Ensure torchvision operators are added in C++ (#2798)

* Ensure torchvision operators are registered in C++ via weak symbols

* Add note to README on how to ensure that torchvision operators are available in C++

* Fix dllimport/dllexport on windows, format files

* Factor out common macros in single file

* Expose cuda_version in the API, use it to avoid pruning of ops initializer

README.rst

@@ -123,6 +123,11 @@ so make sure that it is also available to cmake via the ``CMAKE_PREFIX_PATH``.
 
 For an example setup, take a look at ``examples/cpp/hello_world``.
 
+TorchVision Operators
+---------------------
+In order to get the torchvision operators registered with torch (eg. for the JIT), all you need to do is to ensure that you
+:code:`#include <torchvision/vision.h>` in your project.
+
 Documentation
 =============
 You can find the API documentation on the pytorch website: https://pytorch.org/docs/stable/torchvision/index.html

torchvision/csrc/cpu/vision_cpu.h

 #pragma once
 #include <torch/extension.h>
-
-#ifdef _WIN32
-#if defined(torchvision_EXPORTS)
-#define VISION_API __declspec(dllexport)
-#else
-#define VISION_API __declspec(dllimport)
-#endif
-#else
-#define VISION_API
-#endif
+#include "../macros.h"
 
 VISION_API std::tuple<at::Tensor, at::Tensor> ROIPool_forward_cpu(
     const at::Tensor& input,

torchvision/csrc/cuda/vision_cuda.h

 #pragma once
 #include <torch/extension.h>
-
-#ifdef _WIN32
-#if defined(torchvision_EXPORTS)
-#define VISION_API __declspec(dllexport)
-#else
-#define VISION_API __declspec(dllimport)
-#endif
-#else
-#define VISION_API
-#endif
+#include "../macros.h"
 
 VISION_API at::Tensor ROIAlign_forward_cuda(
     const at::Tensor& input,

torchvision/csrc/macros.h

+#ifndef TORCHVISION_MACROS_H
+#define TORCHVISION_MACROS_H
+
+#ifdef _WIN32
+#if defined(torchvision_EXPORTS)
+#define VISION_API __declspec(dllexport)
+#else
+#define VISION_API __declspec(dllimport)
+#endif
+#else
+#define VISION_API
+#endif
+
+#if (defined __cpp_inline_variables) || __cplusplus >= 201703L
+#define VISION_INLINE_VARIABLE inline
+#else
+#ifdef _MSC_VER
+#define VISION_INLINE_VARIABLE __declspec(selectany)
+#else
+#define VISION_INLINE_VARIABLE __attribute__((weak))
+#endif
+#endif
+
+#endif // TORCHVISION_MACROS_H

torchvision/csrc/vision.cpp

@@ -34,13 +34,15 @@ PyMODINIT_FUNC PyInit__C(void) {
 #endif
 #endif
 
-int64_t _cuda_version() {
+namespace vision {
+int64_t cuda_version() noexcept {
 #ifdef WITH_CUDA
   return CUDA_VERSION;
 #else
   return -1;
 #endif
 }
+} // namespace vision
 
 TORCH_LIBRARY(torchvision, m) {
   m.def("nms(Tensor dets, Tensor scores, float iou_threshold) -> Tensor");
@@ -53,7 +55,7 @@ TORCH_LIBRARY(torchvision, m) {
   m.def("ps_roi_align", &ps_roi_align);
   m.def("ps_roi_pool", &ps_roi_pool);
   m.def("deform_conv2d", &deform_conv2d);
-  m.def("_cuda_version", &_cuda_version);
+  m.def("_cuda_version", &vision::cuda_version);
 }
 
 TORCH_LIBRARY_IMPL(torchvision, CPU, m) {

torchvision/csrc/vision.h

@@ -2,5 +2,18 @@
 #define VISION_H
 
 #include <torchvision/models/models.h>
+#include <cstdint>
+#include "macros.h"
+
+namespace vision {
+VISION_API int64_t cuda_version() noexcept;
+
+namespace detail {
+// Dummy variable to reference a symbol from vision.cpp.
+// This ensures that the torchvision library and the ops registration
+// initializers are not pruned.
+VISION_INLINE_VARIABLE int64_t _cuda_version = cuda_version();
+} // namespace detail
+} // namespace vision
 
 #endif // VISION_H