[ET-VK][build] Fix Windows Vulkan/Android build and llama ETDump profiling

Summary:

Three additive, backward-compatible fixes needed to build the ExecuTorch Vulkan
backend and profile the llama runner from a Windows host. Each is guarded so it
is a no-op on non-Windows / non-profiling builds.

- backends/vulkan/runtime/gen_vulkan_spv.py: normalize CRLF -> LF at the top of
  preprocess() before the trailing-backslash macro-continuation regex. GLSL
  templates checked out with CRLF otherwise leave a lone backslash that escape()
  turns into an unterminated Python string literal
  ("SyntaxError: unterminated string literal") during shader codegen.

- third-party/CMakeLists.txt: allow importing a prebuilt host flatc/flatcc via
  FLATC_EXECUTABLE / FLATCC_EXECUTABLE instead of building them from source. The
  existing WIN32 host-tool build only supports the Visual Studio generator
  (VS-only CMAKE_GENERATOR_PLATFORM, .exe-less BUILD_BYPRODUCTS), which fails
  under the Ninja generator required by the Android NDK cross-compile.

- examples/models/llama/CMakeLists.txt: link etdump and define
  ET_EVENT_TRACER_ENABLED when the etdump target exists (i.e. core libs built
  with EXECUTORCH_BUILD_DEVTOOLS=ON + EXECUTORCH_ENABLE_EVENT_TRACER=ON). This
  enables ETDump collection and makes the Vulkan delegate emit per-dispatch GPU
  timestamps for per-shader profiling.

Test Plan:

Cross-compiled ExecuTorch core libs + llama_main for Android arm64 with the
Vulkan backend enabled from a Windows host (NDK r29, Ninja generator, imported
host flatc/flatcc). Confirmed libvulkan_backend.a is linked and llama_main runs
on a Galaxy S24 (Adreno 750): Qwen3 0.6B ~80 tok/s decode, Llama 3.2 1B 8da4w
~60 tok/s decode. With DEVTOOLS/EVENT_TRACER on, collected an ETDump containing
per-shader Vulkan events and analyzed prefill/decode breakdown via
custom_inspector.py --mode llm.

Reviewers:

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: SS-JIA

backends/vulkan/runtime/gen_vulkan_spv.py

@@ -641,6 +641,12 @@ def escape(line: str) -> str:
 def preprocess(
     input_text: str, variables: Dict[str, Any], input_path: str = "codegen"
 ) -> str:
+    # Normalize line endings first. Templates checked out with CRLF (common on
+    # Windows) otherwise break the trailing-backslash handling below: in
+    # re.MULTILINE, $ matches immediately before \n, so a CR would sit between a
+    # trailing \ and the line end and defeat the r"\\$" match, leaving a lone
+    # backslash that escape() turns into an unterminated Python string literal.
+    input_text = input_text.replace("\r\n", "\n").replace("\r", "\n")
     # Workaround to handle source files using \ to extend mecros to a new line
     input_text = re.sub(r"\\$", r"\\\\", input_text, flags=re.MULTILINE)
 

examples/models/llama/CMakeLists.txt

@@ -241,8 +241,21 @@ if(ANDROID)
   list(APPEND link_libraries log)
 endif()
 
+# ETDump profiling support (OSS build). The llama runner does not link etdump or
+# define ET_EVENT_TRACER_ENABLED by default. When the core libs are built with
+# EXECUTORCH_BUILD_DEVTOOLS=ON + EXECUTORCH_ENABLE_EVENT_TRACER=ON, find_package
+# imports the etdump target; linking it and defining the macro enables ETDump
+# collection and makes the Vulkan delegate emit per-dispatch GPU timestamps.
+if(TARGET etdump)
+  list(APPEND link_libraries etdump)
+endif()
+
 add_executable(llama_main ${_srcs})
 
+if(TARGET etdump)
+  target_compile_definitions(llama_main PRIVATE ET_EVENT_TRACER_ENABLED)
+endif()
+
 # Copy MLX metallib for runtime if MLX delegate is enabled
 if(TARGET mlxdelegate)
   executorch_target_copy_mlx_metallib(llama_main)

third-party/CMakeLists.txt

@@ -46,6 +46,15 @@ else()
   set(_flatbuffers_ep_additional_args)
 endif()
 
+# Allow reusing a prebuilt host flatc instead of building it from source. This is
+# required when cross-compiling with the Ninja generator on Windows: the WIN32
+# branch above only supports the Visual Studio generator (it passes the VS-only
+# CMAKE_GENERATOR_PLATFORM), and the flatbuffers_ep host sub-build otherwise
+# fails under Ninja. Pass -DFLATC_EXECUTABLE=<path to a host flatc(.exe)>.
+if(FLATC_EXECUTABLE)
+  add_executable(flatc IMPORTED GLOBAL)
+  set_target_properties(flatc PROPERTIES IMPORTED_LOCATION "${FLATC_EXECUTABLE}")
+else()
 # We use ExternalProject to build flatc from source to force it target the host.
 # Otherwise, flatc will target the project's toolchain (i.e. iOS, or Android).
 ExternalProject_Add(
@@ -86,6 +95,7 @@ else()
     flatc PROPERTIES IMPORTED_LOCATION ${INSTALL_DIR}/bin/flatc
   )
 endif()
+endif()
 
 # TODO: re-enable once flatbuffers is added as a subdirectory.
 # set(FLATBUFFERS_BUILD_FLATC OFF) set(FLATBUFFERS_INSTALL OFF)
@@ -104,6 +114,14 @@ else()
   set(_flatcc_extra_cmake_args)
 endif()
 
+# Allow reusing a prebuilt host flatcc (see the flatc note above). Needed for the
+# Ninja-on-Windows cross-compile, where the flatcc_ep host sub-build fails. Pass
+# -DFLATCC_EXECUTABLE=<path to a host flatcc(.exe)>. flatcc_cli is only consumed
+# by devtools/etdump, so this imported target is unused when devtools is off.
+if(FLATCC_EXECUTABLE)
+  add_executable(flatcc_cli IMPORTED GLOBAL)
+  set_target_properties(flatcc_cli PROPERTIES IMPORTED_LOCATION "${FLATCC_EXECUTABLE}")
+else()
 # Similar to flatbuffers, we want to build flatcc for the host. See inline
 # comments in the flatbuffers ExternalProject_Add for more details.
 ExternalProject_Add(
@@ -142,6 +160,7 @@ else()
     flatcc_cli PROPERTIES IMPORTED_LOCATION ${INSTALL_DIR}/bin/flatcc
   )
 endif()
+endif()
 
 set(FLATCC_RTONLY
     ON