Warning: In function my_gen, (b)float16 type operation is emulated, which is likely to slow down the performance. If your target supports native (b)float16 operations, it could be improved by adding Target feature to enable it.

#include "Halide.h"

using namespace Halide;

class HalfFloatMatMul : public Halide::Generator<HalfFloatMatMul> {
public:
    Input<Buffer<float16_t, 2>> A{"A"};
    Input<Buffer<float16_t, 2>> B{"B"};

    Output<Buffer<float, 2>> output{"C"};

    void generate() {
        RDom r(0, A.dim(0).extent(), "r");
    	matmul(x, y) += cast<float>(A(r, y)) * cast<float>(B(r, x));
        output(x, y) = matmul(x, y);
    }

    void schedule() {
        if (using_autoscheduler()) {
            A.set_estimates({{0, 1024}, {0, 1024}});
            B.set_estimates({{0, 1024}, {0, 1024}});
            output.set_estimates({{0, 1024}, {0, 1024}});
        } else {
            matmul.compute_root();
        }
    }

private:
    Func matmul;
    Var x{"x"}, y{"y"};
};

HALIDE_REGISTER_GENERATOR(HalfFloatMatMul, my_gen)