#include "vulkan_device.h"

#include "core/log.h"

#define GLFW_INCLUDE_VULKAN
#include <GLFW/glfw3.h>

#include <imgui.h>
#include <imgui_impl_glfw.h>
#include <imgui_impl_vulkan.h>

#include <glm/glm.hpp>
#include <glm/gtc/matrix_transform.hpp>
#include "stb_image.h"

#include <vector>
#include <array>
#include <algorithm>
#include <cstring>
#include <cstdlib>
#include <fstream>

namespace Donut::RHI
{
    namespace
    {
        constexpr int MAX_FRAMES_IN_FLIGHT = 2;
        constexpr uint32_t MAX_BINDINGS = 8;

        #define VKD_CHECK(expr)                                            \
            do {                                                           \
                VkResult _r = (expr);                                      \
                if (_r != VK_SUCCESS) {                                    \
                    DONUT_ERROR("Vulkan RHI: {} failed ({})", #expr, (int)_r); \
                    return false;                                          \
                }                                                          \
            } while (0)

        auto vk_format(Format f) -> VkFormat
        {
            switch (f) {
                case Format::RGBA16F: return VK_FORMAT_R16G16B16A16_SFLOAT;
                case Format::D32:     return VK_FORMAT_D32_SFLOAT;
                default:              return VK_FORMAT_R8G8B8A8_UNORM;
            }
        }
        auto vk_attr_format(uint32_t comps) -> VkFormat
        {
            switch (comps) {
                case 1:  return VK_FORMAT_R32_SFLOAT;
                case 2:  return VK_FORMAT_R32G32_SFLOAT;
                case 3:  return VK_FORMAT_R32G32B32_SFLOAT;
                default: return VK_FORMAT_R32G32B32A32_SFLOAT;
            }
        }
        auto vk_topology(Topology t) -> VkPrimitiveTopology
        { return t == Topology::Lines ? VK_PRIMITIVE_TOPOLOGY_LINE_LIST : VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST; }
        auto vk_compare(CompareOp o) -> VkCompareOp
        { return o == CompareOp::Always ? VK_COMPARE_OP_ALWAYS : o == CompareOp::LessEqual ? VK_COMPARE_OP_LESS_OR_EQUAL : VK_COMPARE_OP_LESS; }
        auto vk_filter(Filter f) -> VkFilter { return f == Filter::Nearest ? VK_FILTER_NEAREST : VK_FILTER_LINEAR; }
        auto vk_cull(CullMode c) -> VkCullModeFlags
        { return c == CullMode::None ? VK_CULL_MODE_NONE : c == CullMode::Back ? VK_CULL_MODE_BACK_BIT : VK_CULL_MODE_FRONT_BIT; }

        class VulkanDevice;

        // ---- Buffer: host-visible + coherent, persistently mapped -----------
        class VkBufferR : public Buffer
        {
        public:
            VkBufferR(VkDevice d, VkBuffer b, VkDeviceMemory m, void* mapped, size_t size)
                : m_device(d), m_buf(b), m_mem(m), m_mapped(mapped), m_size(size) {}
            ~VkBufferR() override
            {
                if (m_mapped) vkUnmapMemory(m_device, m_mem);
                if (m_buf) vkDestroyBuffer(m_device, m_buf, nullptr);
                if (m_mem) vkFreeMemory(m_device, m_mem, nullptr);
            }
            auto update(const void* data, size_t size) -> void override
            { if (m_mapped) std::memcpy(m_mapped, data, std::min(size, m_size)); }

            VkDevice m_device; VkBuffer m_buf; VkDeviceMemory m_mem; void* m_mapped; size_t m_size;
        };

        // ---- Texture: sampled image (2D or cube). Owns its handles unless it is
        //      a borrowed wrapper around a render-target view. ------------------
        class VkTextureR : public Texture
        {
        public:
            VkTextureR() = default;
            ~VkTextureR() override
            {
                if (!m_owns) return;
                if (m_sampler) vkDestroySampler(m_device, m_sampler, nullptr);
                if (m_view)    vkDestroyImageView(m_device, m_view, nullptr);
                if (m_image)   vkDestroyImage(m_device, m_image, nullptr);
                if (m_mem)     vkFreeMemory(m_device, m_mem, nullptr);
            }
            VkDevice      m_device  = VK_NULL_HANDLE;
            VkImage       m_image   = VK_NULL_HANDLE;
            VkDeviceMemory m_mem    = VK_NULL_HANDLE;
            VkImageView   m_view    = VK_NULL_HANDLE;
            VkSampler     m_sampler = VK_NULL_HANDLE;
            bool          m_owns    = true;
        };

        // ---- RenderTarget: off-screen colour image + framebuffer ------------
        class VkRenderTargetR : public RenderTarget
        {
        public:
            ~VkRenderTargetR() override
            {
                if (m_fb)      vkDestroyFramebuffer(m_device, m_fb, nullptr);
                if (m_sampler) vkDestroySampler(m_device, m_sampler, nullptr);
                if (m_view)    vkDestroyImageView(m_device, m_view, nullptr);
                if (m_image)   vkDestroyImage(m_device, m_image, nullptr);
                if (m_mem)     vkFreeMemory(m_device, m_mem, nullptr);
            }
            auto width()  const -> int override { return m_w; }
            auto height() const -> int override { return m_h; }
            auto color_texture() -> Texture* override { return &m_color; }

            VkDevice       m_device  = VK_NULL_HANDLE;
            int            m_w = 0, m_h = 0;
            VkImage        m_image   = VK_NULL_HANDLE;
            VkDeviceMemory m_mem     = VK_NULL_HANDLE;
            VkImageView    m_view    = VK_NULL_HANDLE;
            VkSampler      m_sampler = VK_NULL_HANDLE;
            VkFramebuffer  m_fb      = VK_NULL_HANDLE;
            VkTextureR     m_color;   // borrowed wrapper (view+sampler) for sampling
        };

        // ---- Pipeline -------------------------------------------------------
        class VkPipelineR : public Pipeline
        {
        public:
            ~VkPipelineR() override
            {
                if (m_pipeline)   vkDestroyPipeline(m_device, m_pipeline, nullptr);
                if (m_layout)     vkDestroyPipelineLayout(m_device, m_layout, nullptr);
                if (m_set_layout) vkDestroyDescriptorSetLayout(m_device, m_set_layout, nullptr);
            }
            VkDevice              m_device      = VK_NULL_HANDLE;
            VkPipeline            m_pipeline    = VK_NULL_HANDLE;
            VkPipelineLayout      m_layout      = VK_NULL_HANDLE;
            VkDescriptorSetLayout m_set_layout  = VK_NULL_HANDLE;
            std::vector<ResourceSlot> m_resources;
        };

        // ---- CommandList ----------------------------------------------------
        class VkCommandListR : public CommandList
        {
        public:
            auto begin_render_pass(RenderTarget* target, const glm::vec4& clear) -> void override
            {
                VkClearValue cvs[2]{};
                cvs[0].color = { { clear.r, clear.g, clear.b, clear.a } };
                cvs[1].depthStencil = { 1.0f, 0 };
                VkRenderPassBeginInfo rpbi{ VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO };
                if (target)
                {
                    auto* rt = static_cast<VkRenderTargetR*>(target);
                    rpbi.renderPass = m_offscreen_rp; rpbi.framebuffer = rt->m_fb;
                    rpbi.renderArea = { { 0, 0 }, { (uint32_t)rt->m_w, (uint32_t)rt->m_h } };
                    rpbi.clearValueCount = 1; rpbi.pClearValues = cvs;
                }
                else
                {
                    rpbi.renderPass = m_swapchain_rp; rpbi.framebuffer = m_swapchain_fb;
                    rpbi.renderArea = { { 0, 0 }, m_extent };
                    rpbi.clearValueCount = 2; rpbi.pClearValues = cvs;
                }
                vkCmdBeginRenderPass(m_cmd, &rpbi, VK_SUBPASS_CONTENTS_INLINE);
            }
            auto end_render_pass() -> void override { vkCmdEndRenderPass(m_cmd); }

            auto bind_pipeline(Pipeline* p) -> void override
            {
                m_pipe = static_cast<VkPipelineR*>(p);
                vkCmdBindPipeline(m_cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, m_pipe->m_pipeline);
            }
            auto set_viewport(int x, int y, int w, int h, bool flip_y) -> void override
            {
                VkViewport vp{ (float)x, flip_y ? (float)(y + h) : (float)y,
                               (float)w, flip_y ? -(float)h : (float)h, 0.0f, 1.0f };
                VkRect2D sc{ { x, y }, { (uint32_t)w, (uint32_t)h } };
                vkCmdSetViewport(m_cmd, 0, 1, &vp);
                vkCmdSetScissor(m_cmd, 0, 1, &sc);
            }
            auto bind_uniform(uint32_t binding, Buffer* ubo) -> void override
            {
                if (binding >= MAX_BINDINGS) return;
                m_buf_info[binding] = { static_cast<VkBufferR*>(ubo)->m_buf, 0, VK_WHOLE_SIZE };
            }
            auto bind_texture(uint32_t binding, Texture* texture) -> void override
            {
                if (binding >= MAX_BINDINGS) return;
                auto* t = static_cast<VkTextureR*>(texture);
                m_img_info[binding] = { t->m_sampler, t->m_view, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL };
            }
            auto bind_vertex_buffer(Buffer* vb) -> void override
            {
                VkBuffer b = static_cast<VkBufferR*>(vb)->m_buf; VkDeviceSize off = 0;
                vkCmdBindVertexBuffers(m_cmd, 0, 1, &b, &off);
            }
            auto bind_index_buffer(Buffer* ib) -> void override
            { vkCmdBindIndexBuffer(m_cmd, static_cast<VkBufferR*>(ib)->m_buf, 0, VK_INDEX_TYPE_UINT32); }
            auto draw(uint32_t vertex_count) -> void override
            { flush_descriptors(); vkCmdDraw(m_cmd, vertex_count, 1, 0, 0); }
            auto draw_indexed(uint32_t index_count) -> void override
            { flush_descriptors(); vkCmdDrawIndexed(m_cmd, index_count, 1, 0, 0, 0); }

            // Allocate + write + bind a descriptor set for the current pipeline's
            // declared resources, using whatever was bound since bind_pipeline.
            auto flush_descriptors() -> void
            {
                if (!m_pipe || m_pipe->m_resources.empty()) return;
                VkDescriptorSetAllocateInfo ai{ VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO };
                ai.descriptorPool = m_frame_pool; ai.descriptorSetCount = 1; ai.pSetLayouts = &m_pipe->m_set_layout;
                VkDescriptorSet set = VK_NULL_HANDLE;
                if (vkAllocateDescriptorSets(m_device, &ai, &set) != VK_SUCCESS)
                { DONUT_ERROR("Vulkan RHI: descriptor set allocation failed"); return; }

                std::array<VkWriteDescriptorSet, MAX_BINDINGS> writes{};
                uint32_t n = 0;
                for (const auto& r : m_pipe->m_resources)
                {
                    VkWriteDescriptorSet& w = writes[n++];
                    w.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
                    w.dstSet = set; w.dstBinding = r.binding; w.descriptorCount = 1;
                    if (r.kind == ResourceKind::UniformBuffer)
                    { w.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER; w.pBufferInfo = &m_buf_info[r.binding]; }
                    else
                    { w.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER; w.pImageInfo = &m_img_info[r.binding]; }
                }
                vkUpdateDescriptorSets(m_device, n, writes.data(), 0, nullptr);
                vkCmdBindDescriptorSets(m_cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, m_pipe->m_layout, 0, 1, &set, 0, nullptr);
            }

            // Set by the device at begin_frame:
            VkDevice        m_device       = VK_NULL_HANDLE;
            VkCommandBuffer m_cmd          = VK_NULL_HANDLE;
            VkRenderPass    m_swapchain_rp = VK_NULL_HANDLE;
            VkRenderPass    m_offscreen_rp = VK_NULL_HANDLE;
            VkFramebuffer   m_swapchain_fb = VK_NULL_HANDLE;
            VkExtent2D      m_extent{};
            VkDescriptorPool m_frame_pool  = VK_NULL_HANDLE;

            VkPipelineR* m_pipe = nullptr;
            VkDescriptorBufferInfo m_buf_info[MAX_BINDINGS]{};
            VkDescriptorImageInfo  m_img_info[MAX_BINDINGS]{};
        };

        // ---- Device ---------------------------------------------------------
        class VulkanDevice : public Device
        {
        public:
            auto init(void* glfwWindow, int width, int height) -> bool override;
            auto shutdown() -> void override;
            auto resize(int width, int height) -> void override { m_framebuffer_resized = true; m_width = width; m_height = height; }
            auto wait_idle() -> void override { if (m_device) vkDeviceWaitIdle(m_device); }

            auto create_buffer(BufferType type, size_t size, const void* data) -> Ref<Buffer> override;
            auto create_texture(int w, int h, Format format, Filter filter, const void* data) -> Ref<Texture> override;
            auto create_cubemap_from_hdri(const std::string& path) -> Ref<Texture> override;
            auto create_render_target(int w, int h, Format color, bool with_depth, Filter filter, int mips) -> Ref<RenderTarget> override;
            auto create_pipeline(const PipelineDesc& desc) -> Ref<Pipeline> override;

            auto begin_frame(const glm::vec4& clear) -> CommandList* override;
            auto end_frame() -> void override;

            auto init_imgui() -> void override;
            auto imgui_new_frame() -> void override;
            auto imgui_render(CommandList& cmds) -> void override;

            auto device_name() const -> const std::string& override { return m_gpu_name; }

            // --- internals ---
            auto find_memory_type(uint32_t filter, VkMemoryPropertyFlags flags) const -> uint32_t;
            auto create_buffer_raw(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags props, VkBuffer& buf, VkDeviceMemory& mem) const -> bool;
            auto load_spirv(const std::string& path) const -> std::vector<uint32_t>;
            auto create_shader_module(const std::string& path, VkShaderModule& out) const -> bool;

            auto create_instance() -> bool;
            auto pick_physical_and_device() -> bool;
            auto create_swapchain() -> bool;
            auto create_image_views() -> bool;
            auto create_swapchain_render_pass() -> bool;
            auto create_offscreen_render_pass() -> bool;
            auto create_depth_and_framebuffers() -> bool;
            auto create_command_and_sync() -> bool;
            auto recreate_swapchain() -> bool;
            auto cleanup_swapchain() -> void;

            GLFWwindow* m_window = nullptr;
            int m_width = 0, m_height = 0;
            bool m_framebuffer_resized = false;
            std::string m_gpu_name;

            VkInstance       m_instance = VK_NULL_HANDLE;
            VkSurfaceKHR     m_surface  = VK_NULL_HANDLE;
            VkPhysicalDevice m_physical = VK_NULL_HANDLE;
            VkDevice         m_device   = VK_NULL_HANDLE;
            uint32_t         m_graphics_family = 0, m_present_family = 0;
            VkQueue          m_graphics_queue = VK_NULL_HANDLE, m_present_queue = VK_NULL_HANDLE;
            VkPhysicalDeviceMemoryProperties m_mem_props{};

            VkSwapchainKHR m_swapchain = VK_NULL_HANDLE;
            VkFormat       m_swapchain_format = VK_FORMAT_B8G8R8A8_UNORM;
            VkExtent2D     m_extent{};
            std::vector<VkImage>       m_images;
            std::vector<VkImageView>   m_image_views;
            VkRenderPass               m_swapchain_rp = VK_NULL_HANDLE;
            VkRenderPass               m_offscreen_rp = VK_NULL_HANDLE;
            std::vector<VkFramebuffer> m_framebuffers;
            VkImage m_depth_image = VK_NULL_HANDLE; VkDeviceMemory m_depth_mem = VK_NULL_HANDLE; VkImageView m_depth_view = VK_NULL_HANDLE;

            VkCommandPool m_command_pool = VK_NULL_HANDLE;
            std::vector<VkCommandBuffer> m_command_buffers;
            std::vector<VkSemaphore> m_image_available;
            std::vector<VkSemaphore> m_render_finished;
            std::vector<VkFence>     m_in_flight;
            std::vector<VkFence>     m_images_in_flight;
            VkFence m_geo_in_use = VK_NULL_HANDLE;
            uint32_t m_current_frame = 0, m_image_index = 0;

            std::vector<VkDescriptorPool> m_frame_pools;    // one per frame in flight
            VkDescriptorPool m_imgui_pool = VK_NULL_HANDLE;
            bool m_imgui = false;

            VkCommandListR m_cmds;
        };

        // ==================================================================
        auto VulkanDevice::find_memory_type(uint32_t filter, VkMemoryPropertyFlags flags) const -> uint32_t
        {
            for (uint32_t i = 0; i < m_mem_props.memoryTypeCount; ++i)
                if ((filter & (1u << i)) && (m_mem_props.memoryTypes[i].propertyFlags & flags) == flags)
                    return i;
            return UINT32_MAX;
        }

        auto VulkanDevice::create_buffer_raw(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags props,
                                             VkBuffer& buf, VkDeviceMemory& mem) const -> bool
        {
            VkBufferCreateInfo bci{ VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO };
            bci.size = size; bci.usage = usage; bci.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
            if (vkCreateBuffer(m_device, &bci, nullptr, &buf) != VK_SUCCESS) return false;
            VkMemoryRequirements req{}; vkGetBufferMemoryRequirements(m_device, buf, &req);
            VkMemoryAllocateInfo ai{ VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
            ai.allocationSize = req.size; ai.memoryTypeIndex = find_memory_type(req.memoryTypeBits, props);
            if (vkAllocateMemory(m_device, &ai, nullptr, &mem) != VK_SUCCESS) return false;
            vkBindBufferMemory(m_device, buf, mem, 0);
            return true;
        }

        auto VulkanDevice::load_spirv(const std::string& path) const -> std::vector<uint32_t>
        {
            std::ifstream file(path, std::ios::ate | std::ios::binary);
            if (!file.is_open()) return {};
            size_t size = (size_t)file.tellg();
            std::vector<uint32_t> data(size / 4);
            file.seekg(0); file.read(reinterpret_cast<char*>(data.data()), size);
            return data;
        }

        auto VulkanDevice::create_shader_module(const std::string& path, VkShaderModule& out) const -> bool
        {
            auto spv = load_spirv(path);
            if (spv.empty()) { DONUT_ERROR("Vulkan RHI: failed to load SPIR-V {}", path); return false; }
            VkShaderModuleCreateInfo ci{ VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO };
            ci.codeSize = spv.size() * 4; ci.pCode = spv.data();
            return vkCreateShaderModule(m_device, &ci, nullptr, &out) == VK_SUCCESS;
        }

        auto VulkanDevice::create_instance() -> bool
        {
            VkApplicationInfo app{ VK_STRUCTURE_TYPE_APPLICATION_INFO };
            app.pApplicationName = "Donut"; app.apiVersion = VK_API_VERSION_1_2;

            uint32_t glfwExtCount = 0;
            const char** glfwExts = glfwGetRequiredInstanceExtensions(&glfwExtCount);
            if (!glfwExts) { DONUT_ERROR("Vulkan RHI: GLFW reports no surface support"); return false; }
            std::vector<const char*> exts(glfwExts, glfwExts + glfwExtCount);
            exts.push_back(VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME);
            exts.push_back(VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME);

            std::vector<const char*> layers;
            uint32_t layer_count = 0; vkEnumerateInstanceLayerProperties(&layer_count, nullptr);
            std::vector<VkLayerProperties> avail(layer_count);
            vkEnumerateInstanceLayerProperties(&layer_count, avail.data());
            for (const auto& l : avail)
                if (std::strcmp(l.layerName, "VK_LAYER_KHRONOS_validation") == 0)
                    layers.push_back("VK_LAYER_KHRONOS_validation");

            VkInstanceCreateInfo ici{ VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO };
            ici.flags = VK_INSTANCE_CREATE_ENUMERATE_PORTABILITY_BIT_KHR;
            ici.pApplicationInfo = &app;
            ici.enabledExtensionCount = (uint32_t)exts.size(); ici.ppEnabledExtensionNames = exts.data();
            ici.enabledLayerCount = (uint32_t)layers.size(); ici.ppEnabledLayerNames = layers.data();
            VkResult r = vkCreateInstance(&ici, nullptr, &m_instance);
            if (r != VK_SUCCESS && !layers.empty())
            {
                DONUT_WARN("Vulkan RHI: validation layer unavailable, continuing without it");
                ici.enabledLayerCount = 0; ici.ppEnabledLayerNames = nullptr;
                r = vkCreateInstance(&ici, nullptr, &m_instance);
            }
            if (r != VK_SUCCESS) { DONUT_ERROR("Vulkan RHI: vkCreateInstance failed ({})", (int)r); return false; }
            VKD_CHECK(glfwCreateWindowSurface(m_instance, m_window, nullptr, &m_surface));
            DONUT_INFO("Vulkan RHI: instance + surface created (validation {})", layers.empty() ? "off" : "on");
            return true;
        }

        auto VulkanDevice::pick_physical_and_device() -> bool
        {
            uint32_t count = 0; vkEnumeratePhysicalDevices(m_instance, &count, nullptr);
            if (count == 0) { DONUT_ERROR("Vulkan RHI: no physical devices"); return false; }
            std::vector<VkPhysicalDevice> devices(count);
            vkEnumeratePhysicalDevices(m_instance, &count, devices.data());
            m_physical = devices[0];

            uint32_t q = 0; vkGetPhysicalDeviceQueueFamilyProperties(m_physical, &q, nullptr);
            std::vector<VkQueueFamilyProperties> qfams(q);
            vkGetPhysicalDeviceQueueFamilyProperties(m_physical, &q, qfams.data());
            bool fg = false, fp = false;
            for (uint32_t i = 0; i < q; ++i)
            {
                if (!fg && (qfams[i].queueFlags & VK_QUEUE_GRAPHICS_BIT)) { m_graphics_family = i; fg = true; }
                VkBool32 present = VK_FALSE; vkGetPhysicalDeviceSurfaceSupportKHR(m_physical, i, m_surface, &present);
                if (!fp && present) { m_present_family = i; fp = true; }
            }
            if (!fg || !fp) { DONUT_ERROR("Vulkan RHI: no graphics/present queue"); return false; }

            std::vector<const char*> dev_exts = { VK_KHR_SWAPCHAIN_EXTENSION_NAME };
            uint32_t dec = 0; vkEnumerateDeviceExtensionProperties(m_physical, nullptr, &dec, nullptr);
            std::vector<VkExtensionProperties> dep(dec);
            vkEnumerateDeviceExtensionProperties(m_physical, nullptr, &dec, dep.data());
            for (const auto& e : dep)
                if (std::strcmp(e.extensionName, "VK_KHR_portability_subset") == 0)
                    dev_exts.push_back("VK_KHR_portability_subset");

            float priority = 1.0f;
            std::vector<VkDeviceQueueCreateInfo> qcis;
            uint32_t families[2] = { m_graphics_family, m_present_family };
            for (uint32_t i = 0; i < (m_graphics_family == m_present_family ? 1u : 2u); ++i)
            {
                VkDeviceQueueCreateInfo qci{ VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO };
                qci.queueFamilyIndex = families[i]; qci.queueCount = 1; qci.pQueuePriorities = &priority;
                qcis.push_back(qci);
            }
            VkDeviceCreateInfo dci{ VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO };
            dci.queueCreateInfoCount = (uint32_t)qcis.size(); dci.pQueueCreateInfos = qcis.data();
            dci.enabledExtensionCount = (uint32_t)dev_exts.size(); dci.ppEnabledExtensionNames = dev_exts.data();
            VKD_CHECK(vkCreateDevice(m_physical, &dci, nullptr, &m_device));
            vkGetDeviceQueue(m_device, m_graphics_family, 0, &m_graphics_queue);
            vkGetDeviceQueue(m_device, m_present_family, 0, &m_present_queue);

            VkPhysicalDeviceProperties props{}; vkGetPhysicalDeviceProperties(m_physical, &props);
            vkGetPhysicalDeviceMemoryProperties(m_physical, &m_mem_props);
            m_gpu_name = props.deviceName;
            DONUT_INFO("Vulkan RHI device: {}", m_gpu_name);
            return true;
        }

        auto VulkanDevice::create_swapchain() -> bool
        {
            VkSurfaceCapabilitiesKHR caps{};
            vkGetPhysicalDeviceSurfaceCapabilitiesKHR(m_physical, m_surface, &caps);
            uint32_t fc = 0; vkGetPhysicalDeviceSurfaceFormatsKHR(m_physical, m_surface, &fc, nullptr);
            std::vector<VkSurfaceFormatKHR> formats(fc);
            vkGetPhysicalDeviceSurfaceFormatsKHR(m_physical, m_surface, &fc, formats.data());
            VkSurfaceFormatKHR chosen = formats[0];
            for (const auto& f : formats)
                if (f.format == VK_FORMAT_B8G8R8A8_UNORM && f.colorSpace == VK_COLOR_SPACE_SRGB_NONLINEAR_KHR) chosen = f;
            m_swapchain_format = chosen.format;

            if (caps.currentExtent.width != UINT32_MAX) m_extent = caps.currentExtent;
            else {
                m_extent.width  = std::clamp((uint32_t)m_width,  caps.minImageExtent.width,  caps.maxImageExtent.width);
                m_extent.height = std::clamp((uint32_t)m_height, caps.minImageExtent.height, caps.maxImageExtent.height);
            }
            uint32_t image_count = caps.minImageCount + 1;
            if (caps.maxImageCount > 0 && image_count > caps.maxImageCount) image_count = caps.maxImageCount;

            VkSwapchainCreateInfoKHR sci{ VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR };
            sci.surface = m_surface; sci.minImageCount = image_count;
            sci.imageFormat = chosen.format; sci.imageColorSpace = chosen.colorSpace;
            sci.imageExtent = m_extent; sci.imageArrayLayers = 1;
            sci.imageUsage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
            sci.preTransform = caps.currentTransform; sci.compositeAlpha = VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR;
            sci.presentMode = VK_PRESENT_MODE_FIFO_KHR; sci.clipped = VK_TRUE;
            uint32_t fam[2] = { m_graphics_family, m_present_family };
            if (m_graphics_family != m_present_family)
            { sci.imageSharingMode = VK_SHARING_MODE_CONCURRENT; sci.queueFamilyIndexCount = 2; sci.pQueueFamilyIndices = fam; }
            else sci.imageSharingMode = VK_SHARING_MODE_EXCLUSIVE;
            VKD_CHECK(vkCreateSwapchainKHR(m_device, &sci, nullptr, &m_swapchain));
            uint32_t n = 0; vkGetSwapchainImagesKHR(m_device, m_swapchain, &n, nullptr);
            m_images.resize(n); vkGetSwapchainImagesKHR(m_device, m_swapchain, &n, m_images.data());
            return true;
        }

        auto VulkanDevice::create_image_views() -> bool
        {
            m_image_views.resize(m_images.size());
            for (size_t i = 0; i < m_images.size(); ++i)
            {
                VkImageViewCreateInfo vci{ VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
                vci.image = m_images[i]; vci.viewType = VK_IMAGE_VIEW_TYPE_2D; vci.format = m_swapchain_format;
                vci.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
                VKD_CHECK(vkCreateImageView(m_device, &vci, nullptr, &m_image_views[i]));
            }
            return true;
        }

        // Swapchain pass: colour + depth. Scene geometry uses the depth; the
        // black-hole present + ImGui simply don't test it.
        auto VulkanDevice::create_swapchain_render_pass() -> bool
        {
            VkAttachmentDescription atts[2]{};
            atts[0].format = m_swapchain_format; atts[0].samples = VK_SAMPLE_COUNT_1_BIT;
            atts[0].loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR; atts[0].storeOp = VK_ATTACHMENT_STORE_OP_STORE;
            atts[0].stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; atts[0].stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
            atts[0].initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; atts[0].finalLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
            atts[1].format = VK_FORMAT_D32_SFLOAT; atts[1].samples = VK_SAMPLE_COUNT_1_BIT;
            atts[1].loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR; atts[1].storeOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
            atts[1].stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; atts[1].stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
            atts[1].initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; atts[1].finalLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
            VkAttachmentReference color_ref{ 0, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL };
            VkAttachmentReference depth_ref{ 1, VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL };
            VkSubpassDescription subpass{}; subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS;
            subpass.colorAttachmentCount = 1; subpass.pColorAttachments = &color_ref; subpass.pDepthStencilAttachment = &depth_ref;
            VkSubpassDependency dep{};
            dep.srcSubpass = VK_SUBPASS_EXTERNAL; dep.dstSubpass = 0;
            dep.srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
            dep.dstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
            dep.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
            VkRenderPassCreateInfo rpci{ VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO };
            rpci.attachmentCount = 2; rpci.pAttachments = atts; rpci.subpassCount = 1; rpci.pSubpasses = &subpass;
            rpci.dependencyCount = 1; rpci.pDependencies = &dep;
            VKD_CHECK(vkCreateRenderPass(m_device, &rpci, nullptr, &m_swapchain_rp));
            return true;
        }

        // Off-screen colour pass (RGBA8), leaving the image SHADER_READ_ONLY so the
        // present pass can sample it. Shared by every render target.
        auto VulkanDevice::create_offscreen_render_pass() -> bool
        {
            VkAttachmentDescription color{};
            color.format = VK_FORMAT_R8G8B8A8_UNORM; color.samples = VK_SAMPLE_COUNT_1_BIT;
            color.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR; color.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
            color.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; color.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
            color.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; color.finalLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
            VkAttachmentReference ref{ 0, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL };
            VkSubpassDescription subpass{}; subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS;
            subpass.colorAttachmentCount = 1; subpass.pColorAttachments = &ref;
            VkSubpassDependency deps[2]{};
            deps[0].srcSubpass = VK_SUBPASS_EXTERNAL; deps[0].dstSubpass = 0;
            deps[0].srcStageMask = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; deps[0].srcAccessMask = VK_ACCESS_SHADER_READ_BIT;
            deps[0].dstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT; deps[0].dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
            deps[1].srcSubpass = 0; deps[1].dstSubpass = VK_SUBPASS_EXTERNAL;
            deps[1].srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT; deps[1].srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
            deps[1].dstStageMask = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; deps[1].dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
            VkRenderPassCreateInfo rpci{ VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO };
            rpci.attachmentCount = 1; rpci.pAttachments = &color; rpci.subpassCount = 1; rpci.pSubpasses = &subpass;
            rpci.dependencyCount = 2; rpci.pDependencies = deps;
            VKD_CHECK(vkCreateRenderPass(m_device, &rpci, nullptr, &m_offscreen_rp));
            return true;
        }

        auto VulkanDevice::create_depth_and_framebuffers() -> bool
        {
            VkImageCreateInfo dici{ VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO };
            dici.imageType = VK_IMAGE_TYPE_2D; dici.format = VK_FORMAT_D32_SFLOAT;
            dici.extent = { m_extent.width, m_extent.height, 1 };
            dici.mipLevels = 1; dici.arrayLayers = 1; dici.samples = VK_SAMPLE_COUNT_1_BIT;
            dici.tiling = VK_IMAGE_TILING_OPTIMAL; dici.usage = VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
            VKD_CHECK(vkCreateImage(m_device, &dici, nullptr, &m_depth_image));
            VkMemoryRequirements dreq{}; vkGetImageMemoryRequirements(m_device, m_depth_image, &dreq);
            VkMemoryAllocateInfo dai{ VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
            dai.allocationSize = dreq.size; dai.memoryTypeIndex = find_memory_type(dreq.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
            VKD_CHECK(vkAllocateMemory(m_device, &dai, nullptr, &m_depth_mem));
            VKD_CHECK(vkBindImageMemory(m_device, m_depth_image, m_depth_mem, 0));
            VkImageViewCreateInfo dvci{ VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
            dvci.image = m_depth_image; dvci.viewType = VK_IMAGE_VIEW_TYPE_2D; dvci.format = VK_FORMAT_D32_SFLOAT;
            dvci.subresourceRange = { VK_IMAGE_ASPECT_DEPTH_BIT, 0, 1, 0, 1 };
            VKD_CHECK(vkCreateImageView(m_device, &dvci, nullptr, &m_depth_view));

            m_framebuffers.resize(m_image_views.size());
            for (size_t i = 0; i < m_image_views.size(); ++i)
            {
                VkImageView att[2] = { m_image_views[i], m_depth_view };
                VkFramebufferCreateInfo fbci{ VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO };
                fbci.renderPass = m_swapchain_rp; fbci.attachmentCount = 2; fbci.pAttachments = att;
                fbci.width = m_extent.width; fbci.height = m_extent.height; fbci.layers = 1;
                VKD_CHECK(vkCreateFramebuffer(m_device, &fbci, nullptr, &m_framebuffers[i]));
            }
            return true;
        }

        auto VulkanDevice::create_command_and_sync() -> bool
        {
            VkCommandPoolCreateInfo pci{ VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO };
            pci.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT; pci.queueFamilyIndex = m_graphics_family;
            VKD_CHECK(vkCreateCommandPool(m_device, &pci, nullptr, &m_command_pool));
            m_command_buffers.resize(MAX_FRAMES_IN_FLIGHT);
            VkCommandBufferAllocateInfo cbai{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO };
            cbai.commandPool = m_command_pool; cbai.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY; cbai.commandBufferCount = MAX_FRAMES_IN_FLIGHT;
            VKD_CHECK(vkAllocateCommandBuffers(m_device, &cbai, m_command_buffers.data()));

            m_image_available.resize(MAX_FRAMES_IN_FLIGHT);
            m_in_flight.resize(MAX_FRAMES_IN_FLIGHT);
            m_render_finished.resize(m_images.size());
            m_images_in_flight.assign(m_images.size(), VK_NULL_HANDLE);
            VkSemaphoreCreateInfo sci{ VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO };
            VkFenceCreateInfo fci{ VK_STRUCTURE_TYPE_FENCE_CREATE_INFO }; fci.flags = VK_FENCE_CREATE_SIGNALED_BIT;
            for (int i = 0; i < MAX_FRAMES_IN_FLIGHT; ++i)
            { VKD_CHECK(vkCreateSemaphore(m_device, &sci, nullptr, &m_image_available[i])); VKD_CHECK(vkCreateFence(m_device, &fci, nullptr, &m_in_flight[i])); }
            for (size_t i = 0; i < m_images.size(); ++i)
                VKD_CHECK(vkCreateSemaphore(m_device, &sci, nullptr, &m_render_finished[i]));

            m_frame_pools.resize(MAX_FRAMES_IN_FLIGHT);
            VkDescriptorPoolSize sizes[2] = {
                { VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, 512 },
                { VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 256 },
            };
            for (int i = 0; i < MAX_FRAMES_IN_FLIGHT; ++i)
            {
                VkDescriptorPoolCreateInfo dpci{ VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO };
                dpci.maxSets = 256; dpci.poolSizeCount = 2; dpci.pPoolSizes = sizes;
                VKD_CHECK(vkCreateDescriptorPool(m_device, &dpci, nullptr, &m_frame_pools[i]));
            }
            return true;
        }

        auto VulkanDevice::init(void* glfwWindow, int width, int height) -> bool
        {
            m_window = (GLFWwindow*)glfwWindow; m_width = width; m_height = height;
            if (!create_instance())              return false;
            if (!pick_physical_and_device())     return false;
            if (!create_swapchain())             return false;
            if (!create_image_views())           return false;
            if (!create_swapchain_render_pass()) return false;
            if (!create_offscreen_render_pass()) return false;
            if (!create_depth_and_framebuffers())return false;
            if (!create_command_and_sync())      return false;
            DONUT_INFO("Vulkan RHI device ready: {} swapchain images, {}x{}", (int)m_images.size(), m_extent.width, m_extent.height);
            return true;
        }

        auto VulkanDevice::cleanup_swapchain() -> void
        {
            for (auto fb : m_framebuffers) vkDestroyFramebuffer(m_device, fb, nullptr);
            m_framebuffers.clear();
            if (m_depth_view)  { vkDestroyImageView(m_device, m_depth_view, nullptr); m_depth_view = VK_NULL_HANDLE; }
            if (m_depth_image) { vkDestroyImage(m_device, m_depth_image, nullptr); m_depth_image = VK_NULL_HANDLE; }
            if (m_depth_mem)   { vkFreeMemory(m_device, m_depth_mem, nullptr); m_depth_mem = VK_NULL_HANDLE; }
            for (auto iv : m_image_views) vkDestroyImageView(m_device, iv, nullptr);
            m_image_views.clear();
            if (m_swapchain) { vkDestroySwapchainKHR(m_device, m_swapchain, nullptr); m_swapchain = VK_NULL_HANDLE; }
        }

        auto VulkanDevice::recreate_swapchain() -> bool
        {
            int w = 0, h = 0; glfwGetFramebufferSize(m_window, &w, &h);
            while (w == 0 || h == 0) { glfwGetFramebufferSize(m_window, &w, &h); glfwWaitEvents(); }
            m_width = w; m_height = h;
            vkDeviceWaitIdle(m_device);
            cleanup_swapchain();
            if (!create_swapchain())             return false;
            if (!create_image_views())           return false;
            if (!create_depth_and_framebuffers())return false;
            m_images_in_flight.assign(m_images.size(), VK_NULL_HANDLE);
            return true;
        }

        auto VulkanDevice::create_buffer(BufferType type, size_t size, const void* data) -> Ref<Buffer>
        {
            VkBufferUsageFlags usage = type == BufferType::Index ? VK_BUFFER_USAGE_INDEX_BUFFER_BIT
                                     : type == BufferType::Uniform ? VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
                                     : VK_BUFFER_USAGE_VERTEX_BUFFER_BIT;
            VkBuffer buf = VK_NULL_HANDLE; VkDeviceMemory mem = VK_NULL_HANDLE;
            create_buffer_raw(size, usage, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, buf, mem);
            void* mapped = nullptr; vkMapMemory(m_device, mem, 0, size, 0, &mapped);
            if (data && mapped) std::memcpy(mapped, data, size);
            return create_ref<VkBufferR>(m_device, buf, mem, mapped, size);
        }

        auto VulkanDevice::create_texture(int w, int h, Format format, Filter filter, const void* data) -> Ref<Texture>
        {
            auto tex = create_ref<VkTextureR>(); tex->m_device = m_device;
            VkFormat fmt = vk_format(format);
            VkImageCreateInfo ici{ VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO };
            ici.imageType = VK_IMAGE_TYPE_2D; ici.format = fmt; ici.extent = { (uint32_t)w, (uint32_t)h, 1 };
            ici.mipLevels = 1; ici.arrayLayers = 1; ici.samples = VK_SAMPLE_COUNT_1_BIT;
            ici.tiling = VK_IMAGE_TILING_OPTIMAL; ici.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
            vkCreateImage(m_device, &ici, nullptr, &tex->m_image);
            VkMemoryRequirements req{}; vkGetImageMemoryRequirements(m_device, tex->m_image, &req);
            VkMemoryAllocateInfo ai{ VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
            ai.allocationSize = req.size; ai.memoryTypeIndex = find_memory_type(req.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
            vkAllocateMemory(m_device, &ai, nullptr, &tex->m_mem);
            vkBindImageMemory(m_device, tex->m_image, tex->m_mem, 0);

            size_t bpp = format == Format::RGBA16F ? 8 : 4;
            VkDeviceSize sz = (VkDeviceSize)w * h * bpp;
            VkBuffer staging = VK_NULL_HANDLE; VkDeviceMemory staging_mem = VK_NULL_HANDLE;
            create_buffer_raw(sz, VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, staging, staging_mem);
            void* mp = nullptr; vkMapMemory(m_device, staging_mem, 0, sz, 0, &mp);
            if (data) std::memcpy(mp, data, sz); else std::memset(mp, 0, sz);
            vkUnmapMemory(m_device, staging_mem);

            VkCommandBufferAllocateInfo cbai{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO };
            cbai.commandPool = m_command_pool; cbai.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY; cbai.commandBufferCount = 1;
            VkCommandBuffer cmd; vkAllocateCommandBuffers(m_device, &cbai, &cmd);
            VkCommandBufferBeginInfo bi{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO }; bi.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
            vkBeginCommandBuffer(cmd, &bi);
            VkImageMemoryBarrier b{ VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER };
            b.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED; b.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
            b.image = tex->m_image; b.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
            b.srcAccessMask = 0; b.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
            vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &b);
            VkBufferImageCopy copy{}; copy.imageSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 }; copy.imageExtent = { (uint32_t)w, (uint32_t)h, 1 };
            vkCmdCopyBufferToImage(cmd, staging, tex->m_image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &copy);
            VkImageMemoryBarrier r = b; r.oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL; r.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
            r.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT; r.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
            vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &r);
            vkEndCommandBuffer(cmd);
            VkSubmitInfo si{ VK_STRUCTURE_TYPE_SUBMIT_INFO }; si.commandBufferCount = 1; si.pCommandBuffers = &cmd;
            vkQueueSubmit(m_graphics_queue, 1, &si, VK_NULL_HANDLE); vkQueueWaitIdle(m_graphics_queue);
            vkFreeCommandBuffers(m_device, m_command_pool, 1, &cmd);
            vkDestroyBuffer(m_device, staging, nullptr); vkFreeMemory(m_device, staging_mem, nullptr);

            VkImageViewCreateInfo vci{ VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
            vci.image = tex->m_image; vci.viewType = VK_IMAGE_VIEW_TYPE_2D; vci.format = fmt;
            vci.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
            vkCreateImageView(m_device, &vci, nullptr, &tex->m_view);
            VkSamplerCreateInfo smci{ VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO };
            smci.magFilter = vk_filter(filter); smci.minFilter = vk_filter(filter);
            smci.addressModeU = smci.addressModeV = smci.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
            vkCreateSampler(m_device, &smci, nullptr, &tex->m_sampler);
            return tex;
        }

        auto VulkanDevice::create_render_target(int w, int h, Format color, bool /*with_depth*/, Filter filter, int /*mips*/) -> Ref<RenderTarget>
        {
            auto rt = create_ref<VkRenderTargetR>();
            rt->m_device = m_device; rt->m_w = w; rt->m_h = h;
            VkFormat fmt = vk_format(color);
            VkImageCreateInfo ici{ VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO };
            ici.imageType = VK_IMAGE_TYPE_2D; ici.format = fmt; ici.extent = { (uint32_t)w, (uint32_t)h, 1 };
            ici.mipLevels = 1; ici.arrayLayers = 1; ici.samples = VK_SAMPLE_COUNT_1_BIT;
            ici.tiling = VK_IMAGE_TILING_OPTIMAL; ici.usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
            vkCreateImage(m_device, &ici, nullptr, &rt->m_image);
            VkMemoryRequirements req{}; vkGetImageMemoryRequirements(m_device, rt->m_image, &req);
            VkMemoryAllocateInfo ai{ VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
            ai.allocationSize = req.size; ai.memoryTypeIndex = find_memory_type(req.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
            vkAllocateMemory(m_device, &ai, nullptr, &rt->m_mem);
            vkBindImageMemory(m_device, rt->m_image, rt->m_mem, 0);
            VkImageViewCreateInfo vci{ VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
            vci.image = rt->m_image; vci.viewType = VK_IMAGE_VIEW_TYPE_2D; vci.format = fmt;
            vci.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
            vkCreateImageView(m_device, &vci, nullptr, &rt->m_view);
            VkSamplerCreateInfo smci{ VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO };
            smci.magFilter = vk_filter(filter); smci.minFilter = vk_filter(filter);
            smci.addressModeU = smci.addressModeV = smci.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
            vkCreateSampler(m_device, &smci, nullptr, &rt->m_sampler);
            VkFramebufferCreateInfo fbci{ VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO };
            fbci.renderPass = m_offscreen_rp; fbci.attachmentCount = 1; fbci.pAttachments = &rt->m_view;
            fbci.width = w; fbci.height = h; fbci.layers = 1;
            vkCreateFramebuffer(m_device, &fbci, nullptr, &rt->m_fb);

            rt->m_color.m_device = m_device; rt->m_color.m_view = rt->m_view; rt->m_color.m_sampler = rt->m_sampler; rt->m_color.m_owns = false;
            return rt;
        }

        // Builds an environment cubemap from an equirectangular HDRI: render the 6
        // faces with the EquirectToCubemap pipeline, then a full mip chain by
        // linear down-blits (so divergence-based LOD reads a blurred sky). Returns
        // a Texture owning the cube image/view/sampler.
        auto VulkanDevice::create_cubemap_from_hdri(const std::string& path) -> Ref<Texture>
        {
            auto tex = create_ref<VkTextureR>(); tex->m_device = m_device;
            const VkMemoryPropertyFlags host_vis = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
            const uint32_t FACE = 1024;
            const VkFormat cube_fmt = VK_FORMAT_R16G16B16A16_SFLOAT;
            uint32_t CUBE_MIPS = 1; for (uint32_t s = FACE; s > 1; s >>= 1) ++CUBE_MIPS;

            VkImageCreateInfo cci{ VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO };
            cci.flags = VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT;
            cci.imageType = VK_IMAGE_TYPE_2D; cci.format = cube_fmt; cci.extent = { FACE, FACE, 1 };
            cci.mipLevels = CUBE_MIPS; cci.arrayLayers = 6; cci.samples = VK_SAMPLE_COUNT_1_BIT;
            cci.tiling = VK_IMAGE_TILING_OPTIMAL;
            cci.usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_SAMPLED_BIT
                      | VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
            vkCreateImage(m_device, &cci, nullptr, &tex->m_image);
            VkMemoryRequirements creq{}; vkGetImageMemoryRequirements(m_device, tex->m_image, &creq);
            VkMemoryAllocateInfo cai{ VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
            cai.allocationSize = creq.size; cai.memoryTypeIndex = find_memory_type(creq.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
            vkAllocateMemory(m_device, &cai, nullptr, &tex->m_mem);
            vkBindImageMemory(m_device, tex->m_image, tex->m_mem, 0);
            VkImageViewCreateInfo cvci{ VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
            cvci.image = tex->m_image; cvci.viewType = VK_IMAGE_VIEW_TYPE_CUBE; cvci.format = cube_fmt;
            cvci.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, CUBE_MIPS, 0, 6 };
            vkCreateImageView(m_device, &cvci, nullptr, &tex->m_view);
            VkSamplerCreateInfo csm{ VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO };
            csm.magFilter = VK_FILTER_LINEAR; csm.minFilter = VK_FILTER_LINEAR;
            csm.mipmapMode = VK_SAMPLER_MIPMAP_MODE_LINEAR; csm.minLod = 0.0f; csm.maxLod = (float)CUBE_MIPS;
            csm.addressModeU = csm.addressModeV = csm.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
            vkCreateSampler(m_device, &csm, nullptr, &tex->m_sampler);

            int w = 0, h = 0, ch = 0;
            float* pixels = stbi_loadf(path.c_str(), &w, &h, &ch, 4);
            if (!pixels)
            {
                DONUT_WARN("Vulkan RHI: HDRI '{}' could not be loaded; using a dark background", path);
                VkCommandBufferAllocateInfo cbai{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO };
                cbai.commandPool = m_command_pool; cbai.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY; cbai.commandBufferCount = 1;
                VkCommandBuffer cmd; vkAllocateCommandBuffers(m_device, &cbai, &cmd);
                VkCommandBufferBeginInfo bi{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO }; bi.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
                vkBeginCommandBuffer(cmd, &bi);
                VkImageMemoryBarrier tb{ VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER };
                tb.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED; tb.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
                tb.image = tex->m_image; tb.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, CUBE_MIPS, 0, 6 };
                tb.srcAccessMask = 0; tb.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
                vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &tb);
                VkClearColorValue dark{}; dark.float32[0] = 0.02f; dark.float32[1] = 0.02f; dark.float32[2] = 0.05f; dark.float32[3] = 1.0f;
                VkImageSubresourceRange rng{ VK_IMAGE_ASPECT_COLOR_BIT, 0, CUBE_MIPS, 0, 6 };
                vkCmdClearColorImage(cmd, tex->m_image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, &dark, 1, &rng);
                VkImageMemoryBarrier rb = tb; rb.oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL; rb.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
                rb.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT; rb.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
                vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &rb);
                vkEndCommandBuffer(cmd);
                VkSubmitInfo si{ VK_STRUCTURE_TYPE_SUBMIT_INFO }; si.commandBufferCount = 1; si.pCommandBuffers = &cmd;
                vkQueueSubmit(m_graphics_queue, 1, &si, VK_NULL_HANDLE); vkQueueWaitIdle(m_graphics_queue);
                vkFreeCommandBuffers(m_device, m_command_pool, 1, &cmd);
                return tex;
            }

            const VkFormat eq_fmt = VK_FORMAT_R16G16B16A16_SFLOAT;
            size_t texel_count = (size_t)w * h * 4;
            VkDeviceSize eq_size = (VkDeviceSize)texel_count * sizeof(uint16_t);
            VkBuffer eq_staging; VkDeviceMemory eq_staging_mem;
            create_buffer_raw(eq_size, VK_BUFFER_USAGE_TRANSFER_SRC_BIT, host_vis, eq_staging, eq_staging_mem);
            void* mp = nullptr; vkMapMemory(m_device, eq_staging_mem, 0, eq_size, 0, &mp);
            uint16_t* dst = (uint16_t*)mp;
            for (size_t i = 0; i < texel_count; ++i) { __fp16 hf = (__fp16)pixels[i]; std::memcpy(&dst[i], &hf, sizeof(uint16_t)); }
            vkUnmapMemory(m_device, eq_staging_mem);
            stbi_image_free(pixels);

            VkImage eq_image; VkDeviceMemory eq_mem;
            VkImageCreateInfo eci{ VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO };
            eci.imageType = VK_IMAGE_TYPE_2D; eci.format = eq_fmt; eci.extent = { (uint32_t)w, (uint32_t)h, 1 };
            eci.mipLevels = 1; eci.arrayLayers = 1; eci.samples = VK_SAMPLE_COUNT_1_BIT;
            eci.tiling = VK_IMAGE_TILING_OPTIMAL; eci.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
            vkCreateImage(m_device, &eci, nullptr, &eq_image);
            VkMemoryRequirements ereq{}; vkGetImageMemoryRequirements(m_device, eq_image, &ereq);
            VkMemoryAllocateInfo eai{ VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
            eai.allocationSize = ereq.size; eai.memoryTypeIndex = find_memory_type(ereq.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
            vkAllocateMemory(m_device, &eai, nullptr, &eq_mem);
            vkBindImageMemory(m_device, eq_image, eq_mem, 0);
            VkImageView eq_view;
            VkImageViewCreateInfo evci{ VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
            evci.image = eq_image; evci.viewType = VK_IMAGE_VIEW_TYPE_2D; evci.format = eq_fmt;
            evci.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
            vkCreateImageView(m_device, &evci, nullptr, &eq_view);
            VkSampler eq_sampler;
            VkSamplerCreateInfo esm{ VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO };
            esm.magFilter = VK_FILTER_LINEAR; esm.minFilter = VK_FILTER_LINEAR;
            esm.addressModeU = VK_SAMPLER_ADDRESS_MODE_REPEAT;
            esm.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
            esm.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
            vkCreateSampler(m_device, &esm, nullptr, &eq_sampler);

            VkImageView face_views[6];
            for (uint32_t i = 0; i < 6; ++i)
            {
                VkImageViewCreateInfo fvci{ VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
                fvci.image = tex->m_image; fvci.viewType = VK_IMAGE_VIEW_TYPE_2D; fvci.format = cube_fmt;
                fvci.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, i, 1 };
                vkCreateImageView(m_device, &fvci, nullptr, &face_views[i]);
            }

            VkAttachmentDescription color{};
            color.format = cube_fmt; color.samples = VK_SAMPLE_COUNT_1_BIT;
            color.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR; color.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
            color.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; color.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
            color.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; color.finalLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
            VkAttachmentReference ref{ 0, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL };
            VkSubpassDescription subpass{}; subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS; subpass.colorAttachmentCount = 1; subpass.pColorAttachments = &ref;
            VkSubpassDependency dep{}; dep.srcSubpass = 0; dep.dstSubpass = VK_SUBPASS_EXTERNAL;
            dep.srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT; dep.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
            dep.dstStageMask = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; dep.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
            VkRenderPass rp;
            VkRenderPassCreateInfo rpci{ VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO };
            rpci.attachmentCount = 1; rpci.pAttachments = &color; rpci.subpassCount = 1; rpci.pSubpasses = &subpass; rpci.dependencyCount = 1; rpci.pDependencies = &dep;
            vkCreateRenderPass(m_device, &rpci, nullptr, &rp);
            VkFramebuffer face_fb[6];
            for (uint32_t i = 0; i < 6; ++i)
            {
                VkFramebufferCreateInfo fbci{ VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO };
                fbci.renderPass = rp; fbci.attachmentCount = 1; fbci.pAttachments = &face_views[i]; fbci.width = FACE; fbci.height = FACE; fbci.layers = 1;
                vkCreateFramebuffer(m_device, &fbci, nullptr, &face_fb[i]);
            }

            VkDescriptorSetLayoutBinding binds[2]{};
            binds[0].binding = 0; binds[0].descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER; binds[0].descriptorCount = 1; binds[0].stageFlags = VK_SHADER_STAGE_VERTEX_BIT;
            binds[1].binding = 1; binds[1].descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER; binds[1].descriptorCount = 1; binds[1].stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT;
            VkDescriptorSetLayout set_layout;
            VkDescriptorSetLayoutCreateInfo dslci{ VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO }; dslci.bindingCount = 2; dslci.pBindings = binds;
            vkCreateDescriptorSetLayout(m_device, &dslci, nullptr, &set_layout);
            VkDescriptorPoolSize psizes[2] = { { VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, 6 }, { VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 6 } };
            VkDescriptorPool pool;
            VkDescriptorPoolCreateInfo dpci{ VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO }; dpci.maxSets = 6; dpci.poolSizeCount = 2; dpci.pPoolSizes = psizes;
            vkCreateDescriptorPool(m_device, &dpci, nullptr, &pool);

            VkShaderModule vmod, fmod;
            create_shader_module("assets/shaders/generated/equirect_to_cubemap.vertexMain.spv", vmod);
            create_shader_module("assets/shaders/generated/equirect_to_cubemap.fragmentMain.spv", fmod);
            VkPipelineLayout playout;
            VkPipelineLayoutCreateInfo plci{ VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO }; plci.setLayoutCount = 1; plci.pSetLayouts = &set_layout;
            vkCreatePipelineLayout(m_device, &plci, nullptr, &playout);
            VkPipelineShaderStageCreateInfo stages[2]{};
            stages[0].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; stages[0].stage = VK_SHADER_STAGE_VERTEX_BIT;   stages[0].module = vmod; stages[0].pName = "main";
            stages[1].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; stages[1].stage = VK_SHADER_STAGE_FRAGMENT_BIT; stages[1].module = fmod; stages[1].pName = "main";
            VkVertexInputBindingDescription vib{ 0, 12, VK_VERTEX_INPUT_RATE_VERTEX };
            VkVertexInputAttributeDescription via{ 0, 0, VK_FORMAT_R32G32B32_SFLOAT, 0 };
            VkPipelineVertexInputStateCreateInfo vin{ VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO };
            vin.vertexBindingDescriptionCount = 1; vin.pVertexBindingDescriptions = &vib; vin.vertexAttributeDescriptionCount = 1; vin.pVertexAttributeDescriptions = &via;
            VkPipelineInputAssemblyStateCreateInfo ia{ VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO }; ia.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
            VkViewport vp{ 0, 0, (float)FACE, (float)FACE, 0, 1 }; VkRect2D sc{ { 0, 0 }, { FACE, FACE } };
            VkPipelineViewportStateCreateInfo vps{ VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO }; vps.viewportCount = 1; vps.pViewports = &vp; vps.scissorCount = 1; vps.pScissors = &sc;
            VkPipelineRasterizationStateCreateInfo rs{ VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO }; rs.polygonMode = VK_POLYGON_MODE_FILL; rs.cullMode = VK_CULL_MODE_NONE; rs.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE; rs.lineWidth = 1.0f;
            VkPipelineMultisampleStateCreateInfo ms{ VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO }; ms.rasterizationSamples = VK_SAMPLE_COUNT_1_BIT;
            VkPipelineColorBlendAttachmentState cba{}; cba.colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT | VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT;
            VkPipelineColorBlendStateCreateInfo cb{ VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO }; cb.attachmentCount = 1; cb.pAttachments = &cba;
            VkPipeline pipeline;
            VkGraphicsPipelineCreateInfo gpci{ VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO };
            gpci.stageCount = 2; gpci.pStages = stages; gpci.pVertexInputState = &vin; gpci.pInputAssemblyState = &ia; gpci.pViewportState = &vps;
            gpci.pRasterizationState = &rs; gpci.pMultisampleState = &ms; gpci.pColorBlendState = &cb; gpci.layout = playout; gpci.renderPass = rp; gpci.subpass = 0;
            vkCreateGraphicsPipelines(m_device, VK_NULL_HANDLE, 1, &gpci, nullptr, &pipeline);
            vkDestroyShaderModule(m_device, vmod, nullptr); vkDestroyShaderModule(m_device, fmod, nullptr);

            float cube_verts[] = {
                -1,1,-1, -1,-1,-1, 1,-1,-1, 1,-1,-1, 1,1,-1, -1,1,-1,
                -1,-1,1, -1,-1,-1, -1,1,-1, -1,1,-1, -1,1,1, -1,-1,1,
                 1,-1,-1, 1,-1,1, 1,1,1, 1,1,1, 1,1,-1, 1,-1,-1,
                -1,-1,1, -1,1,1, 1,1,1, 1,1,1, 1,-1,1, -1,-1,1,
                -1,1,-1, 1,1,-1, 1,1,1, 1,1,1, -1,1,1, -1,1,-1,
                -1,-1,-1, -1,-1,1, 1,-1,-1, 1,-1,-1, -1,-1,1, 1,-1,1,
            };
            VkBuffer cube_vb; VkDeviceMemory cube_vb_mem;
            create_buffer_raw(sizeof(cube_verts), VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, host_vis, cube_vb, cube_vb_mem);
            vkMapMemory(m_device, cube_vb_mem, 0, sizeof(cube_verts), 0, &mp); std::memcpy(mp, cube_verts, sizeof(cube_verts)); vkUnmapMemory(m_device, cube_vb_mem);

            glm::mat4 proj = glm::perspective(glm::radians(90.0f), 1.0f, 0.1f, 10.0f);
            proj[1][1] *= -1.0f;
            glm::mat4 views[6] = {
                glm::lookAt(glm::vec3(0), glm::vec3( 1, 0, 0), glm::vec3(0, -1, 0)),
                glm::lookAt(glm::vec3(0), glm::vec3(-1, 0, 0), glm::vec3(0, -1, 0)),
                glm::lookAt(glm::vec3(0), glm::vec3( 0, 1, 0), glm::vec3(0, 0, 1)),
                glm::lookAt(glm::vec3(0), glm::vec3( 0, -1, 0), glm::vec3(0, 0, -1)),
                glm::lookAt(glm::vec3(0), glm::vec3( 0, 0, 1), glm::vec3(0, -1, 0)),
                glm::lookAt(glm::vec3(0), glm::vec3( 0, 0, -1), glm::vec3(0, -1, 0)),
            };
            VkBuffer ubo[6]; VkDeviceMemory ubo_mem[6]; VkDescriptorSet sets[6];
            for (uint32_t i = 0; i < 6; ++i)
            {
                create_buffer_raw(128, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT, host_vis, ubo[i], ubo_mem[i]);
                glm::mat4 mats[2] = { glm::transpose(proj), glm::transpose(views[i]) };
                vkMapMemory(m_device, ubo_mem[i], 0, 128, 0, &mp); std::memcpy(mp, mats, 128); vkUnmapMemory(m_device, ubo_mem[i]);
                VkDescriptorSetAllocateInfo dsai{ VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO }; dsai.descriptorPool = pool; dsai.descriptorSetCount = 1; dsai.pSetLayouts = &set_layout;
                vkAllocateDescriptorSets(m_device, &dsai, &sets[i]);
                VkDescriptorBufferInfo buf_info{ ubo[i], 0, VK_WHOLE_SIZE };
                VkDescriptorImageInfo img_info{ eq_sampler, eq_view, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL };
                VkWriteDescriptorSet ws[2]{};
                ws[0].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; ws[0].dstSet = sets[i]; ws[0].dstBinding = 0; ws[0].descriptorCount = 1; ws[0].descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER; ws[0].pBufferInfo = &buf_info;
                ws[1].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; ws[1].dstSet = sets[i]; ws[1].dstBinding = 1; ws[1].descriptorCount = 1; ws[1].descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER; ws[1].pImageInfo = &img_info;
                vkUpdateDescriptorSets(m_device, 2, ws, 0, nullptr);
            }

            VkCommandBufferAllocateInfo cbai{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO };
            cbai.commandPool = m_command_pool; cbai.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY; cbai.commandBufferCount = 1;
            VkCommandBuffer cmd; vkAllocateCommandBuffers(m_device, &cbai, &cmd);
            VkCommandBufferBeginInfo bi{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO }; bi.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
            vkBeginCommandBuffer(cmd, &bi);
            VkImageMemoryBarrier to_dst{ VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER };
            to_dst.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED; to_dst.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
            to_dst.image = eq_image; to_dst.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
            to_dst.srcAccessMask = 0; to_dst.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
            vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &to_dst);
            VkBufferImageCopy copy{}; copy.imageSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 }; copy.imageExtent = { (uint32_t)w, (uint32_t)h, 1 };
            vkCmdCopyBufferToImage(cmd, eq_staging, eq_image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &copy);
            VkImageMemoryBarrier to_read = to_dst; to_read.oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL; to_read.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
            to_read.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT; to_read.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
            vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &to_read);

            VkClearValue clear{}; clear.color = { { 0, 0, 0, 1 } };
            for (uint32_t i = 0; i < 6; ++i)
            {
                VkRenderPassBeginInfo rpbi{ VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO };
                rpbi.renderPass = rp; rpbi.framebuffer = face_fb[i]; rpbi.renderArea = { { 0, 0 }, { FACE, FACE } }; rpbi.clearValueCount = 1; rpbi.pClearValues = &clear;
                vkCmdBeginRenderPass(cmd, &rpbi, VK_SUBPASS_CONTENTS_INLINE);
                vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
                vkCmdBindDescriptorSets(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, playout, 0, 1, &sets[i], 0, nullptr);
                VkDeviceSize off = 0; vkCmdBindVertexBuffers(cmd, 0, 1, &cube_vb, &off);
                vkCmdDraw(cmd, 36, 1, 0, 0);
                vkCmdEndRenderPass(cmd);
            }

            {
                VkImageMemoryBarrier src0{ VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER };
                src0.image = tex->m_image; src0.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 6 };
                src0.oldLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; src0.newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
                src0.srcAccessMask = VK_ACCESS_SHADER_READ_BIT; src0.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
                vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &src0);
                int32_t mipW = (int32_t)FACE, mipH = (int32_t)FACE;
                for (uint32_t m = 1; m < CUBE_MIPS; ++m)
                {
                    int32_t nW = mipW > 1 ? mipW / 2 : 1, nH = mipH > 1 ? mipH / 2 : 1;
                    VkImageMemoryBarrier bd{ VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER };
                    bd.image = tex->m_image; bd.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, m, 1, 0, 6 };
                    bd.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED; bd.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
                    bd.srcAccessMask = 0; bd.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
                    vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &bd);
                    VkImageBlit blit{};
                    blit.srcOffsets[1] = { mipW, mipH, 1 }; blit.srcSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, m - 1, 0, 6 };
                    blit.dstOffsets[1] = { nW, nH, 1 };     blit.dstSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, m, 0, 6 };
                    vkCmdBlitImage(cmd, tex->m_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, tex->m_image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &blit, VK_FILTER_LINEAR);
                    VkImageMemoryBarrier bs{ VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER };
                    bs.image = tex->m_image; bs.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, m, 1, 0, 6 };
                    bs.oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL; bs.newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
                    bs.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT; bs.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
                    vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &bs);
                    mipW = nW; mipH = nH;
                }
                VkImageMemoryBarrier fin{ VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER };
                fin.image = tex->m_image; fin.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, CUBE_MIPS, 0, 6 };
                fin.oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL; fin.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
                fin.srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT; fin.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
                vkCmdPipelineBarrier(cmd, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &fin);
            }

            vkEndCommandBuffer(cmd);
            VkSubmitInfo si{ VK_STRUCTURE_TYPE_SUBMIT_INFO }; si.commandBufferCount = 1; si.pCommandBuffers = &cmd;
            vkQueueSubmit(m_graphics_queue, 1, &si, VK_NULL_HANDLE); vkQueueWaitIdle(m_graphics_queue);

            vkFreeCommandBuffers(m_device, m_command_pool, 1, &cmd);
            for (uint32_t i = 0; i < 6; ++i) { vkDestroyBuffer(m_device, ubo[i], nullptr); vkFreeMemory(m_device, ubo_mem[i], nullptr); vkDestroyFramebuffer(m_device, face_fb[i], nullptr); vkDestroyImageView(m_device, face_views[i], nullptr); }
            vkDestroyBuffer(m_device, cube_vb, nullptr); vkFreeMemory(m_device, cube_vb_mem, nullptr);
            vkDestroyPipeline(m_device, pipeline, nullptr); vkDestroyPipelineLayout(m_device, playout, nullptr);
            vkDestroyDescriptorPool(m_device, pool, nullptr); vkDestroyDescriptorSetLayout(m_device, set_layout, nullptr);
            vkDestroyRenderPass(m_device, rp, nullptr);
            vkDestroySampler(m_device, eq_sampler, nullptr); vkDestroyImageView(m_device, eq_view, nullptr);
            vkDestroyImage(m_device, eq_image, nullptr); vkFreeMemory(m_device, eq_mem, nullptr);
            vkDestroyBuffer(m_device, eq_staging, nullptr); vkFreeMemory(m_device, eq_staging_mem, nullptr);
            DONUT_INFO("Vulkan RHI: HDRI cubemap built from {} ({}x{} equirect -> {}^2 cube)", path, w, h, (int)FACE);
            return tex;
        }

        auto VulkanDevice::create_pipeline(const PipelineDesc& desc) -> Ref<Pipeline>
        {
            auto p = create_ref<VkPipelineR>(); p->m_device = m_device; p->m_resources = desc.resources;

            std::vector<VkDescriptorSetLayoutBinding> binds;
            for (const auto& r : desc.resources)
            {
                VkDescriptorSetLayoutBinding b{};
                b.binding = r.binding; b.descriptorCount = 1;
                b.descriptorType = r.kind == ResourceKind::UniformBuffer ? VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER : VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
                b.stageFlags = VK_SHADER_STAGE_VERTEX_BIT | VK_SHADER_STAGE_FRAGMENT_BIT;
                binds.push_back(b);
            }
            VkDescriptorSetLayoutCreateInfo dslci{ VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO };
            dslci.bindingCount = (uint32_t)binds.size(); dslci.pBindings = binds.data();
            vkCreateDescriptorSetLayout(m_device, &dslci, nullptr, &p->m_set_layout);
            VkPipelineLayoutCreateInfo plci{ VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO };
            plci.setLayoutCount = 1; plci.pSetLayouts = &p->m_set_layout;
            vkCreatePipelineLayout(m_device, &plci, nullptr, &p->m_layout);

            VkShaderModule vmod = VK_NULL_HANDLE, fmod = VK_NULL_HANDLE;
            if (!create_shader_module("assets/shaders/generated/" + desc.shader + ".vertexMain.spv", vmod) ||
                !create_shader_module("assets/shaders/generated/" + desc.shader + ".fragmentMain.spv", fmod))
            { DONUT_ERROR("Vulkan RHI: shader '{}' modules failed", desc.shader); return p; }
            VkPipelineShaderStageCreateInfo stages[2]{};
            stages[0].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; stages[0].stage = VK_SHADER_STAGE_VERTEX_BIT;   stages[0].module = vmod; stages[0].pName = "main";
            stages[1].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; stages[1].stage = VK_SHADER_STAGE_FRAGMENT_BIT; stages[1].module = fmod; stages[1].pName = "main";

            VkVertexInputBindingDescription vib{ 0, desc.vertex_layout.stride, VK_VERTEX_INPUT_RATE_VERTEX };
            std::vector<VkVertexInputAttributeDescription> vias;
            for (const auto& a : desc.vertex_layout.attributes)
                vias.push_back({ a.location, 0, vk_attr_format(a.components), a.offset });
            VkPipelineVertexInputStateCreateInfo vin{ VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO };
            vin.vertexBindingDescriptionCount = desc.vertex_layout.stride ? 1 : 0; vin.pVertexBindingDescriptions = &vib;
            vin.vertexAttributeDescriptionCount = (uint32_t)vias.size(); vin.pVertexAttributeDescriptions = vias.data();

            VkPipelineInputAssemblyStateCreateInfo ia{ VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO }; ia.topology = vk_topology(desc.topology);
            VkPipelineViewportStateCreateInfo vps{ VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO }; vps.viewportCount = 1; vps.scissorCount = 1;
            VkDynamicState dyn[2] = { VK_DYNAMIC_STATE_VIEWPORT, VK_DYNAMIC_STATE_SCISSOR };
            VkPipelineDynamicStateCreateInfo dsci{ VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO }; dsci.dynamicStateCount = 2; dsci.pDynamicStates = dyn;
            VkPipelineRasterizationStateCreateInfo rs{ VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO };
            rs.polygonMode = VK_POLYGON_MODE_FILL; rs.cullMode = vk_cull(desc.cull); rs.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE; rs.lineWidth = 1.0f;
            VkPipelineMultisampleStateCreateInfo ms{ VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO }; ms.rasterizationSamples = VK_SAMPLE_COUNT_1_BIT;
            VkPipelineDepthStencilStateCreateInfo ds{ VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO };
            ds.depthTestEnable = desc.depth_test ? VK_TRUE : VK_FALSE; ds.depthWriteEnable = desc.depth_write ? VK_TRUE : VK_FALSE; ds.depthCompareOp = vk_compare(desc.depth_op);
            VkPipelineColorBlendAttachmentState cba{};
            cba.colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT | VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT;
            if (desc.blend == BlendMode::AlphaBlend)
            {
                cba.blendEnable = VK_TRUE;
                cba.srcColorBlendFactor = VK_BLEND_FACTOR_SRC_ALPHA; cba.dstColorBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA; cba.colorBlendOp = VK_BLEND_OP_ADD;
                cba.srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE; cba.dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO; cba.alphaBlendOp = VK_BLEND_OP_ADD;
            }
            VkPipelineColorBlendStateCreateInfo cb{ VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO }; cb.attachmentCount = 1; cb.pAttachments = &cba;

            VkGraphicsPipelineCreateInfo gpci{ VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO };
            gpci.stageCount = 2; gpci.pStages = stages;
            gpci.pVertexInputState = &vin; gpci.pInputAssemblyState = &ia; gpci.pViewportState = &vps;
            gpci.pRasterizationState = &rs; gpci.pMultisampleState = &ms; gpci.pColorBlendState = &cb; gpci.pDynamicState = &dsci;
            if (desc.has_depth) gpci.pDepthStencilState = &ds;
            gpci.layout = p->m_layout;
            gpci.renderPass = desc.has_depth ? m_swapchain_rp : m_offscreen_rp;
            gpci.subpass = 0;
            VkResult pr = vkCreateGraphicsPipelines(m_device, VK_NULL_HANDLE, 1, &gpci, nullptr, &p->m_pipeline);
            vkDestroyShaderModule(m_device, vmod, nullptr); vkDestroyShaderModule(m_device, fmod, nullptr);
            if (pr != VK_SUCCESS) DONUT_ERROR("Vulkan RHI: pipeline '{}' creation failed ({})", desc.shader, (int)pr);
            return p;
        }

        auto VulkanDevice::begin_frame(const glm::vec4&) -> CommandList*
        {
            if (m_device == VK_NULL_HANDLE) return nullptr;
            vkWaitForFences(m_device, 1, &m_in_flight[m_current_frame], VK_TRUE, UINT64_MAX);
            VkResult r = vkAcquireNextImageKHR(m_device, m_swapchain, UINT64_MAX, m_image_available[m_current_frame], VK_NULL_HANDLE, &m_image_index);
            if (r == VK_ERROR_OUT_OF_DATE_KHR) { recreate_swapchain(); return nullptr; }
            if (r != VK_SUCCESS && r != VK_SUBOPTIMAL_KHR) { DONUT_ERROR("Vulkan RHI: acquire failed ({})", (int)r); return nullptr; }
            if (m_images_in_flight[m_image_index] != VK_NULL_HANDLE)
                vkWaitForFences(m_device, 1, &m_images_in_flight[m_image_index], VK_TRUE, UINT64_MAX);
            m_images_in_flight[m_image_index] = m_in_flight[m_current_frame];
            if (m_geo_in_use != VK_NULL_HANDLE)
                vkWaitForFences(m_device, 1, &m_geo_in_use, VK_TRUE, UINT64_MAX);

            vkResetDescriptorPool(m_device, m_frame_pools[m_current_frame], 0);
            VkCommandBuffer cmd = m_command_buffers[m_current_frame];
            vkResetCommandBuffer(cmd, 0);
            VkCommandBufferBeginInfo bi{ VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO };
            vkBeginCommandBuffer(cmd, &bi);

            m_cmds.m_device = m_device; m_cmds.m_cmd = cmd;
            m_cmds.m_swapchain_rp = m_swapchain_rp; m_cmds.m_offscreen_rp = m_offscreen_rp;
            m_cmds.m_swapchain_fb = m_framebuffers[m_image_index]; m_cmds.m_extent = m_extent;
            m_cmds.m_frame_pool = m_frame_pools[m_current_frame]; m_cmds.m_pipe = nullptr;
            return &m_cmds;
        }

        auto VulkanDevice::end_frame() -> void
        {
            VkCommandBuffer cmd = m_command_buffers[m_current_frame];
            vkEndCommandBuffer(cmd);
            VkPipelineStageFlags wait_stage = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
            VkSubmitInfo submit{ VK_STRUCTURE_TYPE_SUBMIT_INFO };
            submit.waitSemaphoreCount = 1; submit.pWaitSemaphores = &m_image_available[m_current_frame]; submit.pWaitDstStageMask = &wait_stage;
            submit.commandBufferCount = 1; submit.pCommandBuffers = &cmd;
            submit.signalSemaphoreCount = 1; submit.pSignalSemaphores = &m_render_finished[m_image_index];
            vkResetFences(m_device, 1, &m_in_flight[m_current_frame]);
            if (vkQueueSubmit(m_graphics_queue, 1, &submit, m_in_flight[m_current_frame]) != VK_SUCCESS)
            { DONUT_ERROR("Vulkan RHI: queue submit failed"); return; }
            m_geo_in_use = m_in_flight[m_current_frame];

            VkPresentInfoKHR present{ VK_STRUCTURE_TYPE_PRESENT_INFO_KHR };
            present.waitSemaphoreCount = 1; present.pWaitSemaphores = &m_render_finished[m_image_index];
            present.swapchainCount = 1; present.pSwapchains = &m_swapchain; present.pImageIndices = &m_image_index;
            VkResult r = vkQueuePresentKHR(m_present_queue, &present);
            if (r == VK_ERROR_OUT_OF_DATE_KHR || r == VK_SUBOPTIMAL_KHR || m_framebuffer_resized)
            { m_framebuffer_resized = false; recreate_swapchain(); }
            m_current_frame = (m_current_frame + 1) % MAX_FRAMES_IN_FLIGHT;
        }

        auto VulkanDevice::init_imgui() -> void
        {
            VkDescriptorPoolSize pool_size{ VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1000 };
            VkDescriptorPoolCreateInfo dpci{ VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO };
            dpci.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT; dpci.maxSets = 1000;
            dpci.poolSizeCount = 1; dpci.pPoolSizes = &pool_size;
            vkCreateDescriptorPool(m_device, &dpci, nullptr, &m_imgui_pool);

            IMGUI_CHECKVERSION(); ImGui::CreateContext();
            ImGui::GetIO().ConfigFlags |= ImGuiConfigFlags_NavEnableKeyboard;
            ImGui::GetIO().ConfigFlags |= ImGuiConfigFlags_DockingEnable;
            ImGui::StyleColorsDark();
            ImGui_ImplGlfw_InitForVulkan(m_window, true);
            ImGui_ImplVulkan_InitInfo info{};
            info.ApiVersion = VK_API_VERSION_1_2; info.Instance = m_instance; info.PhysicalDevice = m_physical;
            info.Device = m_device; info.QueueFamily = m_graphics_family; info.Queue = m_graphics_queue;
            info.DescriptorPool = m_imgui_pool; info.RenderPass = m_swapchain_rp;
            info.MinImageCount = 2; info.ImageCount = (uint32_t)m_images.size(); info.MSAASamples = VK_SAMPLE_COUNT_1_BIT;
            if (!ImGui_ImplVulkan_Init(&info)) { DONUT_ERROR("Vulkan RHI: ImGui_ImplVulkan_Init failed"); return; }
            m_imgui = true;
            DONUT_INFO("Vulkan RHI: ImGui backend initialized");
        }

        auto VulkanDevice::imgui_new_frame() -> void
        {
            if (!m_imgui) return;
            ImGui_ImplVulkan_NewFrame(); ImGui_ImplGlfw_NewFrame(); ImGui::NewFrame();
        }

        auto VulkanDevice::imgui_render(CommandList& cmds) -> void
        {
            if (!m_imgui) return;
            ImGui::Render();
            ImGui_ImplVulkan_RenderDrawData(ImGui::GetDrawData(), static_cast<VkCommandListR&>(cmds).m_cmd);
        }

        auto VulkanDevice::shutdown() -> void
        {
            if (m_device == VK_NULL_HANDLE)
            {
                if (m_instance && m_surface) { vkDestroySurfaceKHR(m_instance, m_surface, nullptr); m_surface = VK_NULL_HANDLE; }
                if (m_instance) { vkDestroyInstance(m_instance, nullptr); m_instance = VK_NULL_HANDLE; }
                return;
            }
            vkDeviceWaitIdle(m_device);
            if (m_imgui) { ImGui_ImplVulkan_Shutdown(); ImGui_ImplGlfw_Shutdown(); ImGui::DestroyContext(); m_imgui = false; }
            if (m_imgui_pool) vkDestroyDescriptorPool(m_device, m_imgui_pool, nullptr);
            for (auto p : m_frame_pools) vkDestroyDescriptorPool(m_device, p, nullptr);
            m_frame_pools.clear();
            for (auto s : m_render_finished) vkDestroySemaphore(m_device, s, nullptr);
            for (auto s : m_image_available) vkDestroySemaphore(m_device, s, nullptr);
            for (auto f : m_in_flight)       vkDestroyFence(m_device, f, nullptr);
            m_render_finished.clear(); m_image_available.clear(); m_in_flight.clear();
            if (m_command_pool) vkDestroyCommandPool(m_device, m_command_pool, nullptr);
            if (m_offscreen_rp) vkDestroyRenderPass(m_device, m_offscreen_rp, nullptr);
            if (m_swapchain_rp) vkDestroyRenderPass(m_device, m_swapchain_rp, nullptr);
            cleanup_swapchain();
            vkDestroyDevice(m_device, nullptr); m_device = VK_NULL_HANDLE;
            if (m_surface)  vkDestroySurfaceKHR(m_instance, m_surface, nullptr);
            if (m_instance) vkDestroyInstance(m_instance, nullptr);
            m_surface = VK_NULL_HANDLE; m_instance = VK_NULL_HANDLE;
        }
    }

    auto create_vulkan_device() -> Scope<Device> { return create_scope<VulkanDevice>(); }

    auto vulkan_prepare_glfw() -> void
    {
#ifdef __APPLE__
        if (!getenv("VK_ICD_FILENAMES"))
            setenv("VK_ICD_FILENAMES", "/opt/homebrew/etc/vulkan/icd.d/MoltenVK_icd.json", 0);
        if (!getenv("VK_LAYER_PATH"))
            setenv("VK_LAYER_PATH", "/opt/homebrew/share/vulkan/explicit_layer.d", 0);
        if (!getenv("DYLD_LIBRARY_PATH"))
            setenv("DYLD_LIBRARY_PATH", "/opt/homebrew/lib", 0);
#endif
        glfwInitVulkanLoader(vkGetInstanceProcAddr);
    }
}