Windows shared build (pytorch#13550)

Summary:
Hi guys,

I'd like to build Caffe2 with more supported options in Windows with Microsoft Visual Studios.
This is the first pull request.
Running scripts/build_windows_shared.bat is able to build Caffe2 with both CMAKE_BUILD_TYPE=Debug and CMAKE_BUILD_TYPE=Release with Visual Studio 14 2015.
CUDA is 9.0, cudnn is 7.0.5, glog, gflags and lmdb are supported on my system.
Python is 3.5, Detectron works from python interface as well.
It was even possible to debug detectron code and step into caffe2_gpu.dll with pdbs built.

What is disappointing, that c10/experimental ops don't build with this Visual Studio generator, I added special option INCLUDE_EXPERIMENTAL_C10_OPS (default ON) to deal with it in build_windows_shared.bat.

After this pull request the next step is to add Visual Studio 2017 support in the script.
Pull Request resolved: pytorch#13550

Reviewed By: ezyang

Differential Revision: D13042597

Pulled By: orionr

fbshipit-source-id: f313f909f599cd582a1d000eff766eef3a9fc4fc

Author: ArutyunovG

CMakeLists.txt

@@ -65,6 +65,7 @@ option(BUILD_DOCS "Build Caffe2 documentation" OFF)
 option(BUILD_CUSTOM_PROTOBUF "Build and use Caffe2's own protobuf under third_party" ON)
 option(BUILD_PYTHON "Build Python binaries" ON)
 option(BUILD_CAFFE2_OPS "Build Caffe2 operators" ON)
+option(BUILD_C10_EXPERIMENTAL_OPS "Build c10 experimental operators" ON)
 option(BUILD_SHARED_LIBS "Build libcaffe2.so" ON)
 cmake_dependent_option(
     CAFFE2_LINK_LOCAL_PROTOBUF "If set, build protobuf inside libcaffe2.so." ON

aten/src/ATen/core/TensorTypeId.cpp

@@ -1,40 +1,2 @@
 #pragma once
-
-#include <iostream>
-#include <string>
-#include "c10/util/IdWrapper.h"
-#include "c10/macros/Macros.h"
-
-namespace at {
-
-namespace details {
-using _tensorTypeId_underlyingType = uint8_t;
-}
-
-/**
- * Dynamic type ID of a Tensor argument.  It represents something like
- * CPUTensor, etc.
- */
-class CAFFE2_API TensorTypeId final
-    : public at::
-          IdWrapper<TensorTypeId, details::_tensorTypeId_underlyingType> {
- public:
-  // Don't use this!
-  // Unfortunately, a default constructor needs to be defined because of
-  // https://reviews.llvm.org/D41223
-  constexpr TensorTypeId() noexcept : IdWrapper(0) {}
-
- private:
-  constexpr explicit TensorTypeId(
-      details::_tensorTypeId_underlyingType id) noexcept
-      : IdWrapper(id) {}
-
-  friend class TensorTypeIdCreator;
-  friend CAFFE2_API std::ostream& operator<<(std::ostream&, TensorTypeId);
-};
-
-CAFFE2_API std::ostream& operator<<(std::ostream&, at::TensorTypeId);
-
-} // namespace at
-
-C10_DEFINE_HASH_FOR_IDWRAPPER(at::TensorTypeId)
+#include <c10/util/TensorTypeId.h>

aten/src/ATen/core/TensorTypeId.h

@@ -1,109 +1,2 @@
 #pragma once
-
-/**
- * To register your own tensor types, do in a header file:
- *   AT_DECLARE_TENSOR_TYPE(MY_TENSOR)
- * and in one (!) cpp file:
- *   AT_DEFINE_TENSOR_TYPE(MY_TENSOR)
- * Both must be in the same namespace.
- */
-
-#include "ATen/core/TensorTypeId.h"
-#include "c10/macros/Macros.h"
-
-#include <atomic>
-#include <mutex>
-#include <unordered_set>
-
-namespace at {
-
-class CAFFE2_API TensorTypeIdCreator final {
- public:
-  TensorTypeIdCreator();
-
-  at::TensorTypeId create();
-
-  static constexpr at::TensorTypeId undefined() noexcept {
-    return TensorTypeId(0);
-  }
-
- private:
-  std::atomic<details::_tensorTypeId_underlyingType> last_id_;
-
-  C10_DISABLE_COPY_AND_ASSIGN(TensorTypeIdCreator);
-};
-
-class CAFFE2_API TensorTypeIdRegistry final {
- public:
-  TensorTypeIdRegistry();
-
-  void registerId(at::TensorTypeId id);
-  void deregisterId(at::TensorTypeId id);
-
- private:
-  std::unordered_set<at::TensorTypeId> registeredTypeIds_;
-  std::mutex mutex_;
-
-  C10_DISABLE_COPY_AND_ASSIGN(TensorTypeIdRegistry);
-};
-
-class CAFFE2_API TensorTypeIds final {
- public:
-  static TensorTypeIds& singleton();
-
-  at::TensorTypeId createAndRegister();
-  void deregister(at::TensorTypeId id);
-
-  static constexpr at::TensorTypeId undefined() noexcept;
-
- private:
-  TensorTypeIds();
-
-  TensorTypeIdCreator creator_;
-  TensorTypeIdRegistry registry_;
-
-  C10_DISABLE_COPY_AND_ASSIGN(TensorTypeIds);
-};
-
-inline constexpr at::TensorTypeId TensorTypeIds::undefined() noexcept {
-  return TensorTypeIdCreator::undefined();
-}
-
-class CAFFE2_API TensorTypeIdRegistrar final {
- public:
-  TensorTypeIdRegistrar();
-  ~TensorTypeIdRegistrar();
-
-  at::TensorTypeId id() const noexcept;
-
- private:
-  at::TensorTypeId id_;
-
-  C10_DISABLE_COPY_AND_ASSIGN(TensorTypeIdRegistrar);
-};
-
-inline at::TensorTypeId TensorTypeIdRegistrar::id() const noexcept {
-  return id_;
-}
-
-#define AT_DECLARE_TENSOR_TYPE(TensorName) \
-  CAFFE2_API at::TensorTypeId TensorName()
-
-#define AT_DEFINE_TENSOR_TYPE(TensorName)           \
-  at::TensorTypeId TensorName() {                   \
-    static TensorTypeIdRegistrar registration_raii; \
-    return registration_raii.id();                  \
-  }
-
-AT_DECLARE_TENSOR_TYPE(UndefinedTensorId);
-AT_DECLARE_TENSOR_TYPE(CPUTensorId); // PyTorch/Caffe2 supported
-AT_DECLARE_TENSOR_TYPE(CUDATensorId); // PyTorch/Caffe2 supported
-AT_DECLARE_TENSOR_TYPE(SparseCPUTensorId); // PyTorch only
-AT_DECLARE_TENSOR_TYPE(SparseCUDATensorId); // PyTorch only
-AT_DECLARE_TENSOR_TYPE(MKLDNNTensorId); // Caffe2 only
-AT_DECLARE_TENSOR_TYPE(OpenGLTensorId); // Caffe2 only
-AT_DECLARE_TENSOR_TYPE(OpenCLTensorId); // Caffe2 only
-AT_DECLARE_TENSOR_TYPE(IDEEPTensorId); // Caffe2 only
-AT_DECLARE_TENSOR_TYPE(HIPTensorId); // Caffe2 only
-
-} // namespace at
+#include "c10/util/TensorTypeIdRegistration.h"

aten/src/ATen/core/TensorTypeIdRegistration.h

@@ -19,3 +19,6 @@ at::DataPtr THCIpcDeleter::makeDataPtr(void* data, int device) {
   auto* context = new THCIpcDeleter(data, device);
   return {data, context, &deleteTHCIpcDeleter, at::Device(at::DeviceType::CUDA, cur_device)};
 }
+
+THCIpcDeleter::THCIpcDeleter(void* data, int device)
+    : data_(data), device_(device) {}

aten/src/THC/THCAllocator.cpp

@@ -8,7 +8,7 @@
 #ifdef __cplusplus
 class CAFFE2_API THCIpcDeleter {
  public:
-  THCIpcDeleter(void* data, int device) : data_(data), device_(device) {};
+  THCIpcDeleter(void* data, int device);
   ~THCIpcDeleter();
   static at::DataPtr makeDataPtr(void* data, int device);
 private:

aten/src/THC/THCAllocator.h

@@ -141,7 +141,7 @@ void loadInput(
         vector<string> input_dims_str = caffe2::split(',', input_dims_list[i]);
         vector<int> input_dims;
         for (const string& s : input_dims_str) {
-          input_dims.push_back(caffe2::stoi(s));
+          input_dims.push_back(c10::stoi(s));
         }
         caffe2::Blob* blob = workspace->GetBlob(input_names[i]);
         if (blob == nullptr) {

binaries/benchmark_helper.cc

@@ -99,9 +99,9 @@ std::vector<float> convertToVector(cv::Mat& img) {
     } else if (step == "normalize") {
       normalize = {255, 255, 255};
     } else if (step == "mean") {
-      mean = {0.406, 0.456, 0.485};
+      mean = {0.406f, 0.456f, 0.485f};
     } else if (step == "std") {
-      std = {0.225, 0.224, 0.229};
+      std = {0.225f, 0.224f, 0.229f};
     } else if (step == "bgrtorgb") {
       bgrtorgb = true;
     } else {
@@ -143,9 +143,14 @@ std::vector<float> convertOneImage(std::string& filename) {
   assert(filename[0] != '~');
 
   std::cout << "Converting " << filename << std::endl;
+
   // Load image
   cv::Mat img = cv::imread(
+#if CV_MAJOR_VERSION <= 3
       filename, FLAGS_color ? CV_LOAD_IMAGE_COLOR : CV_LOAD_IMAGE_GRAYSCALE);
+#else
+      filename, FLAGS_color ? cv::IMREAD_COLOR : cv::IMREAD_GRAYSCALE);
+#endif
 
   cv::Mat crop = cropToSquare(img);
 

binaries/convert_image_to_tensor.cc

@@ -127,7 +127,7 @@ int main(int argc, char** argv) {
         vector<string> input_dims_str = caffe2::split(',', input_dims_list[i]);
         vector<int> input_dims;
         for (const string& s : input_dims_str) {
-          input_dims.push_back(caffe2::stoi(s));
+          input_dims.push_back(c10::stoi(s));
         }
         caffe2::Blob* blob = workspace->GetBlob(input_names[i]);
         if (blob == nullptr) {

binaries/speed_benchmark.cc

@@ -23,7 +23,7 @@ namespace c10 {
  * for you, given the underlying type supports it.
  */
 template <class ConcreteType, class UnderlyingType>
-class CAFFE2_API IdWrapper {
+class C10_API IdWrapper {
  public:
   using underlying_type = UnderlyingType;
   using concrete_type = ConcreteType;

c10/util/IdWrapper.h

@@ -53,7 +53,7 @@ static inline uint64_t NextPowerOf2(uint64_t A) {
 } // namespace detail
 
 /// This is all the non-templated stuff common to all SmallVectors.
-class CAFFE2_API SmallVectorBase {
+class C10_API SmallVectorBase {
  protected:
   void *BeginX, *EndX, *CapacityX;
 

c10/util/SmallVector.h

@@ -2,6 +2,7 @@
 #define C10_UTIL_STRINGUTIL_H_
 
 #include <c10/macros/Macros.h>
+#include <c10/util/string_utils.h>
 
 #include <cstddef>
 #include <ostream>
@@ -73,28 +74,6 @@ struct C10_API SourceLocation {
 
 std::ostream& operator<<(std::ostream& out, const SourceLocation& loc);
 
-/// Portable implementation of std::stoi, which works for Android builds.
-///
-/// TODO: You won't be able to call this unqualified, because ADL means that it
-/// will be ambiguous with std::stoi.  Maybe we should fix this by giving
-/// our version a different name.
-inline int stoi(const std::string& str) {
-#if defined(__ANDROID__)
-  std::stringstream ss;
-  int n = 0;
-  ss << str;
-  ss >> n;
-  return n;
-#else
-  return std::stoi(str);
-#endif // defined(__ANDROID__)
-}
-
 } // namespace c10
 
-// TODO: Remove me when namespace unification occurs
-namespace at {
-using c10::stoi;
-}
-
 #endif // C10_UTIL_STRINGUTIL_H_

c10/util/StringUtil.h

@@ -0,0 +1,10 @@
+#include "c10/util/TensorTypeId.h"
+#include "c10/util/string_utils.h"
+
+namespace c10 {
+
+std::ostream& operator<<(std::ostream& str, c10::TensorTypeId rhs) {
+  return str << c10::to_string(rhs.underlyingId());
+}
+
+} // namespace c10

c10/util/TensorTypeId.cpp

@@ -0,0 +1,43 @@
+#ifndef TENSOR_TYPE_ID_H_
+#define TENSOR_TYPE_ID_H_
+
+#include <iostream>
+#include <string>
+#include "c10/macros/Macros.h"
+#include "c10/util/IdWrapper.h"
+
+namespace c10 {
+
+namespace details {
+using _tensorTypeId_underlyingType = uint8_t;
+}
+
+/**
+ * Dynamic type ID of a Tensor argument.  It represents something like
+ * CPUTensor, etc.
+ */
+class C10_API TensorTypeId final
+    : public at::
+          IdWrapper<TensorTypeId, details::_tensorTypeId_underlyingType> {
+ public:
+  // Don't use this!
+  // Unfortunately, a default constructor needs to be defined because of
+  // https://reviews.llvm.org/D41223
+  constexpr TensorTypeId() noexcept : IdWrapper(0) {}
+
+ private:
+  constexpr explicit TensorTypeId(
+      details::_tensorTypeId_underlyingType id) noexcept
+      : IdWrapper(id) {}
+
+  friend class TensorTypeIdCreator;
+  friend C10_API std::ostream& operator<<(std::ostream&, TensorTypeId);
+};
+
+C10_API std::ostream& operator<<(std::ostream&, c10::TensorTypeId);
+
+} // namespace c10
+
+C10_DEFINE_HASH_FOR_IDWRAPPER(c10::TensorTypeId)
+
+#endif // TENSOR_TYPE_ID_H_