Get pointer to underlying data in mex C++ API

I would like to get a pointer to the underlying data of an input array, using the C++ mex API, while avoiding copies of the data.

This was partially answered in

https://www.mathworks.com/matlabcentral/answers/436916-how-to-access-raw-data-via-the-mex-c-api

eg. I tried

matlab::data::TypedArray<double> ar(std::move(inputs[0]));
double* a = ar.release().get();

and

matlab::data::TypedArray<double> ar(inputs[0]);
double* a = ar.release().get();

However, I've compared that to the result with the C-API,

double *p = mxGetPr(prhs[i]);

and the results are different.

That is, the C++ API must be copying the data. I get that this probably supports safety, but I still want to get at the underlying data, without unnecessary memory copying.

My use case is that I want to call libraries written with xtensor [1], using xt::adapt. I need a pointer to do that.

[1] https://xtensor.readthedocs.io/en/latest/

Answer 1

[Andrew Janke]

That mxGetPr() call is copying the data. The MEX API changed in R2018a when Matlab switched to the "interleaved complex data model". See https://www.mathworks.com/help/matlab/matlab_external/matlab-support-for-interleaved-complex.html. Now, when you call parts of the MEX API that use the old "separate complex" data model, it does a memory copy to convert the data to the old format. mxGetPr is one of those legacy functions.

Use the new mxGetDoubles() or similar functions instead. https://www.mathworks.com/help/matlab/apiref/mxgetdoubles.html. Those will get you direct access to the new interleaved-complex representation without a copy step.