simple method of forcing single precision floating point operations

There are some frustrating things that occur when trying to generate single precision math using embedded coder. I only have single precision hardware available so it's important to use single precision operations and variables.

In a simple demonstration function where the inputs and outputs have been set to single in the model explorer:

function y = fcn(u)
y = u;
for i = single(0:1000000-1)
    y = i*2+i^2;
end

the bounds of the loop must be specifically set to single or else an error is raised. There are lots of little problems like this that come up that make editing preexisting code/simulink projects pretty frustrating to debug.

Also when using trig and square root functions, the generated c code uses sin and cos instead of sinf and cosf etc. despite the fact that the arguments are single precision. The generated code needs manual editing from what I can tell to get consistent single precision math throughout.

Is there a better way?

[dpb]

I'd guess the "better way" unless TMW fixes such issues which I wouldn't expect any time soon if ever for the latter problem would be to use #define to do the translation of the library function names.

I unfortunately don't have any suggestions for the other issue; I agree that lacking a declaration facility with Matlab makes dealing with such issues a pit(proverbial)a[ppendage]

Answer 1

[Adam Barber]

Hey Jonah,

I'm not sure if you tried this already, but you should be able to specify single precision by following the steps here:

http://www.mathworks.com/help/rtw/ug/specify-single-precision.html

with an example here:

http://www.mathworks.com/help/ecoder/ug/use-single-data-type-as-default-for-underspecified-types.html

If you don't explicitly set data types anywhere, then Simulink should only generate code for single precision. If you do find inconsistencies I would definitely bring those up with Technical Support.

Hope that helps,

-Adam

[Jonah Caplan]

Hey Adam,

I hadn't tried it out but it didn't work with the for loop example above with the single() removed. It's definitely easier than manually setting each variable to single but there's still the same problem building.

I'm assuming the problem with sin instead of sinf isn't affected either.

This assignment writes a 'double' value into a 'single' type. Code generation does not support changing types through assignment. Check preceding assignments or input type specifications for type mismatches.

Function 'ForLoop' (#24.61.62), line 4, column 5:
"y"

wrapping the line is also not effective:

y = single(i*2+i^2);

results in c code:

y = (real32_T)((((real_T)i) * 2.0) + (((real_T)i) * ((real_T)i)));

which is still not single precision operations.

[Jonah Caplan]

also, things get really ugly if the loop bounds are variable because you can't cast the type anymore, which means each individual instance of i in the code needs to have a single() around it.

function y = fcn(u,loopLimit)
y = u;
for i = single(0:loopLimit)
    y = i*2+i^2;
end

gives the error:

   FOR loop index expressions of unknown size are only supported if they are of the form A:B or A:B:C.

Function 'ForLoop' (#24.37.85), line 3, column 1:
"for i = single(0:loopLimit)"
Launch diagnostic report.

so you need to do:

function y = fcn(u,loopLimit)
y = u;
for i = 0:loopLimit
    y = i*single(2)+single(i)^2;
end

which is getting a bit cumbersome

[Jonah Caplan]

Better yet, the single() can be eliminated by combining the specify single precision option and ensuring all constants are set to Inherit: Inherit via back propagation rather than Inherit: Inherit from 'Constant value'

phew!

Answer 2

[Amish Ghadiya]

How matlab Embedded function support fixed point data type Help me if any one knows the solution.

*Below mentioned Embedded matlab function code* 
function [Out1, Out2] = CheckCount(int1,int2)
%#eml

persistent var1 CNT;
if isempty(var1 ); var1 =[0 0 0 0 0]; end
if isempty(CNT); CNT=0 ; end

Out1 = 0;
Out2 = 1;

var1= [ int1, var1 (1:4) ];

if sum(var1) > 20
    CNT = CNT + 1 ; 
    if CNT >= 5 ; CNT=5 ; end
    Out2 = 1 - (CNT*0.2);
    Out1 = 1;
end
if int2< 7.2
    Out2 = 0;
end

input and output data type are same 
ex: fixdt(0,16,0.1,0)

When i try to compile below mentioned error is displayed

*This assignment writes a 'double' value into a 'uint8' type. Code generation does not support changing types through assignment. Check preceding assignments or input type specifications for type mismatches.

Function 'SCA__USSDIST_Embedded MATLAB ' (#1474.182.186), line 8, column 1: "Out1" Launch diagnostic report.*

[Arun Badigannavar]

Even I am facing the same error, with stateflow

Answer 3

[leong]

the code replacement library is intended for such use.. although it might not be as easy..