On Tuesday, August 18, 2020 at 1:44:10 PM UTC-7, gtwrek wrote:

(snip, I wrote)

> >Many algorithms are much better done in fixed point, but teaching it 
> >seems to be a lost art. Part of the reason is that most high-level languages 
> >don't make it easy to do. 

(snip)

> >Maybe some will have some good counterexamples. More DSP chips support 
> >floating point, and it isn't all that hard to do now. But often enough, fixed 
> >point is the right choice.

> "Doing DSP" in FPGAs (more and more common) is almost exclusively fixed 
> point. Floating point in FPGAs makes almost no sense at all. Even 
> though FPGAs vendors keep offering more and more "High-level" tools that 
> make floating point more easily accessable within FPGAs, it's mostly 
> (drumroll...) pointless. 

It does seem that there are some people doing floating point in FPGAs, 
and for some scientific (non-DSP) problems it might be useful.

It isn't so bad, until you get to pre- and post-normalization, which takes
a huge amount of logic for add/subtract, not so much for multiply and
divide.  

Newer FPGA families with 6 input LUTs, instead of the 4 input LUTs
of previous generations should be better.

> Other than padding the FPGA Vendor pockets by doing more and more 
> useless things on a FPGA and driving up the resources consumed. This 
> causes bigger FPGAs to be (needlessly) selected.

Over 50 years ago, IBM design System/360 with a hexadecimal floating
point format.  That is, the exponent is of 16 instead of 2. This is
especially convenient for fast hardware (that is, not microprogrammed)
implementations as it simplifies the barrel shifter needed. 

Numerically, though, it was not so nice, and some new numerical
analysis methods had to be found to work with it.  It might not 
be so bad in an FPGA, though.

In article <12ce8e9c-1c39-4871-a46b-093da51e84bbn@googlegroups.com>,
ga...@u.washington.edu <gah4@u.washington.edu> wrote:
>On Saturday, August 15, 2020 at 5:00:47 AM UTC-7, blo...@columbus.rr.com wrote:
>> DO many people still do DSP in fixed point or are the processors mostly 
>> floating point and the DSP designer no longer needs to work in fixed point?
>
>Many algorithms are much better done in fixed point, but teaching it
>seems to be a lost art.  Part of the reason is that most high-level languages
>don't make it easy to do. 
>
>D.E.Knuth says that finance and typesetting should be done in fixed point,
>but it might be that not so many people even know that.
>
>In numerical terms, values that have an absolute uncertainty should be done in
>fixed point, and relative uncertainty in floating point. 
>
>For most DSP algorithms, using extra bits for more precision is more useful
>than using them for exponents.  
>
>Maybe some will have some good counterexamples.  More DSP chips support
>floating point, and it isn't all that hard to do now.  But often enough, fixed
>point is the right choice.

"Doing DSP" in FPGAs (more and more common) is almost exclusively fixed
point.  Floating point in FPGAs makes almost no sense at all.  Even
though FPGAs vendors keep offering more and more "High-level" tools that
make floating point more easily accessable within FPGAs, it's mostly 
(drumroll...) pointless. 

Other than padding the FPGA Vendor pockets by doing more and more
useless things on a FPGA and driving up the resources consumed.  This
causes bigger FPGAs to be (needlessly) selected.

On Saturday, August 15, 2020 at 5:00:47 AM UTC-7, blo...@columbus.rr.com wrote:
> DO many people still do DSP in fixed point or are the processors mostly 
> floating point and the DSP designer no longer needs to work in fixed point?

Many algorithms are much better done in fixed point, but teaching it
seems to be a lost art.  Part of the reason is that most high-level languages
don't make it easy to do. 

D.E.Knuth says that finance and typesetting should be done in fixed point,
but it might be that not so many people even know that.

In numerical terms, values that have an absolute uncertainty should be done in
fixed point, and relative uncertainty in floating point. 

For most DSP algorithms, using extra bits for more precision is more useful
than using them for exponents.  

Maybe some will have some good counterexamples.  More DSP chips support
floating point, and it isn't all that hard to do now.  But often enough, fixed
point is the right choice.

On Sat, 15 Aug 2020 07:59:33 -0500, Richard Owlett
<rowlett@cloud85.net> wrote:

>On 08/15/2020 07:00 AM, blocher@columbus.rr.com wrote:
>> DO many people still do DSP in fixed point or are the processors mostly > floating point and the DSP designer no longer needs to work in fixed 
>point?
>> 
>
>ROFL - You are *not* asking the right question ;}
>
>Consider;
>   are you recognizing a two tone sequence?
>   are you doing real-time speech recognition?
>
>P.S. Haven't needed to do any Signal Processing since being a BSEE 
>student in 60's ;/
>
>


I make pretty good use of the internal 32 bit FPU in my STM32F4
processor where I would do things fixed point before.  Then convert
back to fixed point to use that output.   Had to write some ASM code
so that the IAR compiler would use the correct rounding instruction
though.

Not doing much DSP really in there.

On 08/15/2020 07:00 AM, blocher@columbus.rr.com wrote:
> DO many people still do DSP in fixed point or are the processors mostly > floating point and the DSP designer no longer needs to work in fixed 
point?
> 

ROFL - You are *not* asking the right question ;}

Consider;
   are you recognizing a two tone sequence?
   are you doing real-time speech recognition?

P.S. Haven't needed to do any Signal Processing since being a BSEE 
student in 60's ;/

DO many people still do DSP in fixed point or are the processors mostly floating point and the DSP designer no longer needs to work in fixed point?