you mean? which operations and compared to what?

It is a rather open ended question.. if specific operations are of concern please let known like divide , multiply etc? 

Also what kind of application are your working on?

I'm making PMSM motor driver with drv8312 kit, and I made it almost, except of motion profile.
by using CLA, to do paralel processing, I'm trying to program real time motion profile generator in CLA.
so, when I calculate the motion profile with floating point, end positon has error. (CLA support floating point, not 64bit double)
then, I used IQmath, but it didn't work in CLA.

because CLA doesn't support the function with more than two parameters.

IQmath function need three parameters for multiplication.

#define   _IQmpy(A,B)    __IQmpy(A,B,GLOBAL_Q)

I hope that I can use IQmath in CLA , someday !

Regards,

-Bang Hyun-

 Bill

I can't understand the relation between 10 and 51. 

plz describe it in detail.

Regards

-Bang Hyun-