Fix issue with debug VS2022 build when python bindings are enabled (#9794)

* Add intermediate header between the ORT code and pybind11 to workaround an issue with VS2022 debug builds by making sure corecrt.h is included first.

This avoids the _STL_ASSERT macro being defined in an incompatible way for a debug build by pybind including the python headers with _DEBUG temporarily undefined .

See #9735 for details.

onnxruntime/python/onnxruntime_pybind.h

@@ -0,0 +1,18 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+
+// If performing a debug build with VS2022 (_MSC_VER == 1930) we need to include corecrt.h before pybind
+// so that the _STL_ASSERT macro is defined in a compatible way.
+//
+// pybind11/pybind11.h includes pybind11/detail/common.h, which undefines _DEBUG whilst including the Python headers
+// (which in turn include corecrt.h). This alters how the _STL_ASSERT macro is defined and causes the build to fail.
+//
+// see https://github.com/microsoft/onnxruntime/issues/9735
+//
+#if defined(_MSC_VER) && defined(_DEBUG) && _MSC_VER >= 1930
+#include <corecrt.h>
+#endif
+
+#include <pybind11/pybind11.h>

onnxruntime/python/onnxruntime_pybind_exceptions.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include <pybind11/pybind11.h>
+#include "onnxruntime_pybind.h"  // must use this for the include of <pybind11/pybind11.h>
 #include <stdexcept>
 #include "core/common/status.h"
 

onnxruntime/python/onnxruntime_pybind_mlvalue.h

@@ -4,8 +4,8 @@
 #pragma once
 #define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
 
+#include "onnxruntime_pybind.h"  // must use this for the include of <pybind11/pybind11.h>
 #include <pybind11/iostream.h>
-#include <pybind11/pybind11.h>
 #include <pybind11/numpy.h>
 #include <pybind11/stl.h>
 

onnxruntime/python/onnxruntime_pybind_module.cc

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include <pybind11/pybind11.h>
+#include "onnxruntime_pybind.h"  // must use this for the include of <pybind11/pybind11.h>
 #include <pybind11/stl.h>
 #include "core/providers/get_execution_providers.h"
 
@@ -20,5 +20,5 @@ PYBIND11_MODULE(onnxruntime_pybind11_state, m) {
       "The order of elements represents the default priority order of Execution Providers "
       "from highest to lowest.");
 }
-}
-}
+}  // namespace python
+}  // namespace onnxruntime

onnxruntime/python/onnxruntime_pybind_state.h

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include <pybind11/pybind11.h>
+#include "onnxruntime_pybind.h"  // must use this for the include of <pybind11/pybind11.h>
 
 namespace onnxruntime {
 namespace python {
@@ -10,6 +10,5 @@ void addGlobalMethods(py::module& m, Environment& env);
 void addObjectMethods(py::module& m, Environment& env);
 void addOrtValueMethods(pybind11::module& m);
 
-
-}
-}
+}  // namespace python
+}  // namespace onnxruntime

onnxruntime/python/onnxruntime_pybind_state_common.h

@@ -15,7 +15,7 @@
 #include "core/dlpack/dlpack_converter.h"
 #endif
 
-#include <pybind11/pybind11.h>
+#include "onnxruntime_pybind.h"  // must use this for the include of <pybind11/pybind11.h>
 
 // execution provider factory creator headers
 struct OrtStatus {
@@ -207,7 +207,7 @@ class SparseTensor;
 #endif
 namespace python {
 
-using ExecutionProviderRegistrationFn = std::function<void(InferenceSession*, 
+using ExecutionProviderRegistrationFn = std::function<void(InferenceSession*,
                                                            const std::vector<std::string>&,
                                                            const ProviderOptionsMap&)>;
 
@@ -313,7 +313,7 @@ inline AllocatorPtr& GetAllocator() {
 //   so incoming arrays do not disappear. To this end we create an instance of SparseTensor
 //   on top of the user provided numpy arrays and create a duplicate of py::objects for those
 //   numpy array for ref-counting purposes and store it here.
-// 
+//
 // - to be able to expose SparseTensor returned from run method. We get an OrtValue from run()
 //   and store a copy of it in ort_value_. The OrtValue shared_ptr ref-counting will make sure
 //   the memory stays around.
@@ -375,7 +375,6 @@ class PySparseTensor {
   std::unique_ptr<OrtValue> AsOrtValue() const;
 
  private:
-
   //  instance_ represents data that comes as input. Thus we depend on numpy
   // arrays that own the underlying memory to stay around. We store copies
   // of py::objects for those arrays in backing_storage_ as an extra ref-count.
@@ -480,5 +479,5 @@ std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_Nnapi(
 std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_Rknpu();
 std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_CoreML(uint32_t flags);
 
-constexpr const char* kDefaultExecutionProviderEntry = "GetProvider"; 
+constexpr const char* kDefaultExecutionProviderEntry = "GetProvider";
 }  // namespace onnxruntime