Data layout

[cianciosa]

I'm trying to save the arrays to a netcdf file. But I'm discovering that data is sometimes inconsistent with the shape. Consider the reduced example here.

#include <iostream>

#include <mlx/mlx.h>

int main(int argc, const char * argv[]) {
    mlx::core::array a({1.0,1.0,1.0,0.0,0.0,0.0},{1,2,3});
    std::cout << a.shape() << std::endl;
    std::cout << a << std::endl;

    mlx::core::array b = mlx::core::ones({3,3});
    mlx::core::array c = mlx::core::ones({2,3})/3.0;

    mlx::core::array adotb = mlx::core::einsum("...ij,ij,jk->...ik", {a,c,b});
    std::cout << adotb.shape() << std::endl;
    std::cout << adotb << std::endl;

    float *ap = a.data<float> ();
    float *adotbp = adotb.data<float> ();

    for (size_t i = 0; i < 6; i++) {
        std::cout << ap[i] << " " << adotbp[i] << std::endl;
    }
    
    return 0;
}

If I run this, the arrays print to the screen in the expected order but the data layout is inconsistent.

(1,2,3)
array([[[1, 1, 1],
        [0, 0, 0]]], dtype=float32)
(1,2,3)
array([[[1, 1, 1],
        [0, 0, 0]]], dtype=float32)
1 1
1 0
1 1
0 0
0 1
0 0

Is there a way to get a pointer to memory laid out consistent with the shape using the C++ api?

[awni]

The underlying data is not necessarily row- contiguous. If you transpose and array for example then you can't access it linearly and expect the values to be correct. You have two options:

1. Use the strides and shape to correctly access the underlying data
2. Ensure the data is contiguous before you access it. E.g. a = contiguous(a)

In either case you should be extra careful to make sure the array is evaluated before you access it's raw pointer as there are no checks for that.