Arm backend: Fixes for VGF models (#13681)

* Models with constants/weight data no longer crash
* Missing intermediate allocation prevented correct output
* The emulation layer required a number of extensions
* A latent bug on indexing for VGF with >1 segment
* Memory barriers for non-unified memory Vulkan targets

Change-Id: Ide2aea329981a37b7787c9c676d6a2d1fd4ff3c8

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218

Signed-off-by: Rob Elliott <robert.elliott@arm.com>

Author: robell

backends/arm/runtime/VGFBackend.cpp

@@ -249,7 +249,6 @@ VkResult vkml_allocate_basics(
     VkDevice* device,
     VkQueue* queue,
     VkCommandPool* command_pool) {
-  const char* dev_exts[] = {"VK_ARM_tensors", "VK_ARM_data_graph"};
   VkResult result;
 
   if (VK_SUCCESS != volkInitialize()) {
@@ -372,11 +371,103 @@ VkResult vkml_allocate_basics(
       .pQueuePriorities = &qp,
   };
 
+  // Query features
+  VkPhysicalDeviceVulkan12Features available_12 = {
+      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
+      .pNext = NULL,
+  };
+  VkPhysicalDeviceVulkan11Features available_11 = {
+      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
+      .pNext = &available_12,
+  };
+  VkPhysicalDeviceFeatures2 available_2 = {
+      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
+      .pNext = &available_11,
+  };
+  vkGetPhysicalDeviceFeatures2(*physical_device, &available_2);
+
+  // Select features
+  VkPhysicalDeviceShaderReplicatedCompositesFeaturesEXT features_c{
+      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_REPLICATED_COMPOSITES_FEATURES_EXT,
+      nullptr};
+  features_c.shaderReplicatedComposites = true;
+  VkPhysicalDeviceVulkan13Features features_13{
+      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES, nullptr};
+  features_13.synchronization2 = true;
+  features_13.maintenance4 = true;
+  features_13.pipelineCreationCacheControl = true;
+  features_13.pNext = &features_c;
+  VkPhysicalDeviceVulkan12Features features_12{
+      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES, nullptr};
+  features_12.hostQueryReset = true;
+  features_12.storageBuffer8BitAccess = true;
+  features_12.uniformAndStorageBuffer8BitAccess =
+      available_12.uniformAndStorageBuffer8BitAccess;
+  features_12.shaderInt8 = true;
+  features_12.shaderFloat16 = available_12.shaderFloat16;
+  features_12.vulkanMemoryModel = true;
+  features_12.vulkanMemoryModelDeviceScope =
+      available_12.vulkanMemoryModelDeviceScope;
+  features_12.pNext = &features_13;
+  VkPhysicalDeviceVulkan11Features features_11{
+      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES, nullptr};
+  features_11.storageBuffer16BitAccess = available_11.storageBuffer16BitAccess;
+  features_11.uniformAndStorageBuffer16BitAccess =
+      available_11.uniformAndStorageBuffer16BitAccess;
+  features_11.pNext = &features_12;
+  VkPhysicalDeviceTensorFeaturesARM features_tensor{
+      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TENSOR_FEATURES_ARM, nullptr};
+  features_tensor.shaderTensorAccess = true;
+  features_tensor.tensors = true;
+  features_tensor.pNext = &features_11;
+  VkPhysicalDeviceDataGraphFeaturesARM features_graph{
+      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DATA_GRAPH_FEATURES_ARM, nullptr};
+  features_graph.dataGraph = true;
+  features_graph.pNext = &features_tensor;
+
+  VkPhysicalDeviceFeatures device_features = {};
+  device_features.shaderInt16 = VK_TRUE;
+  device_features.shaderInt64 = VK_TRUE;
+
+  // Extension strings to enable
+  auto dev_exts = {
+      "VK_ARM_tensors",
+      "VK_ARM_data_graph",
+      "VK_KHR_maintenance4",
+      "VK_KHR_maintenance5",
+      "VK_KHR_deferred_host_operations",
+      "VK_EXT_shader_replicated_composites"};
+
+  uint32_t exts = 0;
+  vkEnumerateDeviceExtensionProperties(
+      *physical_device, nullptr, &exts, nullptr);
+  vector<VkExtensionProperties> available(exts);
+  vkEnumerateDeviceExtensionProperties(
+      *physical_device, nullptr, &exts, available.data());
+
+  vector<const char*> requested_exts;
+  for (auto& ext : dev_exts) {
+    bool found = false;
+    for (auto const& ext_avail : available) {
+      if (strcmp(ext, ext_avail.extensionName) == 0) {
+        found = true;
+        requested_exts.push_back(ext);
+      }
+    }
+    if (found == false) {
+      ET_LOG(Info, "Failed to find extension %s", ext);
+    }
+  }
+
+  // Create the device with our subset of features
   VkDeviceCreateInfo dci{VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO, nullptr};
   dci.queueCreateInfoCount = 1;
   dci.pQueueCreateInfos = &queue_info;
-  dci.enabledExtensionCount = 2;
-  dci.ppEnabledExtensionNames = dev_exts;
+  dci.enabledExtensionCount = requested_exts.size();
+  dci.ppEnabledExtensionNames = requested_exts.data();
+  ;
+  dci.pEnabledFeatures = &device_features;
+  dci.pNext = &features_graph;
   result = vkCreateDevice(*physical_device, &dci, nullptr, device);
   if (result != VK_SUCCESS) {
     ET_LOG(Error, "Failed to create VkDevice");