DSPRelated.com
Forums
   Forums    More Forums    speechcoding  

Fixed point conversion

Started by pai_raviin ●April 30, 2005
Hi all,

Can anybody help me with the fixed point conversion of following
code:

while (iCoef < iNextBandR)
{
*pfltXformSubBark = 0.0;
i = 0;
while (i < iWidth)
{
*pfltXformSubBark += *pfltOuterEarXform * *pfltOuterEarXform;
i++;
}
iCoef += iWidth;
pfltXformSubBark++;
iWidth = (1<<iDownFactor);
}

I am taking the values of pfltOuterEarXform in as Q30 format in
fixed point format (because the values of pfltOuterEarXform in
between 1.02527 & 0.021747) and doing the squaring of
pfltOuterEarXform and accumulating with pfltXformSubBark if and only
if i < iWidth.

And I observed the range of pfltXformSubBark values in between
0.000945910 and 33.6377.

When I do the multiplication there is overflow problem(?) i.e. it's
exceeding Q30 format. How to convert this piece of code into fixed
point??

P. Ravichandra Reddy


Reply by bhagawan reddy ●May 3, 20052005-05-03
Dear Ravichandra,
 
      Here you are doing Multiply and Accumulate(MAC)  operation.Hence you cannot expect *pfltXformSubBark  to be in Q30 format.
 
Declare this as 40-bit or 64-bit, whatever your processor supports, and balance betweeen the Integer and fractional bits you need.
 
Regards,
Bhagawan


pai_raviin <p...@yahoo.co.in> wrote:
Hi all,

Can anybody help me with the fixed point conversion of following
code:

while (iCoef < iNextBandR)
{
*pfltXformSubBark = 0.0;
i = 0;
while (i < iWidth)
{
*pfltXformSubBark += *pfltOuterEarXform * *pfltOuterEarXform;
i++;
}
iCoef += iWidth;
pfltXformSubBark++;
iWidth = (1< }

I am taking the values of pfltOuterEarXform in as Q30 format in
fixed point format (because the values of pfltOuterEarXform in
between 1.02527 & 0.021747) and doing the squaring of
pfltOuterEarXform and accumulating with pfltXformSubBark if and only
if i < iWidth.

And I observed the range of pfltXformSubBark values in between
0.000945910 and 33.6377.

When I do the multiplication there is overflow problem(?) i.e. it's
exceeding Q30 format. How to convert this piece of code into fixed
point??

P. Ravichandra Reddy

------------------------ Yahoo! Groups Sponsor --------------------~-->
In low income neighborhoods, 84% do not own computers.
At Network for Good, help bridge the Digital Divide!
http://us.click.yahoo.com/EA3HyD/3MnJAA/79vVAA/GP4qlB/TM
--------------------------------~-
<*> To visit your group on the web, go to:
http://groups.yahoo.com/group/speechcoding/

<*> To unsubscribe from this group, send an email to:
s...@yahoogroups.com

<*

Yahoo! India Matrimony: Find your life partner online.

Reply by venkat ramanan ●May 3, 20052005-05-03

Dear Ravichandra,
As bhagawan said,
Multiplying two variables of x bits, u will get 2x bits. We have shift the 2x bits by 1 bit left (to eliminate the effect of one of the sign bit I guess).
So your right hand side variable have to be greater than q30 format (nearly double) and finally you can round of to Q30 format instead of keeping the right hand side variable in Q30 format. This will be more accurate than the later.

Regards,
Venkat
  On Tue, 03 May 2005 bhagawan reddy wrote :
>Dear Ravichandra,
>
>      Here you are doing Multiply and Accumulate(MAC)  operation.Hence you cannot expect *pfltXformSubBark  to be in Q30 format.
>
>Declare this as 40-bit or 64-bit, whatever your processor supports, and balance betweeen the Integer and fractional bits you need.
>
>Regards,
>Bhagawan >pai_raviin <p...@yahoo.co.in> wrote:
>Hi all,
>
>Can anybody help me with the fixed point conversion of following
>code:
>
>while (iCoef < iNextBandR)
>{
>*pfltXformSubBark = 0.0;
>i = 0;
>while (i < iWidth)
>{
>*pfltXformSubBark += *pfltOuterEarXform * *pfltOuterEarXform;
>i++;
>}
>iCoef += iWidth;
>pfltXformSubBark++;
>iWidth = (1< }
>
>I am taking the values of pfltOuterEarXform in as Q30 format in
>fixed point format (because the values of pfltOuterEarXform in
>between 1.02527 & 0.021747) and doing the squaring of
>pfltOuterEarXform and accumulating with pfltXformSubBark if and only
>if i < iWidth.
>
>And I observed the range of pfltXformSubBark values in between
>0.000945910 and 33.6377.
>
>When I do the multiplication there is overflow problem(?) i.e. it's
>exceeding Q30 format. How to convert this piece of code into fixed
>point??
>
>P. Ravichandra Reddy >Yahoo! India Matrimony: Find your life partneronline.

Mistakes are not end of the world but repeating them is
Reply by Raj Baretto ●May 5, 20052005-05-05


You need 6 bits to take care of the overflow according to your accumulated size(but this depends again on iWidth). As you are using a accumulator normally of 4 bits it would be fine when you are not doing any shifting in the loop.But while storing it in the result array or using this after MAC operations you need to shift or use an appropriate assembly instructions so that it stores a part of your higher accumulated value.Also remember if you are multiplying in the fractional mode you dont need take care of the shifts to compensate for the format.The only care care you need to take is the overflow occuring due to acccumulation. If you are using addition or scaling in your code with this functionality then you can use the same along with the final shifting operations.

Rgrds,
Rajendra.
On Tue, 03 May 2005 venkat ramanan wrote :
>Dear Ravichandra,
>As bhagawan said,
>Multiplying two variables of x bits, u will get 2x bits. We have shift the 2x bits by 1 bit left (to eliminate the effect of one of the sign bit I guess).
>So your right hand side variable have to be greater than q30 format (nearly double) and finally you can round of to Q30 format instead of keeping the right hand side variable in Q30 format. This will be more accurate than the later.
>
>Regards,
>Venkat >
>On Tue, 03 May 2005 bhagawan reddy wrote :
> >Dear Ravichandra,
> >
> > Here you are doing Multiply and Accumulate(MAC) operation.Hence you cannot expect *pfltXformSubBark to be in Q30 format.
> >
> >Declare this as 40-bit or 64-bit, whatever your processor supports, and balance betweeen the Integer and fractional bits you need.
> >
> >Regards,
> >Bhagawan
> >
> >
> >pai_raviin <pai_raviin@pai_...> wrote:
> >Hi all,
> >
> >Can anybody help me with the fixed point conversion of following
> >code:
> >
> >while (iCoef < iNextBandR)
> >{
> >*pfltXformSubBark = 0.0;
> >i = 0;
> >while (i < iWidth)
> >{
> >*pfltXformSubBark += *pfltOuterEarXform * *pfltOuterEarXform;
> >i++;
> >}
> >iCoef += iWidth;
> >pfltXformSubBark++;
> >iWidth = (1< }
> >
> >I am taking the values of pfltOuterEarXform in as Q30 format in
> >fixed point format (because the values of pfltOuterEarXform in
> >between 1.02527 & 0.021747) and doing the squaring of
> >pfltOuterEarXform and accumulating with pfltXformSubBark if and only
> >if i < iWidth.
> >
> >And I observed the range of pfltXformSubBark values in between
> >0.000945910 and 33.6377.
> >
> >When I do the multiplication there is overflow problem(?) i.e. it's
> >exceeding Q30 format. How to convert this piece of code into fixed
> >point??
> >
> >P. Ravichandra Reddy
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >Yahoo! India Matrimony: Find your life partneronline. >Mistakes are not end of the world but repeating them is