lovr/src/core/gpu_vk.c

#include "gpu.h"
#include <string.h>
#define VK_NO_PROTOTYPES
#include <vulkan/vulkan.h>
#ifdef _WIN32
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#else
#include <dlfcn.h>
#endif

typedef struct gpu_memory gpu_memory;

// Objects

struct gpu_buffer {
  VkBuffer handle;
  gpu_memory* memory;
};

size_t gpu_sizeof_buffer() { return sizeof(gpu_buffer); }

// Pools

struct gpu_memory {
  VkDeviceMemory handle;
  void* pointer;
  uint32_t refs;
};

typedef enum {
  GPU_MEMORY_BUFFER_GPU,
  GPU_MEMORY_COUNT
} gpu_memory_type;

typedef struct {
  gpu_memory* block;
  uint32_t cursor;
  uint16_t memoryType;
  uint16_t memoryFlags;
} gpu_allocator;

typedef struct {
  void* handle;
  VkObjectType type;
  uint32_t tick;
} gpu_victim;

typedef struct {
  uint32_t head;
  uint32_t tail;
  gpu_victim data[256];
} gpu_morgue;

typedef struct {
  VkCommandPool pool;
  VkSemaphore semaphores[2];
  VkFence fence;
} gpu_tick;

// State

static struct {
  void* library;
  gpu_config config;
  VkInstance instance;
  VkPhysicalDevice adapter;
  VkDevice device;
  VkQueue queue;
  uint32_t queueFamilyIndex;
  VkDebugUtilsMessengerEXT messenger;
  gpu_allocator allocators[GPU_MEMORY_COUNT];
  uint8_t allocatorLookup[GPU_MEMORY_COUNT];
  gpu_memory memory[256];
  gpu_morgue morgue;
  gpu_tick ticks[2];
  uint32_t tick[2];
} state;

// Helpers

enum { CPU, GPU };

#define MIN(a, b) (a < b ? a : b)
#define MAX(a, b) (a > b ? a : b)
#define COUNTOF(x) (sizeof(x) / sizeof(x[0]))
#define ALIGN(p, n) (((uintptr_t) (p) + (n - 1)) & ~(n - 1))
#define VK(f, s) if (!vcheck(f, s))
#define CHECK(c, s) if (!check(c, s))
#define TICK_MASK (COUNTOF(state.ticks) - 1)
#define MORGUE_MASK (COUNTOF(state.morgue.data) - 1)

static gpu_memory* gpu_allocate(gpu_memory_type type, VkMemoryRequirements info, VkDeviceSize* offset);
static void gpu_release(gpu_memory* memory);
static void condemn(void* handle, VkObjectType type);
static void expunge(void);
static VkBool32 relay(VkDebugUtilsMessageSeverityFlagBitsEXT severity, VkDebugUtilsMessageTypeFlagsEXT flags, const VkDebugUtilsMessengerCallbackDataEXT* data, void* userdata);
static void nickname(void* object, VkObjectType type, const char* name);
static bool vcheck(VkResult result, const char* message);
static bool check(bool condition, const char* message);

// Loader

// Functions that don't require an instance
#define GPU_FOREACH_ANONYMOUS(X)\
  X(vkCreateInstance)

// Functions that require an instance but don't require a device
#define GPU_FOREACH_INSTANCE(X)\
  X(vkDestroyInstance)\
  X(vkCreateDebugUtilsMessengerEXT)\
  X(vkDestroyDebugUtilsMessengerEXT)\
  X(vkDestroySurfaceKHR)\
  X(vkEnumeratePhysicalDevices)\
  X(vkGetPhysicalDeviceProperties2)\
  X(vkGetPhysicalDeviceFeatures2)\
  X(vkGetPhysicalDeviceMemoryProperties)\
  X(vkGetPhysicalDeviceFormatProperties)\
  X(vkGetPhysicalDeviceQueueFamilyProperties)\
  X(vkGetPhysicalDeviceSurfaceSupportKHR)\
  X(vkGetPhysicalDeviceSurfaceCapabilitiesKHR)\
  X(vkGetPhysicalDeviceSurfaceFormatsKHR)\
  X(vkCreateDevice)\
  X(vkDestroyDevice)\
  X(vkGetDeviceQueue)\
  X(vkGetDeviceProcAddr)

// Functions that require a device
#define GPU_FOREACH_DEVICE(X)\
  X(vkSetDebugUtilsObjectNameEXT)\
  X(vkCmdBeginDebugUtilsLabelEXT)\
  X(vkCmdEndDebugUtilsLabelEXT)\
  X(vkDeviceWaitIdle)\
  X(vkQueueSubmit)\
  X(vkQueuePresentKHR)\
  X(vkCreateSwapchainKHR)\
  X(vkDestroySwapchainKHR)\
  X(vkGetSwapchainImagesKHR)\
  X(vkAcquireNextImageKHR)\
  X(vkCreateCommandPool)\
  X(vkDestroyCommandPool)\
  X(vkResetCommandPool)\
  X(vkAllocateCommandBuffers)\
  X(vkBeginCommandBuffer)\
  X(vkEndCommandBuffer)\
  X(vkCreateFence)\
  X(vkDestroyFence)\
  X(vkResetFences)\
  X(vkGetFenceStatus)\
  X(vkWaitForFences)\
  X(vkCreateSemaphore)\
  X(vkDestroySemaphore)\
  X(vkCmdPipelineBarrier)\
  X(vkCreateQueryPool)\
  X(vkDestroyQueryPool)\
  X(vkCmdResetQueryPool)\
  X(vkCmdWriteTimestamp)\
  X(vkCmdCopyQueryPoolResults)\
  X(vkCreateBuffer)\
  X(vkDestroyBuffer)\
  X(vkGetBufferMemoryRequirements)\
  X(vkBindBufferMemory)\
  X(vkCreateImage)\
  X(vkDestroyImage)\
  X(vkGetImageMemoryRequirements)\
  X(vkBindImageMemory)\
  X(vkCmdCopyBuffer)\
  X(vkCmdCopyImage)\
  X(vkCmdBlitImage)\
  X(vkCmdCopyBufferToImage)\
  X(vkCmdCopyImageToBuffer)\
  X(vkCmdFillBuffer)\
  X(vkCmdClearColorImage)\
  X(vkAllocateMemory)\
  X(vkFreeMemory)\
  X(vkMapMemory)\
  X(vkCreateSampler)\
  X(vkDestroySampler)\
  X(vkCreateRenderPass)\
  X(vkDestroyRenderPass)\
  X(vkCmdBeginRenderPass)\
  X(vkCmdEndRenderPass)\
  X(vkCreateImageView)\
  X(vkDestroyImageView)\
  X(vkCreateFramebuffer)\
  X(vkDestroyFramebuffer)\
  X(vkCreateShaderModule)\
  X(vkDestroyShaderModule)\
  X(vkCreateDescriptorSetLayout)\
  X(vkDestroyDescriptorSetLayout)\
  X(vkCreatePipelineLayout)\
  X(vkDestroyPipelineLayout)\
  X(vkCreateDescriptorPool)\
  X(vkDestroyDescriptorPool)\
  X(vkAllocateDescriptorSets)\
  X(vkResetDescriptorPool)\
  X(vkUpdateDescriptorSets)\
  X(vkCreatePipelineCache)\
  X(vkDestroyPipelineCache)\
  X(vkCreateGraphicsPipelines)\
  X(vkCreateComputePipelines)\
  X(vkDestroyPipeline)\
  X(vkCmdSetViewport)\
  X(vkCmdSetScissor)\
  X(vkCmdPushConstants)\
  X(vkCmdBindPipeline)\
  X(vkCmdBindDescriptorSets)\
  X(vkCmdBindVertexBuffers)\
  X(vkCmdBindIndexBuffer)\
  X(vkCmdDraw)\
  X(vkCmdDrawIndexed)\
  X(vkCmdDrawIndirect)\
  X(vkCmdDrawIndexedIndirect)\
  X(vkCmdDispatch)\
  X(vkCmdDispatchIndirect)

// Used to load/declare Vulkan functions without lots of clutter
#define GPU_LOAD_ANONYMOUS(fn) fn = (PFN_##fn) vkGetInstanceProcAddr(NULL, #fn);
#define GPU_LOAD_INSTANCE(fn) fn = (PFN_##fn) vkGetInstanceProcAddr(state.instance, #fn);
#define GPU_LOAD_DEVICE(fn) fn = (PFN_##fn) vkGetDeviceProcAddr(state.device, #fn);
#define GPU_DECLARE(fn) static PFN_##fn fn;

// Declare function pointers
GPU_FOREACH_ANONYMOUS(GPU_DECLARE)
GPU_FOREACH_INSTANCE(GPU_DECLARE)
GPU_FOREACH_DEVICE(GPU_DECLARE)

// Buffer

bool gpu_buffer_init(gpu_buffer* buffer, gpu_buffer_info* info) {
  if (info->handle) {
    buffer->handle = (VkBuffer) info->handle;
    return true;
  }

  VkBufferCreateInfo createInfo = {
    .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
    .size = info->size,
    .usage =
      VK_BUFFER_USAGE_VERTEX_BUFFER_BIT |
      VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
      VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
      VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
      VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT |
      VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
      VK_BUFFER_USAGE_TRANSFER_DST_BIT
  };

  VK(vkCreateBuffer(state.device, &createInfo, NULL, &buffer->handle), "Could not create buffer") return false;
  nickname(buffer->handle, VK_OBJECT_TYPE_BUFFER, info->label);

  VkDeviceSize offset;
  VkMemoryRequirements requirements;
  vkGetBufferMemoryRequirements(state.device, buffer->handle, &requirements);
  buffer->memory = gpu_allocate(GPU_MEMORY_BUFFER_GPU, requirements, &offset);

  VK(vkBindBufferMemory(state.device, buffer->handle, buffer->memory->handle, offset), "Could not bind buffer memory") {
    vkDestroyBuffer(state.device, buffer->handle, NULL);
    gpu_release(buffer->memory);
    return false;
  }

  if (info->pointer) {
    *info->pointer = buffer->memory->pointer ? (char*) buffer->memory->pointer + offset : NULL;
  }

  return true;
}

void gpu_buffer_destroy(gpu_buffer* buffer) {
  condemn(buffer->handle, VK_OBJECT_TYPE_BUFFER);
  gpu_release(buffer->memory);
}

// Entry

bool gpu_init(gpu_config* config) {
  state.config = *config;

  // Load
#ifdef _WIN32
  state.library = LoadLibraryA("vulkan-1.dll");
  CHECK(state.library, "Failed to load vulkan library") return gpu_destroy(), false;
  PFN_vkGetInstanceProcAddr vkGetInstanceProcAddr = (PFN_vkGetInstanceProcAddr) GetProcAddress(state.library, "vkGetInstanceProcAddr");
#elif __APPLE__
  state.library = dlopen("libvulkan.1.dylib", RTLD_NOW | RTLD_LOCAL);
  CHECK(state.library, "Failed to load vulkan library") return gpu_destroy(), false;
  PFN_vkGetInstanceProcAddr vkGetInstanceProcAddr = (PFN_vkGetInstanceProcAddr) dlsym(state.library, "vkGetInstanceProcAddr");
#else
  state.library = dlopen("libvulkan.so", RTLD_NOW | RTLD_LOCAL);
  CHECK(state.library, "Failed to load vulkan library") return gpu_destroy(), false;
  PFN_vkGetInstanceProcAddr vkGetInstanceProcAddr = (PFN_vkGetInstanceProcAddr) dlsym(state.library, "vkGetInstanceProcAddr");
#endif
  GPU_FOREACH_ANONYMOUS(GPU_LOAD_ANONYMOUS);

  { // Instance
    VkInstanceCreateInfo instanceInfo = {
      .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
      .pApplicationInfo = &(VkApplicationInfo) {
        .sType = VK_STRUCTURE_TYPE_APPLICATION_INFO,
        .pEngineName = config->engineName,
        .engineVersion = VK_MAKE_VERSION(config->engineVersion[0], config->engineVersion[1], config->engineVersion[3]),
        .apiVersion = VK_MAKE_VERSION(1, 1, 0)
      },
      .enabledLayerCount = state.config.debug ? 1 : 0,
      .ppEnabledLayerNames = (const char*[]) { "VK_LAYER_KHRONOS_validation" },
      .enabledExtensionCount = state.config.debug ? 1 : 0,
      .ppEnabledExtensionNames = (const char*[]) { "VK_EXT_debug_utils" }
    };

    VK(vkCreateInstance(&instanceInfo, NULL, &state.instance), "Instance creation failed") return gpu_destroy(), false;

    GPU_FOREACH_INSTANCE(GPU_LOAD_INSTANCE);

    if (state.config.debug) {
      VkDebugUtilsMessengerCreateInfoEXT messengerInfo = {
        .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
        .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT | VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT,
        .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT | VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT,
        .pfnUserCallback = relay
      };

      VK(vkCreateDebugUtilsMessengerEXT(state.instance, &messengerInfo, NULL, &state.messenger), "Debug hook setup failed") return gpu_destroy(), false;
    }
  }

  { // Device
    uint32_t deviceCount = 1;
    VK(vkEnumeratePhysicalDevices(state.instance, &deviceCount, &state.adapter), "Physical device enumeration failed") return gpu_destroy(), false;

    VkPhysicalDeviceMultiviewProperties multiviewProperties = { .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES };
    VkPhysicalDeviceSubgroupProperties subgroupProperties = { .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES, .pNext = &multiviewProperties };
    VkPhysicalDeviceProperties2 properties2 = { .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2, .pNext = &subgroupProperties };
    vkGetPhysicalDeviceProperties2(state.adapter, &properties2);

    if (config->device) {
      VkPhysicalDeviceProperties* properties = &properties2.properties;
      config->device->deviceId = properties->deviceID;
      config->device->vendorId = properties->vendorID;
      memcpy(config->device->deviceName, properties->deviceName, MIN(sizeof(config->device->deviceName), sizeof(properties->deviceName)));
      config->device->renderer = "Vulkan";
      config->device->subgroupSize = subgroupProperties.subgroupSize;
      config->device->discrete = properties->deviceType == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU;
    }

    if (config->limits) {
      VkPhysicalDeviceLimits* limits = &properties2.properties.limits;
      config->limits->textureSize2D = MIN(limits->maxImageDimension2D, UINT16_MAX);
      config->limits->textureSize3D = MIN(limits->maxImageDimension3D, UINT16_MAX);
      config->limits->textureSizeCube = MIN(limits->maxImageDimensionCube, UINT16_MAX);
      config->limits->textureLayers = MIN(limits->maxImageArrayLayers, UINT16_MAX);
      config->limits->renderSize[0] = limits->maxFramebufferWidth;
      config->limits->renderSize[1] = limits->maxFramebufferHeight;
      config->limits->renderSize[2] = multiviewProperties.maxMultiviewViewCount;
      config->limits->uniformBufferRange = limits->maxUniformBufferRange;
      config->limits->storageBufferRange = limits->maxStorageBufferRange;
      config->limits->uniformBufferAlign = limits->minUniformBufferOffsetAlignment;
      config->limits->storageBufferAlign = limits->minStorageBufferOffsetAlignment;
      config->limits->vertexAttributes = limits->maxVertexInputAttributes;
      config->limits->vertexBuffers = limits->maxVertexInputBindings;
      config->limits->vertexBufferStride = MIN(limits->maxVertexInputBindingStride, UINT16_MAX);
      config->limits->vertexShaderOutputs = limits->maxVertexOutputComponents;
      config->limits->clipDistances = limits->maxClipDistances;
      config->limits->cullDistances = limits->maxCullDistances;
      config->limits->clipAndCullDistances = limits->maxCombinedClipAndCullDistances;
      config->limits->computeDispatchCount[0] = limits->maxComputeWorkGroupCount[0];
      config->limits->computeDispatchCount[1] = limits->maxComputeWorkGroupCount[1];
      config->limits->computeDispatchCount[2] = limits->maxComputeWorkGroupCount[2];
      config->limits->computeWorkgroupSize[0] = limits->maxComputeWorkGroupSize[0];
      config->limits->computeWorkgroupSize[1] = limits->maxComputeWorkGroupSize[1];
      config->limits->computeWorkgroupSize[2] = limits->maxComputeWorkGroupSize[2];
      config->limits->computeWorkgroupVolume = limits->maxComputeWorkGroupInvocations;
      config->limits->computeSharedMemory = limits->maxComputeSharedMemorySize;
      config->limits->pushConstantSize = limits->maxPushConstantsSize;
      config->limits->indirectDrawCount = limits->maxDrawIndirectCount;
      config->limits->instances = multiviewProperties.maxMultiviewInstanceIndex;
      config->limits->anisotropy = limits->maxSamplerAnisotropy;
      config->limits->pointSize = limits->pointSizeRange[1];
    }

    VkPhysicalDeviceShaderDrawParameterFeatures shaderDrawParameterFeatures = {
      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
    };

    VkPhysicalDeviceMultiviewFeatures multiviewFeatures = {
      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES,
      .pNext = &shaderDrawParameterFeatures
    };

    VkPhysicalDeviceFeatures2 enabledFeatures = {
      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
      .pNext = &multiviewFeatures
    };

    if (config->features) {
      VkPhysicalDeviceFeatures2 features2 = { .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2 };
      VkPhysicalDeviceFeatures* enable = &enabledFeatures.features;
      VkPhysicalDeviceFeatures* supports = &features2.features;
      vkGetPhysicalDeviceFeatures2(state.adapter, &features2);

      // Required features
      enable->fullDrawIndexUint32 = true;
      multiviewFeatures.multiview = true;
      shaderDrawParameterFeatures.shaderDrawParameters = true;

      // Internal features (exposed as limits)
      enable->samplerAnisotropy = supports->samplerAnisotropy;
      enable->multiDrawIndirect = supports->multiDrawIndirect;
      enable->shaderClipDistance = supports->shaderClipDistance;
      enable->shaderCullDistance = supports->shaderCullDistance;
      enable->largePoints = supports->largePoints;

      // Optional features (currently always enabled when supported)
      config->features->textureBC = enable->textureCompressionBC = supports->textureCompressionBC;
      config->features->textureASTC = enable->textureCompressionASTC_LDR = supports->textureCompressionASTC_LDR;
      config->features->wireframe = enable->fillModeNonSolid = supports->fillModeNonSolid;
      config->features->depthClamp = enable->depthClamp = supports->depthClamp;
      config->features->indirectDrawFirstInstance = enable->drawIndirectFirstInstance = supports->drawIndirectFirstInstance;
      config->features->float64 = enable->shaderFloat64 = supports->shaderFloat64;
      config->features->int64 = enable->shaderInt64 = supports->shaderInt64;
      config->features->int16 = enable->shaderInt16 = supports->shaderInt16;
    }

    state.queueFamilyIndex = ~0u;
    VkQueueFamilyProperties queueFamilies[8];
    uint32_t queueFamilyCount = COUNTOF(queueFamilies);
    vkGetPhysicalDeviceQueueFamilyProperties(state.adapter, &queueFamilyCount, queueFamilies);
    for (uint32_t i = 0; i < queueFamilyCount; i++) {
      uint32_t mask = VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT;
      if ((queueFamilies[i].queueFlags & mask) == mask) {
        state.queueFamilyIndex = i;
        break;
      }
    }
    CHECK(state.queueFamilyIndex != ~0u, "Queue selection failed") return gpu_destroy(), false;

    VkDeviceCreateInfo deviceInfo = {
      .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
      .pNext = config->features ? &enabledFeatures : NULL,
      .queueCreateInfoCount = 1,
      .pQueueCreateInfos = &(VkDeviceQueueCreateInfo) {
        .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
        .queueFamilyIndex = state.queueFamilyIndex,
        .pQueuePriorities = &(float) { 1.f },
        .queueCount = 1
      }
    };

    VK(vkCreateDevice(state.adapter, &deviceInfo, NULL, &state.device), "Device creation failed") return gpu_destroy(), false;
    vkGetDeviceQueue(state.device, state.queueFamilyIndex, 0, &state.queue);
    GPU_FOREACH_DEVICE(GPU_LOAD_DEVICE);
  }

  { // Allocators
    VkPhysicalDeviceMemoryProperties memoryProperties;
    vkGetPhysicalDeviceMemoryProperties(state.adapter, &memoryProperties);
    VkMemoryType* memoryTypes = memoryProperties.memoryTypes;

    struct { VkBufferUsageFlags usage; VkMemoryPropertyFlags flags; } bufferFlags[] = {
      [GPU_MEMORY_BUFFER_GPU] = {
        .usage =
          VK_BUFFER_USAGE_VERTEX_BUFFER_BIT |
          VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
          VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
          VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
          VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT |
          VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
          VK_BUFFER_USAGE_TRANSFER_DST_BIT,
        .flags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
      }
    };

    // Without VK_KHR_maintenance4, you have to create objects to figure out memory requirements
    for (uint32_t i = 0; i < COUNTOF(bufferFlags); i++) {
      gpu_allocator* allocator = &state.allocators[i];

      VkBufferCreateInfo createInfo = {
        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
        .usage = bufferFlags[i].usage,
        .size = 4
      };

      VkBuffer buffer;
      VkMemoryRequirements requirements;
      vkCreateBuffer(state.device, &createInfo, NULL, &buffer);
      vkGetBufferMemoryRequirements(state.device, buffer, &requirements);
      vkDestroyBuffer(state.device, buffer, NULL);

      uint32_t fallback = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;

      for (uint32_t j = 0; j < memoryProperties.memoryTypeCount; j++) {
        if (~requirements.memoryTypeBits & (1 << i)) {
          continue;
        }

        if ((memoryTypes[i].propertyFlags & fallback) == fallback) {
          allocator->memoryFlags = memoryTypes[j].propertyFlags;
          allocator->memoryType = j;
          continue;
        }

        if ((memoryTypes[i].propertyFlags & bufferFlags[i].flags) == bufferFlags[i].flags) {
          allocator->memoryFlags = memoryTypes[j].propertyFlags;
          allocator->memoryType = j;
          break;
        }
      }
    }
  }

  return true;
}

void gpu_destroy(void) {
  if (state.device) vkDeviceWaitIdle(state.device);
  if (state.device) vkDestroyDevice(state.device, NULL);
  if (state.messenger) vkDestroyDebugUtilsMessengerEXT(state.instance, state.messenger, NULL);
  if (state.instance) vkDestroyInstance(state.instance, NULL);
#ifdef _WIN32
  if (state.library) FreeLibrary(state.library);
#else
  if (state.library) dlclose(state.library);
#endif
  memset(&state, 0, sizeof(state));
}

// Helpers

static gpu_memory* gpu_allocate(gpu_memory_type type, VkMemoryRequirements info, VkDeviceSize* offset) {
  gpu_allocator* allocator = &state.allocators[state.allocatorLookup[type]];

  static const uint32_t blockSizes[] = {
    [GPU_MEMORY_BUFFER_GPU] = 1 << 26
  };

  uint32_t blockSize = blockSizes[type];
  uint32_t cursor = ALIGN(allocator->cursor, info.alignment);

  if (allocator->block && cursor + info.size <= blockSize) {
    allocator->cursor = cursor + info.size;
    allocator->block->refs++;
    *offset = cursor;
    return allocator->block;
  }

  // If there wasn't an active block or it overflowed, find an empty block to allocate
  for (uint32_t i = 0; i < COUNTOF(state.memory); i++) {
    if (!state.memory[i].handle) {
      gpu_memory* memory = &state.memory[i];

      VkMemoryAllocateInfo memoryInfo = {
        .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
        .allocationSize = MAX(blockSize, info.size),
        .memoryTypeIndex = allocator->memoryType
      };

      VK(vkAllocateMemory(state.device, &memoryInfo, NULL, &memory->handle), "Failed to allocate GPU memory") {
        allocator->block = NULL;
        return NULL;
      }

      if (allocator->memoryFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
        VK(vkMapMemory(state.device, memory->handle, 0, VK_WHOLE_SIZE, 0, &memory->pointer), "Failed to map memory") {
          vkFreeMemory(state.device, memory->handle, NULL);
          memory->handle = NULL;
          return NULL;
        }
      }

      allocator->block = memory;
      allocator->cursor = info.size;
      *offset = 0;
      return memory;
    }
  }

  check(false, "Out of space for memory blocks");
  return NULL;
}

static void gpu_release(gpu_memory* memory) {
  if (--memory->refs == 0) {
    condemn(memory->handle, VK_OBJECT_TYPE_DEVICE_MEMORY);
  }
}

static void condemn(void* handle, VkObjectType type) {
  if (!handle) return;
  gpu_morgue* morgue = &state.morgue;

  // If the morgue is full, perform an emergency expunge
  if (morgue->head - morgue->tail >= COUNTOF(morgue->data)) {
    vkDeviceWaitIdle(state.device);
    state.tick[GPU] = state.tick[CPU];
    expunge();
  }

  morgue->data[morgue->head++ & MORGUE_MASK] = (gpu_victim) { handle, type, state.tick[CPU] };
}

static void expunge() {
  gpu_morgue* morgue = &state.morgue;
  while (morgue->tail != morgue->head && state.tick[GPU] >= morgue->data[morgue->tail & MORGUE_MASK].tick) {
    gpu_victim* victim = &morgue->data[morgue->tail++ & MORGUE_MASK];
    switch (victim->type) {
      case VK_OBJECT_TYPE_BUFFER: vkDestroyBuffer(state.device, victim->handle, NULL); break;
      case VK_OBJECT_TYPE_DEVICE_MEMORY: vkFreeMemory(state.device, victim->handle, NULL); break;
      default: break;
    }
  }
}

static VkBool32 relay(VkDebugUtilsMessageSeverityFlagBitsEXT severity, VkDebugUtilsMessageTypeFlagsEXT flags, const VkDebugUtilsMessengerCallbackDataEXT* data, void* userdata) {
  if (state.config.callback) {
    bool severe = severity & VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT;
    state.config.callback(state.config.userdata, data->pMessage, severe);
  }
  return VK_FALSE;
}

static void nickname(void* handle, VkObjectType type, const char* name) {
  if (name && state.config.debug) {
    union { uint64_t u64; void* p; } pointer = { .p = handle };

    VkDebugUtilsObjectNameInfoEXT info = {
      .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_OBJECT_NAME_INFO_EXT,
      .objectType = type,
      .objectHandle = pointer.u64,
      .pObjectName = name
    };

    VK(vkSetDebugUtilsObjectNameEXT(state.device, &info), "Nickname failed") {}
  }
}

static bool vcheck(VkResult result, const char* message) {
  if (result >= 0) return true;
  if (!state.config.callback) return false;
#define CASE(x) case x: state.config.callback(state.config.userdata, "Vulkan error: " #x, false); break;
  switch (result) {
    CASE(VK_ERROR_OUT_OF_HOST_MEMORY);
    CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY);
    CASE(VK_ERROR_INITIALIZATION_FAILED);
    CASE(VK_ERROR_DEVICE_LOST);
    CASE(VK_ERROR_MEMORY_MAP_FAILED);
    CASE(VK_ERROR_LAYER_NOT_PRESENT);
    CASE(VK_ERROR_EXTENSION_NOT_PRESENT);
    CASE(VK_ERROR_FEATURE_NOT_PRESENT);
    CASE(VK_ERROR_TOO_MANY_OBJECTS);
    CASE(VK_ERROR_FORMAT_NOT_SUPPORTED);
    CASE(VK_ERROR_FRAGMENTED_POOL);
    CASE(VK_ERROR_OUT_OF_POOL_MEMORY);
    default: break;
  }
#undef CASE
  state.config.callback(state.config.userdata, message, true);
  return false;
}

static bool check(bool condition, const char* message) {
  if (!condition && state.config.callback) {
    state.config.callback(state.config.userdata, message, true);
  }
  return condition;
}