feat: make numpy.h compatible with both NumPy 1.x and 2.x (#5050)

* API: Make `numpy.h` compatible with both NumPy 1.x and 2.x

* TST: Update numpy dtype flags test to not covert flags to char

* API: Add `numpy2.h` instead and make `numpy.h` safe

This means that users of `numpy.h` cannot be broken, but need to
update to `numpy2.h` if they want to compile for NumPy 2.

Using Macros simply and didn't bother to try to remove unnecessary
code paths.

* API: Rather than `numpy2.h` use a define for the user.

* Thread `PYBIND11_NUMPY2_SUPPORT` through things and try to adept test matrix

* Small fixups (shouldn't matter)?

* Fixup.  Does upgrading scipy help?  (it shouldn't?)

(Some other small fixup)

* Use NumPy 2 nightlies for ubuntu-latest job also

* BUG: Fix numpy.bool check

* TST: Fix complexwarning

* BUG: Fix the fact that only the 50 slot is filled with the copy alias

(There were 3 functions all doing the same, only this slot survived 2.x)

* TST: One more test tweak

* TST: Use "long" name for long, since it changed on windows

* TST: Apparently we didn't always have ulong, so just use `L`

* TST: Enforce dtype='l' for test as default isn't long anymore on windows

* Rename macro and invert logic to PYBIND11_NUMPY_1_ONLY

* PYBIND11_INTERNAL_NUMPY_1_ONLY_DETECTED

* Test and code comment expansion

* CI: Use pre-releases of numpy/scipy from pip via explicit version

* CI: NumPy 2 only available on almalinux (as it is Python >=3.9)

* MAINT: Match name more exactly and adopt error phrasing

* MAINT: Pushed early, move helper to be private member

* fix error message compilation when using NumPy 1.x-only backcompat

* silence name shadowing warning

* chore: minor optimization

Signed-off-by: Henry Schreiner <henryschreineriii@gmail.com>

---------

Signed-off-by: Henry Schreiner <henryschreineriii@gmail.com>
Co-authored-by: Ralf W. Grosse-Kunstleve <rwgk@google.com>
Co-authored-by: Henry Schreiner <henryschreineriii@gmail.com>

.github/workflows/ci.yml

@@ -108,12 +108,14 @@ jobs:
       run: python -m pip install pytest-github-actions-annotate-failures
 
     # First build - C++11 mode and inplace
-    # More-or-less randomly adding -DPYBIND11_SIMPLE_GIL_MANAGEMENT=ON here.
+    # More-or-less randomly adding -DPYBIND11_SIMPLE_GIL_MANAGEMENT=ON here
+    # (same for PYBIND11_NUMPY_1_ONLY, but requires a NumPy 1.x at runtime).
     - name: Configure C++11 ${{ matrix.args }}
       run: >
         cmake -S . -B .
         -DPYBIND11_WERROR=ON
         -DPYBIND11_SIMPLE_GIL_MANAGEMENT=ON
+        -DPYBIND11_NUMPY_1_ONLY=ON
         -DDOWNLOAD_CATCH=ON
         -DDOWNLOAD_EIGEN=ON
         -DCMAKE_CXX_STANDARD=11
@@ -138,11 +140,13 @@ jobs:
 
     # Second build - C++17 mode and in a build directory
     # More-or-less randomly adding -DPYBIND11_SIMPLE_GIL_MANAGEMENT=OFF here.
+    # (same for PYBIND11_NUMPY_1_ONLY, but requires a NumPy 1.x at runtime).
     - name: Configure C++17
       run: >
         cmake -S . -B build2
         -DPYBIND11_WERROR=ON
         -DPYBIND11_SIMPLE_GIL_MANAGEMENT=OFF
+        -DPYBIND11_NUMPY_1_ONLY=ON
         -DDOWNLOAD_CATCH=ON
         -DDOWNLOAD_EIGEN=ON
         -DCMAKE_CXX_STANDARD=17
@@ -660,6 +664,11 @@ jobs:
       run: |
         python3 -m pip install cmake -r tests/requirements.txt
 
+    - name: Ensure NumPy 2 is used (required Python >= 3.9)
+      if: matrix.container == 'almalinux:9'
+      run: |
+        python3 -m pip install 'numpy>=2.0.0b1' 'scipy>=1.13.0rc1'
+
     - name: Configure
       shell: bash
       run: >
@@ -895,8 +904,10 @@ jobs:
         python-version: ${{ matrix.python }}
 
     - name: Prepare env
+      # Ensure use of NumPy 2 (via NumPy nightlies but can be changed soon)
       run: |
         python3 -m pip install -r tests/requirements.txt
+        python3 -m pip install 'numpy>=2.0.0b1' 'scipy>=1.13.0rc1'
 
     - name: Update CMake
       uses: jwlawson/actions-setup-cmake@v2.0

CMakeLists.txt

@@ -109,13 +109,18 @@ option(PYBIND11_TEST "Build pybind11 test suite?" ${PYBIND11_MASTER_PROJECT})
 option(PYBIND11_NOPYTHON "Disable search for Python" OFF)
 option(PYBIND11_SIMPLE_GIL_MANAGEMENT
        "Use simpler GIL management logic that does not support disassociation" OFF)
+option(PYBIND11_NUMPY_1_ONLY
+       "Disable NumPy 2 support to avoid changes to previous pybind11 versions." OFF)
 set(PYBIND11_INTERNALS_VERSION
     ""
     CACHE STRING "Override the ABI version, may be used to enable the unstable ABI.")
 
 if(PYBIND11_SIMPLE_GIL_MANAGEMENT)
   add_compile_definitions(PYBIND11_SIMPLE_GIL_MANAGEMENT)
 endif()
+if(PYBIND11_NUMPY_1_ONLY)
+  add_compile_definitions(PYBIND11_NUMPY_1_ONLY)
+endif()
 
 cmake_dependent_option(
   USE_PYTHON_INCLUDE_DIR

include/pybind11/cast.h

@@ -327,8 +327,9 @@ class type_caster<bool> {
             value = false;
             return true;
         }
-        if (convert || (std::strcmp("numpy.bool_", Py_TYPE(src.ptr())->tp_name) == 0)) {
-            // (allow non-implicit conversion for numpy booleans)
+        if (convert || is_numpy_bool(src)) {
+            // (allow non-implicit conversion for numpy booleans), use strncmp
+            // since NumPy 1.x had an additional trailing underscore.
 
             Py_ssize_t res = -1;
             if (src.is_none()) {
@@ -360,6 +361,15 @@ class type_caster<bool> {
         return handle(src ? Py_True : Py_False).inc_ref();
     }
     PYBIND11_TYPE_CASTER(bool, const_name("bool"));
+
+private:
+    // Test if an object is a NumPy boolean (without fetching the type).
+    static inline bool is_numpy_bool(handle object) {
+        const char *type_name = Py_TYPE(object.ptr())->tp_name;
+        // Name changed to `numpy.bool` in NumPy 2, `numpy.bool_` is needed for 1.x support
+        return std::strcmp("numpy.bool", type_name) == 0
+               || std::strcmp("numpy.bool_", type_name) == 0;
+    }
 };
 
 // Helper class for UTF-{8,16,32} C++ stl strings:

include/pybind11/detail/common.h

@@ -296,6 +296,10 @@ PYBIND11_WARNING_DISABLE_MSVC(4505)
 #    undef copysign
 #endif
 
+#if defined(PYBIND11_NUMPY_1_ONLY)
+#    define PYBIND11_INTERNAL_NUMPY_1_ONLY_DETECTED
+#endif
+
 #if defined(PYPY_VERSION) && !defined(PYBIND11_SIMPLE_GIL_MANAGEMENT)
 #    define PYBIND11_SIMPLE_GIL_MANAGEMENT
 #endif

include/pybind11/numpy.h

@@ -29,10 +29,15 @@
 #include <utility>
 #include <vector>
 
+#if defined(PYBIND11_NUMPY_1_ONLY) && !defined(PYBIND11_INTERNAL_NUMPY_1_ONLY_DETECTED)
+#    error PYBIND11_NUMPY_1_ONLY must be defined before any pybind11 header is included.
+#endif
+
 /* This will be true on all flat address space platforms and allows us to reduce the
    whole npy_intp / ssize_t / Py_intptr_t business down to just ssize_t for all size
    and dimension types (e.g. shape, strides, indexing), instead of inflicting this
-   upon the library user. */
+   upon the library user.
+   Note that NumPy 2 now uses ssize_t for `npy_intp` to simplify this. */
 static_assert(sizeof(::pybind11::ssize_t) == sizeof(Py_intptr_t), "ssize_t != Py_intptr_t");
 static_assert(std::is_signed<Py_intptr_t>::value, "Py_intptr_t must be signed");
 // We now can reinterpret_cast between py::ssize_t and Py_intptr_t (MSVC + PyPy cares)
@@ -53,7 +58,8 @@ struct handle_type_name<array> {
 template <typename type, typename SFINAE = void>
 struct npy_format_descriptor;
 
-struct PyArrayDescr_Proxy {
+/* NumPy 1 proxy (always includes legacy fields) */
+struct PyArrayDescr1_Proxy {
     PyObject_HEAD
     PyObject *typeobj;
     char kind;
@@ -68,6 +74,43 @@ struct PyArrayDescr_Proxy {
     PyObject *names;
 };
 
+#ifndef PYBIND11_NUMPY_1_ONLY
+struct PyArrayDescr_Proxy {
+    PyObject_HEAD
+    PyObject *typeobj;
+    char kind;
+    char type;
+    char byteorder;
+    char _former_flags;
+    int type_num;
+    /* Additional fields are NumPy version specific. */
+};
+#else
+/* NumPy 1.x only, we can expose all fields */
+using PyArrayDescr_Proxy = PyArrayDescr1_Proxy;
+#endif
+
+/* NumPy 2 proxy, including legacy fields */
+struct PyArrayDescr2_Proxy {
+    PyObject_HEAD
+    PyObject *typeobj;
+    char kind;
+    char type;
+    char byteorder;
+    char _former_flags;
+    int type_num;
+    std::uint64_t flags;
+    ssize_t elsize;
+    ssize_t alignment;
+    PyObject *metadata;
+    Py_hash_t hash;
+    void *reserved_null[2];
+    /* The following fields only exist if 0 <= type_num < 2056 */
+    char *subarray;
+    PyObject *fields;
+    PyObject *names;
+};
+
 struct PyArray_Proxy {
     PyObject_HEAD
     char *data;
@@ -131,6 +174,14 @@ PYBIND11_NOINLINE module_ import_numpy_core_submodule(const char *submodule_name
     object numpy_version = numpy_lib.attr("NumpyVersion")(version_string);
     int major_version = numpy_version.attr("major").cast<int>();
 
+#ifdef PYBIND11_NUMPY_1_ONLY
+    if (major_version >= 2) {
+        throw std::runtime_error(
+            "This extension was built with PYBIND11_NUMPY_1_ONLY defined, "
+            "but NumPy 2 is used in this process. For NumPy2 compatibility, "
+            "this extension needs to be rebuilt without the PYBIND11_NUMPY_1_ONLY define.");
+    }
+#endif
     /* `numpy.core` was renamed to `numpy._core` in NumPy 2.0 as it officially
         became a private module. */
     std::string numpy_core_path = major_version >= 2 ? "numpy._core" : "numpy.core";
@@ -203,6 +254,8 @@ struct npy_api {
             NPY_ULONG_, NPY_ULONGLONG_, NPY_UINT_),
     };
 
+    unsigned int PyArray_RUNTIME_VERSION_;
+
     struct PyArray_Dims {
         Py_intptr_t *ptr;
         int len;
@@ -241,6 +294,7 @@ struct npy_api {
     PyObject *(*PyArray_FromAny_)(PyObject *, PyObject *, int, int, int, PyObject *);
     int (*PyArray_DescrConverter_)(PyObject *, PyObject **);
     bool (*PyArray_EquivTypes_)(PyObject *, PyObject *);
+#ifdef PYBIND11_NUMPY_1_ONLY
     int (*PyArray_GetArrayParamsFromObject_)(PyObject *,
                                              PyObject *,
                                              unsigned char,
@@ -249,6 +303,7 @@ struct npy_api {
                                              Py_intptr_t *,
                                              PyObject **,
                                              PyObject *);
+#endif
     PyObject *(*PyArray_Squeeze_)(PyObject *);
     // Unused. Not removed because that affects ABI of the class.
     int (*PyArray_SetBaseObject_)(PyObject *, PyObject *);
@@ -266,7 +321,8 @@ struct npy_api {
         API_PyArray_DescrFromScalar = 57,
         API_PyArray_FromAny = 69,
         API_PyArray_Resize = 80,
-        API_PyArray_CopyInto = 82,
+        // CopyInto was slot 82 and 50 was effectively an alias. NumPy 2 removed 82.
+        API_PyArray_CopyInto = 50,
         API_PyArray_NewCopy = 85,
         API_PyArray_NewFromDescr = 94,
         API_PyArray_DescrNewFromType = 96,
@@ -275,7 +331,9 @@ struct npy_api {
         API_PyArray_View = 137,
         API_PyArray_DescrConverter = 174,
         API_PyArray_EquivTypes = 182,
+#ifdef PYBIND11_NUMPY_1_ONLY
         API_PyArray_GetArrayParamsFromObject = 278,
+#endif
         API_PyArray_SetBaseObject = 282
     };
 
@@ -290,7 +348,8 @@ struct npy_api {
         npy_api api;
 #define DECL_NPY_API(Func) api.Func##_ = (decltype(api.Func##_)) api_ptr[API_##Func];
         DECL_NPY_API(PyArray_GetNDArrayCFeatureVersion);
-        if (api.PyArray_GetNDArrayCFeatureVersion_() < 0x7) {
+        api.PyArray_RUNTIME_VERSION_ = api.PyArray_GetNDArrayCFeatureVersion_();
+        if (api.PyArray_RUNTIME_VERSION_ < 0x7) {
             pybind11_fail("pybind11 numpy support requires numpy >= 1.7.0");
         }
         DECL_NPY_API(PyArray_Type);
@@ -309,7 +368,9 @@ struct npy_api {
         DECL_NPY_API(PyArray_View);
         DECL_NPY_API(PyArray_DescrConverter);
         DECL_NPY_API(PyArray_EquivTypes);
+#ifdef PYBIND11_NUMPY_1_ONLY
         DECL_NPY_API(PyArray_GetArrayParamsFromObject);
+#endif
         DECL_NPY_API(PyArray_SetBaseObject);
 
 #undef DECL_NPY_API
@@ -331,6 +392,14 @@ inline const PyArrayDescr_Proxy *array_descriptor_proxy(const PyObject *ptr) {
     return reinterpret_cast<const PyArrayDescr_Proxy *>(ptr);
 }
 
+inline const PyArrayDescr1_Proxy *array_descriptor1_proxy(const PyObject *ptr) {
+    return reinterpret_cast<const PyArrayDescr1_Proxy *>(ptr);
+}
+
+inline const PyArrayDescr2_Proxy *array_descriptor2_proxy(const PyObject *ptr) {
+    return reinterpret_cast<const PyArrayDescr2_Proxy *>(ptr);
+}
+
 inline bool check_flags(const void *ptr, int flag) {
     return (flag == (array_proxy(ptr)->flags & flag));
 }
@@ -610,10 +679,32 @@ class dtype : public object {
     }
 
     /// Size of the data type in bytes.
+#ifdef PYBIND11_NUMPY_1_ONLY
     ssize_t itemsize() const { return detail::array_descriptor_proxy(m_ptr)->elsize; }
+#else
+    ssize_t itemsize() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return detail::array_descriptor1_proxy(m_ptr)->elsize;
+        }
+        return detail::array_descriptor2_proxy(m_ptr)->elsize;
+    }
+#endif
 
     /// Returns true for structured data types.
+#ifdef PYBIND11_NUMPY_1_ONLY
     bool has_fields() const { return detail::array_descriptor_proxy(m_ptr)->names != nullptr; }
+#else
+    bool has_fields() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return detail::array_descriptor1_proxy(m_ptr)->names != nullptr;
+        }
+        const auto *proxy = detail::array_descriptor2_proxy(m_ptr);
+        if (proxy->type_num < 0 || proxy->type_num >= 2056) {
+            return false;
+        }
+        return proxy->names != nullptr;
+    }
+#endif
 
     /// Single-character code for dtype's kind.
     /// For example, floating point types are 'f' and integral types are 'i'.
@@ -639,11 +730,29 @@ class dtype : public object {
     /// Single character for byteorder
     char byteorder() const { return detail::array_descriptor_proxy(m_ptr)->byteorder; }
 
-    /// Alignment of the data type
+/// Alignment of the data type
+#ifdef PYBIND11_NUMPY_1_ONLY
     int alignment() const { return detail::array_descriptor_proxy(m_ptr)->alignment; }
+#else
+    ssize_t alignment() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return detail::array_descriptor1_proxy(m_ptr)->alignment;
+        }
+        return detail::array_descriptor2_proxy(m_ptr)->alignment;
+    }
+#endif
 
-    /// Flags for the array descriptor
+/// Flags for the array descriptor
+#ifdef PYBIND11_NUMPY_1_ONLY
     char flags() const { return detail::array_descriptor_proxy(m_ptr)->flags; }
+#else
+    std::uint64_t flags() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return (unsigned char) detail::array_descriptor1_proxy(m_ptr)->flags;
+        }
+        return detail::array_descriptor2_proxy(m_ptr)->flags;
+    }
+#endif
 
 private:
     static object &_dtype_from_pep3118() {
@@ -810,9 +919,7 @@ class array : public buffer {
     }
 
     /// Byte size of a single element
-    ssize_t itemsize() const {
-        return detail::array_descriptor_proxy(detail::array_proxy(m_ptr)->descr)->elsize;
-    }
+    ssize_t itemsize() const { return dtype().itemsize(); }
 
     /// Total number of bytes
     ssize_t nbytes() const { return size() * itemsize(); }

tests/conftest.py

@@ -218,4 +218,5 @@ def pytest_report_header(config):
         f" {pybind11_tests.cpp_std}"
         f" {pybind11_tests.PYBIND11_INTERNALS_ID}"
         f" PYBIND11_SIMPLE_GIL_MANAGEMENT={pybind11_tests.PYBIND11_SIMPLE_GIL_MANAGEMENT}"
+        f" PYBIND11_NUMPY_1_ONLY={pybind11_tests.PYBIND11_NUMPY_1_ONLY}"
     )

tests/pybind11_tests.cpp

@@ -95,6 +95,12 @@ PYBIND11_MODULE(pybind11_tests, m) {
 #else
         false;
 #endif
+    m.attr("PYBIND11_NUMPY_1_ONLY") =
+#if defined(PYBIND11_NUMPY_1_ONLY)
+        true;
+#else
+        false;
+#endif
 
     bind_ConstructorStats(m);
 

tests/test_eigen_matrix.py

@@ -608,7 +608,9 @@ def test_both_ref_mutators():
 def test_nocopy_wrapper():
     # get_elem requires a column-contiguous matrix reference, but should be
     # callable with other types of matrix (via copying):
-    int_matrix_colmajor = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], order="F")
+    int_matrix_colmajor = np.array(
+        [[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype="l", order="F"
+    )
     dbl_matrix_colmajor = np.array(
         int_matrix_colmajor, dtype="double", order="F", copy=True
     )