Qualcomm AI Engine Direct - Support Windows native build (pytorch#20052)

## Summary

This PR updates the CMake files to enable native Windows builds for the
QNN ExecuTorch components.
Specifically, it adds support for building both the QNN backend as a
Windows dynamic library (.dll) and the corresponding runner executable
(.exe).

## Core Changes

### 1. CMake Enhancements

- Extend existing CMake to support native Windows builds.
- Enable generation of `qnn_executorch_backend.dll` and
`qnn_executor_runner.exe`.

### 2. Build Script Support

- Provide a dedicated build script to simplify and standardize the
Windows build process.
- The script configures the appropriate CMake options and toolchain
settings required for Windows.

## Build instructions

1. `.\install_executorch.bat` to build `PyQnnManagerAdaptor*.pyd` and
its static library as dependencies.
2. `.\backends\qualcomm\scripts\build.ps1 -SkipArm64Windows -Release` to
build Windows x86_64 AOT on Windows x86_64.
3. `.\backends\qualcomm\scripts\build.ps1 -SkipX86Windows -Release` to
cross compile and build WoA Runtime on Windows x86_64.

## Follow-up
- Add PyTorch documentation change.

---------

Co-authored-by: Cheng-Hsin Weng <chenweng@qti.qualcomm.com>

Author: zhaoxul-qti

backends/qualcomm/CMakeLists.txt

@@ -77,7 +77,11 @@ if(${ANDROID})
   find_library(android_log log)
 endif()
 
-add_compile_options("-Wall" "-Werror" "-fvisibility=hidden")
+if(MSVC)
+  add_compile_options("/wd4996")
+else()
+  add_compile_options("-Wall" "-Werror" "-fvisibility=hidden")
+endif()
 add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 
 # GNU emits warning for ignored attributes Unfortunately, we use
@@ -89,21 +93,26 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
 endif()
 
 if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-  # strip symbols
-  add_link_options(LINKER:-s,--gc-sections)
-  if(${CMAKE_SYSTEM_PROCESSOR} MATCHES Hexagon)
-    add_compile_options(
-      "-Os"
-      "-ffunction-sections"
-      "-fdata-sections"
-      "-frtti"
-      "-fno-exceptions"
-      "-fomit-frame-pointer"
-      "-fno-asynchronous-unwind-tables"
-    )
+  if(MSVC)
+    add_compile_options("/O2")
   else()
-
-    add_compile_options("-O3" "-ffunction-sections" "-fdata-sections" "-frtti")
+    # strip symbols
+    add_link_options(LINKER:-s,--gc-sections)
+    if(${CMAKE_SYSTEM_PROCESSOR} MATCHES Hexagon)
+      add_compile_options(
+        "-Os"
+        "-ffunction-sections"
+        "-fdata-sections"
+        "-frtti"
+        "-fno-exceptions"
+        "-fomit-frame-pointer"
+        "-fno-asynchronous-unwind-tables"
+      )
+    else()
+      add_compile_options(
+        "-O3" "-ffunction-sections" "-fdata-sections" "-frtti"
+      )
+    endif()
   endif()
 endif()
 
@@ -275,16 +284,27 @@ if(${CMAKE_SYSTEM_PROCESSOR} MATCHES Hexagon)
   )
 endif()
 
-set_target_properties(
-  qnn_executorch_backend PROPERTIES LINK_FLAGS "-Wl,-rpath='$ORIGIN'"
-)
+if(MSVC)
+  target_compile_definitions(
+    qnn_executorch_backend PRIVATE QNN_EXECUTORCH_BUILDING_DLL
+  )
+  set_target_properties(
+    qnn_executorch_backend PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON
+  )
+else()
+  set_target_properties(
+    qnn_executorch_backend PROPERTIES LINK_FLAGS "-Wl,-rpath='$ORIGIN'"
+  )
+endif()
 target_link_libraries(
   shared_buffer PRIVATE qnn_executorch_logging ${CMAKE_DL_LIBS}
 )
 #
 # add linker option
 #
-executorch_target_link_options_shared_lib(qnn_executorch_backend)
+if(NOT MSVC)
+  executorch_target_link_options_shared_lib(qnn_executorch_backend)
+endif()
 
 #
 # add sources
@@ -316,7 +336,7 @@ if(${CMAKE_SYSTEM_PROCESSOR} MATCHES Hexagon)
 endif()
 
 # QNN pybind
-if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
+if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64|AMD64")
   add_subdirectory(
     ${EXECUTORCH_SOURCE_DIR}/third-party/pybind11
     ${CMAKE_CURRENT_BINARY_DIR}/pybind11
@@ -350,9 +370,13 @@ if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
 
   if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
     # need to allow exceptions in pybind
-    set(_pybind_compile_options -Wno-deprecated-declarations -fPIC -frtti
-                                -fexceptions
-    )
+    if(MSVC)
+      set(_pybind_compile_options /wd4996 /EHsc)
+    else()
+      set(_pybind_compile_options -Wno-deprecated-declarations -fPIC -frtti
+                                  -fexceptions
+      )
+    endif()
     target_compile_options(
       PyQnnManagerAdaptor PUBLIC ${_pybind_compile_options}
     )

backends/qualcomm/aot/wrappers/CMakeLists.txt

@@ -7,9 +7,9 @@
 # wrappers
 target_sources(
   wrappers
-  PUBLIC ${CMAKE_CURRENT_LIST_DIR}/TensorWrapper.cpp
-         ${CMAKE_CURRENT_LIST_DIR}/TensorWrapper.h
-  PRIVATE ${CMAKE_CURRENT_LIST_DIR}/QuantizeParamsWrapper.cpp
+  PUBLIC ${CMAKE_CURRENT_LIST_DIR}/TensorWrapper.h
+  PRIVATE ${CMAKE_CURRENT_LIST_DIR}/TensorWrapper.cpp
+          ${CMAKE_CURRENT_LIST_DIR}/QuantizeParamsWrapper.cpp
           ${CMAKE_CURRENT_LIST_DIR}/QuantizeParamsWrapper.h
           ${CMAKE_CURRENT_LIST_DIR}/OpWrapper.cpp
           ${CMAKE_CURRENT_LIST_DIR}/OpWrapper.h

backends/qualcomm/runtime/backends/CMakeLists.txt

@@ -44,7 +44,7 @@ target_sources(
 )
 
 set(platform target)
-if(${CMAKE_SYSTEM_PROCESSOR} MATCHES x86_64)
+if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64|AMD64")
   set(platform host)
 endif()