// --- Arithmetic --- .def("__add__", [](const Array &a, const Array &b) { return a + b; }) .def("__add__", [](const Array &a, double b) { return a + b; }) .def("__radd__", [](const Array &a, double b) { return b + a; }) .def("__sub__", [](const Array &a, const Array &b) { return a - b; }) .def("__sub__", [](const Array &a, double b) { return a - b; }) .def("__rsub__", [](const Array &a, double b) { return b - a; }) .def("__mul__", [](const Array &a, const Array &b) { return a * b; }) .def("__mul__", [](const Array &a, double b) { return a * b; }) .def("__rmul__", [](const Array &a, double b) { return b * a; }) .def("__truediv__", [](const Array &a, const Array &b) { return a / b; }) .def("__truediv__", [](const Array &a, double b) { return a / b; }) .def("__rtruediv__", [](const Array &a, double b) { return b / a; }) .def("__floordiv__", [](const Array &a, const Array &b) { return floor(a / b); }) .def("__floordiv__", [](const Array &a, double b) { return floor(a / b); }) .def("__rfloordiv__", [](const Array &a, double b) { return floor(b / a); }) .def("__mod__", [](const Array &a, const Array &b) { return a % b; }) .def("__mod__", [](const Array &a, double b) { return a % b; }) .def("__pow__", [](const Array &a, const Array &b) { return pow(a, b); }) .def("__pow__", [](const Array &a, double b) { return pow(a, full(a.shape(), b, a.dtype(), a.place())); }) .def("__rpow__", [](const Array &a, double b) { return pow(full(a.shape(), b, a.dtype(), a.place()), a); }) .def("__matmul__", [](const Array &a, const Array &b) { return matmul(a, b); }) // --- Unary --- .def("__neg__", [](const Array &a) { return -a; }) .def("__pos__", [](const Array &a) { return +a; }) .def("__abs__", [](const Array &a) { return ins::abs(a); }) .def("__invert__", [](const Array &a) { return ~a; }) // --- Bitwise --- .def("__and__", [](const Array &a, const Array &b) { return a & b; }) .def("__and__", [](const Array &a, int64_t b) { return a & b; }) .def("__rand__", [](const Array &a, int64_t b) { return b & a; }) .def("__or__", [](const Array &a, const Array &b) { return a | b; }) .def("__or__", [](const Array &a, int64_t b) { return a | b; }) .def("__ror__", [](const Array &a, int64_t b) { return b | a; }) .def("__xor__", [](const Array &a, const Array &b) { return a ^ b; }) .def("__xor__", [](const Array &a, int64_t b) { return a ^ b; }) .def("__rxor__", [](const Array &a, int64_t b) { return b ^ a; }) .def("__lshift__", [](const Array &a, const Array &b) { return a << b; }) .def("__lshift__", [](const Array &a, int64_t b) { return a << b; }) .def("__rshift__", [](const Array &a, const Array &b) { return a >> b; }) .def("__rshift__", [](const Array &a, int64_t b) { return a >> b; }) // --- Comparison --- .def("__eq__", [](const Array &a, const Array &b) { return equal(a, b); }) .def("__ne__", [](const Array &a, const Array &b) { return not_equal(a, b); }) .def("__lt__", [](const Array &a, const Array &b) { return less(a, b); }) .def("__le__", [](const Array &a, const Array &b) { return less_equal(a, b); }) .def("__gt__", [](const Array &a, const Array &b) { return greater(a, b); }) .def("__ge__", [](const Array &a, const Array &b) { return greater_equal(a, b); }) // --- Conversion --- .def("__bool__", [](const Array &a) { return static_cast<bool>(a); }) .def("__len__", [](const Array &a) { INS_CHECK(a.defined(), "Array is not initialized"); if (a.shape().ndim() == 0) throw py::type_error("len() of unsized object"); return a.shape().dim(0); }) .def("__int__", [](const Array &a) { INS_CHECK(a.numel() == 1, "int() only works for scalar arrays"); Array cpu = a.to(CPUPlace()); switch (cpu.dtype()) { case DType::F64: return py::cast(static_cast<int64_t>(cpu.data<double>()[0])); case DType::F32: return py::cast(static_cast<int64_t>(cpu.data<float>()[0])); case DType::I64: return py::cast(cpu.data<int64_t>()[0]); case DType::I32: return py::cast(static_cast<int64_t>(cpu.data<int32_t>()[0])); case DType::BOOL: return py::cast(static_cast<int64_t>(cpu.data<bool>()[0])); default: throw py::type_error("unsupported dtype for int()"); } }) .def("__float__", [](const Array &a) { INS_CHECK(a.numel() == 1, "float() only works for scalar arrays"); Array cpu = a.to(CPUPlace()); switch (cpu.dtype()) { case DType::F64: return py::cast(cpu.data<double>()[0]); case DType::F32: return py::cast(static_cast<double>(cpu.data<float>()[0])); case DType::I64: return py::cast(static_cast<double>(cpu.data<int64_t>()[0])); case DType::I32: return py::cast(static_cast<double>(cpu.data<int32_t>()[0])); default: throw py::type_error("unsupported dtype for float()"); } })

How to apply

Add all operators to bindings/python/insight_py.cpp Array class

Build: cmake --build . --target insight_python

Test: python3 -m pytest tests/cpu/python/test_operators.py -v

The test file covers all operators with CPU+GPU, Array+scalar, broadcasting, edge cases

Notes

__matmul__ needs matmul() from linalg ops

__floordiv__ uses floor(a/b) — no native C++ floor division operator

__pow__ scalar variant uses full(a.shape(), b, ...) to broadcast scalar

__int__/__float__ must copy to CPU first (.to(CPUPlace()))

Bitwise operators use int64_t for scalar (not double)

Python Dunder Operators for pybind11 Array

Complete operator set

Add to bindings/python/insight_py.cpp in the Array class definition: