Test out float128

python/pybind11/BUILD.bazel

@@ -98,6 +98,12 @@ cc_binary(
     deps = [":include"],
 )
 
+cc_binary(
+    name = "test_numpy_issue1785",
+    srcs = ["custom_tests/test_numpy_issue1785.cc"],
+    deps = [":include"],
+)
+
 cc_binary(
     name = "test_pybind_issue_repro",
     srcs = ["custom_tests/test_pybind_issue_repro.cc"],

python/pybind11/custom_tests/test_numpy_issue1785.cc

@@ -0,0 +1,60 @@
+// https://github.com/pybind/pybind11/issues/1785
+#include <Eigen/Dense>
+
+#include <pybind11/embed.h>
+#include <pybind11/eval.h>
+// Using Eigen, 'cause I don't want to worry about ownership with capsules
+// or buffers.
+#include <pybind11/eigen.h>
+#include <pybind11/pybind11.h>
+
+namespace py = pybind11;
+
+using float128 = long double;
+static_assert(sizeof(float128) == 16, "Bad size");
+
+template <typename T, int Dim>
+using Vector = Eigen::Matrix<T, Dim, 1>;
+
+void init_module(py::module m) {
+  m.def("make_array", []() {
+    return Vector<float128, 3>(1, 2, 3);
+  });
+  m.def("sum_array", [](Vector<float128, -1> x) {
+    // Return as 1x1 - I don't know how to easily return NumPy scalars...
+    return Vector<float128, 1>(x.array().sum());
+  });
+}
+
+int main(int, char**) {
+  py::scoped_interpreter guard{};
+
+  py::module m("test_module");
+  init_module(m);
+  py::globals()["m"] = m;
+
+  py::print("[ Eval ]");
+  py::exec(R"""(
+import numpy as np
+
+def main():
+    x = np.array([1, 2], dtype=np.float128)
+    print(repr(m.sum_array(x)))
+    print(repr(m.make_array()))
+
+main()
+)""");
+
+  py::print("[ Done ]");
+
+  return 0;
+}
+
+/* Output:
+
+[ Eval ]
+array([ 3.0], dtype=float128)
+array([ 1.0,  2.0,  3.0], dtype=float128)
+[ Done ]
+
+*/