Merge branch 'master' into master

.github/workflows/ci.yml

@@ -1019,7 +1019,6 @@ jobs:
           git
           mingw-w64-${{matrix.env}}-gcc
           mingw-w64-${{matrix.env}}-python-pip
-          mingw-w64-${{matrix.env}}-python-numpy
           mingw-w64-${{matrix.env}}-cmake
           mingw-w64-${{matrix.env}}-make
           mingw-w64-${{matrix.env}}-python-pytest
@@ -1031,7 +1030,7 @@ jobs:
       with:
         msystem: ${{matrix.sys}}
         install: >-
-          git
+          mingw-w64-${{matrix.env}}-python-numpy
           mingw-w64-${{matrix.env}}-python-scipy
           mingw-w64-${{matrix.env}}-eigen3
 

include/pybind11/numpy.h

@@ -215,6 +215,7 @@ constexpr int platform_lookup(int I, Ints... Is) {
 }
 
 struct npy_api {
+    // If you change this code, please review `normalized_dtype_num` below.
     enum constants {
         NPY_ARRAY_C_CONTIGUOUS_ = 0x0001,
         NPY_ARRAY_F_CONTIGUOUS_ = 0x0002,
@@ -486,6 +487,74 @@ DECL_NPY_SCALAR(std::complex<long double>, NPY_CLONGDOUBLE);
 
 #undef DECL_NPY_SCALAR
 
+// This table normalizes typenums by mapping NPY_INT_, NPY_LONG, ... to NPY_INT32_, NPY_INT64, ...
+// This is needed to correctly handle situations where multiple typenums map to the same type,
+// e.g. NPY_LONG_ may be equivalent to NPY_INT_ or NPY_LONGLONG_ despite having a different
+// typenum. The normalized typenum should always match the values used in npy_format_descriptor.
+// If you change this code, please review `enum constants` above.
+static constexpr int normalized_dtype_num[npy_api::NPY_VOID_ + 1] = {
+    // NPY_BOOL_ =>
+    npy_api::NPY_BOOL_,
+    // NPY_BYTE_ =>
+    npy_api::NPY_BYTE_,
+    // NPY_UBYTE_ =>
+    npy_api::NPY_UBYTE_,
+    // NPY_SHORT_ =>
+    npy_api::NPY_INT16_,
+    // NPY_USHORT_ =>
+    npy_api::NPY_UINT16_,
+    // NPY_INT_ =>
+    sizeof(int) == sizeof(std::int16_t)   ? npy_api::NPY_INT16_
+    : sizeof(int) == sizeof(std::int32_t) ? npy_api::NPY_INT32_
+    : sizeof(int) == sizeof(std::int64_t) ? npy_api::NPY_INT64_
+                                          : npy_api::NPY_INT_,
+    // NPY_UINT_ =>
+    sizeof(unsigned int) == sizeof(std::uint16_t)   ? npy_api::NPY_UINT16_
+    : sizeof(unsigned int) == sizeof(std::uint32_t) ? npy_api::NPY_UINT32_
+    : sizeof(unsigned int) == sizeof(std::uint64_t) ? npy_api::NPY_UINT64_
+                                                    : npy_api::NPY_UINT_,
+    // NPY_LONG_ =>
+    sizeof(long) == sizeof(std::int16_t)   ? npy_api::NPY_INT16_
+    : sizeof(long) == sizeof(std::int32_t) ? npy_api::NPY_INT32_
+    : sizeof(long) == sizeof(std::int64_t) ? npy_api::NPY_INT64_
+                                           : npy_api::NPY_LONG_,
+    // NPY_ULONG_ =>
+    sizeof(unsigned long) == sizeof(std::uint16_t)   ? npy_api::NPY_UINT16_
+    : sizeof(unsigned long) == sizeof(std::uint32_t) ? npy_api::NPY_UINT32_
+    : sizeof(unsigned long) == sizeof(std::uint64_t) ? npy_api::NPY_UINT64_
+                                                     : npy_api::NPY_ULONG_,
+    // NPY_LONGLONG_ =>
+    sizeof(long long) == sizeof(std::int16_t)   ? npy_api::NPY_INT16_
+    : sizeof(long long) == sizeof(std::int32_t) ? npy_api::NPY_INT32_
+    : sizeof(long long) == sizeof(std::int64_t) ? npy_api::NPY_INT64_
+                                                : npy_api::NPY_LONGLONG_,
+    // NPY_ULONGLONG_ =>
+    sizeof(unsigned long long) == sizeof(std::uint16_t)   ? npy_api::NPY_UINT16_
+    : sizeof(unsigned long long) == sizeof(std::uint32_t) ? npy_api::NPY_UINT32_
+    : sizeof(unsigned long long) == sizeof(std::uint64_t) ? npy_api::NPY_UINT64_
+                                                          : npy_api::NPY_ULONGLONG_,
+    // NPY_FLOAT_ =>
+    npy_api::NPY_FLOAT_,
+    // NPY_DOUBLE_ =>
+    npy_api::NPY_DOUBLE_,
+    // NPY_LONGDOUBLE_ =>
+    npy_api::NPY_LONGDOUBLE_,
+    // NPY_CFLOAT_ =>
+    npy_api::NPY_CFLOAT_,
+    // NPY_CDOUBLE_ =>
+    npy_api::NPY_CDOUBLE_,
+    // NPY_CLONGDOUBLE_ =>
+    npy_api::NPY_CLONGDOUBLE_,
+    // NPY_OBJECT_ =>
+    npy_api::NPY_OBJECT_,
+    // NPY_STRING_ =>
+    npy_api::NPY_STRING_,
+    // NPY_UNICODE_ =>
+    npy_api::NPY_UNICODE_,
+    // NPY_VOID_ =>
+    npy_api::NPY_VOID_,
+};
+
 inline PyArray_Proxy *array_proxy(void *ptr) { return reinterpret_cast<PyArray_Proxy *>(ptr); }
 
 inline const PyArray_Proxy *array_proxy(const void *ptr) {
@@ -833,6 +902,13 @@ public:
         return detail::npy_format_descriptor<typename std::remove_cv<T>::type>::dtype();
     }
 
+    /// Return the type number associated with a C++ type.
+    /// This is the constexpr equivalent of `dtype::of<T>().num()`.
+    template <typename T>
+    static constexpr int num_of() {
+        return detail::npy_format_descriptor<typename std::remove_cv<T>::type>::value;
+    }
+
     /// Size of the data type in bytes.
 #ifdef PYBIND11_NUMPY_1_ONLY
     ssize_t itemsize() const { return detail::array_descriptor_proxy(m_ptr)->elsize; }
@@ -874,7 +950,9 @@ public:
         return detail::array_descriptor_proxy(m_ptr)->type;
     }
 
-    /// type number of dtype.
+    /// Type number of dtype. Note that different values may be returned for equivalent types,
+    /// e.g. even though ``long`` may be equivalent to ``int`` or ``long long``, they still have
+    /// different type numbers. Consider using `normalized_num` to avoid this.
     int num() const {
         // Note: The signature, `dtype::num` follows the naming of NumPy's public
         // Python API (i.e., ``dtype.num``), rather than its internal
@@ -882,6 +960,17 @@ public:
         return detail::array_descriptor_proxy(m_ptr)->type_num;
     }
 
+    /// Type number of dtype, normalized to match the return value of `num_of` for equivalent
+    /// types. This function can be used to write switch statements that correctly handle
+    /// equivalent types with different type numbers.
+    int normalized_num() const {
+        int value = num();
+        if (value >= 0 && value <= detail::npy_api::NPY_VOID_) {
+            return detail::normalized_dtype_num[value];
+        }
+        return value;
+    }
+
     /// Single character for byteorder
     char byteorder() const { return detail::array_descriptor_proxy(m_ptr)->byteorder; }
 
@@ -1545,7 +1634,11 @@ public:
 };
 
 template <typename T>
-struct npy_format_descriptor<T, enable_if_t<is_same_ignoring_cvref<T, PyObject *>::value>> {
+struct npy_format_descriptor<
+    T,
+    enable_if_t<is_same_ignoring_cvref<T, PyObject *>::value
+                || ((std::is_same<T, handle>::value || std::is_same<T, object>::value)
+                    && sizeof(T) == sizeof(PyObject *))>> {
     static constexpr auto name = const_name("object");
 
     static constexpr int value = npy_api::NPY_OBJECT_;

tests/test_numpy_array.cpp

@@ -156,6 +156,55 @@ py::handle auxiliaries(T &&r, T2 &&r2) {
     return l.release();
 }
 
+template <typename PyObjectType>
+PyObjectType convert_to_pyobjecttype(py::object obj);
+
+template <>
+PyObject *convert_to_pyobjecttype<PyObject *>(py::object obj) {
+    return obj.release().ptr();
+}
+
+template <>
+py::handle convert_to_pyobjecttype<py::handle>(py::object obj) {
+    return obj.release();
+}
+
+template <>
+py::object convert_to_pyobjecttype<py::object>(py::object obj) {
+    return obj;
+}
+
+template <typename PyObjectType>
+std::string pass_array_return_sum_str_values(const py::array_t<PyObjectType> &objs) {
+    std::string sum_str_values;
+    for (const auto &obj : objs) {
+        sum_str_values += py::str(obj.attr("value"));
+    }
+    return sum_str_values;
+}
+
+template <typename PyObjectType>
+py::list pass_array_return_as_list(const py::array_t<PyObjectType> &objs) {
+    return objs;
+}
+
+template <typename PyObjectType>
+py::array_t<PyObjectType> return_array_cpp_loop(const py::list &objs) {
+    py::size_t arr_size = py::len(objs);
+    py::array_t<PyObjectType> arr_from_list(static_cast<py::ssize_t>(arr_size));
+    PyObjectType *data = arr_from_list.mutable_data();
+    for (py::size_t i = 0; i < arr_size; i++) {
+        assert(!data[i]);
+        data[i] = convert_to_pyobjecttype<PyObjectType>(objs[i].attr("value"));
+    }
+    return arr_from_list;
+}
+
+template <typename PyObjectType>
+py::array_t<PyObjectType> return_array_from_list(const py::list &objs) {
+    return objs;
+}
+
 // note: declaration at local scope would create a dangling reference!
 static int data_i = 42;
 
@@ -520,28 +569,21 @@ TEST_SUBMODULE(numpy_array, sm) {
     sm.def("round_trip_float", [](double d) { return d; });
 
     sm.def("pass_array_pyobject_ptr_return_sum_str_values",
-           [](const py::array_t<PyObject *> &objs) {
+           pass_array_return_sum_str_values<PyObject *>);
-               std::string sum_str_values;
+    sm.def("pass_array_handle_return_sum_str_values",
-               for (const auto &obj : objs) {
+           pass_array_return_sum_str_values<py::handle>);
-                   sum_str_values += py::str(obj.attr("value"));
+    sm.def("pass_array_object_return_sum_str_values",
-               }
+           pass_array_return_sum_str_values<py::object>);
-               return sum_str_values;
-           });
 
-    sm.def("pass_array_pyobject_ptr_return_as_list",
+    sm.def("pass_array_pyobject_ptr_return_as_list", pass_array_return_as_list<PyObject *>);
-           [](const py::array_t<PyObject *> &objs) -> py::list { return objs; });
+    sm.def("pass_array_handle_return_as_list", pass_array_return_as_list<py::handle>);
+    sm.def("pass_array_object_return_as_list", pass_array_return_as_list<py::object>);
 
-    sm.def("return_array_pyobject_ptr_cpp_loop", [](const py::list &objs) {
+    sm.def("return_array_pyobject_ptr_cpp_loop", return_array_cpp_loop<PyObject *>);
-        py::size_t arr_size = py::len(objs);
+    sm.def("return_array_handle_cpp_loop", return_array_cpp_loop<py::handle>);
-        py::array_t<PyObject *> arr_from_list(static_cast<py::ssize_t>(arr_size));
+    sm.def("return_array_object_cpp_loop", return_array_cpp_loop<py::object>);
-        PyObject **data = arr_from_list.mutable_data();
-        for (py::size_t i = 0; i < arr_size; i++) {
-            assert(data[i] == nullptr);
-            data[i] = py::cast<PyObject *>(objs[i].attr("value"));
-        }
-        return arr_from_list;
-    });
 
-    sm.def("return_array_pyobject_ptr_from_list",
+    sm.def("return_array_pyobject_ptr_from_list", return_array_from_list<PyObject *>);
-           [](const py::list &objs) -> py::array_t<PyObject *> { return objs; });
+    sm.def("return_array_handle_from_list", return_array_from_list<py::handle>);
+    sm.def("return_array_object_from_list", return_array_from_list<py::object>);
 }

tests/test_numpy_array.py

@@ -629,45 +629,61 @@ def UnwrapPyValueHolder(vhs):
     return [vh.value for vh in vhs]
 
 
-def test_pass_array_pyobject_ptr_return_sum_str_values_ndarray():
+PASS_ARRAY_PYOBJECT_RETURN_SUM_STR_VALUES_FUNCTIONS = [
+    m.pass_array_pyobject_ptr_return_sum_str_values,
+    m.pass_array_handle_return_sum_str_values,
+    m.pass_array_object_return_sum_str_values,
+]
+
+
+@pytest.mark.parametrize(
+    "pass_array", PASS_ARRAY_PYOBJECT_RETURN_SUM_STR_VALUES_FUNCTIONS
+)
+def test_pass_array_object_return_sum_str_values_ndarray(pass_array):
     # Intentionally all temporaries, do not change.
     assert (
-        m.pass_array_pyobject_ptr_return_sum_str_values(
+        pass_array(np.array(WrapWithPyValueHolder(-3, "four", 5.0), dtype=object))
-            np.array(WrapWithPyValueHolder(-3, "four", 5.0), dtype=object)
-        )
         == "-3four5.0"
     )
 
 
-def test_pass_array_pyobject_ptr_return_sum_str_values_list():
+@pytest.mark.parametrize(
+    "pass_array", PASS_ARRAY_PYOBJECT_RETURN_SUM_STR_VALUES_FUNCTIONS
+)
+def test_pass_array_object_return_sum_str_values_list(pass_array):
     # Intentionally all temporaries, do not change.
-    assert (
+    assert pass_array(WrapWithPyValueHolder(2, "three", -4.0)) == "2three-4.0"
-        m.pass_array_pyobject_ptr_return_sum_str_values(
-            WrapWithPyValueHolder(2, "three", -4.0)
-        )
-        == "2three-4.0"
-    )
 
 
-def test_pass_array_pyobject_ptr_return_as_list():
+@pytest.mark.parametrize(
+    "pass_array",
+    [
+        m.pass_array_pyobject_ptr_return_as_list,
+        m.pass_array_handle_return_as_list,
+        m.pass_array_object_return_as_list,
+    ],
+)
+def test_pass_array_object_return_as_list(pass_array):
     # Intentionally all temporaries, do not change.
     assert UnwrapPyValueHolder(
-        m.pass_array_pyobject_ptr_return_as_list(
+        pass_array(np.array(WrapWithPyValueHolder(-1, "two", 3.0), dtype=object))
-            np.array(WrapWithPyValueHolder(-1, "two", 3.0), dtype=object)
-        )
     ) == [-1, "two", 3.0]
 
 
 @pytest.mark.parametrize(
-    ("return_array_pyobject_ptr", "unwrap"),
+    ("return_array", "unwrap"),
     [
         (m.return_array_pyobject_ptr_cpp_loop, list),
+        (m.return_array_handle_cpp_loop, list),
+        (m.return_array_object_cpp_loop, list),
         (m.return_array_pyobject_ptr_from_list, UnwrapPyValueHolder),
+        (m.return_array_handle_from_list, UnwrapPyValueHolder),
+        (m.return_array_object_from_list, UnwrapPyValueHolder),
     ],
 )
-def test_return_array_pyobject_ptr_cpp_loop(return_array_pyobject_ptr, unwrap):
+def test_return_array_object_cpp_loop(return_array, unwrap):
     # Intentionally all temporaries, do not change.
-    arr_from_list = return_array_pyobject_ptr(WrapWithPyValueHolder(6, "seven", -8.0))
+    arr_from_list = return_array(WrapWithPyValueHolder(6, "seven", -8.0))
     assert isinstance(arr_from_list, np.ndarray)
     assert arr_from_list.dtype == np.dtype("O")
     assert unwrap(arr_from_list) == [6, "seven", -8.0]

tests/test_numpy_dtypes.cpp

@@ -11,6 +11,9 @@
 
 #include "pybind11_tests.h"
 
+#include <cstdint>
+#include <stdexcept>
+
 #ifdef __GNUC__
 #    define PYBIND11_PACKED(cls) cls __attribute__((__packed__))
 #else
@@ -297,6 +300,15 @@ py::list test_dtype_ctors() {
     return list;
 }
 
+template <typename T>
+py::array_t<T> dispatch_array_increment(py::array_t<T> arr) {
+    py::array_t<T> res(arr.shape(0));
+    for (py::ssize_t i = 0; i < arr.shape(0); ++i) {
+        res.mutable_at(i) = T(arr.at(i) + 1);
+    }
+    return res;
+}
+
 struct A {};
 struct B {};
 
@@ -496,6 +508,98 @@ TEST_SUBMODULE(numpy_dtypes, m) {
         }
         return list;
     });
+    m.def("test_dtype_num_of", []() -> py::list {
+        py::list res;
+#define TEST_DTYPE(T) res.append(py::make_tuple(py::dtype::of<T>().num(), py::dtype::num_of<T>()));
+        TEST_DTYPE(bool)
+        TEST_DTYPE(char)
+        TEST_DTYPE(unsigned char)
+        TEST_DTYPE(short)
+        TEST_DTYPE(unsigned short)
+        TEST_DTYPE(int)
+        TEST_DTYPE(unsigned int)
+        TEST_DTYPE(long)
+        TEST_DTYPE(unsigned long)
+        TEST_DTYPE(long long)
+        TEST_DTYPE(unsigned long long)
+        TEST_DTYPE(float)
+        TEST_DTYPE(double)
+        TEST_DTYPE(long double)
+        TEST_DTYPE(std::complex<float>)
+        TEST_DTYPE(std::complex<double>)
+        TEST_DTYPE(std::complex<long double>)
+        TEST_DTYPE(int8_t)
+        TEST_DTYPE(uint8_t)
+        TEST_DTYPE(int16_t)
+        TEST_DTYPE(uint16_t)
+        TEST_DTYPE(int32_t)
+        TEST_DTYPE(uint32_t)
+        TEST_DTYPE(int64_t)
+        TEST_DTYPE(uint64_t)
+#undef TEST_DTYPE
+        return res;
+    });
+    m.def("test_dtype_normalized_num", []() -> py::list {
+        py::list res;
+#define TEST_DTYPE(NT, T)                                                                         \
+    res.append(py::make_tuple(py::dtype(py::detail::npy_api::NT).normalized_num(),                \
+                              py::dtype::num_of<T>()));
+        TEST_DTYPE(NPY_BOOL_, bool)
+        TEST_DTYPE(NPY_BYTE_, char);
+        TEST_DTYPE(NPY_UBYTE_, unsigned char);
+        TEST_DTYPE(NPY_SHORT_, short);
+        TEST_DTYPE(NPY_USHORT_, unsigned short);
+        TEST_DTYPE(NPY_INT_, int);
+        TEST_DTYPE(NPY_UINT_, unsigned int);
+        TEST_DTYPE(NPY_LONG_, long);
+        TEST_DTYPE(NPY_ULONG_, unsigned long);
+        TEST_DTYPE(NPY_LONGLONG_, long long);
+        TEST_DTYPE(NPY_ULONGLONG_, unsigned long long);
+        TEST_DTYPE(NPY_FLOAT_, float);
+        TEST_DTYPE(NPY_DOUBLE_, double);
+        TEST_DTYPE(NPY_LONGDOUBLE_, long double);
+        TEST_DTYPE(NPY_CFLOAT_, std::complex<float>);
+        TEST_DTYPE(NPY_CDOUBLE_, std::complex<double>);
+        TEST_DTYPE(NPY_CLONGDOUBLE_, std::complex<long double>);
+        TEST_DTYPE(NPY_INT8_, int8_t);
+        TEST_DTYPE(NPY_UINT8_, uint8_t);
+        TEST_DTYPE(NPY_INT16_, int16_t);
+        TEST_DTYPE(NPY_UINT16_, uint16_t);
+        TEST_DTYPE(NPY_INT32_, int32_t);
+        TEST_DTYPE(NPY_UINT32_, uint32_t);
+        TEST_DTYPE(NPY_INT64_, int64_t);
+        TEST_DTYPE(NPY_UINT64_, uint64_t);
+#undef TEST_DTYPE
+        return res;
+    });
+    m.def("test_dtype_switch", [](const py::array &arr) -> py::array {
+        switch (arr.dtype().normalized_num()) {
+            case py::dtype::num_of<int8_t>():
+                return dispatch_array_increment<int8_t>(arr);
+            case py::dtype::num_of<uint8_t>():
+                return dispatch_array_increment<uint8_t>(arr);
+            case py::dtype::num_of<int16_t>():
+                return dispatch_array_increment<int16_t>(arr);
+            case py::dtype::num_of<uint16_t>():
+                return dispatch_array_increment<uint16_t>(arr);
+            case py::dtype::num_of<int32_t>():
+                return dispatch_array_increment<int32_t>(arr);
+            case py::dtype::num_of<uint32_t>():
+                return dispatch_array_increment<uint32_t>(arr);
+            case py::dtype::num_of<int64_t>():
+                return dispatch_array_increment<int64_t>(arr);
+            case py::dtype::num_of<uint64_t>():
+                return dispatch_array_increment<uint64_t>(arr);
+            case py::dtype::num_of<float>():
+                return dispatch_array_increment<float>(arr);
+            case py::dtype::num_of<double>():
+                return dispatch_array_increment<double>(arr);
+            case py::dtype::num_of<long double>():
+                return dispatch_array_increment<long double>(arr);
+            default:
+                throw std::runtime_error("Unsupported dtype");
+        }
+    });
     m.def("test_dtype_methods", []() {
         py::list list;
         auto dt1 = py::dtype::of<int32_t>();

tests/test_numpy_dtypes.py

@@ -188,6 +188,28 @@ def test_dtype(simple_dtype):
             chr(np.dtype(ch).flags) for ch in expected_chars
         ]
 
+    for a, b in m.test_dtype_num_of():
+        assert a == b
+
+    for a, b in m.test_dtype_normalized_num():
+        assert a == b
+
+    arr = np.array([4, 84, 21, 36])
+    # Note: "ulong" does not work in NumPy 1.x, so we use "L"
+    assert (m.test_dtype_switch(arr.astype("byte")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("ubyte")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("short")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("ushort")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("intc")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("uintc")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("long")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("L")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("longlong")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("ulonglong")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("single")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("double")) == arr + 1).all()
+    assert (m.test_dtype_switch(arr.astype("longdouble")) == arr + 1).all()
+
 
 def test_recarray(simple_dtype, packed_dtype):
     elements = [(False, 0, 0.0, -0.0), (True, 1, 1.5, -2.5), (False, 2, 3.0, -5.0)]