vulkan: use memory budget extension to read memory usage (ggml-org#15545)

* vulkan: use memory budget extension to read memory usage

* fix: formatting and names

* formatting

* fix: detect and cache memory budget extension availability on init

* fix: read `budgetprops.heapBudget` instead of `heap.size` when memory budget extension is available

* style: lints

Author: giladgd

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1424,6 +1424,7 @@ struct vk_instance_t {
     PFN_vkGetCalibratedTimestampsEXT pfn_vkGetCalibratedTimestampsEXT = {};
 
     std::vector<size_t> device_indices;
+    std::vector<bool>   device_supports_membudget;
     vk_device devices[GGML_VK_MAX_DEVICES];
 };
 
@@ -4431,7 +4432,6 @@ static void ggml_vk_instance_init() {
         vk_instance.pfn_vkCmdBeginDebugUtilsLabelEXT = (PFN_vkCmdBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdBeginDebugUtilsLabelEXT");
         vk_instance.pfn_vkCmdEndDebugUtilsLabelEXT =   (PFN_vkCmdEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdEndDebugUtilsLabelEXT");
         vk_instance.pfn_vkCmdInsertDebugUtilsLabelEXT = (PFN_vkCmdInsertDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdInsertDebugUtilsLabelEXT");
-
     }
 
     vk_perf_logger_enabled    = getenv("GGML_VK_PERF_LOGGER")    != nullptr;
@@ -4448,10 +4448,12 @@ static void ggml_vk_instance_init() {
         }
     }
 
+    std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
+
     // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan
     char * devices_env = getenv("GGML_VK_VISIBLE_DEVICES");
     if (devices_env != nullptr) {
-        size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
+        size_t num_available_devices = devices.size();
 
         std::string devices(devices_env);
         std::replace(devices.begin(), devices.end(), ',', ' ');
@@ -4466,8 +4468,6 @@ static void ggml_vk_instance_init() {
             vk_instance.device_indices.push_back(tmp);
         }
     } else {
-        std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
-
         // If no vulkan devices are found, return early
         if (devices.empty()) {
             GGML_LOG_INFO("ggml_vulkan: No devices found.\n");
@@ -4572,6 +4572,19 @@ static void ggml_vk_instance_init() {
     GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size());
 
     for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
+        vk::PhysicalDevice vkdev = devices[vk_instance.device_indices[i]];
+        std::vector<vk::ExtensionProperties> extensionprops = vkdev.enumerateDeviceExtensionProperties();
+
+        bool membudget_supported = false;
+        for (const auto & ext : extensionprops) {
+            if (strcmp(VK_EXT_MEMORY_BUDGET_EXTENSION_NAME, ext.extensionName) == 0) {
+                membudget_supported = true;
+                break;
+            }
+        }
+
+        vk_instance.device_supports_membudget.push_back(membudget_supported);
+
         ggml_vk_print_gpu_info(i);
     }
 }
@@ -11881,15 +11894,29 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
 
 void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
     GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+    GGML_ASSERT(device < (int) vk_instance.device_supports_membudget.size());
 
     vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
+    vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops;
+    vk::PhysicalDeviceMemoryProperties2 memprops = {};
+    bool membudget_supported = vk_instance.device_supports_membudget[device];
+
+    if (membudget_supported) {
+        memprops.pNext = &budgetprops;
+    }
+    vkdev.getMemoryProperties2(&memprops);
 
-    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+    for (uint32_t i = 0; i < memprops.memoryProperties.memoryHeapCount; ++i) {
+        const vk::MemoryHeap & heap = memprops.memoryProperties.memoryHeaps[i];
 
-    for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
         if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
             *total = heap.size;
-            *free = heap.size;
+
+            if (membudget_supported && i < budgetprops.heapUsage.size()) {
+                *free = budgetprops.heapBudget[i] - budgetprops.heapUsage[i];
+            } else {
+                *free = heap.size;
+            }
             break;
         }
     }