Arm backend: Enable host memory capabilities with fixed buck deps.

backends/arm/runtime/VGFBackend.cpp

@@ -103,9 +103,13 @@ VkResult vkml_allocate_basics(
     VkCommandPool* command_pool,
     uint32_t* queue_family_index,
     bool request_neural_statistics,
-    bool* neural_statistics_device_enabled);
+    bool* neural_statistics_device_enabled,
+    bool request_host_memory_import,
+    VgfHostMemoryImportCapabilities* host_memory_import_capabilities);
 
 // Helper functions to dump VGF Delegate Boundary Inputs
+constexpr const char* kVgfHostMemoryImportEnableEnv =
+    "EXECUTORCH_VGF_ENABLE_HOST_MEMORY_IMPORT";
 constexpr const char* kVgfDumpInputsDirEnv = "EXECUTORCH_VGF_DUMP_INPUTS_DIR";
 constexpr const char* kVgfDumpInputsAndExitEnv =
     "EXECUTORCH_VGF_DUMP_INPUTS_AND_EXIT";
@@ -418,8 +422,11 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
 
     VkResult result;
     neural_statistics_config_ = get_vgf_neural_statistics_runtime_config();
+    const bool request_host_memory_import =
+        env_flag_enabled(kVgfHostMemoryImportEnableEnv);
 
-    // Fetch basic vulkan objects once
+    // Fetch basic Vulkan objects once. Device extensions must be selected here,
+    // before vkCreateDevice; VgfRepr only receives already-created handles.
     result = vkml_allocate_basics(
         &vk_instance,
         &vk_physical_device,
@@ -428,7 +435,9 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
         &vk_command_pool,
         &vk_queue_family_index,
         neural_statistics_config_.requested,
-        &neural_statistics_device_enabled_);
+        &neural_statistics_device_enabled_,
+        request_host_memory_import,
+        &host_memory_import_capabilities_);
     if (result != VK_SUCCESS) {
       ET_LOG(
           Error, "Failed to initialize the Vulkan device error 0x%08X", result);
@@ -535,7 +544,8 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
         vk_queue_family_index,
         neural_statistics_config_.requested,
         neural_statistics_device_enabled_,
-        neural_statistics_config_.mode_index);
+        neural_statistics_config_.mode_index,
+        host_memory_import_capabilities_);
 
 #ifdef ET_EVENT_TRACER_ENABLED
     event_tracer_end_profiling_delegate(event_tracer, allocate_repr_event);
@@ -865,6 +875,7 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
     vk_queue_family_index = UINT32_MAX;
     neural_statistics_config_ = {};
     neural_statistics_device_enabled_ = false;
+    host_memory_import_capabilities_ = {};
     is_initialized_ = false;
     // Do not call volkFinalize(): the Vulkan backend shares the loader.
   }
@@ -879,6 +890,7 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
   uint32_t vk_queue_family_index = UINT32_MAX;
   VgfNeuralStatisticsRuntimeConfig neural_statistics_config_{};
   bool neural_statistics_device_enabled_ = false;
+  VgfHostMemoryImportCapabilities host_memory_import_capabilities_{};
   bool is_initialized_ = false;
 };
 
@@ -896,12 +908,17 @@ VkResult vkml_allocate_basics(
     VkCommandPool* command_pool,
     uint32_t* queue_family_index,
     bool request_neural_statistics,
-    bool* neural_statistics_device_enabled) {
+    bool* neural_statistics_device_enabled,
+    bool request_host_memory_import,
+    VgfHostMemoryImportCapabilities* host_memory_import_capabilities) {
   VkResult result;
 
   if (neural_statistics_device_enabled != nullptr) {
     *neural_statistics_device_enabled = false;
   }
+  if (host_memory_import_capabilities != nullptr) {
+    *host_memory_import_capabilities = {};
+  }
 
   if (VK_SUCCESS != volkInitialize()) {
     ET_LOG(Error, "Volk failed to initialize");
@@ -1156,6 +1173,59 @@ VkResult vkml_allocate_basics(
 
   vector<const char*> requested_exts;
 
+  bool enable_host_memory_import_device = false;
+#if defined(VK_EXT_external_memory_host)
+  const bool host_memory_import_advertised = std::any_of(
+      available.begin(), available.end(), [](const auto& ext_avail) {
+        return std::strcmp(
+                   VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME,
+                   ext_avail.extensionName) == 0;
+      });
+
+  VkDeviceSize min_imported_host_pointer_alignment = 0;
+  if (host_memory_import_advertised) {
+    VkPhysicalDeviceExternalMemoryHostPropertiesEXT host_memory_properties{
+        .sType =
+            VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT,
+        .pNext = nullptr,
+    };
+    VkPhysicalDeviceProperties2 properties_2{
+        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2,
+        .pNext = &host_memory_properties,
+    };
+    vkGetPhysicalDeviceProperties2(*physical_device, &properties_2);
+    min_imported_host_pointer_alignment =
+        host_memory_properties.minImportedHostPointerAlignment;
+  }
+
+  if (host_memory_import_capabilities != nullptr) {
+    host_memory_import_capabilities->physical_device_advertised =
+        host_memory_import_advertised;
+    host_memory_import_capabilities->min_imported_host_pointer_alignment =
+        min_imported_host_pointer_alignment;
+  }
+
+  enable_host_memory_import_device = vgf_host_memory_import_should_be_enabled(
+      request_host_memory_import, host_memory_import_advertised);
+  if (enable_host_memory_import_device) {
+    requested_exts.push_back(VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME);
+  } else if (request_host_memory_import) {
+    ET_LOG(
+        Info,
+        "%s was requested but the Vulkan physical device does not expose %s",
+        kVgfHostMemoryImportEnableEnv,
+        VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME);
+  }
+#else
+  if (request_host_memory_import) {
+    ET_LOG(
+        Info,
+        "%s was requested but Vulkan headers do not expose "
+        "VK_EXT_external_memory_host",
+        kVgfHostMemoryImportEnableEnv);
+  }
+#endif
+
   const bool bfloat16_extension_available = std::any_of(
       available.begin(), available.end(), [](const auto& ext_avail) {
         return std::strcmp(
@@ -1263,12 +1333,36 @@ VkResult vkml_allocate_basics(
   // Load the device with volk and populate function pointers
   volkLoadDevice(*device);
 
+  if (host_memory_import_capabilities != nullptr) {
+    // Only a successful vkCreateDevice makes extension enablement
+    // authoritative. Physical-device advertisement alone is insufficient for
+    // later zero-copy code to use VK_EXT_external_memory_host device
+    // functionality.
+    host_memory_import_capabilities->logical_device_enabled =
+        enable_host_memory_import_device;
+    ET_LOG(
+        Info,
+        "VGF host memory import: requested=%d advertised=%d enabled=%d "
+        "minImportedHostPointerAlignment=%llu",
+        static_cast<int>(request_host_memory_import),
+        static_cast<int>(
+            host_memory_import_capabilities->physical_device_advertised),
+        static_cast<int>(
+            host_memory_import_capabilities->logical_device_enabled),
+        static_cast<unsigned long long>(
+            host_memory_import_capabilities
+                ->min_imported_host_pointer_alignment));
+  }
+
   vkGetDeviceQueue(*device, qf, 0, queue);
 
   VkCommandPoolCreateInfo poolInfo{
       .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
       .pNext = nullptr,
-      .flags = 0,
+      // VGF records a persistent per-repr command buffer. Host-memory import
+      // may later rebind descriptors and re-record only that command buffer.
+      // Keep the legacy flags=0 behavior unless the extension is truly enabled.
+      .flags = vgf_command_pool_flags(enable_host_memory_import_device),
       .queueFamilyIndex = qf,
   };
   result = vkCreateCommandPool(*device, &poolInfo, nullptr, command_pool);

backends/arm/runtime/VGFSetup.h

@@ -22,6 +22,7 @@ using executorch::runtime::CompileSpec;
 #include <executorch/backends/vulkan/runtime/vk_api/vk_api.h>
 
 #include <executorch/backends/arm/runtime/VGFNeuralStatistics.h>
+#include <executorch/backends/arm/runtime/VGFVulkanFeatures.h>
 
 namespace executorch {
 namespace backends {
@@ -108,7 +109,8 @@ class VgfRepr {
       uint32_t queue_family_index = UINT32_MAX,
       bool neural_statistics_requested = false,
       bool neural_statistics_device_enabled = false,
-      int neural_statistics_mode_index = 1)
+      int neural_statistics_mode_index = 1,
+      VgfHostMemoryImportCapabilities host_memory_import_capabilities = {})
       : vk_instance(inst),
         vk_physical(phys),
         vk_device(dev),
@@ -117,7 +119,8 @@ class VgfRepr {
         vk_queue_family_index(queue_family_index),
         neural_statistics_requested_(neural_statistics_requested),
         neural_statistics_device_enabled_(neural_statistics_device_enabled),
-        neural_statistics_mode_index_(neural_statistics_mode_index) {}
+        neural_statistics_mode_index_(neural_statistics_mode_index),
+        host_memory_import_capabilities_(host_memory_import_capabilities) {}
 
   /*
    * Process a VGF ready for execution, allocate necessary Vulkan objects.
@@ -175,6 +178,23 @@ class VgfRepr {
     return neural_statistics_requested_;
   }
 
+  bool host_memory_import_advertised() const {
+    return host_memory_import_capabilities_.physical_device_advertised;
+  }
+
+  bool host_memory_import_enabled() const {
+    return host_memory_import_capabilities_.logical_device_enabled;
+  }
+
+  VkDeviceSize min_imported_host_pointer_alignment() const {
+    return host_memory_import_capabilities_.min_imported_host_pointer_alignment;
+  }
+
+  const VgfHostMemoryImportCapabilities& host_memory_import_capabilities()
+      const {
+    return host_memory_import_capabilities_;
+  }
+
   ~VgfRepr() {
     free_vgf();
     if (vk_pipeline_cache != VK_NULL_HANDLE) {
@@ -203,6 +223,7 @@ class VgfRepr {
   bool neural_statistics_requested_ = false;
   bool neural_statistics_device_enabled_ = false;
   int neural_statistics_mode_index_ = 1;
+  VgfHostMemoryImportCapabilities host_memory_import_capabilities_{};
 
   bool timestamp_queries_enabled = false;
   uint32_t timestamp_valid_bits = 0;

backends/arm/runtime/VGFVulkanFeatures.h

@@ -13,6 +13,29 @@ namespace executorch {
 namespace backends {
 namespace vgf {
 
+// Snapshot of VK_EXT_external_memory_host state for the exact Vulkan device
+// used by VGF. Physical-device advertisement and logical-device enablement are
+// deliberately separate because Vulkan does not provide a post-creation query
+// for the list of extensions enabled at vkCreateDevice().
+struct VgfHostMemoryImportCapabilities {
+  bool physical_device_advertised = false;
+  bool logical_device_enabled = false;
+  VkDeviceSize min_imported_host_pointer_alignment = 0;
+};
+
+inline bool vgf_host_memory_import_should_be_enabled(
+    bool requested,
+    bool physical_device_advertised) {
+  return requested && physical_device_advertised;
+}
+
+inline VkCommandPoolCreateFlags vgf_command_pool_flags(
+    bool host_memory_import_enabled) {
+  return host_memory_import_enabled
+      ? VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT
+      : 0;
+}
+
 inline VkPhysicalDeviceDataGraphFeaturesARM make_vgf_data_graph_features(
     void* p_next) {
   VkPhysicalDeviceDataGraphFeaturesARM features{};

backends/arm/test/vgf_vulkan_features_test.cpp

@@ -42,6 +42,60 @@ TEST(VgfVulkanFeaturesTest, RequiresDataGraphShaderModuleSupport) {
   EXPECT_TRUE(vgf_data_graph_features_supported(available));
 }
 
+TEST(VgfVulkanFeaturesTest, EnablesHostMemoryImportWhenRequestedAndAdvertised) {
+  EXPECT_TRUE(vgf_host_memory_import_should_be_enabled(
+      /*requested=*/true, /*physical_device_advertised=*/true));
+}
+
+TEST(
+    VgfVulkanFeaturesTest,
+    DoesNotEnableHostMemoryImportWhenAdvertisedButNotRequested) {
+  EXPECT_FALSE(vgf_host_memory_import_should_be_enabled(
+      /*requested=*/false, /*physical_device_advertised=*/true));
+}
+
+TEST(
+    VgfVulkanFeaturesTest,
+    DoesNotEnableHostMemoryImportWhenRequestedButNotAdvertised) {
+  EXPECT_FALSE(vgf_host_memory_import_should_be_enabled(
+      /*requested=*/true, /*physical_device_advertised=*/false));
+}
+
+TEST(VgfVulkanFeaturesTest, LegacyHostMemoryImportPathRemainsDisabled) {
+  EXPECT_FALSE(vgf_host_memory_import_should_be_enabled(
+      /*requested=*/false, /*physical_device_advertised=*/false));
+}
+
+TEST(VgfVulkanFeaturesTest, HostImportCommandPoolSupportsIndividualReset) {
+  EXPECT_EQ(
+      vgf_command_pool_flags(/*host_memory_import_enabled=*/true),
+      VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT);
+}
+
+TEST(VgfVulkanFeaturesTest, LegacyCommandPoolFlagsRemainUnchanged) {
+  EXPECT_EQ(vgf_command_pool_flags(/*host_memory_import_enabled=*/false), 0u);
+}
+
+TEST(VgfVulkanFeaturesTest, HostMemoryImportCapabilitiesDefaultToUnavailable) {
+  const VgfHostMemoryImportCapabilities capabilities{};
+  EXPECT_FALSE(capabilities.physical_device_advertised);
+  EXPECT_FALSE(capabilities.logical_device_enabled);
+  EXPECT_EQ(capabilities.min_imported_host_pointer_alignment, 0u);
+}
+
+TEST(
+    VgfVulkanFeaturesTest,
+    HostMemoryImportCapabilitiesDistinguishAdvertisedFromEnabled) {
+  VgfHostMemoryImportCapabilities capabilities{};
+  capabilities.physical_device_advertised = true;
+  capabilities.logical_device_enabled = false;
+  capabilities.min_imported_host_pointer_alignment = 4096;
+
+  EXPECT_TRUE(capabilities.physical_device_advertised);
+  EXPECT_FALSE(capabilities.logical_device_enabled);
+  EXPECT_EQ(capabilities.min_imported_host_pointer_alignment, 4096u);
+}
+
 } // namespace
 } // namespace vgf
 } // namespace backends