#include "gpu.h"

#include <assert.h>
#include <stdio.h>
#include <string.h>

#define QUEUE_FAMILY_PROP_BUFFER_SIZE 16
#define EXT_PROP_BUFFER_SIZE 256
#define SURFACE_FMT_BUFFER_SIZE 64
#define PRESENT_MODE_BUFFER_SIZE 8
#define SWAPCHAIN_IMG_BUFFER_SIZE 4

#define NELEMS(arr) (sizeof(arr) / sizeof(arr[0]))

#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define CLAMP(x, min, max) MAX((min), MIN(x, (max)))

static const char *layers[] = { "VK_LAYER_KHRONOS_validation" };
static const char *extensions[] = {
	VK_KHR_SWAPCHAIN_EXTENSION_NAME,
};

static constexpr size_t extension_count = NELEMS(extensions);

static const char vert_shader[] = {
#embed VERT_SHADER_SPV
};

static const char frag_shader[] = {
#embed FRAG_SHADER_SPV
};

static VkShaderModule
load_shader_module(VkDevice dev, const char *code, size_t code_size)
{
	const VkShaderModuleCreateInfo create_info = {
		.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
		.codeSize = code_size,
		.pCode = (uint32_t *)code,
	};
	VkShaderModule mod;
	const VkResult result
	    = vkCreateShaderModule(dev, &create_info, nullptr, &mod);
	assert(result == VK_SUCCESS);

	return mod;
}

static void alloc_dev_mem(
    const gpu_ctx_t *ctx, VkMemoryRequirements reqs,
    VkMemoryPropertyFlags props, VkDeviceMemory *mem_out)
{
	// Find appropriate memory type
	VkPhysicalDeviceMemoryProperties mem_props;
	vkGetPhysicalDeviceMemoryProperties(ctx->physical_dev, &mem_props);
	uint32_t mem_type = UINT32_MAX;
	for (unsigned i = 0; i < mem_props.memoryTypeCount; ++i) {
		if ((reqs.memoryTypeBits & (1 << i)) == 0)
			continue;
		if ((mem_props.memoryTypes[i].propertyFlags & props) == props) {
			mem_type = i;
			break;
		}
	}
	assert(mem_type != UINT32_MAX);

	// Allocate memory
	VkMemoryAllocateInfo alloc_config = {
		.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
		.allocationSize = reqs.size,
		.memoryTypeIndex = mem_type,
	};
	const VkResult result
	    = vkAllocateMemory(ctx->dev, &alloc_config, nullptr, mem_out);
	assert(result == VK_SUCCESS);
}

static void create_buffer(
    const gpu_ctx_t *ctx, VkDeviceSize size, VkBufferUsageFlags usage,
    VkMemoryPropertyFlags props, VkBuffer *buf_out, VkDeviceMemory *mem_out)
{
	// Create buffer
	const VkBufferCreateInfo buf_config = {
		.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
		.size = size,
		.usage = usage,
		.sharingMode = VK_SHARING_MODE_EXCLUSIVE,
	};
	const VkResult result
	    = vkCreateBuffer(ctx->dev, &buf_config, nullptr, buf_out);
	assert(result == VK_SUCCESS);

	// Allocate and bind memory
	VkMemoryRequirements mem_reqs;
	vkGetBufferMemoryRequirements(ctx->dev, *buf_out, &mem_reqs);
	alloc_dev_mem(ctx, mem_reqs, props, mem_out);
	vkBindBufferMemory(ctx->dev, *buf_out, *mem_out, 0);
}

static void load_buffer(
    const gpu_ctx_t *ctx, VkBuffer dest_buf, const void *data, size_t size,
    VkBuffer staging_buf, VkDeviceMemory staging_buf_mem,
    VkCommandPool cmd_pool)
{
	VkResult result;

	// Copy data to staging buffer
	void *staging_buf_ptr;
	vkMapMemory(ctx->dev, staging_buf_mem, 0, size, 0, &staging_buf_ptr);
	memcpy(staging_buf_ptr, data, size);
	vkUnmapMemory(ctx->dev, staging_buf_mem);

	// Allocate command buffer
	const VkCommandBufferAllocateInfo cmd_buf_config = {
		.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
		.commandPool = cmd_pool,
		.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
		.commandBufferCount = 1,
	};
	VkCommandBuffer cmd_buf;
	result = vkAllocateCommandBuffers(ctx->dev, &cmd_buf_config, &cmd_buf);
	assert(result == VK_SUCCESS);

	// Record copy command
	VkCommandBufferBeginInfo cmd_begin = {
		.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
		.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
	};
	vkBeginCommandBuffer(cmd_buf, &cmd_begin);
	VkBufferCopy cmd_copy = { .size = size };
	vkCmdCopyBuffer(cmd_buf, staging_buf, dest_buf, 1, &cmd_copy);
	vkEndCommandBuffer(cmd_buf);

	// Send copy command
	VkSubmitInfo copy_submit = {
		.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
		.commandBufferCount = 1,
		.pCommandBuffers = &cmd_buf,
	};
	vkQueueSubmit(ctx->queue, 1, &copy_submit, VK_NULL_HANDLE);
	vkQueueWaitIdle(ctx->queue);

	vkFreeCommandBuffers(ctx->dev, cmd_pool, 1, &cmd_buf);
}

void gpu_init(gpu_ctx_t *ctx, const gpu_config_t *config, model_t model)
{
	// Initialise Vulkan instance.
	const VkApplicationInfo app_info = {
		.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO,
		.pApplicationName = config->application_name,
		.applicationVersion = VK_MAKE_VERSION(1, 0, 0),
		.pEngineName = "No Engine",
		.engineVersion = VK_MAKE_VERSION(1, 0, 0),
		.apiVersion = VK_API_VERSION_1_0,
	};
	uint32_t glfw_ext_count = 0;
	const char **glfw_ext_names
	    = glfwGetRequiredInstanceExtensions(&glfw_ext_count);
	const VkInstanceCreateInfo inst_create_info = {
		.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
		.pApplicationInfo = &app_info,
		.enabledExtensionCount = glfw_ext_count,
		.ppEnabledExtensionNames = glfw_ext_names,
		.enabledLayerCount = sizeof(layers) / sizeof(layers[0]),
		.ppEnabledLayerNames = layers,
	};
	VkResult result
	    = vkCreateInstance(&inst_create_info, nullptr, &ctx->inst);
	assert(result == VK_SUCCESS);

	// Create surface
	result = glfwCreateWindowSurface(
	    ctx->inst, config->window, nullptr, &ctx->surface);
	assert(result == VK_SUCCESS);

	// Select a physical device
	uint32_t dev_count = 1;
	result = vkEnumeratePhysicalDevices(
	    ctx->inst, &dev_count, &ctx->physical_dev);
	assert(result == VK_SUCCESS || result == VK_INCOMPLETE);

	// Select queue family
	uint32_t queue_family_count = QUEUE_FAMILY_PROP_BUFFER_SIZE;
	VkQueueFamilyProperties
	    queue_family_props[QUEUE_FAMILY_PROP_BUFFER_SIZE];
	vkGetPhysicalDeviceQueueFamilyProperties(
	    ctx->physical_dev, &queue_family_count, queue_family_props);
	uint32_t queue_family;
	bool found_queue_family = false;
	for (unsigned i = 0; i < queue_family_count; ++i) {
		if (queue_family_props[i].queueFlags & VK_QUEUE_GRAPHICS_BIT) {
			VkBool32 surface_support = false;
			vkGetPhysicalDeviceSurfaceSupportKHR(
			    ctx->physical_dev, i, ctx->surface, &surface_support);
			if (surface_support) {
				found_queue_family = true;
				queue_family = i;
				break;
			}
		}
	}
	assert(found_queue_family);

	// Check that the physical device supports all extensions
	uint32_t ext_count = EXT_PROP_BUFFER_SIZE;
	VkExtensionProperties ext_props[EXT_PROP_BUFFER_SIZE];
	result = vkEnumerateDeviceExtensionProperties(
	    ctx->physical_dev, nullptr, &ext_count, ext_props);
	assert(result == VK_SUCCESS);
	bool supported[extension_count] = {};
	for (unsigned i = 0; i < ext_count; ++i) {
		for (unsigned j = 0; j < extension_count; ++j) {
			if (strcmp(extensions[j], ext_props[i].extensionName) == 0) {
				supported[j] = true;
				break;
			}
		}
	}
	for (unsigned j = 0; j < extension_count; ++j) {
		if (!supported[j]) {
			fprintf(stderr, "Unsupported: %s\n", extensions[j]);
			assert(false);
		}
	}

	// Create the logical device and get the queue handle.
	float queue_priority = 1.0;
	const VkDeviceQueueCreateInfo queue_create_info = {
		.sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
		.queueFamilyIndex = queue_family,
		.queueCount = 1,
		.pQueuePriorities = &queue_priority,
	};
	const VkDeviceCreateInfo dev_create_info = {
		.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
		.pQueueCreateInfos = &queue_create_info,
		.queueCreateInfoCount = 1,
		.enabledExtensionCount = extension_count,
		.ppEnabledExtensionNames = extensions,
	};
	result = vkCreateDevice(
	    ctx->physical_dev, &dev_create_info, nullptr, &ctx->dev);
	assert(result == VK_SUCCESS);
	vkGetDeviceQueue(ctx->dev, queue_family, 0, &ctx->queue);

	// Select a surface format, preferring R8G8B8A8_SRGB with
	// SRGB_NONLINEAR colour space.
	uint32_t surface_fmt_count = SURFACE_FMT_BUFFER_SIZE;
	VkSurfaceFormatKHR surface_fmts[SURFACE_FMT_BUFFER_SIZE];
	result = vkGetPhysicalDeviceSurfaceFormatsKHR(
	    ctx->physical_dev, ctx->surface, &surface_fmt_count, surface_fmts);
	assert(result == VK_SUCCESS);
	assert(surface_fmt_count > 0);
	VkSurfaceFormatKHR surface_fmt = surface_fmts[0];
	for (unsigned i = 0; i < surface_fmt_count; ++i) {
		if (surface_fmts[i].format == VK_FORMAT_R8G8B8A8_SRGB
		    && surface_fmts[i].colorSpace
		        == VK_COLOR_SPACE_SRGB_NONLINEAR_KHR) {
			surface_fmt = surface_fmts[i];
		}
	}

	// Query surface capabilities
	VkSurfaceCapabilitiesKHR surface_caps;
	result = vkGetPhysicalDeviceSurfaceCapabilitiesKHR(
	    ctx->physical_dev, ctx->surface, &surface_caps);
	assert(result == VK_SUCCESS);

	// Select swap chain extent
	if (surface_caps.currentExtent.width == UINT32_MAX) {
		int width, height;
		glfwGetFramebufferSize(config->window, &width, &height);
		ctx->swapchain_extent.width = CLAMP(
		    (uint32_t)width, surface_caps.minImageExtent.width,
		    surface_caps.maxImageExtent.width);
		ctx->swapchain_extent.height = CLAMP(
		    (uint32_t)height, surface_caps.minImageExtent.height,
		    surface_caps.maxImageExtent.height);
	} else {
		ctx->swapchain_extent = surface_caps.currentExtent;
	}

	// Select image count
	uint32_t swapchain_img_count = surface_caps.minImageCount + 1;
	assert(swapchain_img_count <= SWAPCHAIN_IMG_BUFFER_SIZE);
	assert(
	    surface_caps.maxImageCount == 0
	    || swapchain_img_count <= surface_caps.maxImageCount);

	// Create the swap chain
	const VkSwapchainCreateInfoKHR swapchain_create_info = {
		.sType = VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR,
		.surface = ctx->surface,
		.minImageCount = swapchain_img_count,
		.imageFormat = surface_fmt.format,
		.imageColorSpace = surface_fmt.colorSpace,
		.imageExtent = ctx->swapchain_extent,
		.imageArrayLayers = 1,
		.imageUsage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
		.imageSharingMode = VK_SHARING_MODE_EXCLUSIVE,
		.preTransform = surface_caps.currentTransform,
		.compositeAlpha = VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR,
		.presentMode = VK_PRESENT_MODE_FIFO_KHR,
		.clipped = VK_TRUE,
		.oldSwapchain = VK_NULL_HANDLE,
	};
	result = vkCreateSwapchainKHR(
	    ctx->dev, &swapchain_create_info, nullptr, &ctx->swapchain);
	assert(result == VK_SUCCESS);

	// Retreive swapchain images
	VkImage swapchain_imgs[SWAPCHAIN_IMG_BUFFER_SIZE];
	ctx->swapchain_img_count = SWAPCHAIN_IMG_BUFFER_SIZE;
	result = vkGetSwapchainImagesKHR(
	    ctx->dev, ctx->swapchain, &swapchain_img_count, swapchain_imgs);
	assert(result == VK_SUCCESS);

	// Create swapchain image views
	VkImageViewCreateInfo img_view_create_info = {
		.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
		.viewType = VK_IMAGE_VIEW_TYPE_2D,
		.format = surface_fmt.format,
		.components = {
			.r = VK_COMPONENT_SWIZZLE_IDENTITY,
			.g = VK_COMPONENT_SWIZZLE_IDENTITY,
			.b = VK_COMPONENT_SWIZZLE_IDENTITY,
			.a = VK_COMPONENT_SWIZZLE_IDENTITY,
		},
		.subresourceRange = {
			.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
			.baseMipLevel = 0,
			.levelCount = 1,
			.baseArrayLayer = 0,
			.layerCount = 1,
		},
	};
	for (unsigned i = 0; i < swapchain_img_count; ++i) {
		img_view_create_info.image = swapchain_imgs[i];
		result = vkCreateImageView(
		    ctx->dev, &img_view_create_info, nullptr,
		    ctx->swapchain_img_views + i);
		assert(result == VK_SUCCESS);
	}

	// Set up depth image
	const VkFormat depth_img_fmt = VK_FORMAT_D32_SFLOAT;
	const VkImageCreateInfo depth_img_config = {
		.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
		.imageType = VK_IMAGE_TYPE_2D,
		.extent = {
			.width = ctx->swapchain_extent.width,
			.height = ctx->swapchain_extent.height,
			.depth = 1,
		},
		.mipLevels = 1,
		.arrayLayers = 1,
		.format = depth_img_fmt,
		.tiling = VK_IMAGE_TILING_OPTIMAL,
		.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
		.usage = VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT,
		.samples = VK_SAMPLE_COUNT_1_BIT,
		.sharingMode = VK_SHARING_MODE_EXCLUSIVE,
	};
	result = vkCreateImage(
	    ctx->dev, &depth_img_config, nullptr, &ctx->depth_img);
	assert(result == VK_SUCCESS);
	VkMemoryRequirements mem_reqs;
	vkGetImageMemoryRequirements(ctx->dev, ctx->depth_img, &mem_reqs);
	alloc_dev_mem(
	    ctx, mem_reqs, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
	    &ctx->depth_img_mem);
	vkBindImageMemory(ctx->dev, ctx->depth_img, ctx->depth_img_mem, 0);

	// Create depth image view
	const VkImageViewCreateInfo depth_img_view_config = {
		.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
		.image = ctx->depth_img,
		.viewType = VK_IMAGE_VIEW_TYPE_2D,
		.format = depth_img_fmt,
		.subresourceRange = {
			.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT,
			.levelCount = 1,
			.layerCount = 1,
		},
	};
	result = vkCreateImageView(
	    ctx->dev, &depth_img_view_config, nullptr, &ctx->depth_img_view);
	assert(result == VK_SUCCESS);

	// Load vertex and fragment shader modules
	ctx->vert_shader
	    = load_shader_module(ctx->dev, vert_shader, sizeof(vert_shader));
	ctx->frag_shader
	    = load_shader_module(ctx->dev, frag_shader, sizeof(frag_shader));
	const VkPipelineShaderStageCreateInfo shader_stages[] = {
		{
		    .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
		    .stage = VK_SHADER_STAGE_VERTEX_BIT,
		    .module = ctx->vert_shader,
		    .pName = "main",
		},
		{
		    .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
		    .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
		    .module = ctx->frag_shader,
		    .pName = "main",
		},
	};

	// Create descriptor set layout
	const VkDescriptorSetLayoutCreateInfo desc_set_layout_config = {
		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
		.bindingCount = 1,
		.pBindings = &(VkDescriptorSetLayoutBinding) {
			.binding = 0,
			.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
			.descriptorCount = 1,
			.stageFlags = VK_SHADER_STAGE_VERTEX_BIT,
		},
	};
	result = vkCreateDescriptorSetLayout(
	    ctx->dev, &desc_set_layout_config, nullptr, &ctx->desc_set_layout);
	assert(result == VK_SUCCESS);

	// Create descriptor pool
	const VkDescriptorPoolCreateInfo desc_pool_config = {
		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
		.poolSizeCount = 1,
		.pPoolSizes = &(VkDescriptorPoolSize) {
			.type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
			.descriptorCount = swapchain_img_count,
		},
		.maxSets = swapchain_img_count,
	};
	result = vkCreateDescriptorPool(
	    ctx->dev, &desc_pool_config, nullptr, &ctx->desc_pool);
	assert(result == VK_SUCCESS);

	// Create and configure descriptor sets
	VkDescriptorSetLayout desc_set_layouts[swapchain_img_count];
	for (unsigned i = 0; i < swapchain_img_count; ++i)
		desc_set_layouts[i] = ctx->desc_set_layout;
	const VkDescriptorSetAllocateInfo desc_sets_alloc_config = {
		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
		.descriptorPool = ctx->desc_pool,
		.descriptorSetCount = swapchain_img_count,
		.pSetLayouts = desc_set_layouts,
	};
	result = vkAllocateDescriptorSets(
	    ctx->dev, &desc_sets_alloc_config, ctx->desc_sets);
	assert(result == VK_SUCCESS);

	// Create and map MVP buffer
	const size_t mvp_buffer_size = swapchain_img_count * sizeof(mvp_t);
	create_buffer(
	    ctx, mvp_buffer_size, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT,
	    VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
	        | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
	    &ctx->mvp_buffer, &ctx->mvp_buffer_mem);
	vkMapMemory(
	    ctx->dev, ctx->mvp_buffer_mem, 0, mvp_buffer_size, 0,
	    (void **)&ctx->mvps);

	// Configure descriptor sets to use MVP buffer
	for (unsigned i = 0; i < swapchain_img_count; ++i) {
		const VkWriteDescriptorSet desc_write = {
			.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
			.dstSet = ctx->desc_sets[i],
			.dstBinding = 0,
			.dstArrayElement = 0,
			.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
			.descriptorCount = 1,
			.pBufferInfo = &(VkDescriptorBufferInfo) {
				.buffer = ctx->mvp_buffer,
				.offset = i * sizeof(mvp_t),
				.range = sizeof(mvp_t),
			},
		};
		vkUpdateDescriptorSets(ctx->dev, 1, &desc_write, 0, nullptr);
	}

	// Set up input configuration
	const VkVertexInputBindingDescription input_binding_descs[] = {
		{
		    .binding = 0,
		    .stride = sizeof(vertex_t),
		    .inputRate = VK_VERTEX_INPUT_RATE_VERTEX,
		},
	};
	const VkVertexInputAttributeDescription input_attribute_descs[] = {
		{
		    .binding = 0,
		    .location = 0,
		    .format = VK_FORMAT_R32G32B32_SFLOAT,
		    .offset = offsetof(vertex_t, position),
		},
		{
		    .binding = 0,
		    .location = 1,
		    .format = VK_FORMAT_R32G32B32_SFLOAT,
		    .offset = offsetof(vertex_t, normal),
		},
	};
	const VkPipelineVertexInputStateCreateInfo vertex_input_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
		.vertexBindingDescriptionCount = NELEMS(input_binding_descs),
		.pVertexBindingDescriptions = input_binding_descs,
		.vertexAttributeDescriptionCount = NELEMS(input_attribute_descs),
		.pVertexAttributeDescriptions = input_attribute_descs,
	};
	const VkPipelineInputAssemblyStateCreateInfo input_assembly_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
		.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST,
	};

	// Set up viewport configuration
	const VkViewport viewport = {
		.width = (float)ctx->swapchain_extent.width,
		.height = (float)ctx->swapchain_extent.height,
		.minDepth = 0.0,
		.maxDepth = 1.0,
	};
	const VkRect2D scissor = { .extent = ctx->swapchain_extent };
	VkPipelineViewportStateCreateInfo viewport_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
		.viewportCount = 1,
		.pViewports = &viewport,
		.scissorCount = 1,
		.pScissors = &scissor,
	};

	// Populate rasterizer configuration
	const VkPipelineRasterizationStateCreateInfo rasterizer_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
		.depthClampEnable = VK_FALSE,
		.rasterizerDiscardEnable = VK_FALSE,
		.polygonMode = VK_POLYGON_MODE_FILL,
		.lineWidth = 1.0f,
		.cullMode = VK_CULL_MODE_NONE,
	};

	// Populate multisampling configuration
	const VkPipelineMultisampleStateCreateInfo multisampling_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
		.sampleShadingEnable = VK_FALSE,
		.rasterizationSamples = VK_SAMPLE_COUNT_1_BIT,
	};

	// Set up colour blending configuration
	const VkPipelineColorBlendAttachmentState colour_blend_attatchment = {
		.colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT
		    | VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT,
		.blendEnable = VK_FALSE,
	};
	const VkPipelineColorBlendStateCreateInfo colour_blend_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
		.logicOpEnable = VK_FALSE,
		.attachmentCount = 1,
		.pAttachments = &colour_blend_attatchment,
	};

	// Create pipeline layout
	const VkPipelineLayoutCreateInfo pipeline_layout_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
		.setLayoutCount = 1,
		.pSetLayouts = &ctx->desc_set_layout,
	};
	result = vkCreatePipelineLayout(
	    ctx->dev, &pipeline_layout_config, nullptr, &ctx->pipeline_layout);
	assert(result == VK_SUCCESS);

	// Set up colour attachment
	const VkAttachmentDescription colour_attachment = {
		.format = surface_fmt.format,
		.samples = VK_SAMPLE_COUNT_1_BIT,
		.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR,
		.storeOp = VK_ATTACHMENT_STORE_OP_STORE,
		.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE,
		.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE,
		.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
		.finalLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
	};
	const VkAttachmentReference colour_attachment_ref = {
		.attachment = 0,
		.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
	};

	// Set up depth attatchment
	const VkAttachmentDescription depth_attatchment = {
		.format = depth_img_fmt,
		.samples = VK_SAMPLE_COUNT_1_BIT,
		.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR,
		.storeOp = VK_ATTACHMENT_STORE_OP_DONT_CARE,
		.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE,
		.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE,
		.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
		.finalLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
	};
	VkAttachmentReference depth_attachment_ref = {
		.attachment = 1,
		.layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
	};

	// Create render pass
	const VkAttachmentDescription attachment_descs[] = {
		colour_attachment,
		depth_attatchment,
	};
	const VkSubpassDependency dependency = {
		.srcSubpass = VK_SUBPASS_EXTERNAL,
		.srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT
		    | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT,
		.srcAccessMask = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
		.dstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT
		    | VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT,
		.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
		    | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
	};
	const VkRenderPassCreateInfo render_pass_config = {
		.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
		.attachmentCount = 2,
		.pAttachments = attachment_descs,
		.subpassCount = 1,
		.pSubpasses = &(VkSubpassDescription){
			.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
			.colorAttachmentCount = 1,
			.pColorAttachments = &colour_attachment_ref,
			.pDepthStencilAttachment = &depth_attachment_ref,
		},
		.dependencyCount = 1,
		.pDependencies = &dependency,
	};
	result = vkCreateRenderPass(
	    ctx->dev, &render_pass_config, nullptr, &ctx->render_pass);
	assert(result == VK_SUCCESS);

	// Create pipeline
	const VkPipelineDepthStencilStateCreateInfo depth_stencil_config = {
		.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
		.depthTestEnable = VK_TRUE,
		.depthWriteEnable = VK_TRUE,
		.depthCompareOp = VK_COMPARE_OP_LESS,
		.depthBoundsTestEnable = VK_FALSE,
		.stencilTestEnable = VK_FALSE,
	};
	const VkGraphicsPipelineCreateInfo pipeline_config = {
		.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
		.stageCount = 2,
		.pStages = shader_stages,
		.pVertexInputState = &vertex_input_config,
		.pInputAssemblyState = &input_assembly_config,
		.pViewportState = &viewport_config,
		.pRasterizationState = &rasterizer_config,
		.pMultisampleState = &multisampling_config,
		.pColorBlendState = &colour_blend_config,
		.layout = ctx->pipeline_layout,
		.renderPass = ctx->render_pass,
		.subpass = 0,
		.pDepthStencilState = &depth_stencil_config,
	};
	result = vkCreateGraphicsPipelines(
	    ctx->dev, VK_NULL_HANDLE, 1, &pipeline_config, nullptr,
	    &ctx->pipeline);
	assert(result == VK_SUCCESS);

	// Create framebuffers
	for (unsigned i = 0; i < swapchain_img_count; ++i) {
		VkImageView attachments[] = {
			ctx->swapchain_img_views[i],
			ctx->depth_img_view,
		};
		const VkFramebufferCreateInfo framebuffer_config = {
			.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
			.renderPass = ctx->render_pass,
			.attachmentCount = 2,
			.pAttachments = attachments,
			.width = ctx->swapchain_extent.width,
			.height = ctx->swapchain_extent.height,
			.layers = 1,
		};
		result = vkCreateFramebuffer(
		    ctx->dev, &framebuffer_config, nullptr, ctx->framebuffers + i);
		assert(result == VK_SUCCESS);
	}

	// Create command pool
	const VkCommandPoolCreateInfo cmd_pool_config = {
		.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
		.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
		.queueFamilyIndex = queue_family,
	};
	result = vkCreateCommandPool(
	    ctx->dev, &cmd_pool_config, nullptr, &ctx->cmd_pool);
	assert(result == VK_SUCCESS);

	// Allocate command buffer
	const VkCommandBufferAllocateInfo cmd_buf_config = {
		.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
		.commandPool = ctx->cmd_pool,
		.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
		.commandBufferCount = 1,
	};
	result
	    = vkAllocateCommandBuffers(ctx->dev, &cmd_buf_config, &ctx->cmd_buf);
	assert(result == VK_SUCCESS);

	// Create syncronisation primitives
	VkSemaphoreCreateInfo semaphore_config = {
		.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
	};
	result = vkCreateSemaphore(
	    ctx->dev, &semaphore_config, nullptr, &ctx->img_avail);
	assert(result == VK_SUCCESS);
	result = vkCreateSemaphore(
	    ctx->dev, &semaphore_config, nullptr, &ctx->render_fin);
	assert(result == VK_SUCCESS);
	VkFenceCreateInfo fence_config = {
		.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
		.flags = VK_FENCE_CREATE_SIGNALED_BIT,
	};
	result
	    = vkCreateFence(ctx->dev, &fence_config, nullptr, &ctx->in_flight);
	assert(result == VK_SUCCESS);

	// Set up staging, vertex and index buffers
	const size_t vertex_data_size = model.vertex_count * sizeof(vertex_t);
	const size_t index_data_size = model.index_count * sizeof(vertex_t);
	ctx->index_count = model.index_count;
	VkBuffer staging_buf;
	VkDeviceMemory staging_buf_mem;
	const size_t staging_buf_size = MAX(vertex_data_size, index_data_size);
	create_buffer(
	    ctx, staging_buf_size, VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
	    VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
	        | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
	    &staging_buf, &staging_buf_mem);
	create_buffer(
	    ctx, vertex_data_size,
	    VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_VERTEX_BUFFER_BIT,
	    VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, &ctx->vertex_buf,
	    &ctx->vertex_buf_mem);
	create_buffer(
	    ctx, index_data_size,
	    VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT,
	    VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, &ctx->index_buf,
	    &ctx->index_buf_mem);

	// Create transient command pool
	const VkCommandPoolCreateInfo transient_cmd_pool_config = {
		.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
		.flags = VK_COMMAND_POOL_CREATE_TRANSIENT_BIT,
		.queueFamilyIndex = queue_family,
	};
	VkCommandPool transient_cmd_pool;
	result = vkCreateCommandPool(
	    ctx->dev, &transient_cmd_pool_config, nullptr, &transient_cmd_pool);
	assert(result == VK_SUCCESS);

	// Load vertex and index data
	load_buffer(
	    ctx, ctx->vertex_buf, model.vertices, vertex_data_size, staging_buf,
	    staging_buf_mem, transient_cmd_pool);
	load_buffer(
	    ctx, ctx->index_buf, model.indices, index_data_size, staging_buf,
	    staging_buf_mem, transient_cmd_pool);

	// Clean up staging buffer and transient command pool
	vkDestroyCommandPool(ctx->dev, transient_cmd_pool, nullptr);
	vkDestroyBuffer(ctx->dev, staging_buf, nullptr);
	vkFreeMemory(ctx->dev, staging_buf_mem, nullptr);
}

void gpu_cleanup(const gpu_ctx_t *ctx)
{
	vkDeviceWaitIdle(ctx->dev);

	vkDestroyBuffer(ctx->dev, ctx->index_buf, nullptr);
	vkFreeMemory(ctx->dev, ctx->index_buf_mem, nullptr);
	vkDestroyBuffer(ctx->dev, ctx->vertex_buf, nullptr);
	vkFreeMemory(ctx->dev, ctx->vertex_buf_mem, nullptr);

	vkDestroySemaphore(ctx->dev, ctx->img_avail, nullptr);
	vkDestroySemaphore(ctx->dev, ctx->render_fin, nullptr);
	vkDestroyFence(ctx->dev, ctx->in_flight, nullptr);

	vkDestroyCommandPool(ctx->dev, ctx->cmd_pool, nullptr);

	vkDestroyPipeline(ctx->dev, ctx->pipeline, nullptr);
	vkDestroyRenderPass(ctx->dev, ctx->render_pass, nullptr);
	vkDestroyPipelineLayout(ctx->dev, ctx->pipeline_layout, nullptr);

	vkUnmapMemory(ctx->dev, ctx->mvp_buffer_mem);
	vkDestroyBuffer(ctx->dev, ctx->mvp_buffer, nullptr);
	vkFreeMemory(ctx->dev, ctx->mvp_buffer_mem, nullptr);

	vkDestroyDescriptorPool(ctx->dev, ctx->desc_pool, nullptr);
	vkDestroyDescriptorSetLayout(ctx->dev, ctx->desc_set_layout, nullptr);

	vkDestroyShaderModule(ctx->dev, ctx->vert_shader, nullptr);
	vkDestroyShaderModule(ctx->dev, ctx->frag_shader, nullptr);

	vkDestroyImageView(ctx->dev, ctx->depth_img_view, nullptr);
	vkDestroyImage(ctx->dev, ctx->depth_img, nullptr);
	vkFreeMemory(ctx->dev, ctx->depth_img_mem, nullptr);

	for (unsigned i = 0; i < ctx->swapchain_img_count; ++i) {
		vkDestroyFramebuffer(ctx->dev, ctx->framebuffers[i], nullptr);
		vkDestroyImageView(ctx->dev, ctx->swapchain_img_views[i], nullptr);
	}
	vkDestroySwapchainKHR(ctx->dev, ctx->swapchain, nullptr);

	vkDestroyDevice(ctx->dev, nullptr);
	vkDestroySurfaceKHR(ctx->inst, ctx->surface, nullptr);
	vkDestroyInstance(ctx->inst, nullptr);
}

void gpu_render(const gpu_ctx_t *ctx, const mvp_t *mvp)
{
	VkResult result;

	// Wait for previous frame to finish
	vkWaitForFences(ctx->dev, 1, &ctx->in_flight, VK_TRUE, UINT64_MAX);
	vkResetFences(ctx->dev, 1, &ctx->in_flight);

	// Aquire image from swapchain
	uint32_t img_index;
	result = vkAcquireNextImageKHR(
	    ctx->dev, ctx->swapchain, UINT64_MAX, ctx->img_avail, VK_NULL_HANDLE,
	    &img_index);
	assert(result == VK_SUCCESS);

	// Copy MVP matricies into uniform buffer for this image
	memcpy(ctx->mvps + img_index, mvp, sizeof(mvp_t));

	// Reset command buffer, then begin writing
	vkResetCommandBuffer(ctx->cmd_buf, 0);
	const VkCommandBufferBeginInfo cmd_begin = {
		.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
	};
	result = vkBeginCommandBuffer(ctx->cmd_buf, &cmd_begin);
	assert(result == VK_SUCCESS);

	// Start render pass
	VkClearValue clear_values[] = {
		[0] = { .color = { { 0.0, 0.0, 0.0, 1.0 } } },
		[1] = { .depthStencil = { 1.0, 0.0 } },
	};
	VkRenderPassBeginInfo render_pass_begin = {
		.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
		.renderPass = ctx->render_pass,
		.framebuffer = ctx->framebuffers[img_index],
		.renderArea.extent = ctx->swapchain_extent,
		.clearValueCount = 2,
		.pClearValues = clear_values,
	};
	vkCmdBeginRenderPass(
	    ctx->cmd_buf, &render_pass_begin, VK_SUBPASS_CONTENTS_INLINE);
	vkCmdBindPipeline(
	    ctx->cmd_buf, VK_PIPELINE_BIND_POINT_GRAPHICS, ctx->pipeline);

	// Bind vertex and index buffers
	vkCmdBindVertexBuffers(
	    ctx->cmd_buf, 0, 1, &ctx->vertex_buf, &(VkDeviceSize) { 0 });
	vkCmdBindIndexBuffer(
	    ctx->cmd_buf, ctx->index_buf, 0, VK_INDEX_TYPE_UINT16);

	// Bind descriptor set for this frame
	vkCmdBindDescriptorSets(
	    ctx->cmd_buf, VK_PIPELINE_BIND_POINT_GRAPHICS, ctx->pipeline_layout,
	    0, 1, ctx->desc_sets + img_index, 0, nullptr);

	// Draw
	vkCmdDrawIndexed(ctx->cmd_buf, ctx->index_count, 1, 0, 0, 0);

	// Finish render pass and end writing to command buffer
	vkCmdEndRenderPass(ctx->cmd_buf);
	result = vkEndCommandBuffer(ctx->cmd_buf);
	assert(result == VK_SUCCESS);

	// Submit command buffer
	VkPipelineStageFlags wait_stage
	    = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
	const VkSubmitInfo submit = {
		.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
		.waitSemaphoreCount = 1,
		.pWaitSemaphores = &ctx->img_avail,
		.pWaitDstStageMask = &wait_stage,
		.commandBufferCount = 1,
		.pCommandBuffers = &ctx->cmd_buf,
		.signalSemaphoreCount = 1,
		.pSignalSemaphores = &ctx->render_fin,
	};
	result = vkQueueSubmit(ctx->queue, 1, &submit, ctx->in_flight);
	assert(result == VK_SUCCESS);

	// Present frame
	const VkPresentInfoKHR present_info = {
		.sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
		.waitSemaphoreCount = 1,
		.pWaitSemaphores = &ctx->render_fin,
		.swapchainCount = 1,
		.pSwapchains = &ctx->swapchain,
		.pImageIndices = &img_index,
	};
	result = vkQueuePresentKHR(ctx->queue, &present_info);
	assert(result == VK_SUCCESS);
}