Enable CTC decoder in Windows (#2587)

Summary:
This commit enables CTC decoder on Windows.

The functionality seems to work fine.
The tests are passing, the decoding tutorial runs fine.

The only difference to the Linux/macOS version is that
loading model in XZ compression format is not supported.

![289961785_399620772041679_7768117002438616376_n](https://user-images.githubusercontent.com/855818/181420923-cfbd8402-20de-4e63-b9e4-e39f9aa9fc50.png)

Pull Request resolved: https://github.com/pytorch/audio/pull/2587

Reviewed By: carolineechen, nateanl

Differential Revision: D38276490

Pulled By: mthrok

fbshipit-source-id: f2203b2235c5bbb0220fe560aaaf0e1d5530347a

.circleci/config.yml

@@ -703,7 +703,6 @@ jobs:
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_MFCC_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_SOX: true
-              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CTC_DECODER: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CUDA: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_KALDI: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX: true
@@ -749,7 +748,6 @@ jobs:
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_MFCC_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_SOX: true
-              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CTC_DECODER: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_KALDI: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310: true

.circleci/config.yml.in

@@ -703,7 +703,6 @@ jobs:
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_MFCC_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_SOX: true
-              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CTC_DECODER: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CUDA: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_KALDI: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX: true
@@ -749,7 +748,6 @@ jobs:
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_MFCC_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_SOX: true
-              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CTC_DECODER: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_KALDI: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310: true

CMakeLists.txt

@@ -131,7 +131,12 @@ if(USE_OPENMP)
 endif()
 
 # TORCH_CXX_FLAGS contains the same -D_GLIBCXX_USE_CXX11_ABI value as PyTorch
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall ${TORCH_CXX_FLAGS}")
+if (MSVC)
+  set(warning_flags /W4)
+else()
+  set(warning_flags -Wall)
+endif()
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${warning_flags} ${TORCH_CXX_FLAGS}")
 
 add_subdirectory(third_party)
 add_subdirectory(torchaudio/csrc)

third_party/CMakeLists.txt

 set(CMAKE_CXX_VISIBILITY_PRESET hidden)
 
+file(MAKE_DIRECTORY install/include)
+file(MAKE_DIRECTORY install/lib)
+
 ################################################################################
 # sox
 ################################################################################
@@ -15,12 +18,14 @@ if (BUILD_KALDI)
 endif()
 
 ################################################################################
-# KenLM
+# Decoder
 ################################################################################
 if (BUILD_CTC_DECODER)
   add_subdirectory(zlib)
   add_subdirectory(bzip2)
+  if (NOT MSVC)
     add_subdirectory(lzma)
+  endif()
   add_subdirectory(kenlm)
   add_subdirectory(flashlight-text)
 endif()

third_party/bzip2/CMakeLists.txt

@@ -3,12 +3,36 @@ include(ExternalProject)
 set(INSTALL_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../install)
 set(ARCHIVE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../archives)
 
-set(
+if (MSVC)
+  set(
+    BZIP_LIBRARIES
+    ${INSTALL_DIR}/lib/libbz2.lib
+  )
+  ExternalProject_Add(bzip2-
+    PREFIX ${CMAKE_CURRENT_BINARY_DIR}
+    DOWNLOAD_DIR ${ARCHIVE_DIR}
+    URL https://sourceware.org/pub/bzip2/bzip2-1.0.8.tar.gz
+    URL_HASH SHA256=ab5a03176ee106d3f0fa90e381da478ddae405918153cca248e682cd0c4a2269
+    BUILD_BYPRODUCTS ${BZIP_LIBRARIES}
+    BUILD_IN_SOURCE 1
+    CONFIGURE_COMMAND nmake -f makefile.msc
+    BUILD_COMMAND cp bzlib.h ${INSTALL_DIR}/include
+    INSTALL_COMMAND cp libbz2.* ${INSTALL_DIR}/lib
+    DOWNLOAD_NO_PROGRESS ON
+    LOG_DOWNLOAD ON
+    LOG_UPDATE ON
+    LOG_CONFIGURE ON
+    LOG_BUILD ON
+    LOG_INSTALL ON
+    LOG_MERGED_STDOUTERR ON
+    LOG_OUTPUT_ON_FAILURE ON
+  )
+else()
+  set(
     BZIP_LIBRARIES
     ${INSTALL_DIR}/lib/libbz2.a
   )
-
-ExternalProject_Add(bzip2-
+  ExternalProject_Add(bzip2-
     PREFIX ${CMAKE_CURRENT_BINARY_DIR}
     DOWNLOAD_DIR ${ARCHIVE_DIR}
     URL https://sourceware.org/pub/bzip2/bzip2-1.0.8.tar.gz
@@ -26,7 +50,8 @@ ExternalProject_Add(bzip2-
     LOG_INSTALL ON
     LOG_MERGED_STDOUTERR ON
     LOG_OUTPUT_ON_FAILURE ON
-)
+  )
+endif(MSVC)
 
 add_library(bzip2 INTERFACE)
 add_dependencies(bzip2 bzip2-)

third_party/flashlight-text/CMakeLists.txt

@@ -11,6 +11,12 @@
 
 set(CMAKE_CXX_VISIBILITY_PRESET default)
 
+# the following line is added in order to export symbols when building on Windows
+# this approach has some limitations as documented in https://github.com/pytorch/pytorch/pull/3650
+if (MSVC)
+  set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+endif()
+
 set(
   libflashlight_src
   submodule/flashlight/lib/text/decoder/Utils.cpp

third_party/kenlm/CMakeLists.txt

+# kenlm uses std::binary_function, which had BC breaking change in C++17.
+# On Windows + GPU, torchaudio is compiled with C++ 17 globally.
+# We use C++ 14 for KenLM
+set(CMAKE_CXX_STANDARD 14)
+
 set(
   KENLM_UTIL_SOURCES
   kenlm/util/bit_packing.cc
@@ -54,18 +59,45 @@ target_include_directories(
   PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}"
   )
 
-target_compile_definitions(
-  kenlm
-  PUBLIC KENLM_MAX_ORDER=6
+set(
+  kenlm_deps
+  zlib
+  bzip2
+  )
+
+set(
+  kenlm_compiler_definitions
+  KENLM_MAX_ORDER=6
   HAVE_ZLIB
   HAVE_BZLIB
+  )
+
+if (MSVC)
+  list(
+    APPEND
+    kenlm_compiler_definitions
+    # To avoid warning C4003: not enough arguments for function-like macro invocation 'max'
+    NOMINMAX
+    )
+endif()
+
+if(TARGET lzma)
+  list(APPEND kenlm_deps lzma)
+  list(
+    APPEND
+    kenlm_compiler_definitions
     HAVE_XZLIB
     )
+endif()
+
+target_compile_definitions(
+  kenlm
+  PUBLIC
+  ${kenlm_compiler_definitions}
+  )
 
 target_link_libraries(
   kenlm
   PRIVATE
-  zlib
-  bzip2
-  lzma
+  ${kenlm_deps}
 )

third_party/zlib/CMakeLists.txt

@@ -3,23 +3,48 @@ include(ExternalProject)
 set(INSTALL_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../install)
 set(ARCHIVE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../archives)
 
-# To pass custom environment variables to ExternalProject_Add command,
-# we need to do `${CMAKE_COMMAND} -E env ${envs} <COMMANAD>`.
-# https://stackoverflow.com/a/62437353
-# We constrcut the custom environment variables here
-set(envs
+if (MSVC)
+  set(
+    ZLIB_LIBRARIES
+    ${INSTALL_DIR}/lib/zlib.lib
+  )
+
+  ExternalProject_Add(zlib-
+    PREFIX ${CMAKE_CURRENT_BINARY_DIR}
+    DOWNLOAD_DIR ${ARCHIVE_DIR}
+    URL https://github.com/madler/zlib/archive/refs/tags/v1.2.12.tar.gz
+    URL_HASH SHA256=d8688496ea40fb61787500e863cc63c9afcbc524468cedeb478068924eb54932
+    BUILD_BYPRODUCTS ${ZLIB_LIBRARIES}
+    BUILD_IN_SOURCE 1
+    CONFIGURE_COMMAND nmake -f win32/Makefile.msc zlib.lib
+    BUILD_COMMAND cp zlib.lib ${INSTALL_DIR}/lib/
+    INSTALL_COMMAND cp zlib.h zconf.h ${INSTALL_DIR}/include/
+    LOG_DOWNLOAD ON
+    LOG_UPDATE ON
+    LOG_CONFIGURE ON
+    LOG_BUILD ON
+    LOG_INSTALL ON
+    LOG_MERGED_STDOUTERR ON
+    LOG_OUTPUT_ON_FAILURE ON
+  )
+else()
+  # To pass custom environment variables to ExternalProject_Add command,
+  # we need to do `${CMAKE_COMMAND} -E env ${envs} <COMMANAD>`.
+  # https://stackoverflow.com/a/62437353
+  # We constrcut the custom environment variables here
+  set(envs
     "PKG_CONFIG_PATH=${INSTALL_DIR}/lib/pkgconfig"
     "LDFLAGS=-L${INSTALL_DIR}/lib $ENV{LDFLAGS}"
     "CFLAGS=-I${INSTALL_DIR}/include -fvisibility=hidden -fPIC $ENV{CFLAGS}"
     "prefix=${INSTALL_DIR}"
-)
+  )
 
-set(
+  set(
     ZLIB_LIBRARIES
     ${INSTALL_DIR}/lib/libz.a
   )
 
-ExternalProject_Add(zlib-
+  ExternalProject_Add(zlib-
     PREFIX ${CMAKE_CURRENT_BINARY_DIR}
     DOWNLOAD_DIR ${ARCHIVE_DIR}
     URL https://github.com/madler/zlib/archive/refs/tags/v1.2.12.tar.gz
@@ -34,7 +59,8 @@ ExternalProject_Add(zlib-
     LOG_INSTALL ON
     LOG_MERGED_STDOUTERR ON
     LOG_OUTPUT_ON_FAILURE ON
-)
+  )
+endif()
 
 add_library(zlib INTERFACE)
 add_dependencies(zlib zlib-)

tools/setup_helpers/extension.py

@@ -35,7 +35,7 @@ def _get_build(var, default=False):
 _BUILD_SOX = False if platform.system() == "Windows" else _get_build("BUILD_SOX", True)
 _BUILD_KALDI = False if platform.system() == "Windows" else _get_build("BUILD_KALDI", True)
 _BUILD_RNNT = _get_build("BUILD_RNNT", True)
-_BUILD_CTC_DECODER = False if platform.system() == "Windows" else _get_build("BUILD_CTC_DECODER", True)
+_BUILD_CTC_DECODER = _get_build("BUILD_CTC_DECODER", True)
 _USE_FFMPEG = _get_build("USE_FFMPEG", False)
 _USE_ROCM = _get_build("USE_ROCM", torch.cuda.is_available() and torch.version.hip is not None)
 _USE_CUDA = _get_build("USE_CUDA", torch.cuda.is_available() and torch.version.hip is None)