kompute_tests/test3/main.c

#include <stdio.h>
#include <stdlib.h>
#include <stddef.h>
#include <stdarg.h>
#include <string.h>
#include <errno.h>

#include <vulkan/vulkan.h>
#include <vulkan/vulkan_core.h>

// check for half precision floating point support, for x86 this is equivalent to
// checking for SSE2
#define SUPPORTS_NATIVE_FP16 (__x86_64__ == 1 && __SSE2__ == 1)
// print debug messages
#define DEBUG                1
#define VERBOSE              0

// define half precision floating point
#if SUPPORTS_NATIVE_FP16
// extension is needed due to -pedantic
__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;

// FIXME: including vulkan/vk_enum_string_helper.h does not compile
extern const char *vk_Result_to_str(VkResult input);

// like printf but on stderr
int err(const char *fmt, ...)
{
	va_list ap;
	va_start(ap, fmt);
	int ret = vfprintf(stderr, fmt, ap);
	va_end(ap);
	return ret;
}

// print out all the instance extensions
// NOTE: these are different from device and shader extensions
int vk_enumerate_instance_extensions(void)
{
	uint32_t ex_no = 0;
#if VERBOSE > 0
	vkEnumerateInstanceExtensionProperties(NULL, &ex_no, NULL);
	VkExtensionProperties *ex_arr =
	    malloc(sizeof(VkExtensionProperties) * ex_no);
	if (ex_arr == NULL) {
		err("ERROR: in %s: %s\n", __func__, strerror(errno));
		return -1;
	}
	vkEnumerateInstanceExtensionProperties(NULL, &ex_no, ex_arr);
	printf("Available Properties: \n");
	for (uint32_t i = 0; i < ex_no; i++) {
		printf("\t%s\n", ex_arr[i].extensionName);
	}
	free(ex_arr);
#endif
	return ex_no;
}

// on debug check for support of validation layers and activate one, a validation
// layer is useful to do more error checking at runtime like ckecking for invalid
// arguments, validation layers are available only if vulkan-sdk is installed
// (vulkan-devel on arch)
int vk_activate_validation_layer(VkInstanceCreateInfo *cinfo)
{
	uint32_t prop_no = 0;
#if DEBUG > 0
	vkEnumerateInstanceLayerProperties(&prop_no, NULL);

	VkLayerProperties *prop_arr = malloc(sizeof(VkLayerProperties) * prop_no);
	if (prop_arr == NULL) {
		err("ERROR: in %s: %s\n", __func__, strerror(errno));
		return -1;
	}
	vkEnumerateInstanceLayerProperties(&prop_no, prop_arr);

	for (uint32_t i = 0; i < prop_no; i++) {
		if (strcmp(prop_arr[i].layerName, vk_validation_layer[0]) == 0) {
			cinfo->enabledLayerCount   = vk_validation_layer_no;
			cinfo->ppEnabledLayerNames = vk_validation_layer;
			free(prop_arr);
			return 0;
		}
	}
	free(prop_arr);
	return 1;
#endif
	return 0;
}

VkInstance vk_init(void)
{
	// create a vulkan instance and fill it with the application data
	VkResult   res;
	VkInstance vk_instance = VK_NULL_HANDLE;

	VkApplicationInfo vk_appinfo = {
	    .sType              = VK_STRUCTURE_TYPE_APPLICATION_INFO,
	    .pNext              = NULL,
	    .pApplicationName   = __FILE__,
	    .applicationVersion = VK_MAKE_VERSION(0, 1, 0),
	    .pEngineName        = "no engine",
	    .engineVersion      = VK_MAKE_VERSION(0, 0, 0),
	    .apiVersion =
		VK_API_VERSION_1_2, // api version 1.2 is more widely available
	};

	vk_enumerate_instance_extensions();

	// TODO: check for extension availability
	// TODO: does the lifetime of VkInstanceCreateInfo has to be the same as the
	//       lifetime of VkInstance?
	const char *vk_instance_extensions[] = {
	    VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME,
	};
	const uint32_t vk_instance_extensions_no =
	    (uint32_t)(sizeof(vk_instance_extensions) / sizeof(char *));

	VkInstanceCreateInfo vk_instanceinfo = {
	    .sType            = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
	    .pApplicationInfo = &vk_appinfo,
	    .flags            = VK_INSTANCE_CREATE_ENUMERATE_PORTABILITY_BIT_KHR,
	    .enabledExtensionCount   = vk_instance_extensions_no,
	    .ppEnabledExtensionNames = vk_instance_extensions,
	    .enabledLayerCount       = 0,
	};

	int e = 0;
	if ((e = vk_activate_validation_layer(&vk_instanceinfo))) {
		err("Could not activate validation layers%s\n",
		    e > 0 ? ": No validation layers found" : "");
	}

	res = vkCreateInstance(&vk_instanceinfo, NULL, &vk_instance);
	if (res != VK_SUCCESS) {
		err("ERROR: Could not create vulkan instance %s",
		    vk_Result_to_str(res));
		return VK_NULL_HANDLE;
	} else {
#if VERBOSE > 0
		printf("Created vulkan instance\n");
#endif
	}
	return vk_instance;
}

void vk_destroy(VkInstance vk_instance)
{
	// ...
	vkDestroyInstance(vk_instance, NULL);
}

VkPhysicalDevice vk_physical_device_get(VkInstance vk_instance)
{
	// get the physical devices list
	VkPhysicalDevice vk_phydev = VK_NULL_HANDLE;

	uint32_t          vk_phydevs_no = 0;
	VkPhysicalDevice *vk_phydevs;
	vkEnumeratePhysicalDevices(vk_instance, &vk_phydevs_no, NULL);

	if (vk_phydevs_no == 0) {
		return vk_phydev;
	}

	vk_phydevs = malloc(sizeof(VkPhysicalDevice) * vk_phydevs_no);
	if (vk_phydevs == NULL) {
		err("ERROR: in %s: %s\n", __func__, strerror(errno));
		return NULL;
	}

	vkEnumeratePhysicalDevices(vk_instance, &vk_phydevs_no, vk_phydevs);

	// print out information about each device
	printf("Available Physical Devices: \n");
	for (uint32_t i = 0; i < vk_phydevs_no; i++) {
		VkPhysicalDevice                 dev = vk_phydevs[i];
		VkPhysicalDeviceProperties       dev_properties;
		VkPhysicalDeviceFeatures         dev_features;
		VkPhysicalDeviceMemoryProperties dev_memory;

		vkGetPhysicalDeviceProperties(dev, &dev_properties);
		vkGetPhysicalDeviceFeatures(dev, &dev_features);
		vkGetPhysicalDeviceMemoryProperties(dev, &dev_memory);

		printf(
		    "\tDevice %d: %s, Discrete: %s\n",
		    i,
		    dev_properties.deviceName,
		    dev_properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
			? "true"
			: "false"
		);

		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     is_local =
			    TEST_BIT(mem_flags, VK_MEMORY_HEAP_DEVICE_LOCAL_BIT);
			printf(
			    "\t\tHeap %.2d: local: %d, size: %.3f MiB\n",
			    x,
			    is_local,
			    (float)mem_size / (1024.0 * 1024.0)
			);
		}
	}

	// TODO: find the most suitable physical device, but for now every vulkan
	//       device has to be compatible with compute shaders
	vk_phydev = vk_phydevs[0];

	free(vk_phydevs);
	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) {
		// ...
	}
}

// return the index of the first queue family that supports compute on the device,
// returns a negative index on error
// 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                      qfamily_idx = -1;

	vkGetPhysicalDeviceQueueFamilyProperties(vk_phydev, &vk_qfamilies_no, NULL);

	vk_qfamilies = malloc(sizeof(VkQueueFamilyProperties) * vk_qfamilies_no);
	if (vk_qfamilies == NULL) {
		err("ERROR: in %s: %s\n", __func__, strerror(errno));
		return -1;
	}

	vkGetPhysicalDeviceQueueFamilyProperties(
	    vk_phydev, &vk_qfamilies_no, vk_qfamilies
	);

	for (uint32_t i = 0; i < vk_qfamilies_no; i++) {
		if (TEST_BIT(vk_qfamilies[i].queueFlags, VK_QUEUE_COMPUTE_BIT)) {
			qfamily_idx = i;
		}
	}

	free(vk_qfamilies);
	return qfamily_idx;
}

VkDevice vk_logical_device_create(VkPhysicalDevice vk_phydev, int qfamily_idx)
{
	VkResult res;
	VkDevice vk_logdev         = VK_NULL_HANDLE;
	float    vk_queue_priority = 1.0f;

	// specify which command queues to use for the physical device
	VkDeviceQueueCreateInfo vk_queueinfo = {
	    .sType            = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
	    .pNext            = NULL,
	    .flags            = 0,
	    .queueFamilyIndex = qfamily_idx,
	    .queueCount       = 1,
	    .pQueuePriorities = &vk_queue_priority,
	};

	// specify which device features to use
	// TODO: this
	VkPhysicalDeviceFeatures vk_phydev_features = {0};

	// actually create the logical device
	// TODO: figure out what device extensions are
	// FIXME: here validation layers are ignored but it is still better to define
	//        them for compatibility
	VkDeviceCreateInfo vk_createinfo = {
	    .sType                   = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
	    .pQueueCreateInfos       = &vk_queueinfo,
	    .queueCreateInfoCount    = 1,
	    .pEnabledFeatures        = &vk_phydev_features,
	    .ppEnabledExtensionNames = NULL,
	    .enabledExtensionCount   = 0,
	    .ppEnabledLayerNames     = NULL,
	    .enabledLayerCount       = 0,
	};

	res = vkCreateDevice(vk_phydev, &vk_createinfo, NULL, &vk_logdev);
	if (res != VK_SUCCESS) {
		err("ERROR: Could not create vulkan logical device %s",
		    vk_Result_to_str(res));
		return VK_NULL_HANDLE;
	} else {
#if VERBOSE > 0
		printf("Created vulkan logical device\n");
#endif
	}

	return vk_logdev;
}

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;
	vkGetDeviceQueue(vk_logdev, qfamily_idx, 0, &vk_queue);
	return vk_queue;
}

int main(void)
{
#if VERBOSE > 0
	if (SUPPORTS_NATIVE_FP16) {
		printf("Processor supports half precision floating point\n");
	} else {
		printf("Processor doesn't support half precision floating point\n");
		return EXIT_FAILURE;
	}
#endif

	VkInstance vk_instance = vk_init();
	if (vk_instance == VK_NULL_HANDLE) {
		exit(EXIT_FAILURE);
	}
	VkPhysicalDevice vk_phydev   = vk_physical_device_get(vk_instance);
	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);

	return EXIT_SUCCESS;
}