#![deny(unsafe_code)]
use crate::error::*;
use crate::graphics::permanent::{
  PermanentGraphicsState, GraphicsStateForReinit, QueueFamilyIndices,
  EnableAnisotropy
};
use crate::graphics::util::{
  allocate_buffer, stage_in_buffer,
  pick_memory_type,
  begin_transient_commands, end_transient_commands
};
use crate::shader_data::UniformBlock;

use std::collections::BTreeSet;
use std::io::Cursor;
use std::mem::size_of;

use png::Decoder;
use vulkanalia::{ Device, Instance };
use vulkanalia::vk::{ self, Handle, HasBuilder, InstanceV1_0, DeviceV1_0,
                      KhrSwapchainExtensionDeviceCommands };
use winit::window::Window;


// TODO: use VK_KHR_swapchain_maintenance1 to put a fence on the presentation
// operation. doing that will remove the requirement that we have more
// simultaneous frames than images.
pub const N_SIMULTANEOUS_FRAMES: usize = 5;


//   The WindowDressing collects the Vulkan graphics objects which need to be
// regenerated or modified when the window changes in certain ways, such as
// resizing, but are not needed during rendering. The ones which don't need to
// be regenerated are collected in PermanentGraphicsState. The ones which are
// needed during rendering are collected in RenderState, below.
#[derive(Debug)]
pub struct WindowDressing {
  pub swapchain: Swapchain,

  color_image: vk::Image,
  color_image_memory: vk::DeviceMemory,
  pub color_image_view: vk::ImageView,

  depth_image: vk::Image,
  depth_image_memory: vk::DeviceMemory,
  pub depth_image_view: vk::ImageView,
  pub depth_format: vk::Format,

  pub primary_command_pool: vk::CommandPool,
  pub transient_command_pool: vk::CommandPool,

  texture_image: vk::Image,
  texture_image_memory: vk::DeviceMemory,
  pub texture_image_view: vk::ImageView,
  pub sampler: vk::Sampler,

  pub uniform_buffers: Vec<vk::Buffer>,
  pub uniform_buffer_memory: Vec<vk::DeviceMemory>,

  pub descriptor_pool: vk::DescriptorPool,

  pub concurrency: Concurrency,
}


//   A swapchain is the generalized facility that is used to implement
// double buffering, triple buffering, rendering passes that feed into each
// other, and other things of that nature. It's a first-class thing but for
// now, we use at most one of it. We also support running without one.
#[derive(Debug)]
pub struct Swapchain {
  pub swapchain: vk::SwapchainKHR,
  pub images: Vec<vk::Image>,
  pub image_views: Vec<vk::ImageView>,
  pub format: vk::Format,
  pub extent: vk::Extent2D,
}

#[derive(Debug)]
pub struct Concurrency {
  pub image_available_semaphores: Vec<vk::Semaphore>,
  pub rendering_finished_semaphores: Vec<vk::Semaphore>,

  //   Okay, the lifetime management on the fences is really subtle. There is
  // one fence for each frame, and frame_fences holds the authoritative
  // reference to it.
  //
  //   There is one entry in image_fences for each image. The number of images
  // is not directly related to the number of frames; it will likely be
  // larger, but may be smaller or the same. At the start of execution, the
  // entries are all nulls. Each time an image is acquired from the swapchain,
  // the corresponding entry in image_fences is overwritten with a duplicate
  // of the frame fence. This happens during rendering of the frame, so the
  // frame fence is in the "signaled" state. It will be reset right before
  // submitting the queue, then signaled again when the submission completes.
  pub frame_fences: Vec<vk::Fence>,
  pub image_fences: Vec<vk::Fence>,
}


impl WindowDressing {
  pub fn new(permanent: &PermanentGraphicsState,
             for_reinit: &GraphicsStateForReinit,
             enable_anisotropy: EnableAnisotropy)
      -> Result<Self>
  {
    let window = &permanent.window;
    let instance = &permanent.instance;
    let surface = &permanent.surface;
    let device = &permanent.device;
    let graphics_queue = &permanent.graphics_queue;
    let physical_device = &for_reinit.physical_device;
    let sample_count = for_reinit.sample_count;
    let indices = &for_reinit.indices;

    let swapchain = init_swapchain(
            window, instance, surface, &physical_device, device, &indices)?;

    let (color_image, color_image_memory, color_image_view)
            = init_color(instance, &physical_device, device,
                         &swapchain.extent, sample_count, swapchain.format)?;

    let (depth_image, depth_image_memory, depth_image_view, depth_format)
            = init_depth(instance, &physical_device, device,
                         &swapchain.extent, sample_count)?;

    let (primary_command_pool, transient_command_pool)
            = init_command_pools(device, indices)?;

    let (texture_image, texture_image_memory, texture_image_view, mip_count)
            = init_texture(instance, physical_device, device,
                           graphics_queue, &transient_command_pool)?;

    let sampler = init_sampler(&device, &enable_anisotropy, mip_count)?;

    let (uniform_buffers, uniform_buffer_memory)
            = init_uniform_buffers(instance, physical_device, device,
                                   swapchain.images.len())?;

    let descriptor_pool
            = init_descriptor_pool(device, swapchain.images.len())?;

    let concurrency = init_concurrency(device, &swapchain.images)?;

    Ok(WindowDressing {
      swapchain,
      color_image,
      color_image_memory,
      color_image_view,
      depth_image,
      depth_image_memory,
      depth_image_view,
      depth_format,
      texture_image,
      texture_image_memory,
      texture_image_view,
      sampler,
      uniform_buffers,
      uniform_buffer_memory,
      descriptor_pool,
      primary_command_pool,
      transient_command_pool,
      concurrency,
    })
  }


  #[allow(unsafe_code)]
  pub fn reinit(&mut self, permanent: &PermanentGraphicsState,
                for_reinit: &GraphicsStateForReinit)
      -> Result<()>
  {
    let window = &permanent.window;
    let instance = &permanent.instance;
    let surface = &permanent.surface;
    let device = &permanent.device;
    let physical_device = &for_reinit.physical_device;
    let sample_count = for_reinit.sample_count;
    let indices = &for_reinit.indices;

    unsafe { device.device_wait_idle() }.unwrap();

    self.destroy_replaceable(device);

    let swapchain = init_swapchain(
            window, instance, surface, &physical_device, device, &indices)?;

    let (color_image, color_image_memory, color_image_view)
            = init_color(instance, &physical_device, device,
                         &swapchain.extent, sample_count, swapchain.format)?;

    let (depth_image, depth_image_memory, depth_image_view, depth_format)
            = init_depth(instance, &physical_device, device,
                         &swapchain.extent, sample_count)?;

    let (uniform_buffers, uniform_buffer_memory)
            = init_uniform_buffers(instance, physical_device, device,
                                   swapchain.images.len())?;

    // Notice that we did NOT reuse the descriptor pool.
    let descriptor_pool
            = init_descriptor_pool(device, swapchain.images.len())?;

    self.concurrency.image_fences.resize(swapchain.images.len(),
                                         vk::Fence::null());

    self.swapchain = swapchain;
    self.color_image = color_image;
    self.color_image_memory = color_image_memory;
    self.color_image_view = color_image_view;
    self.depth_image = depth_image;
    self.depth_image_memory = depth_image_memory;
    self.depth_image_view = depth_image_view;
    self.depth_format = depth_format;
    self.uniform_buffers = uniform_buffers;
    self.uniform_buffer_memory = uniform_buffer_memory;
    self.descriptor_pool = descriptor_pool;

    Ok(())
  }


  //   This relies on its caller to have already waited for the device to be
  // idle.
  #[allow(unsafe_code)]
  pub fn destroy(mut self, device: &Device) {
    self.destroy_replaceable(device);

    unsafe { device.destroy_image(self.texture_image, None) };
    unsafe { device.free_memory(self.texture_image_memory, None) };
    unsafe { device.destroy_image_view(self.texture_image_view, None) };
    unsafe { device.destroy_sampler(self.sampler, None) };

    for semaphore in self.concurrency.image_available_semaphores {
      unsafe { device.destroy_semaphore(semaphore, None) };
    }

    for semaphore in self.concurrency.rendering_finished_semaphores {
      unsafe { device.destroy_semaphore(semaphore, None) };
    }

    for fence in self.concurrency.frame_fences {
      unsafe { device.destroy_fence(fence, None) };
    }

    //   Notice that destroy_replaceable() freed the buffers in the pools, but
    // did not destroy the pools.
    unsafe { device.destroy_command_pool(self.primary_command_pool, None) };
    unsafe { device.destroy_command_pool(self.transient_command_pool, None) };
  }


  #[allow(unsafe_code)]
  fn destroy_replaceable(&mut self, device: &Device) {
    //   While the descriptor pool is also a pool, it has a preallocated size
    // which will be different next time. So, we destroy it all the way.
    unsafe { device.destroy_descriptor_pool(self.descriptor_pool, None) };

    //   Notice that, unlike the vertex and index buffers, we destroy and
    // re-create these on every reinitialization. That's because the number of
    // them depends on how many images the swapchain has.
    for buffer in &self.uniform_buffers {
      unsafe { device.destroy_buffer(*buffer, None) };
    }
    for memory in &self.uniform_buffer_memory {
      unsafe { device.free_memory(*memory, None) };
    }

    unsafe { device.destroy_image(self.color_image, None) };
    unsafe { device.free_memory(self.color_image_memory, None) };
    unsafe { device.destroy_image_view(self.color_image_view, None) };
    unsafe { device.destroy_image(self.depth_image, None) };
    unsafe { device.free_memory(self.depth_image_memory, None) };
    unsafe { device.destroy_image_view(self.depth_image_view, None) };

    for view in &self.swapchain.image_views {
      unsafe { device.destroy_image_view(*view, None) };
    }

    unsafe { device.destroy_swapchain_khr(self.swapchain.swapchain, None) };
  }
}


#[allow(unsafe_code)]
fn init_swapchain(window: &Window, instance: &Instance,
                  surface: &vk::SurfaceKHR,
                  physical_device: &vk::PhysicalDevice, device: &Device,
                  indices: &QueueFamilyIndices)
    -> Result<Swapchain>
{
  let (capabilities, formats, presentation_modes)
          = PermanentGraphicsState::find_device_swapchain_features(
                instance, surface, physical_device)?.require()?;

  let format = pick_surface_format(&formats)?;

  let presentation_mode
          = pick_presentation_mode(&presentation_modes)?;
  let extent = pick_image_extent(window, capabilities)?;

  let mut image_count = capabilities.min_image_count + 1;
  if capabilities.max_image_count != 0 {
    image_count
        = image_count.clamp(0, capabilities.max_image_count);
  }

  let mut unique_queue_family_indices = BTreeSet::new();
  unique_queue_family_indices.insert(indices.graphics);
  unique_queue_family_indices.insert(indices.presentation);

  //   If there's only one queue, we use exclusive sharing mode, which
  // will allow things to work without locks. Otherwise we use concurrent
  // mode.
  let (ordered_indices, sharing_mode)
          = if unique_queue_family_indices.len() < 2
  {
    (vec![indices.graphics], vk::SharingMode::EXCLUSIVE)
  } else {
    (vec![indices.graphics, indices.presentation],
     vk::SharingMode::CONCURRENT)
  };

  let swapchain_info = vk::SwapchainCreateInfoKHR::builder()
          .surface(*surface)
          .min_image_count(image_count)
          .image_format(format.format)
          .image_color_space(format.color_space)
          .image_extent(extent)
          .image_array_layers(1)
          .image_usage(vk::ImageUsageFlags::COLOR_ATTACHMENT)
          .image_sharing_mode(sharing_mode)
          .queue_family_indices(&ordered_indices)
          .pre_transform(capabilities.current_transform)
          .composite_alpha(vk::CompositeAlphaFlagsKHR::OPAQUE)
          .present_mode(presentation_mode)
          .clipped(true)
          .old_swapchain(vk::SwapchainKHR::null());

  let swapchain = unsafe {
    device.create_swapchain_khr(&swapchain_info, None)
  }?;

  let images = unsafe {
    device.get_swapchain_images_khr(swapchain)
  }?;

  let mut image_views = Vec::new();
  for image in &images {
    let view = init_image_view(device, image, 1, format.format,
                               vk::ImageAspectFlags::COLOR)?;
    image_views.push(view);
  }

  Ok(Swapchain {
    swapchain, images, image_views,
    format: format.format,
    extent
  })
}


#[allow(unsafe_code)]
fn init_color(instance: &Instance, physical_device: &vk::PhysicalDevice,
              device: &Device, extent: &vk::Extent2D,
              sample_count: vk::SampleCountFlags, format: vk::Format)
    -> Result<(vk::Image, vk::DeviceMemory, vk::ImageView)>
{
  let (image, image_memory)
          = allocate_image(instance, physical_device, device,
                           extent.width, extent.height, 1, sample_count,
                           format,
                           vk::ImageTiling::OPTIMAL,
                           vk::ImageUsageFlags::COLOR_ATTACHMENT
                           | vk::ImageUsageFlags::TRANSIENT_ATTACHMENT,
                           vk::MemoryPropertyFlags::DEVICE_LOCAL)?;

  let image_view = init_image_view(device, &image, 1, format,
                                   vk::ImageAspectFlags::COLOR)?;

  Ok((image, image_memory, image_view))
}


#[allow(unsafe_code)]
fn init_depth(instance: &Instance, physical_device: &vk::PhysicalDevice,
              device: &Device, extent: &vk::Extent2D,
              sample_count: vk::SampleCountFlags)
    -> Result<(vk::Image, vk::DeviceMemory, vk::ImageView, vk::Format)>
{
  let format = pick_depth_format(instance, physical_device)?;

  let (image, image_memory)
          = allocate_image(instance, physical_device, device,
                           extent.width, extent.height, 1, sample_count,
                           format,
                           vk::ImageTiling::OPTIMAL,
                           vk::ImageUsageFlags::DEPTH_STENCIL_ATTACHMENT,
                           vk::MemoryPropertyFlags::DEVICE_LOCAL)?;

  let image_view = init_image_view(device, &image, 1, format,
                                   vk::ImageAspectFlags::DEPTH)?;

  Ok((image, image_memory, image_view, format))
}


#[allow(unsafe_code)]
fn init_texture(instance: &Instance,
                physical_device: &vk::PhysicalDevice, device: &Device,
                queue: &vk::Queue, command_pool: &vk::CommandPool)
    -> Result<(vk::Image, vk::DeviceMemory, vk::ImageView, u32)>
{
  let png = include_bytes!("../../textures/forest_leaves_04_diff.png");

  let decoder = Decoder::new(Cursor::new(png));
  let mut reader = decoder.read_info()?;

  let (width, height) = reader.info().size();

  let format_properties = unsafe {
    instance.get_physical_device_format_properties(*physical_device,
                                                   vk::Format::R8G8B8A8_SRGB)
  };
  let has_linear_filter = format_properties
          .optimal_tiling_features
          .contains(vk::FormatFeatureFlags::SAMPLED_IMAGE_FILTER_LINEAR);
  let mip_count = if has_linear_filter {
    //   This will generate mips all the way down to 1x1. It is not clear
    // whether there's a benefit to that.
    (width.max(height)).ilog2() + 1
  } else {
    1
  };

  let mut pixels = vec![0; reader.info().raw_bytes()];
  reader.next_frame(&mut pixels)?;

  let (staging_buffer, staging_memory, _byte_size)
          = stage_in_buffer(instance, physical_device, device, &pixels)?;

  let (image, image_memory)
          = allocate_image(instance, physical_device, device,
                           width, height, mip_count, vk::SampleCountFlags::_1,
                           vk::Format::R8G8B8A8_SRGB,
                           vk::ImageTiling::OPTIMAL,
                           vk::ImageUsageFlags::SAMPLED
                               | vk::ImageUsageFlags::TRANSFER_SRC
                               | vk::ImageUsageFlags::TRANSFER_DST,
                           vk::MemoryPropertyFlags::DEVICE_LOCAL)?;

  change_image_layout(device, queue, command_pool, &image, mip_count,
                      vk::ImageLayout::UNDEFINED,
                      vk::ImageLayout::TRANSFER_DST_OPTIMAL)?;

  copy_buffer_to_image(device, queue, command_pool, &staging_buffer, &image,
                       width, height)?;

  // This will also change the layout to SHADER_READ_ONLY_OPTIMAL.

  fill_mip_levels(device, queue, command_pool, &image,
                  width, height, mip_count)?;

  let view = init_image_view(device, &image, mip_count,
                             vk::Format::R8G8B8A8_SRGB,
                             vk::ImageAspectFlags::COLOR)?;

  unsafe { device.destroy_buffer(staging_buffer, None) };
  unsafe { device.free_memory(staging_memory, None) };

  Ok((image, image_memory, view, mip_count))
}


fn init_uniform_buffers(instance: &Instance,
                        physical_device: &vk::PhysicalDevice, device: &Device,
                        count: usize)
    -> Result<(Vec<vk::Buffer>, Vec<vk::DeviceMemory>)>
{
  let mut buffers = Vec::new();
  let mut all_memory = Vec::new();

  for _ in 0 .. count {
    let (buffer, memory) = allocate_buffer(
            instance, physical_device, device,
            size_of::<UniformBlock<f32>>() as vk::DeviceSize,
            vk::BufferUsageFlags::UNIFORM_BUFFER,
            vk::MemoryPropertyFlags::HOST_COHERENT
            | vk::MemoryPropertyFlags::HOST_VISIBLE)?;
    buffers.push(buffer);
    all_memory.push(memory);
  }

  Ok((buffers, all_memory))
}


#[allow(unsafe_code)]
fn init_sampler(device: &Device, enable_anisotropy: &EnableAnisotropy,
                mip_count: u32)
    -> Result<vk::Sampler>
{
  let mut sampler_info = vk::SamplerCreateInfo::builder()
          .mag_filter(vk::Filter::LINEAR)
          .min_filter(vk::Filter::LINEAR)
          .address_mode_u(vk::SamplerAddressMode::REPEAT)
          .address_mode_v(vk::SamplerAddressMode::REPEAT)
          .address_mode_w(vk::SamplerAddressMode::REPEAT)
          .border_color(vk::BorderColor::INT_OPAQUE_BLACK)
          .unnormalized_coordinates(false)
          .compare_enable(false)
          .compare_op(vk::CompareOp::ALWAYS)
          .mipmap_mode(vk::SamplerMipmapMode::LINEAR)
          .mip_lod_bias(0.0)
          .min_lod(0.0)
          .max_lod(mip_count as f32);
  sampler_info = if enable_anisotropy.0 {
    sampler_info.anisotropy_enable(true)
                .max_anisotropy(16.0)
  } else {
    sampler_info.anisotropy_enable(false)
                .max_anisotropy(1.0)
  };

  let sampler = unsafe { device.create_sampler(&sampler_info, None) }?;

  Ok(sampler)
}


#[allow(unsafe_code)]
fn init_descriptor_pool(device: &Device, count: usize)
    -> Result<vk::DescriptorPool>
{
  let uniform_block_size = vk::DescriptorPoolSize::builder()
          .type_(vk::DescriptorType::UNIFORM_BUFFER)
          .descriptor_count(count as u32);

  let sampler_size = vk::DescriptorPoolSize::builder()
          .type_(vk::DescriptorType::COMBINED_IMAGE_SAMPLER)
          .descriptor_count(count as u32);

  let sizes = [uniform_block_size, sampler_size];
  let pool_info = vk::DescriptorPoolCreateInfo::builder()
          .pool_sizes(&sizes)
          .max_sets(count as u32);
  let pool = unsafe { device.create_descriptor_pool(&pool_info, None) }?;

  Ok(pool)
}


#[allow(unsafe_code)]
fn init_command_pools(device: &Device, indices: &QueueFamilyIndices)
    -> Result<(vk::CommandPool, vk::CommandPool)>
{
  let command_pool_info = vk::CommandPoolCreateInfo::builder()
          .flags(vk::CommandPoolCreateFlags::TRANSIENT
                 | vk::CommandPoolCreateFlags::RESET_COMMAND_BUFFER)
          .queue_family_index(indices.graphics);

  let primary = unsafe {
    device.create_command_pool(&command_pool_info, None)
  }?;

  command_pool_info.flags(vk::CommandPoolCreateFlags::TRANSIENT);
  let transient = unsafe {
    device.create_command_pool(&command_pool_info, None)
  }?;

  Ok((primary, transient))
}


#[allow(unsafe_code)]
fn init_concurrency(device: &Device,
                    swapchain_images: &Vec<vk::Image>)
    -> Result<Concurrency>
{
  let semaphore_info = vk::SemaphoreCreateInfo::builder();
  let fence_info = vk::FenceCreateInfo::builder()
                       .flags(vk::FenceCreateFlags::SIGNALED);

  let mut image_available_semaphores = Vec::new();
  let mut rendering_finished_semaphores = Vec::new();
  let mut frame_fences = Vec::new();
  for _ in 0 .. N_SIMULTANEOUS_FRAMES {
    image_available_semaphores.push(unsafe {
      device.create_semaphore(&semaphore_info, None)
    }?);

    rendering_finished_semaphores.push(unsafe {
      device.create_semaphore(&semaphore_info, None)
    }?);

    frame_fences.push(unsafe {
      device.create_fence(&fence_info, None)
    }?);
  }

  let mut image_fences = Vec::new();
  for _ in 0 .. swapchain_images.len() {
    image_fences.push(vk::Fence::null());
  }

  Ok(Concurrency {
    image_available_semaphores,
    rendering_finished_semaphores,
    frame_fences,
    image_fences: image_fences,
  })
}


#[allow(unsafe_code)]
fn init_image_view(device: &Device, image: &vk::Image, mip_count: u32,
                   format: vk::Format, aspects: vk::ImageAspectFlags)
    -> Result<vk::ImageView>
{
  //   Component mapping is only for color components (not, for example, depth
  // or stencil components), so we always just want it like this.
  let components = vk::ComponentMapping::builder()
          .r(vk::ComponentSwizzle::IDENTITY)
          .g(vk::ComponentSwizzle::IDENTITY)
          .b(vk::ComponentSwizzle::IDENTITY)
          .a(vk::ComponentSwizzle::IDENTITY);

  let subresource_range = vk::ImageSubresourceRange::builder()
          .aspect_mask(aspects)
          .base_mip_level(0)
          .level_count(mip_count)
          .base_array_layer(0)
          .layer_count(1);

  let view_info = vk::ImageViewCreateInfo::builder()
          .image(*image)
          .view_type(vk::ImageViewType::_2D)
          .format(format)
          .components(components)
          .subresource_range(subresource_range);

  let view = unsafe {
    device.create_image_view(&view_info, None)
  }?;

  Ok(view)
}


fn pick_surface_format(available_formats: &Vec<vk::SurfaceFormatKHR>)
    -> Result<vk::SurfaceFormatKHR>
{
  for format in available_formats {
    if format.format == vk::Format::B8G8R8A8_SRGB
       && format.color_space == vk::ColorSpaceKHR::SRGB_NONLINEAR
    {
      return Ok(format.clone());
    }
  }

  return Ok(available_formats[0].clone());
}


#[allow(unsafe_code)]
fn pick_depth_format(instance: &Instance,
                     physical_device: &vk::PhysicalDevice)
    -> Result<vk::Format>
{
  let required_features = vk::FormatFeatureFlags::DEPTH_STENCIL_ATTACHMENT;

  for format in [vk::Format::D32_SFLOAT,
                 vk::Format::D32_SFLOAT_S8_UINT,
                 vk::Format::D24_UNORM_S8_UINT]
  {
    let properties = unsafe {
      instance.get_physical_device_format_properties(
                   *physical_device, format)
    };

    if properties.optimal_tiling_features.contains(required_features) {
      return Ok(format);
    }
  }

  Err(Error {
    message: "There is no supported depth-buffer sample format.".to_string()
  })
}


fn pick_presentation_mode(_available_modes: &Vec<vk::PresentModeKHR>)
    -> Result<vk::PresentModeKHR>
{
  // It's guaranteed to have this one.
  return Ok(vk::PresentModeKHR::FIFO);
}


fn pick_image_extent(window: &Window,
                     capabilities: vk::SurfaceCapabilitiesKHR)
    -> Result<vk::Extent2D>
{
  if capabilities.current_extent.width != u32::MAX
     && capabilities.current_extent.height != u32::MAX
  {
    Ok(capabilities.current_extent)
  } else {
    let window_size = window.inner_size();

    let width = window_size.width
                           .clamp(capabilities.min_image_extent.width,
                                  capabilities.max_image_extent.width);
    let height = window_size.height
                            .clamp(capabilities.min_image_extent.height,
                                   capabilities.max_image_extent.height);

    Ok(vk::Extent2D::builder().width(width).height(height).build())
  }
}


#[allow(unsafe_code)]
fn allocate_image(instance: &Instance, physical_device: &vk::PhysicalDevice,
                  device: &Device, width: u32, height: u32, mip_count: u32,
                  sample_count: vk::SampleCountFlags, format: vk::Format,
                  tiling: vk::ImageTiling, usage: vk::ImageUsageFlags,
                  memory_flags: vk::MemoryPropertyFlags)
    -> Result<(vk::Image, vk::DeviceMemory)>
{
  let image_info = vk::ImageCreateInfo::builder()
          .image_type(vk::ImageType::_2D)
          .extent(vk::Extent3D { width, height, depth: 1 })
          .mip_levels(mip_count)
          .samples(sample_count)
          .array_layers(1)
          .format(format)
          .tiling(tiling)
          .initial_layout(vk::ImageLayout::UNDEFINED)
          .usage(usage)
          .sharing_mode(vk::SharingMode::EXCLUSIVE)
          .flags(vk::ImageCreateFlags::empty());
  let image = unsafe { device.create_image(&image_info, None) }?;

  let requirements = unsafe { device.get_image_memory_requirements(image) };

  let type_index = pick_memory_type(instance, physical_device,
                                    &memory_flags, &requirements)?;

  let image_memory_info = vk::MemoryAllocateInfo::builder()
          .allocation_size(requirements.size)
          .memory_type_index(type_index);
  let image_memory = unsafe {
    device.allocate_memory(&image_memory_info, None)
  }?;

  unsafe { device.bind_image_memory(image, image_memory, 0) }?;

  Ok((image, image_memory))
}


#[allow(unsafe_code)]
fn copy_buffer_to_image(device: &Device, queue: &vk::Queue,
                        command_pool: &vk::CommandPool, source: &vk::Buffer,
                        destination: &vk::Image, width: u32, height: u32)
    -> Result<()>
{
  let command_buffer = begin_transient_commands(device, command_pool)?;

  let subresource_layers = vk::ImageSubresourceLayers::builder()
          .aspect_mask(vk::ImageAspectFlags::COLOR)
          .mip_level(0)
          .base_array_layer(0)
          .layer_count(1);

  let copy_info = vk::BufferImageCopy::builder()
          .buffer_offset(0)
          .buffer_row_length(0)
          .buffer_image_height(0)
          .image_subresource(subresource_layers)
          .image_offset(vk::Offset3D { x: 0, y: 0, z: 0 })
          .image_extent(vk::Extent3D { width, height, depth: 1 });

  unsafe {
    device.cmd_copy_buffer_to_image(command_buffer, *source, *destination,
                                    vk::ImageLayout::TRANSFER_DST_OPTIMAL,
                                    &[copy_info])
  };

  end_transient_commands(command_buffer, device, queue, command_pool)?;

  Ok(())
}


#[allow(unsafe_code)]
fn change_image_layout(device: &Device, queue: &vk::Queue,
                       command_pool: &vk::CommandPool, image: &vk::Image,
                       mip_count: u32, old: vk::ImageLayout,
                       new: vk::ImageLayout)
    -> Result<()>
{
  let command_buffer = begin_transient_commands(device, command_pool)?;

  let subresource_range = vk::ImageSubresourceRange::builder()
          .aspect_mask(vk::ImageAspectFlags::COLOR)
          .base_mip_level(0)
          .level_count(mip_count)
          .base_array_layer(0)
          .layer_count(1);

  //   Notionally this is a property that our caller is in a better position
  // to know than we are, but in practice the nature of the transition
  // strongly implies a particular phase of the image's lifecycle, so we just
  // compute it here.
  let (source_access, source_stage, destination_access, destination_stage)
          = match (old, new)
  {
    (vk::ImageLayout::UNDEFINED, vk::ImageLayout::TRANSFER_DST_OPTIMAL)
        => (vk::AccessFlags::empty(),
            vk::PipelineStageFlags::TOP_OF_PIPE,
            vk::AccessFlags::TRANSFER_WRITE,
            vk::PipelineStageFlags::TRANSFER),
    (vk::ImageLayout::TRANSFER_DST_OPTIMAL,
     vk::ImageLayout::SHADER_READ_ONLY_OPTIMAL)
        => (vk::AccessFlags::TRANSFER_WRITE,
            vk::PipelineStageFlags::TRANSFER,
            vk::AccessFlags::SHADER_READ,
            vk::PipelineStageFlags::FRAGMENT_SHADER),
    _ => return Err(Error {
      message:
          format!("Don't know how to change from image layout {:?} to {:?}",
                  old, new)
    })
  };

  let barrier_info = vk::ImageMemoryBarrier::builder()
        .image(*image)
        .subresource_range(subresource_range)
        .old_layout(old)
        .new_layout(new)
        .src_queue_family_index(vk::QUEUE_FAMILY_IGNORED)
        .dst_queue_family_index(vk::QUEUE_FAMILY_IGNORED)
        .src_access_mask(source_access)
        .dst_access_mask(destination_access);

  unsafe {
    device.cmd_pipeline_barrier(command_buffer,
                                source_stage, destination_stage,
                                vk::DependencyFlags::empty(),
                                &[] as &[vk::MemoryBarrier],
                                &[] as &[vk::BufferMemoryBarrier],
                                &[barrier_info])
  };

  end_transient_commands(command_buffer, device, queue, command_pool)?;

  Ok(())
}


//   An Image can store multiple mip levels within it, as one of several kinds
// of subresource it has. We deal with this by
#[allow(unsafe_code)]
fn fill_mip_levels(device: &Device, queue: &vk::Queue,
                   command_pool: &vk::CommandPool, image: &vk::Image,
                   original_width: u32, original_height: u32,
                   mip_count: u32)
    -> Result<()>
{
  let command_buffer = begin_transient_commands(device, command_pool)?;

  //   We'll be mutating these two builders as we loop through the mip levels,
  // because we need to construct a lot of similar things. Remember, the
  // builder methods don't mutate in-place, they return a new builder; to
  // avoid confusion we always assign that result back to the same variable.
  let mut barrier_subresource_range = vk::ImageSubresourceRange::builder()
          .aspect_mask(vk::ImageAspectFlags::COLOR)
          .level_count(1)
          .base_array_layer(0)
          .layer_count(1);

  let mut blit_barrier_info = vk::ImageMemoryBarrier::builder()
          .image(*image)
          .src_queue_family_index(vk::QUEUE_FAMILY_IGNORED)
          .dst_queue_family_index(vk::QUEUE_FAMILY_IGNORED);

  //   Now we loop through the mip levels from largest (low numbers) to
  // smallest (high numbers). Conceptually, the only thing we're doing is a
  // blit that copies each mip level from the one immediately before. Recall
  // though that we don't just want to fill in the pixels, we also care about
  // pixel format and memory sharing. There are additional operations to deal
  // with that. These are best done together, as detailed below.
  //
  //   This loop has a lot of code in it, so we make the "paragraphs" a little
  // more dense than usual to make sure the logical grouping is clear.
  let mut source_width = original_width;
  let mut source_height = original_height;
  for destination_mip_level in 1 .. mip_count {
    let source_mip_level = destination_mip_level - 1;
    let destination_width = (source_width / 2).max(1);
    let destination_height = (source_height / 2).max(1);

    //   So. The name pipeline_barrier is a little misleading; it does indeed
    // mean "barrier" in the concurrency sense, but it isn't just initiating
    // a wait, it's also performing any needed mutation. We do one of them
    // here, acting on this iteration's source level, to set it up for
    // reading.
    barrier_subresource_range = barrier_subresource_range
        .base_mip_level(source_mip_level as u32);
    blit_barrier_info = blit_barrier_info
        .old_layout(vk::ImageLayout::TRANSFER_DST_OPTIMAL)
        .new_layout(vk::ImageLayout::TRANSFER_SRC_OPTIMAL)
        .src_access_mask(vk::AccessFlags::TRANSFER_WRITE)
        .dst_access_mask(vk::AccessFlags::TRANSFER_READ)
        .subresource_range(barrier_subresource_range);
    unsafe {
      device.cmd_pipeline_barrier(command_buffer,
                                  vk::PipelineStageFlags::TRANSFER,
                                  vk::PipelineStageFlags::TRANSFER,
                                  vk::DependencyFlags::empty(),
                                  &[] as &[vk::MemoryBarrier],
                                  &[] as &[vk::BufferMemoryBarrier],
                                  &[blit_barrier_info])
    };

    //   Now we do the actual blit. Nice and easy, though specifying the
    // coordinates is a bit verbose.
    let blit_source_layer_info = vk::ImageSubresourceLayers::builder()
            .aspect_mask(vk::ImageAspectFlags::COLOR)
            .mip_level(source_mip_level as u32)
            .base_array_layer(0)
            .layer_count(1);
    let blit_destination_layer_info = vk::ImageSubresourceLayers::builder()
            .aspect_mask(vk::ImageAspectFlags::COLOR)
            .mip_level(destination_mip_level as u32)
            .base_array_layer(0)
            .layer_count(1);
    let blit_info = vk::ImageBlit::builder()
            .src_offsets([vk::Offset3D { x: 0, y: 0, z: 0 },
                          vk::Offset3D {
                            x: source_width as i32,
                            y: source_height as i32,
                            z: 1
                          }])
            .src_subresource(blit_source_layer_info)
            .dst_offsets([vk::Offset3D { x: 0, y: 0, z: 0 },
                          vk::Offset3D {
                            x: destination_width as i32,
                            y: destination_height as i32,
                            z: 1
                          }])
            .dst_subresource(blit_destination_layer_info);
    unsafe {
      device.cmd_blit_image(command_buffer,
                            *image, vk::ImageLayout::TRANSFER_SRC_OPTIMAL,
                            *image, vk::ImageLayout::TRANSFER_DST_OPTIMAL,
                            &[blit_info],
                            vk::Filter::LINEAR)
    };

    //   Now we do another pipeline_barrier. We're still acting on this
    // iteration's source level, not on the destination. We'll never need to
    // use it again except from the shader, so we set it appropriately for
    // that.
    blit_barrier_info = blit_barrier_info
        .old_layout(vk::ImageLayout::TRANSFER_SRC_OPTIMAL)
        .new_layout(vk::ImageLayout::SHADER_READ_ONLY_OPTIMAL)
        .src_access_mask(vk::AccessFlags::TRANSFER_READ)
        .dst_access_mask(vk::AccessFlags::SHADER_READ);
    unsafe {
      device.cmd_pipeline_barrier(command_buffer,
                                  vk::PipelineStageFlags::TRANSFER,
                                  vk::PipelineStageFlags::FRAGMENT_SHADER,
                                  vk::DependencyFlags::empty(),
                                  &[] as &[vk::MemoryBarrier],
                                  &[] as &[vk::BufferMemoryBarrier],
                                  &[blit_barrier_info])
    };

    source_width = destination_width;
    source_height = destination_height;
  }

  let final_mip_level = mip_count - 1;

  //   We need to do one final pipeline_barrier, because the loop didn't do it
  // to the smallest (last) mip level. We change it to have the same settings
  // the loop left the rest of them in. The barrier source properties for this
  // barrier are different from the others because this level was never useds
  // as a blit source, only as a blit destination. The barrier destination
  // properties are the same as the rest, so after this all the subresourcess
  // will be in their fully-ready state.
  barrier_subresource_range = barrier_subresource_range
      .base_mip_level(final_mip_level as u32);
  blit_barrier_info = blit_barrier_info
      .old_layout(vk::ImageLayout::TRANSFER_DST_OPTIMAL)
      .new_layout(vk::ImageLayout::SHADER_READ_ONLY_OPTIMAL)
      .src_access_mask(vk::AccessFlags::TRANSFER_WRITE)
      .dst_access_mask(vk::AccessFlags::SHADER_READ)
      .subresource_range(barrier_subresource_range);
  unsafe {
    device.cmd_pipeline_barrier(command_buffer,
                                vk::PipelineStageFlags::TRANSFER,
                                vk::PipelineStageFlags::FRAGMENT_SHADER,
                                vk::DependencyFlags::empty(),
                                &[] as &[vk::MemoryBarrier],
                                &[] as &[vk::BufferMemoryBarrier],
                                &[blit_barrier_info])
  };

  end_transient_commands(command_buffer, device, queue, command_pool)?;

  Ok(())
}