allocate memory

test3/main.c

@@ -21,6 +21,12 @@
 __extension__ typedef _Float16 half;
 #endif
 
+// useful macros
+#define TEST_BIT(f, b) (!!(f & b))
+#define GIB(x)         ((uint64_t)x * 1024u * 1024u * 1024u)
+#define MIB(x)         ((uint64_t)x * 1024u * 1024u)
+#define KIB(x)         ((uint64_t)x * 1024u)
+
 const char    *vk_validation_layer[]  = {"VK_LAYER_KHRONOS_validation"};
 const uint32_t vk_validation_layer_no = 1;
 
@@ -198,11 +204,12 @@ VkPhysicalDevice vk_physical_device_get(VkInstance vk_instance)
 		for (unsigned x = 0; x < dev_memory.memoryHeapCount; x++) {
 			uint64_t mem_size  = dev_memory.memoryHeaps[x].size;
 			uint32_t mem_flags = dev_memory.memoryHeaps[x].flags;
-			char mem_local = mem_flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
+			char     is_local =
+			    TEST_BIT(mem_flags, VK_MEMORY_HEAP_DEVICE_LOCAL_BIT);
 			printf(
 			    "\t\tHeap %.2d: local: %d, size: %.3f MiB\n",
 			    x,
-			    mem_local,
+			    is_local,
 			    (float)mem_size / (1024.0 * 1024.0)
 			);
 		}
@@ -216,6 +223,63 @@ VkPhysicalDevice vk_physical_device_get(VkInstance vk_instance)
 	return vk_phydev;
 }
 
+// returns the index of a usable memory type in the device that is also backed by
+// a heap with a size of at least min_size bytes
+int vk_device_get_usable_memory_type_index(
+    VkPhysicalDevice vk_phydev, uint64_t min_size
+)
+{
+	int                              memtype_idx = -1;
+	VkPhysicalDeviceMemoryProperties dev_memory;
+	vkGetPhysicalDeviceMemoryProperties(vk_phydev, &dev_memory);
+
+	VkMemoryPropertyFlags flags = 0;
+	uint32_t              idx   = 0;
+	VkMemoryHeap          mem;
+	for (unsigned i = 0; i < dev_memory.memoryTypeCount; i++) {
+		flags = dev_memory.memoryTypes[i].propertyFlags;
+		idx   = dev_memory.memoryTypes[i].heapIndex;
+		mem   = dev_memory.memoryHeaps[idx];
+
+		// TODO: do we need more flags to be set?
+		if (TEST_BIT(flags, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) &&
+		    TEST_BIT(flags, VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) &&
+		    mem.size >= min_size) {
+			// as the name suggests we only care about the memory type
+			// and not the heap itself
+			memtype_idx = i;
+			break;
+		}
+	}
+
+	return memtype_idx;
+}
+
+// do an allocation on the device of size bytes, according to krhonos it is a good
+// idea to do one or few allocations and subdivide them on the host
+// https://github.com/KhronosGroup/Vulkan-Guide/blob/main/chapters/memory_allocation.adoc
+// this memory has to be freed using vkFreeMemory(device, mem, NULL);
+VkDeviceMemory
+vk_allocate_memory(VkDevice vk_logdev, uint32_t memtype_index, uint64_t size)
+{
+	VkMemoryAllocateInfo alloc_info = {
+	    .sType           = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+	    .pNext           = NULL,
+	    .allocationSize  = size,
+	    .memoryTypeIndex = memtype_index,
+	};
+	VkDeviceMemory mem = VK_NULL_HANDLE;
+
+	VkResult res = vkAllocateMemory(vk_logdev, &alloc_info, NULL, &mem);
+	if (res != VK_SUCCESS) {
+		err("Error allocating memory on device: %s\n",
+		    vk_Result_to_str(res));
+		return VK_NULL_HANDLE;
+	}
+
+	return mem;
+}
+
 void vk_physical_device_destroy(VkPhysicalDevice vk_phydev)
 {
 	if (vk_phydev != VK_NULL_HANDLE) {
@@ -225,11 +289,14 @@ void vk_physical_device_destroy(VkPhysicalDevice vk_phydev)
 
 // return the index of the first queue family that supports compute on the device,
 // returns a negative index on error
-int vk_device_compute_queue_index(VkPhysicalDevice vk_phydev)
+// A better approach would be to find a queue that only handled compute workloads
+// (but you need to ignore the transfer bit and for our purposes the sparse binding
+// bit too)
+int vk_device_get_compute_queue_index(VkPhysicalDevice vk_phydev)
 {
 	uint32_t                 vk_qfamilies_no = 0;
 	VkQueueFamilyProperties *vk_qfamilies;
-	int                      supports = -1;
+	int                      qfamily_idx = -1;
 
 	vkGetPhysicalDeviceQueueFamilyProperties(vk_phydev, &vk_qfamilies_no, NULL);
 
@@ -244,13 +311,13 @@ int vk_device_compute_queue_index(VkPhysicalDevice vk_phydev)
 	);
 
 	for (uint32_t i = 0; i < vk_qfamilies_no; i++) {
-		if (vk_qfamilies[i].queueFlags & VK_QUEUE_COMPUTE_BIT) {
+		if (TEST_BIT(vk_qfamilies[i].queueFlags, VK_QUEUE_COMPUTE_BIT)) {
-			supports = i;
+			qfamily_idx = i;
 		}
 	}
 
 	free(vk_qfamilies);
-	return supports;
+	return qfamily_idx;
 }
 
 VkDevice vk_logical_device_create(VkPhysicalDevice vk_phydev, int qfamily_idx)
@@ -307,6 +374,7 @@ void vk_logical_device_destroy(VkDevice vk_logdev)
 	vkDestroyDevice(vk_logdev, NULL);
 }
 
+// get the queue handle from it's index
 VkQueue vk_queue_get(VkDevice vk_logdev, int qfamily_idx)
 {
 	VkQueue vk_queue = VK_NULL_HANDLE;
@@ -330,13 +398,28 @@ int main(void)
 		exit(EXIT_FAILURE);
 	}
 	VkPhysicalDevice vk_phydev   = vk_physical_device_get(vk_instance);
-	int              qfamily_idx = vk_device_compute_queue_index(vk_phydev);
+	int              qfamily_idx = vk_device_get_compute_queue_index(vk_phydev);
 	if (qfamily_idx < 0) {
 		err("The device does not support compute queues\n");
 		exit(EXIT_FAILURE);
 	}
 	VkDevice vk_logdev = vk_logical_device_create(vk_phydev, qfamily_idx);
 
+	int devmem_idx = vk_device_get_usable_memory_type_index(vk_phydev, GIB(1));
+	if (devmem_idx < 0) {
+		err("Could not find a suitable device memory heap\n");
+		exit(EXIT_FAILURE);
+	}
+
+	VkDeviceMemory mem = vk_allocate_memory(vk_logdev, devmem_idx, MIB(256));
+	if (mem == VK_NULL_HANDLE) {
+		exit(EXIT_FAILURE);
+	} else {
+		printf("Successfully allocated memory on device\n");
+	}
+
+	vkFreeMemory(vk_logdev, mem, NULL);
+
 	vk_logical_device_destroy(vk_logdev);
 	vk_physical_device_destroy(vk_phydev);
 	vk_destroy(vk_instance);