I can't find where this behavior of INTERPN is documented. To the contrary, we have:
>> help interpn
interpn N-D interpolation (table lookup).
Vq = interpn(X1,X2,X3,...,V,X1q,X2q,X3q,...) interpolates to find Vq,
the values of the underlying N-D function V at the query points in
arrays X1q,X2q,X3q,etc. For an N-D V, interpn should be called with
2*N+1 arguments. [snip]
I mean, the behavior is obviously useful, but the codegen version is only trying to support documented INTERPN behavior. You can rewrite your example to avoid relying on this behavior by using a nested function instead of an anonymous function:
function matInt = testfun(a,b,M)
% matfun = @(x) squeeze(interpn([-1 0 1],[-1,0,1],M,x(1),x(2)));
function Mq = matfun(x)
mout = size(M,3);
nout = size(M,4);
Mq = zeros(mout,nout,'like',M);
for j = 1:nout
for i = 1:mout
Mq(i,j) = interpn([-1 0 1],[-1,0,1],M(:,:,i,j),x(1),x(2));
end
end
end
matInt = integral(@(s) matfun(a*s+(1-s)*b), 0, 1, 'ArrayValued',true);
end
Seems like there should be another, more clever way, but it's not coming to mind at the moment.
