[libtorch] create a new port

ports/libtorch/fix-c10-glog.patch

@@ -0,0 +1,61 @@
+diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
+index fe74e49..06e09d6 100644
+--- a/c10/util/Logging.cpp
++++ b/c10/util/Logging.cpp
+@@ -4,7 +4,7 @@
+ #ifdef FBCODE_CAFFE2
+ #include <folly/synchronization/SanitizeThread.h>
+ #endif
+-
++#include <glog/logging.h>
+ #include <algorithm>
+ #include <cstdlib>
+ #include <cstring>
+@@ -192,23 +194,13 @@ C10_DEFINE_int(
+     google::GLOG_WARNING,
+     "The minimum log level that caffe2 will output.");
+
+-// Google glog's api does not have an external function that allows one to check
+-// if glog is initialized or not. It does have an internal function - so we are
+-// declaring it here. This is a hack but has been used by a bunch of others too
+-// (e.g. Torch).
+-namespace google {
+-namespace glog_internal_namespace_ {
+-bool IsGoogleLoggingInitialized();
+-} // namespace glog_internal_namespace_
+-} // namespace google
+-
+ namespace c10 {
+ namespace {
+
+ void initGoogleLogging(char const* name) {
+ #if !defined(_MSC_VER)
+   // This trick can only be used on UNIX platforms
+-  if (!::google::glog_internal_namespace_::IsGoogleLoggingInitialized())
++  if (!::google::IsGoogleLoggingInitialized())
+ #endif
+   {
+     ::google::InitGoogleLogging(name);
+diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
+index 1864f17..b958fa0 100644
+--- a/c10/util/Logging.cpp
++++ b/c10/util/Logging.cpp
+@@ -154,18 +154,6 @@ void setLogLevelFlagFromEnv();
+ } // namespace detail
+ } // namespace c10
+
+-#if defined(C10_USE_GFLAGS) && defined(C10_USE_GLOG)
+-// When GLOG depends on GFLAGS, these variables are being defined in GLOG
+-// directly via the GFLAGS definition, so we will use DECLARE_* to declare
+-// them, and use them in Caffe2.
+-// GLOG's minloglevel
+-DECLARE_int32(minloglevel);
+-// GLOG's verbose log value.
+-DECLARE_int32(v);
+-// GLOG's logtostderr value
+-DECLARE_bool(logtostderr);
+-#endif // defined(C10_USE_GFLAGS) && defined(C10_USE_GLOG)
+-
+ #if !defined(C10_USE_GLOG)
+ // This backward compatibility flags are in order to deal with cases where
+ // Caffe2 are not built with glog, but some init flags still pass in these

ports/libtorch/fix-cmake.patch

@@ -0,0 +1,142 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 4bf1e85..48b01ae 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -693,7 +693,7 @@ endif()
+
+ set(BUILD_ONEDNN_GRAPH OFF)
+
+-include(cmake/Dependencies.cmake)
++include(cmake/vcpkg-dependencies.cmake)
+
+ if(USE_CUDA AND (CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 10.2) AND (CMAKE_HOST_SYSTEM_NAME MATCHES "Windows"))
+   # CUDA < 10.2 doesn't support compiling and extracting header dependencies in
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 48b01ae..8e46275 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -660,7 +660,7 @@ if(NOT CMAKE_BUILD_TYPE)
+ endif()
+
+ # The below means we are cross compiling for arm64 or x86_64 on MacOSX
+-if(NOT IOS AND CMAKE_SYSTEM_NAME STREQUAL "Darwin" AND CMAKE_OSX_ARCHITECTURES MATCHES "^(x86_64|arm64)$")
++if(false)
+   set(CROSS_COMPILING_MACOSX TRUE)
+   # We need to compile a universal protoc to not fail protobuf build
+   # We set CMAKE_TRY_COMPILE_TARGET_TYPE to STATIC_LIBRARY (vs executable) to succeed the cmake compiler check for cross-compiling
+@@ -678,6 +678,7 @@ if(NOT IOS AND CMAKE_SYSTEM_NAME STREQUAL "Darwin" AND CMAKE_OSX_ARCHITECTURES M
+   set(PROTOBUF_PROTOC_EXECUTABLE "${PROJECT_SOURCE_DIR}/build_host_protoc/bin/protoc")
+   set(CAFFE2_CUSTOM_PROTOC_EXECUTABLE "${PROJECT_SOURCE_DIR}/build_host_protoc/bin/protoc")
+ endif()
++include(cmake/ProtoBuf.cmake)
+
+ # ---[ Misc checks to cope with various compiler modes
+ include(cmake/MiscCheck.cmake)
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 328210c..c407442 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -112,13 +112,8 @@ endif()
+ # addressed yet.
+
+ if(NOT MSVC AND USE_XNNPACK)
+-  if(NOT TARGET fxdiv)
+-    set(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+-    set(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
+-    add_subdirectory(
+-      "${FXDIV_SOURCE_DIR}"
+-      "${CMAKE_BINARY_DIR}/FXdiv")
+-  endif()
++  find_path(FXDIV_INCLUDE_DIRS "fxdiv.h" REQUIRED)
++  include_directories(${FXDIV_INCLUDE_DIRS})
+ endif()
+
+ add_subdirectory(core)
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index c407442..16fa898 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1011,10 +1011,6 @@ if(USE_CUDA OR USE_ROCM)
+   include(${TORCH_SRC_DIR}/csrc/jit/codegen/cuda/nvfuser.cmake)
+ endif()
+
+-if(NOT MSVC AND USE_XNNPACK)
+-  TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
+-endif()
+-
+ # ==========================================================
+ # formerly-libtorch flags
+ # ==========================================================
+diff --git a/c10/CMakeLists.txt b/c10/CMakeLists.txt
+index 41b1a1a..3b390ec 100644
+--- a/c10/CMakeLists.txt
++++ b/c10/CMakeLists.txt
+@@ -110,7 +110,11 @@ endif()
+ # Note: for now, we will put all export path into one single Caffe2Targets group
+ # to deal with the cmake deployment need. Inside the Caffe2Targets set, the
+ # individual libraries like libc10.so and libcaffe2.so are still self-contained.
+-install(TARGETS c10 EXPORT Caffe2Targets DESTINATION lib)
++install(TARGETS c10 EXPORT Caffe2Targets
++  RUNTIME DESTINATION bin
++  LIBRARY DESTINATION lib
++  ARCHIVE DESTINATION lib
++)
+ install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}
+         DESTINATION include
+         FILES_MATCHING PATTERN "*.h")
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 16fa898..7a343f7 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1547,7 +1547,11 @@ endif()
+
+ caffe2_interface_library(torch torch_library)
+
+-install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
++install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets
++  RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
++  LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
++  ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
++)
+
+ if(USE_CUDA)
+   install(TARGETS torch_cuda torch_cuda_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+@@ -1558,7 +1562,11 @@ if(USE_CUDA)
+ elseif(USE_ROCM)
+   install(TARGETS torch_hip torch_hip_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ endif()
+-install(TARGETS torch torch_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
++install(TARGETS torch torch_library EXPORT Caffe2Targets
++  RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
++  LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
++  ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
++)
+
+ target_link_libraries(torch PUBLIC torch_cpu_library)
+
+@@ -1688,7 +1696,11 @@ if(BUILD_SHARED_LIBS)
+     target_link_libraries(torch_global_deps TBB::tbb)
+   endif()
+
+-  install(TARGETS torch_global_deps DESTINATION "${TORCH_INSTALL_LIB_DIR}")
++  install(TARGETS torch_global_deps
++    RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
++    LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
++    ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
++  )
+ endif()
+
+ # ---[ Caffe2 HIP sources.
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 7a343f7..3390f9b 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1673,7 +1673,7 @@ target_link_libraries(torch_cpu PRIVATE flatbuffers)
+ # namespaces, so libtorch is loaded with all its dependencies in a local scope.
+ # That usually leads to missing symbol errors at run-time, so to avoid a situation like
+ # this we have to preload those libs in a global namespace.
+-if(BUILD_SHARED_LIBS)
++if(BUILD_SHARED_LIBS AND (NOT WIN32))
+   add_library(torch_global_deps SHARED ${TORCH_SRC_DIR}/csrc/empty.c)
+   if(HAVE_SOVERSION)
+     set_target_properties(torch_global_deps PROPERTIES

ports/libtorch/fix-fbgemm-include.patch

@@ -0,0 +1,32 @@
+diff --git a/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp b/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp
+index 224a66f..eb42317 100644
+--- a/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp
++++ b/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp
+@@ -8,7 +8,10 @@
+ #include <torch/library.h>
+
+ #include <c10/util/irange.h>
+-
++#ifdef USE_FBGEMM
++#include <fbgemm/Fbgemm.h>
++#include <fbgemm/FbgemmEmbedding.h>
++#endif
+ int register_embedding_params();
+
+ /*
+diff --git a/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp b/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp
+index f94d27b..299b88b 100644
+--- a/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp
++++ b/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp
+@@ -4,7 +4,10 @@
+ #include <ATen/native/quantized/cpu/fbgemm_utils.h>
+ #include <c10/util/irange.h>
+ #include <torch/library.h>
+-
++#ifdef USE_FBGEMM
++#include <fbgemm/Fbgemm.h>
++#include <fbgemm/FbgemmEmbedding.h>
++#endif
+ int register_embedding_params();
+
+ at::Tensor PackedEmbeddingBagWeight::unpack() {

ports/libtorch/fix-windows.patch

@@ -0,0 +1,100 @@
+diff --git a/torch/csrc/jit/tensorexpr/bounds_overlap.h b/torch/csrc/jit/tensorexpr/bounds_overlap.h
+index c78884d..baf44bb 100644
+--- a/torch/csrc/jit/tensorexpr/bounds_overlap.h
++++ b/torch/csrc/jit/tensorexpr/bounds_overlap.h
+@@ -65,6 +65,10 @@ enum OverlapKind { ContainedOrEqual, Contains, PartialOverlap, NoOverlap };
+ //     operator
+ //     NOT_DETERMINED: Some elements satify the given comparison operator and
+ //     some elements not
++#if defined(_WIN32)
++#undef TRUE
++#undef FALSE
++#endif
+ enum CmpEvalResult { TRUE, FALSE, NOT_DETERMINED };
+
+ // Returns the kind of overlap between Bound A and Bound A in a single
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 3390f9b..578d0bd 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1425,7 +1425,7 @@ target_include_directories(torch_cpu SYSTEM PRIVATE "${Caffe2_DEPENDENCY_INCLUDE
+ torch_set_target_props(torch_cpu)
+
+
+-target_compile_options(torch_cpu PRIVATE "-DCAFFE2_BUILD_MAIN_LIB")
++target_compile_definitions(torch_cpu PRIVATE CAFFE2_BUILD_MAIN_LIB)
+ if(BUILD_SPLIT_CUDA)
+   target_compile_options(torch_cuda_cu PRIVATE "-DTORCH_CUDA_CU_BUILD_MAIN_LIB")
+   target_compile_options(torch_cuda_cpp PRIVATE "-DTORCH_CUDA_CPP_BUILD_MAIN_LIB")
+@@ -1525,12 +1525,12 @@ if(MSVC AND NOT BUILD_SHARED_LIBS)
+   #
+   # NB: This must be target_compile_definitions, not target_compile_options,
+   # as the latter is not respected by nvcc
+-  target_compile_definitions(torch_cpu PUBLIC "AT_CORE_STATIC_WINDOWS=1")
++  target_compile_definitions(torch_cpu PUBLIC AT_CORE_STATIC_WINDOWS=1)
+ endif()
+ if(MSVC AND BUILD_SHARED_LIBS)
+   # ONNX is linked statically and needs to be exported from this library
+   # to be used externally. Make sure that references match the export.
+-  target_compile_options(torch_cpu PRIVATE "-DONNX_BUILD_MAIN_LIB")
++  target_compile_definitions(torch_cpu PRIVATE ONNX_BUILD_MAIN_LIB)
+ endif()
+
+ caffe2_interface_library(torch_cpu torch_cpu_library)
+diff --git a/aten/src/ATen/Parallel.h b/aten/src/ATen/Parallel.h
+index 0b494d9..06b5d88 100644
+--- a/aten/src/ATen/Parallel.h
++++ b/aten/src/ATen/Parallel.h
+@@ -28,13 +28,7 @@ TORCH_API bool in_parallel_region();
+ namespace internal {
+
+ // Initialise num_threads lazily at first parallel call
+-inline TORCH_API void lazy_init_num_threads() {
+-  thread_local bool init = false;
+-  if (C10_UNLIKELY(!init)) {
+-    at::init_num_threads();
+-    init = true;
+-  }
+-}
++TORCH_API void lazy_init_num_threads();
+
+ TORCH_API void set_thread_num(int);
+
+diff --git a/aten/src/ATen/ParallelCommon.cpp b/aten/src/ATen/ParallelCommon.cpp
+index 068c9d3..b869fde 100644
+--- a/aten/src/ATen/ParallelCommon.cpp
++++ b/aten/src/ATen/ParallelCommon.cpp
+@@ -17,6 +17,17 @@
+
+ namespace at {
+
++namespace internal {
++/// @see https://github.com/pytorch/pytorch/issues/87957
++void lazy_init_num_threads() {
++  thread_local bool init = false;
++  if (C10_UNLIKELY(!init)) {
++    at::init_num_threads();
++    init = true;
++  }
++}
++} // namespace internal
++
+ namespace {
+
+ const char* get_env_var(
+diff --git a/c10/CMakeLists.txt b/c10/CMakeLists.txt
+index 3b390ec..7cfa2b0 100644
+--- a/c10/CMakeLists.txt
++++ b/c10/CMakeLists.txt
+@@ -45,7 +45,10 @@ if(HAVE_SOVERSION)
+       VERSION ${TORCH_VERSION} SOVERSION ${TORCH_SOVERSION})
+ endif()
+ # If building shared library, set dllimport/dllexport proper.
+-target_compile_options(c10 PRIVATE "-DC10_BUILD_MAIN_LIB")
++target_compile_definitions(c10 PRIVATE C10_BUILD_MAIN_LIB)
++if(WIN32)
++  target_compile_definitions(c10 PRIVATE NOMINMAX)
++endif()
+ # Enable hidden visibility if compiler supports it.
+ if(${COMPILER_SUPPORTS_HIDDEN_VISIBILITY})
+   target_compile_options(c10 PRIVATE "-fvisibility=hidden")