Arm backend: Add Arm Neural Statistics (pytorch#20662)

This PR adds opt-in VGF Arm Neural Accelerator Statistics collection for
ETDump.

When EXECUTORCH_VGF_ENABLE_NEURAL_STATISTICS=1 is set, the VGF backend
collects VK_ARM_data_graph neural statistics metadata, wraps the opaque
binary blobs in a versioned JSON payload, and attaches it to ETDump
delegate metadata under the VGF_NEURAL_STATISTICS event. The
implementation keeps normal timing-only profiling unchanged by default,
handles unsupported Vulkan/API/bind-point paths gracefully, and adds
Inspector parsing support to recover decoded debug_database,
statistics_info, and statistics_memory bytes.
This option EXECUTORCH_VGF_ENABLE_NEURAL_STATISTICS was added to attach
this payload only when it is requested specifically, otherwise it could
blow up ETDump in some runs.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Signed-off-by: Elena Zhelezina <elena.zhelezina@arm.com>

Author: wwwind

backends/arm/CMakeLists.txt

@@ -184,6 +184,36 @@ if(EXECUTORCH_BUILD_VGF)
       set(LIBVGF_PATH "${_vgf_site_pure}")
     endif()
 
+    if(EXECUTORCH_BUILD_TESTS)
+      add_executable(
+        vgf_neural_statistics_test
+        ${EXECUTORCH_ROOT}/backends/arm/test/vgf_neural_statistics_test.cpp
+        ${EXECUTORCH_ROOT}/backends/arm/runtime/VGFNeuralStatistics.cpp
+      )
+      target_include_directories(
+        vgf_neural_statistics_test
+        PRIVATE ${_common_include_directories} ${VULKAN_HEADERS_PATH}
+                ${VOLK_HEADERS_PATH}
+      )
+      target_compile_options(
+        vgf_neural_statistics_test PRIVATE -DUSE_VULKAN_WRAPPER
+                                           -DUSE_VULKAN_VOLK
+      )
+      target_link_libraries(vgf_neural_statistics_test PRIVATE executorch_core)
+      if(TARGET GTest::gtest_main)
+        target_link_libraries(
+          vgf_neural_statistics_test PRIVATE GTest::gtest_main
+        )
+      else()
+        target_link_libraries(
+          vgf_neural_statistics_test PRIVATE gtest gtest_main
+        )
+      endif()
+      add_test(NAME vgf_neural_statistics_test
+               COMMAND vgf_neural_statistics_test
+      )
+    endif()
+
     set(LIBVGF_STATIC "${LIBVGF_PATH}/lib/libvgf.a")
   endif()
 
@@ -194,8 +224,9 @@ if(EXECUTORCH_BUILD_VGF)
   target_include_directories(vgf INTERFACE "${LIBVGF_INCLUDE}")
 
   # Add backend delegate for VGF
-  set(_vgf_backend_sources backends/arm/runtime/VGFBackend.cpp
-                           backends/arm/runtime/VGFSetup.cpp
+  set(_vgf_backend_sources
+      backends/arm/runtime/VGFBackend.cpp backends/arm/runtime/VGFSetup.cpp
+      backends/arm/runtime/VGFNeuralStatistics.cpp
   )
   if(NOT EXECUTORCH_BUILD_VULKAN)
     list(APPEND _vgf_backend_sources backends/vulkan/third-party/volk/volk.c)

backends/arm/runtime/VGFBackend.cpp

@@ -5,6 +5,7 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#include <algorithm>
 #include <atomic>
 #include <cerrno>
 #include <chrono>
@@ -118,13 +119,21 @@ void vkml_free_basics(
 constexpr const char* kVgfDumpInputsDirEnv = "EXECUTORCH_VGF_DUMP_INPUTS_DIR";
 constexpr const char* kVgfDumpInputsAndExitEnv =
     "EXECUTORCH_VGF_DUMP_INPUTS_AND_EXIT";
+constexpr const char* kVgfNeuralStatisticsEnv =
+    "EXECUTORCH_VGF_ENABLE_NEURAL_STATISTICS";
 
 std::atomic<uint64_t> g_vgf_dump_invocation{0};
 
 bool env_flag_enabled(const char* name) {
   const char* value = std::getenv(name);
-  return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0 &&
-      std::strcmp(value, "false") != 0 && std::strcmp(value, "False") != 0;
+  if (value == nullptr || value[0] == '\0') {
+    return false;
+  }
+
+  return std::strcmp(value, "0") != 0 && std::strcmp(value, "false") != 0 &&
+      std::strcmp(value, "False") != 0 && std::strcmp(value, "FALSE") != 0 &&
+      std::strcmp(value, "off") != 0 && std::strcmp(value, "Off") != 0 &&
+      std::strcmp(value, "OFF") != 0;
 }
 
 const char* descriptor_type_to_string(VkDescriptorType type) {
@@ -692,6 +701,28 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
 #ifdef ET_EVENT_TRACER_ENABLED
     event_tracer_end_profiling_delegate(event_tracer, dispatch_event);
 
+    if (event_tracer != nullptr && env_flag_enabled(kVgfNeuralStatisticsEnv)) {
+      // We attach the neural statistics JSON blob to ETDump as delegate
+      // metadata, when event tracer is active.
+
+      // This is “synthetic” event, which we use as a carrier for metadata
+      EventTracerEntry neural_statistics_event =
+          event_tracer_start_profiling_delegate(
+              event_tracer,
+              kVgfNeuralStatisticsDelegateEventName,
+              /*delegate_debug_id=*/-1);
+
+      // Ask VGF representation for neural accelerator diagnostics
+      std::string neural_statistics_metadata =
+          repr->collect_neural_statistics_metadata();
+
+      event_tracer_end_profiling_delegate(
+          event_tracer,
+          neural_statistics_event,
+          neural_statistics_metadata.data(),
+          neural_statistics_metadata.size());
+    }
+
     EventTracerEntry copy_outputs_event = event_tracer_start_profiling_delegate(
         event_tracer,
         "VGF_COPY_OUTPUTS",