Windows x64 Build with support for xnnpack and llama example

backends/xnnpack/CMakeLists.txt

@@ -73,6 +73,20 @@ foreach(fbs_file ${_xnnpack_schema__srcs})
 endforeach()
 
 # Generate the headers from the .fbs files.
+
+if (WIN32)
+add_custom_command(
+  OUTPUT ${_xnnpack_schema__outputs}
+  COMMAND
+    ${FLATC_EXECUTABLE} --cpp --cpp-std c++11 --scoped-enums -o
+    "${_xnnpack_schema__include_dir}/executorch/backends/xnnpack/serialization"
+    ${_xnnpack_schema__srcs}
+  COMMAND powershell -Command "Move-Item -Path ${_xnnpack_flatbuffer__outputs} -Destination ${_xnnpack_schema__outputs}"
+  WORKING_DIRECTORY ${EXECUTORCH_ROOT}
+  COMMENT "Generating xnnpack_schema headers"
+  VERBATIM
+)
+else()
 add_custom_command(
   OUTPUT ${_xnnpack_schema__outputs}
   COMMAND
@@ -84,6 +98,7 @@ add_custom_command(
   COMMENT "Generating xnnpack_schema headers"
   VERBATIM
 )
+endif()
 
 add_library(xnnpack_schema INTERFACE ${_xnnpack_schema__outputs})
 set_target_properties(xnnpack_schema PROPERTIES LINKER_LANGUAGE CXX)

build/resolve_buck.py

@@ -11,7 +11,6 @@
 import platform
 import stat
 import sys
-import tempfile
 import urllib.request
 
 from dataclasses import dataclass
@@ -85,6 +84,12 @@ class BuckInfo:
         archive_name="buck2-x86_64-apple-darwin.zst",
         target_versions=["3eb1ae97ea963086866b4d2d9ffa966d"],
     ),
+    ("windows", "x86_64"): BuckInfo(
+        archive_name="buck2-x86_64-pc-windows-msvc.exe.zst",
+        target_versions=[
+            "bf1685c4c4ddd9de4592b5a955cb7326fd01e6c4d5f561643422bed961a17401"
+        ],
+    ),
 }
 
 
@@ -135,6 +140,8 @@ def resolve_buck2(args: argparse.Namespace) -> Union[str, int]:
         os_family = "linux"
     elif sys.platform.startswith("darwin"):
         os_family = "darwin"
+    elif sys.platform.startswith("win"):
+        os_family = "windows"
 
     platform_key = (os_family, arch)
     if platform_key not in BUCK_PLATFORM_MAP:
@@ -193,12 +200,12 @@ def resolve_buck2(args: argparse.Namespace) -> Union[str, int]:
 
             buck2_archive_url = f"https://github.com/facebook/buck2/releases/download/{target_buck_version}/{buck_info.archive_name}"
 
-            with tempfile.NamedTemporaryFile() as archive_file:
+            try:
                 print(f"Downloading buck2 from {buck2_archive_url}...", file=sys.stderr)
-                urllib.request.urlretrieve(buck2_archive_url, archive_file.name)
+                archive_file, _ = urllib.request.urlretrieve(buck2_archive_url)
 
                 # Extract and chmod.
-                with open(archive_file.name, "rb") as f:
+                with open(archive_file, "rb") as f:
                     data = f.read()
                     decompressed_bytes = zstd.decompress(data)
 
@@ -207,6 +214,8 @@ def resolve_buck2(args: argparse.Namespace) -> Union[str, int]:
 
                 file_stat = os.stat(buck2_local_path)
                 os.chmod(buck2_local_path, file_stat.st_mode | stat.S_IEXEC)
+            finally:
+                os.remove(archive_file)
 
             return buck2_local_path
 

devtools/CMakeLists.txt

@@ -157,19 +157,45 @@ file(MAKE_DIRECTORY
      ${_program_schema__include_dir}/executorch/devtools/bundled_program
 )
 
+# Note that the flatcc project actually writes its outputs into the source
+# tree instead of under the binary directory, and there's no way to change
+# that behavior.
+if(WIN32)
+  set(_flatcc_bin_path ${_flatcc_source_dir}/bin/${CMAKE_BUILD_TYPE}/flatcc)
+else()
+  set(_flatcc_bin_path ${_flatcc_source_dir}/bin/flatcc)
+endif()
+
+if(WIN32)
+  add_custom_command(
+    OUTPUT ${_etdump_schema__outputs}
+    COMMAND
+      # Note that the flatcc project actually writes its outputs into the source
+      # tree instead of under the binary directory, and there's no way to change
+      # that behavior.
+      ${_flatcc_bin_path} -cwr -o
+      ${_program_schema__include_dir}/executorch/devtools/etdump
+      ${_etdump_schema__srcs}
+    # COMMAND powershell -Command "Remove-Item -Path " ${_etdump_schema_cleanup_paths} " -Force -ErrorAction SilentlyContinue"
+    COMMAND powershell -Command "if (" "'${_etdump_schema_cleanup_paths}'" "-ne '') { Remove-Item -Path " "'${_etdump_schema_cleanup_paths}'" " -Force -ErrorAction SilentlyContinue }"
+    DEPENDS ${_etdump_schema_gen_dep}
+    COMMENT "Generating etdump headers"
+  )
+else()
 add_custom_command(
   OUTPUT ${_etdump_schema__outputs}
   COMMAND
     # Note that the flatcc project actually writes its outputs into the source
     # tree instead of under the binary directory, and there's no way to change
     # that behavior.
-    ${_flatcc_source_dir}/bin/flatcc -cwr -o
+    ${_flatcc_bin_path} -cwr -o
     ${_program_schema__include_dir}/executorch/devtools/etdump
     ${_etdump_schema__srcs}
   COMMAND rm -f ${_etdump_schema_cleanup_paths}
   DEPENDS ${_etdump_schema_gen_dep}
   COMMENT "Generating etdump headers"
 )
+endif()
 
 add_library(
   etdump ${CMAKE_CURRENT_SOURCE_DIR}/etdump/etdump_flatcc.cpp

examples/models/llama/CMakeLists.txt

@@ -80,7 +80,7 @@ find_package(gflags REQUIRED)
 # find `executorch` libraries Same as for gflags
 set(executorch_DIR ${CMAKE_CURRENT_BINARY_DIR}/../../../lib/cmake/ExecuTorch)
 find_package(executorch CONFIG REQUIRED)
-if(CMAKE_TOOLCHAIN_IOS OR ANDROID)
+if(CMAKE_TOOLCHAIN_IOS OR ANDROID OR WIN32 OR APPLE)
   target_link_options_shared_lib(executorch)
 endif()
 
@@ -98,14 +98,21 @@ add_subdirectory(runner)
 set(link_libraries gflags)
 set(_srcs main.cpp)
 
+# in Windows, cpublas and extension_threadpool have duplicated symbols
+if(WIN32 AND TARGET xnnpack_backend)
+  set(CPUBLAS "")
+else()
+  set(CPUBLAS "cpublas")
+endif()
+
 if(EXECUTORCH_BUILD_KERNELS_OPTIMIZED)
   list(
     APPEND
     link_libraries
     optimized_native_cpu_ops_lib
     optimized_kernels
     portable_kernels
-    cpublas
+    ${CPUBLAS}
     eigen_blas
   )
   target_link_options_shared_lib(optimized_native_cpu_ops_lib)
@@ -131,18 +138,26 @@ if(EXECUTORCH_BUILD_TORCHAO)
 endif()
 
 set(XNNPACK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../../backends/xnnpack)
+
+# in Windows, xnnpack_backend and extension_threadpool have duplicated symbols
+if(WIN32 AND TARGET xnnpack_backend)
+  set(EXTENSION_THREADPOOL "")
+else()
+  set(EXTENSION_THREADPOOL "extension_threadpool")
+endif()
+
 # Extra compile option and include dir for pthreadpool
 if(EXECUTORCH_BUILD_PTHREADPOOL)
   list(APPEND _common_compile_options -DET_USE_THREADPOOL)
-  list(APPEND link_libraries extension_threadpool pthreadpool)
+  list(APPEND link_libraries ${EXTENSION_THREADPOOL} pthreadpool)
   list(APPEND _common_include_directories
        ${XNNPACK_ROOT}/third-party/pthreadpool/include
   )
 endif()
 
 # Extra sources for cpuinfo
 if(EXECUTORCH_BUILD_CPUINFO)
-  list(APPEND link_libraries extension_threadpool cpuinfo)
+  list(APPEND link_libraries ${EXTENSION_THREADPOOL} cpuinfo)
   list(APPEND _common_include_directories
        ${XNNPACK_ROOT}/third-party/cpuinfo/include
   )
@@ -216,9 +231,6 @@ target_include_directories(llama_main PUBLIC ${_common_include_directories})
 target_link_libraries(llama_main PUBLIC llama_runner ${link_libraries})
 target_compile_options(llama_main PUBLIC ${_common_compile_options})
 
-if(APPLE)
-  target_link_options_shared_lib(executorch)
-endif()
 
 # Print all summary
 executorch_print_configuration_summary()

examples/models/llama/install_requirements.bat

@@ -0,0 +1,16 @@
+rem Install snakeviz for cProfile flamegraph
+rem Install sentencepiece for llama tokenizer
+pip install snakeviz sentencepiece
+
+rem Install torchao.
+pip install "%~dp0/../../../third-party/ao"
+
+rem Install lm-eval for Model Evaluation with lm-evalution-harness
+rem Install tiktoken for tokenizer
+pip install lm_eval==0.4.5
+pip install tiktoken blobfile
+rem Restore numpy if >= 2.0
+pip install "numpy<2.0"
+
+rem Call the install helper for further setup
+python examples/models/llama/install_requirement_helper.py

examples/models/llama/runner/CMakeLists.txt

@@ -36,6 +36,11 @@ include(${EXECUTORCH_SRCS_FILE})
 
 # build llama_runner library
 list(TRANSFORM _llama_runner__srcs PREPEND "${EXECUTORCH_ROOT}/")
+if (WIN32)
+  set_property(TARGET extension_module PROPERTY
+              IMPORTED_IMPLIB "${EXECUTORCH_ROOT}/cmake-out/lib/extension_module.lib")
+
+endif()
 
 target_include_directories(
   extension_module INTERFACE ${_common_include_directories}
@@ -52,6 +57,7 @@ list(APPEND _llama_runner__srcs
 if(CMAKE_TOOLCHAIN_IOS
    OR ANDROID
    OR APPLE
+   OR WIN32
 )
   # Building a share library on iOS requires code signing On Android we see
   # duplicated registration when using shared lib
@@ -84,4 +90,4 @@ target_link_libraries(llama_runner PUBLIC ${llama_runner_deps})
 target_include_directories(
   llama_runner INTERFACE ${_common_include_directories} ${EXECUTORCH_ROOT}
 )
-target_compile_options(llama_runner PUBLIC ${_preprocessor_flag})
+target_compile_options(llama_runner PUBLIC ${_preprocessor_flag})

extension/data_loader/file_data_loader.cpp

@@ -14,10 +14,10 @@
 #include <cstring>
 #include <limits>
 
+#include <executorch/runtime/platform/unistd.h>
 #include <fcntl.h>
 #include <sys/stat.h>
 #include <sys/types.h>
-#include <unistd.h>
 
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/result.h>
@@ -71,6 +71,9 @@ FileDataLoader::~FileDataLoader() {
   std::free(const_cast<char*>(file_name_));
   // fd_ can be -1 if this instance was moved from, but closing a negative fd is
   // safe (though it will return an error).
+  if (fd_ == -1) {
+    return;
+  }
   ::close(fd_);
 }
 

extension/data_loader/mman.h

@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+// This file ensures that mman.h compatible functions are defined in the global
+// namespace for windows and posix environments.
+
+#pragma once
+
+#include <executorch/runtime/platform/compiler.h>
+
+#ifndef _WIN32
+
+#include <sys/mman.h>
+#include <unistd.h>
+
+ET_INLINE long get_os_page_size(){return sysconf(_SC_PAGESIZE);}
+
+#else
+
+#define NOMINMAX
+#include <windows.h>
+#undef NOMINMAX
+#include <io.h>
+
+#include <executorch/extension/data_loader/mman_windows.h>
+
+ET_INLINE long get_os_page_size() {
+  SYSTEM_INFO si;
+  GetSystemInfo(&si);
+  long pagesize = si.dwAllocationGranularity > si.dwPageSize
+      ? si.dwAllocationGranularity
+      : si.dwPageSize;
+  return pagesize;
+}
+
+#endif