(Matlab Coder)Domain error. To compute complex results from real x, use 'acosd(complex(x))'

When I try to evaluate an acosd value, the generated mex execution is not consistent with the original function and often errors out?

function out = triFun(in)%#codegen
out = real(acosd(in));
end

I use the following statement to generate successfully：

codegen -config:mex triFun.m -args {0.5}

Then test it with the following statement:

for i = 1:10
    a = rand(1,2);
    b = a;
    distA = sqrt(a(1).^2+a(2).^2);
    distB = sqrt(b(1).^2+b(2).^2);

    val = dot(a,b)./(distA.*distB);

    out1 = triFun(val) % always OK
    out2 = triFun_mex(val) % sometimes failed, Domain error. To compute complex results from real x, use 'acosd(complex(x))' ???
end

output error:

test_codegen_mex

out1 =

0

Domain error. To compute complex results from real x, use 'acosd(complex(x))'.

Error in acosd (line 13)

coder.internal.error('Coder:toolbox:ElFunDomainError',mfilename);

Error in triFun (line 2)

out = real(acosd(in));

Error in test_codegen_mex (line 12)

out2 = triFun_mex(val) % sometimes failed, Domain error. To compute complex results from real x, use 'acosd(complex(x))' ???

Question:

Obviously, the a,b I gave are equal and the val value should be equal to 1. There will be no complexity, but out2 will mistake the result for a complex number and out1 will not. How can I modify this to achieve the goal of consistent results?

Run In Matlab R2023a

Answer 1

[Bruno Luong]

"Obviously, the a,b I gave are equal and the val value should be equal to 1"

wrong. Before claming anything, you should first check.

format long
while true
    a = rand(1,2);
    b = a;
    distA = sqrt(a(1).^2+a(2).^2);
    distB = sqrt(b(1).^2+b(2).^2);
    val = dot(a,b)./(distA.*distB);
    if abs(val) > 1
        a
        b
        val
        exceedvalto1 = val-1
        fprintf('val can not be == 1 with numerical error\n')
        break
    end
end

a = 1×2

0.894442198246962 0.268453115545912

b = 1×2

0.894442198246962 0.268453115545912

val = 
   1.000000000000000

exceedvalto1 = 
     2.220446049250313e-16

val can not be == 1 with numerical error

Fix: you need to truncate val by safety

val = max(min(val,1),-1);

before calling acosd

[cui,xingxing]

Is this a precision issue with Floating-Point Numbers（Unexpected Results with Floating-Point Arithmetic）?