Wire Arm VGF backend build and test infrastructure (#18556)

backends/arm/runtime/VGFBackend.cpp

@@ -88,7 +88,14 @@ void vkml_free_basics(
 
 class VGFBackend final : public ::executorch::runtime::BackendInterface {
  public:
-  VGFBackend() {
+  VGFBackend() = default;
+
+  // Lazy Vulkan init — runs on first use, not in the constructor.
+  void ensure_initialized() {
+    if (is_initialized_) {
+      return;
+    }
+
     VkResult result;
 
     // Fetch basic vulkan objects once
@@ -122,6 +129,7 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
 
   bool is_available() const override {
     ET_LOG(Info, "Checking VGFBackend is available");
+    const_cast<VGFBackend*>(this)->ensure_initialized();
     if (!is_initialized_) {
       return false;
     }
@@ -134,6 +142,7 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
       ArrayRef<CompileSpec> compile_specs) const override {
     ET_LOG(Info, "Entered VGF init");
 
+    const_cast<VGFBackend*>(this)->ensure_initialized();
     if (!is_initialized_) {
       ET_LOG(
           Error,
@@ -334,6 +343,16 @@ VkResult vkml_allocate_basics(
   }
   volkLoadInstance(*instance);
 
+  // Bail out if the driver lacks ARM tensor/datagraph extensions.
+  if (!vkCreateTensorARM) {
+    ET_LOG(
+        Error,
+        "Vulkan driver does not support ARM tensor extensions (VK_ARM_tensors)");
+    vkDestroyInstance(*instance, nullptr);
+    *instance = VK_NULL_HANDLE;
+    return VK_ERROR_FEATURE_NOT_PRESENT;
+  }
+
   // Pick first GPU
   uint32_t gpu_count = 0;
   vkEnumeratePhysicalDevices(*instance, &gpu_count, nullptr);

backends/arm/runtime/VGFSetup.cpp

@@ -538,7 +538,7 @@ bool VgfRepr::process_vgf(const char* vgf_data, ArrayRef<CompileSpec> specs) {
       .pNext = nullptr,
       .flags = 0,
       .bindingCount = static_cast<uint32_t>(layout_bindings.size()),
-      layout_bindings.data(),
+      .pBindings = layout_bindings.data(),
   };
   result =
       vkCreateDescriptorSetLayout(vk_device, &layout_info, nullptr, &vk_layout);

backends/arm/runtime/targets.bzl

@@ -32,3 +32,35 @@ def define_common_targets():
             "fbsource//third-party/ethos-u-core-driver:core_driver",
         ],
     )
+    runtime.cxx_library(
+        name = "vgf_backend",
+        srcs = [
+            "VGFBackend.cpp",
+            "VGFSetup.cpp",
+            # Volk must be compiled directly into this target so its global
+            # function-pointer variables live in the same linkage unit.
+            # Linking from a separate static library causes the linker to
+            # drop the symbols when building a shared library.
+            "fbsource//third-party/vulkan-headers-1.4.343/v1.4.343/src:volk_arm_src",
+        ],
+        exported_headers = ["VGFSetup.h"],
+        # @lint-ignore BUCKLINT: Avoid `link_whole=True` (https://fburl.com/avoid-link-whole)
+        link_whole = True,
+        supports_python_dlopen = True,
+        compiler_flags = [
+            "-Wno-global-constructors",
+            "-Wno-header-hygiene",
+            "-Wno-unused-variable",
+            "-Wno-missing-field-initializers",
+            "-DUSE_VULKAN_WRAPPER",
+            "-DUSE_VULKAN_VOLK",
+        ],
+        visibility = ["PUBLIC"],
+        deps = [
+            "//executorch/runtime/backend:interface",
+            "//executorch/runtime/core:core",
+            "fbsource//third-party/arm-vgf-library/v0.8.0/src:vgf",
+            "fbsource//third-party/vulkan-headers-1.4.343/v1.4.343/src:volk_arm",
+            "fbsource//third-party/vulkan-headers-1.4.343/v1.4.343/src:vulkan-headers",
+        ],
+    )

backends/arm/test/runner_utils.py

@@ -35,7 +35,10 @@
 from executorch.backends.arm.tosa.compile_spec import TosaCompileSpec
 from executorch.backends.arm.tosa.specification import Tosa_1_00, TosaSpecification
 from executorch.backends.arm.vgf import VgfCompileSpec
-from executorch.backends.arm.vgf.model_converter import find_model_converter_binary
+from executorch.backends.arm.vgf.model_converter import (
+    find_model_converter_binary,
+    model_converter_env,
+)
 from executorch.exir import ExecutorchProgramManager, ExportedProgram
 from executorch.exir.lowered_backend_module import LoweredBackendModule
 from torch.fx.node import Node
@@ -818,7 +821,7 @@ def model_converter_installed() -> bool:
         return False
 
     try:
-        _run_cmd([model_converter, "--version"], check=True)
+        _run_cmd([model_converter, "--version"], check=True, env=model_converter_env())
     except Exception:
         return False
 

backends/arm/test/targets.bzl

@@ -67,9 +67,21 @@ def define_arm_tests():
             resources = ["conftest.py"],
             compile = "with-source",
             typing = False,
+            env = {} if runtime.is_oss else {
+                "MODEL_CONVERTER_PATH": "$(location fbsource//third-party/pypi/ai-ml-sdk-model-converter/0.8.0:model-converter-bin)",
+                "MODEL_CONVERTER_LIB_DIR": "$(location fbsource//third-party/nvidia-nsight-systems:linux-x86_64)/host-linux-x64",
+                "LAVAPIPE_LIB_PATH": "$(location fbsource//third-party/mesa/src/gallium/frontends/lavapipe:vulkan_lvp)",
+                "EMULATION_LAYER_TENSOR_SO": "$(location fbsource//third-party/arm-ml-emulation-layer/v0.9.0/src:libVkLayer_Tensor)",
+                "EMULATION_LAYER_GRAPH_SO": "$(location fbsource//third-party/arm-ml-emulation-layer/v0.9.0/src:libVkLayer_Graph)",
+                "EMULATION_LAYER_TENSOR_JSON": "$(location fbsource//third-party/arm-ml-emulation-layer/v0.9.0/src:VkLayer_Tensor_json)",
+                "EMULATION_LAYER_GRAPH_JSON": "$(location fbsource//third-party/arm-ml-emulation-layer/v0.9.0/src:VkLayer_Graph_json)",
+            },
             preload_deps = [
                 "//executorch/kernels/quantized:custom_ops_generated_lib",
-            ],
+            ] + ([] if runtime.is_oss else [
+                "fbsource//third-party/khronos:vulkan",
+                "//executorch/backends/arm/runtime:vgf_backend",
+            ]),
             deps = [
                 "//executorch/backends/arm/test:arm_tester" if runtime.is_oss else "//executorch/backends/arm/test/tester/fb:arm_tester_fb",
                 "//executorch/backends/arm/test:conftest",

backends/arm/vgf/backend.py

@@ -28,6 +28,7 @@
     VgfCompileSpec,
 )
 from executorch.backends.arm.vgf.model_converter import (  # type: ignore[import-not-found]
+    model_converter_env,
     require_model_converter_binary,
 )
 from executorch.exir.backend.backend_details import (  # type: ignore[import-not-found]
@@ -157,7 +158,11 @@ def vgf_compile(
         ]
         try:
             subprocess.run(  # nosec B602, B603 - shell invocation constrained to trusted converter binary with trusted inputs
-                conversion_command, shell=False, check=True, capture_output=True
+                conversion_command,
+                shell=False,
+                check=True,
+                capture_output=True,
+                env=model_converter_env(),
             )
         except subprocess.CalledProcessError as process_error:
             conversion_command_str = " ".join(conversion_command)

backends/arm/vgf/model_converter.py

@@ -5,6 +5,7 @@
 
 from __future__ import annotations
 
+import os
 from shutil import which
 from typing import Optional
 
@@ -13,14 +14,33 @@
 
 
 def find_model_converter_binary() -> Optional[str]:
-    """Return the name of the first model converter executable on PATH."""
+    """Return the path/name of the first model converter executable found."""
+    env_path = os.environ.get("MODEL_CONVERTER_PATH")
+    if env_path and os.path.isfile(env_path) and os.access(env_path, os.X_OK):
+        return env_path
 
     for candidate in (MODEL_CONVERTER_BINARY, _MODEL_CONVERTER_FALLBACK_BINARY):
         if which(candidate):
             return candidate
     return None
 
 
+def model_converter_env() -> dict[str, str]:
+    """Return an env dict suitable for running model-converter as a subprocess.
+
+    If MODEL_CONVERTER_LIB_DIR is set, it is prepended to LD_LIBRARY_PATH so the
+    binary can find a compatible libstdc++ (or other shared libs) without
+    polluting the parent process environment.
+
+    """
+    env = dict(os.environ)
+    lib_dir = env.pop("MODEL_CONVERTER_LIB_DIR", None)
+    if lib_dir:
+        existing = env.get("LD_LIBRARY_PATH", "")
+        env["LD_LIBRARY_PATH"] = f"{lib_dir}:{existing}" if existing else lib_dir
+    return env
+
+
 def require_model_converter_binary() -> str:
     """Return a usable model converter executable or raise a helpful error."""