🚀 go-pugleaf

In article <1992Oct27.180855.8049@u.washington.edu> yuhong@u.washington.edu writes:
>I am currently working on coding MPEG audio algorithm on a fixed-point 
>DSP chip.  The original MPEG audio algorithm described in the Committee 
>Draft is a floating-point algorithm.  A lot of constant coefficients 
>are double precision numbers.  And in the psychoacoustic part of the 
>algorithm, the computation of logarithmic and exponential functions (base 10) 
>are required.  Can anybody tell me how to modify the MPEG audio algorithm 
>in order to fit into a fixed-point processor?  How to realize logarithmic 
>and exponential function on a fixed-point processor?  Or the general way 
>to convert a floating-point algorithm to fixed-point one.  Any information
>is appreciated.
>
>Thanks in advance.
>
>-Yuhong
>yuhong@u.washington.edu
> 

Xing Technology Corporation of Arroyo Grande, CA, has ported the decompression
side of the MPEG audio standard to the ADSP-2105 fixed-point DSP processor.
They can be reached at: P.O. Box 950, 456 Carpenter Canyon, Arroyo Grande, CA
93420. Tel: (805) 473-0145. They may be able to provide some insight into
the workings of the MPEG algorithm.

The audio encoder side is a little trickier, but if your application does not
require real-time compression, a fixed-point DSP with the appropriate hooks
for handling double precision math and two-word floating-point data
(such as the ADSP-2100 family) could probably be used with little or no
modifications to the algorithm. The logarithmic and exponential functions
can be calculated by use of Taylor-series expansions, or, if speed is
a requirement, stored in a table. 


-------------------------------------------------------------------------------
Noam Levine                                       email: noam.levine@analog.com
Analog Devices                                    tel: 617.461.3672
DSP Applications                                  fax: 617.461.3010
3 Technology Way, P.O. Box 9106
Norwood, MA  02062-9106
-------------------------------------------------------------------------------

Yuhong Xiong (yuhong@hardy.u.washington.edu) wrote:
: I am currently working on coding MPEG audio algorithm on a fixed-point 
: DSP chip.  The original MPEG audio algorithm described in the Committee 
: Draft is a floating-point algorithm.  A lot of constant coefficients 
: are double precision numbers.  And in the psychoacoustic part of the 
: algorithm, the computation of logarithmic and exponential functions (base 10) 
: are required.  Can anybody tell me how to modify the MPEG audio algorithm 
: in order to fit into a fixed-point processor?  How to realize logarithmic 
: and exponential function on a fixed-point processor?  Or the general way 
: to convert a floating-point algorithm to fixed-point one.  Any information
: is appreciated.
: 
: Thanks in advance.
: 
: -Yuhong
: yuhong@u.washington.edu
:  
	I think it is an interesting problem to convert a floating point
precision program, eg, C-code, into fixed point computations. Since it
is the first step for converting algorithm to architecture. C-language 
do not support fixed point computation. Thus, there exists a big gap from
algorithm to architecture mapping. It is almost impossible to implement
multipliers or adders with floating point computation for real-time ASIC
solution. 
	So, the problem is: who should take the reponsibility of "Precision
Analyis". Architecture designer or algorithm developer? Or there are good
solutions to this problem.

eea78555@twins.ee.nctu.edu.tw