// =============================================
//  Aster: asset_loader.cpp
//  Copyright (c) 2020-2024 Anish Bhobe
// =============================================

#define TINYGLTF_NOEXCEPTION
#define JSON_NOEXCEPTION

#define TINYGLTF_IMPLEMENTATION
#define STB_IMAGE_IMPLEMENTATION
#define STB_IMAGE_WRITE_IMPLEMENTATION

#include "asset_loader.h"

#include "EASTL/fixed_vector.h"
#include "buffer.h"
#include "device.h"
#include "gpu_resource_manager.h"
#include "helpers.h"
#include "image.h"

#include <EASTL/hash_map.h>
#include <glm/gtc/type_ptr.hpp>

#include <tiny_gltf.h>

#if defined(LoadImage)
#undef LoadImage
#endif

constexpr vk::CommandBufferBeginInfo OneTimeCmdBeginInfo = {.flags = vk::CommandBufferUsageFlagBits::eOneTimeSubmit};

vec4
VectorToVec4(const std::vector<double> &vec)
{
    if (vec.empty())
    {
        return vec4{0.0f};
    }
    assert(vec.size() == 4);

    return {vec[0], vec[1], vec[2], vec[3]};
}

vec3
VectorToVec3(const std::vector<double> &vec)
{
    if (vec.empty())
    {
        return vec3{0.0f};
    }
    assert(vec.size() == 3);

    return {vec[0], vec[1], vec[2]};
}

void
AssetLoader::LoadHdrImage(Texture *texture, cstr path, cstr name) const
{
    const Device *pDevice = m_ResourceManager->m_Device;
    ERROR_IF(texture->IsValid(), "Expected invalid image.") THEN_ABORT(-1);

    i32 x, y, nChannels;
    const f32 *data = stbi_loadf(path, &x, &y, &nChannels, 4);
    assert(nChannels == 3);

    ERROR_IF(!data, "Could not load {}", path) THEN_ABORT(-1);

    u32 width = Cast<u32>(x);
    u32 height = Cast<u32>(y);

    StagingBuffer stagingBuffer;
    texture->Init(m_ResourceManager->m_Device, {width, height}, vk::Format::eR32G32B32A32Sfloat, false, path);
    assert(texture->IsValid());
    stagingBuffer.Init(m_ResourceManager->m_Device, (sizeof *data) * x * y * 4, "HDR Staging Buffer");
    stagingBuffer.Write(m_ResourceManager->m_Device, 0, stagingBuffer.GetSize(), data);

#pragma region Setup Copy/Sync primitives
    vk::BufferImageCopy2 copyRegion = {
        .bufferOffset = 0,
        .bufferRowLength = width,
        .bufferImageHeight = height,
        .imageSubresource =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .mipLevel = 0,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
        .imageOffset = {0, 0, 0},
        .imageExtent = texture->m_Extent,
    };
    vk::CopyBufferToImageInfo2 stagingInfo = {
        .srcBuffer = stagingBuffer.m_Buffer,
        .dstImage = texture->m_Image,
        .dstImageLayout = vk::ImageLayout::eTransferDstOptimal,
        .regionCount = 1,
        .pRegions = &copyRegion,
    };
    vk::ImageMemoryBarrier2 readyToStageBarrier = {
        .srcStageMask = vk::PipelineStageFlagBits2::eAllCommands,
        .srcAccessMask = vk::AccessFlagBits2::eNone,
        .dstStageMask = vk::PipelineStageFlagBits2::eAllTransfer,
        .dstAccessMask = vk::AccessFlagBits2::eTransferWrite,
        .oldLayout = vk::ImageLayout::eUndefined,
        .newLayout = vk::ImageLayout::eTransferDstOptimal,
        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .image = texture->m_Image,
        .subresourceRange =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseMipLevel = 0,
                .levelCount = 1,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };
    vk::DependencyInfo readyToStageDependency = {
        .memoryBarrierCount = 0,
        .bufferMemoryBarrierCount = 0,
        .imageMemoryBarrierCount = 1,
        .pImageMemoryBarriers = &readyToStageBarrier,
    };
    vk::ImageMemoryBarrier2 postStagingBarrier = {
        .srcStageMask = vk::PipelineStageFlagBits2::eAllTransfer,
        .srcAccessMask = vk::AccessFlagBits2::eTransferWrite,
        .dstStageMask = vk::PipelineStageFlagBits2::eFragmentShader | vk::PipelineStageFlagBits2::eComputeShader,
        .dstAccessMask = vk::AccessFlagBits2::eShaderRead,
        .oldLayout = vk::ImageLayout::eTransferDstOptimal,
        .newLayout = vk::ImageLayout::eShaderReadOnlyOptimal,
        .srcQueueFamilyIndex = m_TransferQueueIndex,
        .dstQueueFamilyIndex = m_GraphicsQueueIndex,
        .image = texture->m_Image,
        .subresourceRange =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseMipLevel = 0,
                .levelCount = 1,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };
    vk::DependencyInfo postStagingDependency = {
        .memoryBarrierCount = 0,
        .bufferMemoryBarrierCount = 0,
        .imageMemoryBarrierCount = 1,
        .pImageMemoryBarriers = &postStagingBarrier,
    };
#pragma endregion

    AbortIfFailed(m_CommandBuffer.begin(&OneTimeCmdBeginInfo));

#if !defined(ASTER_NDEBUG)
    StackString<128> loadActionName = "Load: ";
    loadActionName += name ? name : path;
    vk::DebugUtilsLabelEXT debugLabel = {
        .pLabelName = loadActionName.c_str(),
        .color = std::array{1.0f, 1.0f, 1.0f, 1.0f},
    };
    m_CommandBuffer.beginDebugUtilsLabelEXT(&debugLabel);
#endif

    m_CommandBuffer.pipelineBarrier2(&readyToStageDependency);
    m_CommandBuffer.copyBufferToImage2(&stagingInfo);
    m_CommandBuffer.pipelineBarrier2(&postStagingDependency);

#if !defined(ASTER_NDEBUG)
    m_CommandBuffer.endDebugUtilsLabelEXT();
#endif

    AbortIfFailed(m_CommandBuffer.end());

    vk::SubmitInfo submitInfo = {
        .waitSemaphoreCount = 0,
        .pWaitDstStageMask = nullptr,
        .commandBufferCount = 1,
        .pCommandBuffers = &m_CommandBuffer,
    };

    vk::Fence fence;
    vk::FenceCreateInfo fenceCreateInfo = {};
    AbortIfFailed(pDevice->m_Device.createFence(&fenceCreateInfo, nullptr, &fence));
    AbortIfFailed(m_TransferQueue.submit(1, &submitInfo, fence));
    AbortIfFailed(pDevice->m_Device.waitForFences(1, &fence, true, MaxValue<u32>));
    pDevice->m_Device.destroy(fence, nullptr);

    AbortIfFailed(pDevice->m_Device.resetCommandPool(m_CommandPool, {}));

    stagingBuffer.Destroy(pDevice);
}

void
GenerateMipMaps(vk::CommandBuffer commandBuffer, Texture *texture, vk::ImageLayout initialLayout,
                vk::ImageLayout finalLayout, vk::PipelineStageFlags2 prevStage, vk::PipelineStageFlags2 finalStage)
{
#if !defined(ASTER_NDEBUG)
    vk::DebugUtilsLabelEXT label = {
        .pLabelName = "Generate Mipmap",
        .color = std::array{0.9f, 0.9f, 0.9f, 1.0f},
    };
    commandBuffer.beginDebugUtilsLabelEXT(&label);
#endif

    vk::ImageMemoryBarrier2 imageStartBarrier = {
        .srcStageMask = prevStage,
        .srcAccessMask = vk::AccessFlagBits2::eNone,
        .dstStageMask = vk::PipelineStageFlagBits2::eTransfer,
        .dstAccessMask = vk::AccessFlagBits2::eTransferRead,
        .oldLayout = initialLayout,
        .newLayout = vk::ImageLayout::eTransferSrcOptimal,
        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .image = texture->m_Image,
        .subresourceRange =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseMipLevel = 0,
                .levelCount = 1,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };
    vk::ImageMemoryBarrier2 mipsStartBarrier = imageStartBarrier;
    mipsStartBarrier.dstAccessMask = vk::AccessFlagBits2::eTransferWrite;
    mipsStartBarrier.oldLayout = vk::ImageLayout::eUndefined;
    mipsStartBarrier.newLayout = vk::ImageLayout::eTransferDstOptimal;
    mipsStartBarrier.subresourceRange = {
        .aspectMask = vk::ImageAspectFlagBits::eColor,
        .baseMipLevel = 1,
        .levelCount = texture->GetMipLevels() - 1,
        .baseArrayLayer = 0,
        .layerCount = 1,
    };
    eastl::fixed_vector<vk::ImageMemoryBarrier2, 2> startBarriers = {
        mipsStartBarrier,
    };
    if (initialLayout != imageStartBarrier.newLayout)
    {
        startBarriers.push_back(imageStartBarrier);
    }

    vk::DependencyInfo imageStartDependency = {
        .imageMemoryBarrierCount = Cast<u32>(startBarriers.size()),
        .pImageMemoryBarriers = startBarriers.data(),
    };

    vk::ImageMemoryBarrier2 nextMipBarrier = {
        .srcStageMask = vk::PipelineStageFlagBits2::eTransfer,
        .srcAccessMask = vk::AccessFlagBits2::eTransferWrite,
        .dstStageMask = vk::PipelineStageFlagBits2::eTransfer,
        .dstAccessMask = vk::AccessFlagBits2::eTransferRead,
        .oldLayout = vk::ImageLayout::eTransferDstOptimal,
        .newLayout = vk::ImageLayout::eTransferSrcOptimal,
        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .image = texture->m_Image,
        .subresourceRange =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseMipLevel = 0,
                .levelCount = 1,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };
    vk::DependencyInfo interMipDependency = {
        .imageMemoryBarrierCount = 1,
        .pImageMemoryBarriers = &nextMipBarrier,
    };

    vk::ImageMemoryBarrier2 imageReadyBarrier = {
        .srcStageMask = vk::PipelineStageFlagBits2::eTransfer,
        .srcAccessMask = vk::AccessFlagBits2::eTransferWrite,
        .dstStageMask = finalStage,
        .dstAccessMask = vk::AccessFlagBits2::eShaderRead,
        .oldLayout = vk::ImageLayout::eTransferSrcOptimal,
        .newLayout = finalLayout,
        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .image = texture->m_Image,
        .subresourceRange =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseMipLevel = 0,
                .levelCount = texture->GetMipLevels(),
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };
    vk::DependencyInfo imageReadyDependency = {
        .imageMemoryBarrierCount = 1,
        .pImageMemoryBarriers = &imageReadyBarrier,
    };

    vk::ImageBlit2 blitRegion = {
        .srcSubresource =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
        .dstSubresource =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };

    vk::BlitImageInfo2 mipBlitInfo = {
        .srcImage = texture->m_Image,
        .srcImageLayout = vk::ImageLayout::eTransferSrcOptimal,
        .dstImage = texture->m_Image,
        .dstImageLayout = vk::ImageLayout::eTransferDstOptimal,
        .regionCount = 1,
        .pRegions = &blitRegion,
        .filter = vk::Filter::eLinear,
    };

    auto calcNextMip = [](i32 prev) { return eastl::max(prev / 2, 1); };

    // Mip Mapping

    commandBuffer.pipelineBarrier2(&imageStartDependency);

    i32 prevMipWidth = Cast<i32>(texture->m_Extent.width);
    i32 prevMipHeight = Cast<i32>(texture->m_Extent.height);

    u32 maxPrevMip = texture->GetMipLevels() - 1;
    for (u32 prevMipLevel = 0; prevMipLevel < maxPrevMip; ++prevMipLevel)
    {
        i32 currentMipWidth = calcNextMip(prevMipWidth);
        i32 currentMipHeight = calcNextMip(prevMipHeight);
        u32 currentMipLevel = prevMipLevel + 1;

        blitRegion.srcSubresource.mipLevel = prevMipLevel;
        blitRegion.srcOffsets = std::array{
            vk::Offset3D{0, 0, 0},
            vk::Offset3D{prevMipWidth, prevMipHeight, 1},
        };
        blitRegion.dstSubresource.mipLevel = currentMipLevel;
        blitRegion.dstOffsets = std::array{
            vk::Offset3D{0, 0, 0},
            vk::Offset3D{currentMipWidth, currentMipHeight, 1},
        };

        nextMipBarrier.subresourceRange.baseMipLevel = currentMipLevel;

        commandBuffer.blitImage2(&mipBlitInfo);
        commandBuffer.pipelineBarrier2(&interMipDependency);

        prevMipHeight = currentMipHeight;
        prevMipWidth = currentMipWidth;
    }

    commandBuffer.pipelineBarrier2(&imageReadyDependency);
#if !defined(ASTER_NDEBUG)
    commandBuffer.endDebugUtilsLabelEXT();
#endif
}

TextureHandle
AssetLoader::LoadImageToGpu(StagingBuffer *stagingBuffer, tinygltf::Image *image, bool isSrgb) const
{
    assert(image->component == 4);
    assert(image->height > 0 && image->width > 0);

    u32 height = Cast<u32>(image->height);
    u32 width = Cast<u32>(image->width);

    vk::Format imageFormat = isSrgb ? vk::Format::eR8G8B8A8Srgb : vk::Format::eR8G8B8A8Unorm;

    Texture texture;

    usize byteSize = image->image.size();
    texture.Init(m_ResourceManager->m_Device, {.width = width, .height = height}, imageFormat, true,
                 image->name.data());
    stagingBuffer->Init(m_ResourceManager->m_Device, byteSize);
    stagingBuffer->Write(m_ResourceManager->m_Device, 0, byteSize, image->image.data());

#if !defined(ASTER_NDEBUG)
    StackString<128> loadActionName = "Load: ";
    loadActionName += image->name.empty() ? "<texture>" : image->name.c_str();
    vk::DebugUtilsLabelEXT debugLabel = {
        .pLabelName = loadActionName.c_str(),
        .color = std::array{1.0f, 1.0f, 1.0f, 1.0f},
    };
    m_CommandBuffer.beginDebugUtilsLabelEXT(&debugLabel);
#endif

#pragma region Barriers and Blits

    vk::ImageMemoryBarrier2 imageStartBarrier = {
        .srcStageMask = vk::PipelineStageFlagBits2::eTopOfPipe,
        .srcAccessMask = vk::AccessFlagBits2::eNone,
        .dstStageMask = vk::PipelineStageFlagBits2::eTransfer,
        .dstAccessMask = vk::AccessFlagBits2::eTransferWrite,
        .oldLayout = vk::ImageLayout::eUndefined,
        .newLayout = vk::ImageLayout::eTransferDstOptimal,
        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .image = texture.m_Image,
        .subresourceRange =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseMipLevel = 0,
                .levelCount = 1,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };
    vk::DependencyInfo imageStartDependency = {
        .memoryBarrierCount = 0,
        .bufferMemoryBarrierCount = 0,
        .imageMemoryBarrierCount = 1,
        .pImageMemoryBarriers = &imageStartBarrier,
    };

    vk::ImageMemoryBarrier2 postStagingBarrier = {
        .srcStageMask = vk::PipelineStageFlagBits2::eAllTransfer,
        .dstStageMask = vk::PipelineStageFlagBits2::eAllTransfer,
        .oldLayout = vk::ImageLayout::eTransferDstOptimal,
        .newLayout = vk::ImageLayout::eTransferSrcOptimal,
        .srcQueueFamilyIndex = m_TransferQueueIndex,
        .dstQueueFamilyIndex = m_GraphicsQueueIndex,
        .image = texture.m_Image,
        .subresourceRange =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .baseMipLevel = 0,
                .levelCount = 1,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
    };
    ;

    vk::DependencyInfo postStagingDependency = {
        .imageMemoryBarrierCount = 1,
        .pImageMemoryBarriers = &postStagingBarrier,
    };

    vk::BufferImageCopy2 imageCopy = {
        .bufferOffset = 0,
        .bufferRowLength = Cast<u32>(image->width),
        .bufferImageHeight = Cast<u32>(image->height),
        .imageSubresource =
            {
                .aspectMask = vk::ImageAspectFlagBits::eColor,
                .mipLevel = 0,
                .baseArrayLayer = 0,
                .layerCount = 1,
            },
        .imageOffset = {},
        .imageExtent = texture.m_Extent,
    };
    vk::CopyBufferToImageInfo2 stagingCopyInfo = {
        .srcBuffer = stagingBuffer->m_Buffer,
        .dstImage = texture.m_Image,
        .dstImageLayout = vk::ImageLayout::eTransferDstOptimal,
        .regionCount = 1,
        .pRegions = &imageCopy,
    };

#pragma endregion

    m_CommandBuffer.pipelineBarrier2(&imageStartDependency);
    m_CommandBuffer.copyBufferToImage2(&stagingCopyInfo);
    m_CommandBuffer.pipelineBarrier2(&postStagingDependency);

    GenerateMipMaps(m_CommandBuffer, &texture, vk::ImageLayout::eTransferSrcOptimal,
                    vk::ImageLayout::eShaderReadOnlyOptimal);

#if !defined(ASTER_NDEBUG)
    m_CommandBuffer.endDebugUtilsLabelEXT();
#endif

    return m_ResourceManager->CommitTexture(&texture);
}

Model
AssetLoader::LoadModelToGpu(cstr path, cstr name)
{
    namespace fs = std::filesystem;
    tinygltf::Model model;
    tinygltf::TinyGLTF loader;

    const Device *pDevice = m_ResourceManager->m_Device;

    const auto fsPath = fs::absolute(path);
    const auto ext = fsPath.extension();
    if (ext == GLTF_ASCII_FILE_EXTENSION)
    {
        std::string err;
        std::string warn;
        if (loader.LoadASCIIFromFile(&model, &err, &warn, fsPath.generic_string()))
        {
            ERROR_IF(!err.empty(), "{}", err)
            ELSE_IF_WARN(!warn.empty(), "{}", warn);
        }
    }
    if (ext == GLTF_BINARY_FILE_EXTENSION)
    {
        std::string err;
        std::string warn;
        if (loader.LoadBinaryFromFile(&model, &err, &warn, fsPath.generic_string()))
        {
            ERROR_IF(!err.empty(), "{}", err)
            ELSE_IF_WARN(!warn.empty(), "{}", warn);
        }
    }

    AbortIfFailed(m_CommandBuffer.begin(&OneTimeCmdBeginInfo));

#if !defined(ASTER_NDEBUG)
    StackString<128> loadActionName = "Load: ";
    loadActionName += name ? name : path;
    vk::DebugUtilsLabelEXT debugLabel = {
        .pLabelName = loadActionName.c_str(),
        .color = std::array{1.0f, 1.0f, 1.0f, 1.0f},
    };
    m_CommandBuffer.beginDebugUtilsLabelEXT(&debugLabel);
#endif

    eastl::vector<StagingBuffer> stagingBuffers;

    eastl::hash_map<i32, TextureHandle> textureHandleMap;

    eastl::vector<Material> materials;
    StorageBuffer materialsBuffer;
    BufferHandle materialsHandle;

    if (!model.materials.empty())
    {
        auto getTextureHandle = [this, &textureHandleMap, &stagingBuffers, &model](i32 index,
                                                                                   bool isSrgb) -> TextureHandle {
            if (index < 0)
            {
                return {};
            }
            const auto iter = textureHandleMap.find(index);
            if (iter != textureHandleMap.end())
            {
                return iter->second;
            }

            auto *image = &model.images[index];
            TextureHandle handle = LoadImageToGpu(&stagingBuffers.push_back(), image, isSrgb);
            textureHandleMap.emplace(index, handle);
            return handle;
        };

        materials.reserve(model.materials.size());
        for (auto &material : model.materials)
        {
            materials.push_back({
                .m_AlbedoFactor = VectorToVec4(material.pbrMetallicRoughness.baseColorFactor),
                .m_EmissionFactor = VectorToVec3(material.emissiveFactor),
                .m_MetalFactor = Cast<f32>(material.pbrMetallicRoughness.metallicFactor),
                .m_RoughFactor = Cast<f32>(material.pbrMetallicRoughness.roughnessFactor),
                .m_AlbedoTex = getTextureHandle(material.pbrMetallicRoughness.baseColorTexture.index, true),
                .m_NormalTex = getTextureHandle(material.normalTexture.index, false),
                .m_MetalRoughTex =
                    getTextureHandle(material.pbrMetallicRoughness.metallicRoughnessTexture.index, false),
                .m_OcclusionTex = getTextureHandle(material.occlusionTexture.index, false),
                .m_EmissionTex = getTextureHandle(material.emissiveTexture.index, true),
            });
        }

        usize materialsByteSize = materials.size() * sizeof materials[0];
        materialsBuffer.Init(pDevice, materialsByteSize, false, name);
        materialsHandle = m_ResourceManager->Commit(&materialsBuffer);

        StagingBuffer &materialStaging = stagingBuffers.push_back();
        materialStaging.Init(pDevice, materialsByteSize);
        materialStaging.Write(pDevice, 0, materialsByteSize, materials.data());

        vk::BufferCopy bufferCopy = {.srcOffset = 0, .dstOffset = 0, .size = materialsByteSize};
        m_CommandBuffer.copyBuffer(materialStaging.m_Buffer, materialsBuffer.m_Buffer, 1, &bufferCopy);
    }

    // TODO: Mesh reordering based on nodes AND OR meshoptimizer
    // TODO: Support scenes

    eastl::vector<vec4> vertexPositions;
    eastl::vector<VertexData> vertexData;
    eastl::vector<u32> indices;
    eastl::vector<MeshPrimitive> meshPrimitives;
    meshPrimitives.reserve(model.meshes.size());

    // Offset, Count
    eastl::vector<eastl::pair<usize, usize>> meshPrimRanges;
    meshPrimRanges.reserve(model.meshes.size());

    u32 vertexOffset = 0;
    u32 indexOffset = 0;

    for (auto &mesh : model.meshes)
    {
        meshPrimRanges.emplace_back(meshPrimitives.size(), mesh.primitives.size());
        for (auto &prim : mesh.primitives)
        {
            u32 vertexCount = 0;
            u32 indexCount = 0;

#pragma region Position
            assert(prim.attributes.contains(APosition));
            assert(prim.mode == TINYGLTF_MODE_TRIANGLES);
            {
                tinygltf::Accessor *posAccessor = &model.accessors[prim.attributes[APosition]];

                assert(posAccessor->count <= MaxValue<u32>);

                tinygltf::BufferView *posBufferView = &model.bufferViews[posAccessor->bufferView];
                tinygltf::Buffer *posBuffer = &model.buffers[posBufferView->buffer];
                usize byteOffset = (posAccessor->byteOffset + posBufferView->byteOffset);

                vertexCount = Cast<u32>(posAccessor->count);
                vertexPositions.reserve(vertexOffset + vertexCount);

                if (posAccessor->type == TINYGLTF_TYPE_VEC4)
                {
                    vec4 *data = Recast<vec4 *>(posBuffer->data.data() + byteOffset);
                    vertexPositions.insert(vertexPositions.end(), data, data + vertexCount);
                }
                else if (posAccessor->type == TINYGLTF_TYPE_VEC3)
                {
                    vec3 *data = Recast<vec3 *>(posBuffer->data.data() + byteOffset);
                    for (u32 i = 0; i < vertexCount; ++i)
                    {
                        vertexPositions.push_back(vec4(data[i], 1.0f));
                    }
                }
                else if (posAccessor->type == TINYGLTF_TYPE_VEC2)
                {
                    vec2 *data = Recast<vec2 *>(posBuffer->data.data() + byteOffset);
                    for (u32 i = 0; i < vertexCount; ++i)
                    {
                        vertexPositions.push_back(vec4(data[i], 0.0f, 1.0f));
                    }
                }
            }
#pragma endregion
#pragma region Vertex Data
            vertexData.resize(vertexPositions.size());

            // Normal Coords
            if (prim.attributes.contains(ANormal))
            {
                tinygltf::Accessor *normAccessor = &model.accessors[prim.attributes[ANormal]];

                assert(normAccessor->count <= MaxValue<u32>);

                tinygltf::BufferView *normBufferView = &model.bufferViews[normAccessor->bufferView];
                tinygltf::Buffer *normBuffer = &model.buffers[normBufferView->buffer];
                usize byteOffset = (normAccessor->byteOffset + normBufferView->byteOffset);

                if (normAccessor->type == TINYGLTF_TYPE_VEC4)
                {
                    vec4 *data = Recast<vec4 *>(normBuffer->data.data() + byteOffset);

                    vec4 *end = data + vertexCount;
                    u32 idx = vertexOffset;
                    vec4 *it = data;
                    while (it != end)
                    {
                        vertexData[idx++].m_Normal = *(it++);
                    }
                }
                else if (normAccessor->type == TINYGLTF_TYPE_VEC3)
                {
                    vec3 *data = Recast<vec3 *>(normBuffer->data.data() + byteOffset);
                    for (u32 i = 0; i < vertexCount; ++i)
                    {
                        auto norm = vec4(data[i], 0.0f);
                        vertexData[vertexOffset + i].m_Normal = norm;
                    }
                }
                else if (normAccessor->type == TINYGLTF_TYPE_VEC2)
                {
                    vec2 *data = Recast<vec2 *>(normBuffer->data.data() + byteOffset);
                    for (u32 i = 0; i < vertexCount; ++i)
                    {
                        auto norm = vec4(data[i], 0.0f, 0.0f);
                        vertexData[vertexOffset + i].m_Normal = norm;
                    }
                }
            }

            // UV0
            if (prim.attributes.contains(ATexCoord0))
            {
                tinygltf::Accessor *uvAccessor = &model.accessors[prim.attributes[ATexCoord0]];

                assert(uvAccessor->count <= MaxValue<u32>);

                tinygltf::BufferView *uvBufferView = &model.bufferViews[uvAccessor->bufferView];
                tinygltf::Buffer *uvBuffer = &model.buffers[uvBufferView->buffer];
                usize byteOffset = (uvAccessor->byteOffset + uvBufferView->byteOffset);

                assert(uvAccessor->type == TINYGLTF_TYPE_VEC2 &&
                       uvAccessor->componentType == TINYGLTF_COMPONENT_TYPE_FLOAT);
                {
                    vec2 *data = Recast<vec2 *>(uvBuffer->data.data() + byteOffset);
                    vec2 *end = data + vertexCount;
                    u32 idx = vertexOffset;
                    vec2 *it = data;
                    while (it != end)
                    {
                        vertexData[idx++].m_TexCoord0 = *(it++);
                    }
                }
            }

            if (prim.attributes.contains(AColor0))
            {
                tinygltf::Accessor *colorAccessor = &model.accessors[prim.attributes[AColor0]];

                assert(colorAccessor->count <= MaxValue<u32>);

                tinygltf::BufferView *colorBufferView = &model.bufferViews[colorAccessor->bufferView];
                tinygltf::Buffer *colorBuffer = &model.buffers[colorBufferView->buffer];
                usize byteOffset = (colorAccessor->byteOffset + colorBufferView->byteOffset);

                if (colorAccessor->type == TINYGLTF_TYPE_VEC4)
                {
                    vec4 *data = Recast<vec4 *>(colorBuffer->data.data() + byteOffset);

                    vec4 *end = data + vertexCount;
                    u32 idx = vertexOffset;
                    vec4 *it = data;
                    while (it != end)
                    {
                        vertexData[idx++].m_Color0 = *(it++);
                    }
                }
                else if (colorAccessor->type == TINYGLTF_TYPE_VEC3)
                {
                    vec3 *data = Recast<vec3 *>(colorBuffer->data.data() + byteOffset);
                    for (u32 i = 0; i < vertexCount; ++i)
                    {
                        auto color = vec4(data[i], 1.0f);
                        vertexData[vertexOffset + i].m_Color0 = color;
                    }
                }
            }
#pragma endregion
#pragma region Indices
            // Indices
            if (prim.indices >= 0)
            {
                tinygltf::Accessor *indexAccessor = &model.accessors[prim.indices];

                assert(indexAccessor->count <= MaxValue<u32>);

                tinygltf::BufferView *indexBufferView = &model.bufferViews[indexAccessor->bufferView];
                tinygltf::Buffer *indexBuffer = &model.buffers[indexBufferView->buffer];
                usize byteOffset = (indexAccessor->byteOffset + indexBufferView->byteOffset);

                indexCount = Cast<u32>(indexAccessor->count);
                indices.reserve(indexOffset + indexCount);

                if (indexAccessor->componentType == TINYGLTF_COMPONENT_TYPE_UNSIGNED_INT)
                {
                    u32 *data = Recast<u32 *>(indexBuffer->data.data() + byteOffset);
                    indices.insert(indices.end(), data, data + indexCount);
                }
                else if (indexAccessor->componentType == TINYGLTF_COMPONENT_TYPE_UNSIGNED_SHORT)
                {
                    u16 *data = Recast<u16 *>(indexBuffer->data.data() + byteOffset);
                    indices.insert(indices.end(), data, data + indexCount);
                }
                else if (indexAccessor->componentType == TINYGLTF_COMPONENT_TYPE_UNSIGNED_BYTE)
                {
                    u8 *data = Recast<u8 *>(indexBuffer->data.data() + byteOffset);
                    indices.insert(indices.end(), data, data + indexCount);
                }
            }
            else
            {
                indexCount = vertexCount;
                indices.reserve(indexOffset + vertexCount);
                for (u32 i = 0; i < indexCount; ++i)
                {
                    indices.push_back(i);
                }
            }
#pragma endregion

            meshPrimitives.push_back({
                .m_VertexOffset = vertexOffset,
                .m_FirstIndex = indexOffset,
                .m_IndexCount = indexCount,
                .m_MaterialIdx = prim.material,
                .m_TransformIdx = -1,
            });

            vertexOffset += vertexCount;
            indexOffset += indexCount;
        }
    }

    Nodes nodes;
    nodes.Add(mat4{1.0f}, -1);
    {
        if (model.defaultScene >= 0)
        {
            eastl::function<void(i32, i32)> processNode = [&processNode, &model, &nodes, &meshPrimRanges,
                                                           &meshPrimitives](i32 idx, i32 parent) -> void {
                const auto *node = &model.nodes[idx];

                vec3 nodeTranslation = vec3{0.0f};
                quat nodeRotation = quat{1.0f, 0.0f, 0.0f, 0.0f};
                vec3 nodeScale = vec3{1.0f};
                mat4 nodeMatrix = mat4{1.0f};
                if (node->translation.size() == 3)
                {
                    nodeTranslation = glm::make_vec3(node->translation.data());
                }
                if (node->rotation.size() == 4)
                {
                    nodeRotation = glm::make_quat(node->rotation.data());
                }
                if (node->scale.size() == 3)
                {
                    // We don't handle the scale 0 special case yet.
                    nodeScale = glm::make_vec3(node->scale.data());
                }
                if (node->matrix.size() == 16)
                {
                    nodeMatrix = glm::make_mat4(node->matrix.data());
                }
                const mat4 transform = translate(mat4(1.0f), nodeTranslation) * mat4_cast(nodeRotation) *
                                       scale(mat4(1.0f), nodeScale) * nodeMatrix;

                const i32 nodeArrayIndex = Cast<i32>(nodes.Add(transform, parent));
                if (node->mesh >= 0)
                {
                    auto [start, count] = meshPrimRanges[node->mesh];
                    const auto end = start + count;
                    for (usize i = start; i != end; ++i)
                    {
                        meshPrimitives[i].m_TransformIdx = nodeArrayIndex;
                    }
                }

                for (const i32 child : node->children)
                {
                    processNode(child, nodeArrayIndex);
                }
            };
            auto *scene = &model.scenes[model.defaultScene];
            for (i32 rootNodeIdx : scene->nodes)
            {
                processNode(rootNodeIdx, 0);
            }
        }
    }

    nodes.Update();

    StorageBuffer nodeBuffer;
    nodeBuffer.Init(pDevice, nodes.GetGlobalTransformByteSize(), true);
    nodeBuffer.Write(pDevice, 0, nodes.GetGlobalTransformByteSize(), nodes.GetGlobalTransformPtr());
    BufferHandle nodeHandle = m_ResourceManager->Commit(&nodeBuffer);

#pragma region Staging / Transfer / Uploads
    BufferHandle positionBufferHandle;
    BufferHandle vertexDataHandle;
    IndexBuffer indexBuffer;

    {
        auto uploadBufferData = [cmd = this->m_CommandBuffer, &stagingBuffers, pDevice](const Buffer *buffer,
                                                                                        const void *data) {
            vk::BufferCopy bufferCopy = {.srcOffset = 0, .dstOffset = 0, .size = buffer->GetSize()};
            StagingBuffer &stagingBuffer = stagingBuffers.push_back();
            stagingBuffer.Init(pDevice, bufferCopy.size);
            stagingBuffer.Write(pDevice, 0, bufferCopy.size, data);
            cmd.copyBuffer(stagingBuffer.m_Buffer, buffer->m_Buffer, 1, &bufferCopy);
        };

        StorageBuffer positionBuffer;
        positionBuffer.Init(pDevice, vertexPositions.size() * sizeof vertexPositions[0], false);
        positionBufferHandle = m_ResourceManager->Commit(&positionBuffer);
        uploadBufferData(&positionBuffer, vertexPositions.data());

        StorageBuffer vertexDataBuffer;
        vertexDataBuffer.Init(pDevice, vertexData.size() * sizeof vertexData[0], false);
        vertexDataHandle = m_ResourceManager->Commit(&vertexDataBuffer);
        uploadBufferData(&vertexDataBuffer, vertexData.data());

        indexBuffer.Init(pDevice, indices.size() * sizeof indices[0]);
        uploadBufferData(&indexBuffer, indices.data());
    }

#pragma endregion

#if !defined(ASTER_NDEBUG)
    m_CommandBuffer.endDebugUtilsLabelEXT();
#endif
    AbortIfFailed(m_CommandBuffer.end());

    vk::SubmitInfo submitInfo = {
        .waitSemaphoreCount = 0,
        .pWaitDstStageMask = nullptr,
        .commandBufferCount = 1,
        .pCommandBuffers = &m_CommandBuffer,
    };

    vk::Fence fence;
    vk::FenceCreateInfo fenceCreateInfo = {};
    AbortIfFailed(pDevice->m_Device.createFence(&fenceCreateInfo, nullptr, &fence));
    AbortIfFailed(m_TransferQueue.submit(1, &submitInfo, fence));
    AbortIfFailed(pDevice->m_Device.waitForFences(1, &fence, true, MaxValue<u32>));
    pDevice->m_Device.destroy(fence, nullptr);

    AbortIfFailed(pDevice->m_Device.resetCommandPool(m_CommandPool, {}));

    for (auto &buffer : stagingBuffers)
    {
        buffer.Destroy(pDevice);
    }

    Model::ModelHandles handles = {
        .m_VertexPositionHandle = positionBufferHandle,
        .m_VertexDataHandle = vertexDataHandle,
        .m_MaterialsHandle = materialsHandle,
        .m_NodeHandle = nodeHandle,
    };

    eastl::vector<TextureHandle> textureHandles;
    textureHandles.reserve(textureHandleMap.size());

    for (auto &[key, val] : textureHandleMap)
    {
        textureHandles.emplace_back(val);
    }

    return Model{
        m_ResourceManager, std::move(textureHandles), std::move(nodes), handles, indexBuffer, meshPrimitives,
    };
}

Model::Model(GpuResourceManager *resourceManager, eastl::vector<TextureHandle> &&textureHandles, Nodes &&nodes,
             const ModelHandles &handles, const IndexBuffer &indexBuffer,
             const eastl::vector<MeshPrimitive> &meshPrimitives)
    : m_ResourceManager(resourceManager)
    , m_TextureHandles(std::move(textureHandles))
    , m_Nodes(std::move(nodes))
    , m_Handles(handles)
    , m_IndexBuffer(indexBuffer)
    , m_MeshPrimitives(meshPrimitives)
{
}

Model::Model(Model &&other) noexcept
    : m_ResourceManager(Take(other.m_ResourceManager))
    , m_TextureHandles(std::move(other.m_TextureHandles))
    , m_Handles(other.m_Handles)
    , m_IndexBuffer(other.m_IndexBuffer)
    , m_MeshPrimitives(std::move(other.m_MeshPrimitives))
{
}

Model &
Model::operator=(Model &&other) noexcept
{
    if (this == &other)
        return *this;
    m_ResourceManager = Take(other.m_ResourceManager);
    m_TextureHandles = std::move(other.m_TextureHandles);
    m_Handles = other.m_Handles;
    m_IndexBuffer = other.m_IndexBuffer;
    m_MeshPrimitives = std::move(other.m_MeshPrimitives);
    return *this;
}

const mat4 &
Model::GetModelTransform() const
{
    return m_Nodes[0];
}

void
Model::SetModelTransform(const mat4 &transform)
{
    m_Nodes.Set(0, transform);
}

Model::~Model()
{
    if (!m_ResourceManager)
        return;

    m_IndexBuffer.Destroy(m_ResourceManager->m_Device);

    m_ResourceManager->Release(m_Handles.m_VertexDataHandle);
    m_ResourceManager->Release(m_Handles.m_NodeHandle);
    m_ResourceManager->Release(m_Handles.m_VertexPositionHandle);
    m_ResourceManager->Release(m_Handles.m_MaterialsHandle);

    for (const TextureHandle &handle : m_TextureHandles)
    {
        m_ResourceManager->Release(handle);
    }
}

void
Model::Update()
{
    if (m_Nodes.Update())
    {
        m_ResourceManager->Write(m_Handles.m_NodeHandle, 0, m_Nodes.GetGlobalTransformByteSize(),
                                 m_Nodes.GetGlobalTransformPtr());
    }
}

AssetLoader::AssetLoader(GpuResourceManager *resourceManager, vk::Queue transferQueue, u32 transferQueueIndex,
                         u32 graphicsQueueIndex)
    : m_ResourceManager(resourceManager)
    , m_TransferQueue(transferQueue)
    , m_TransferQueueIndex(transferQueueIndex)
    , m_GraphicsQueueIndex(graphicsQueueIndex)
{
    const Device *pDevice = resourceManager->m_Device;
    const vk::CommandPoolCreateInfo poolCreateInfo = {
        .flags = vk::CommandPoolCreateFlagBits::eTransient,
        .queueFamilyIndex = transferQueueIndex,
    };
    AbortIfFailedM(pDevice->m_Device.createCommandPool(&poolCreateInfo, nullptr, &m_CommandPool),
                   "Transfer command pool creation failed.");

    pDevice->SetName(m_CommandPool, "Asset Loader Command Pool");

    const vk::CommandBufferAllocateInfo commandBufferAllocateInfo = {
        .commandPool = m_CommandPool,
        .level = vk::CommandBufferLevel::ePrimary,
        .commandBufferCount = 1,
    };
    AbortIfFailed(pDevice->m_Device.allocateCommandBuffers(&commandBufferAllocateInfo, &m_CommandBuffer));

    pDevice->SetName(m_CommandBuffer, "Asset Loader Command Buffer");
}

AssetLoader::~AssetLoader()
{
    if (m_ResourceManager)
    {
        m_ResourceManager->m_Device->m_Device.destroy(m_CommandPool, nullptr);
    }
}

AssetLoader::AssetLoader(AssetLoader &&other) noexcept
    : m_ResourceManager(Take(other.m_ResourceManager))
    , m_CommandPool(other.m_CommandPool)
    , m_CommandBuffer(other.m_CommandBuffer)
    , m_TransferQueue(other.m_TransferQueue)
    , m_TransferQueueIndex(other.m_TransferQueueIndex)
    , m_GraphicsQueueIndex(other.m_GraphicsQueueIndex)
{
}

AssetLoader &
AssetLoader::operator=(AssetLoader &&other) noexcept
{
    if (this == &other)
        return *this;
    m_ResourceManager = Take(other.m_ResourceManager);
    m_CommandPool = other.m_CommandPool;
    m_CommandBuffer = other.m_CommandBuffer;
    m_TransferQueue = other.m_TransferQueue;
    m_TransferQueueIndex = other.m_TransferQueueIndex;
    m_GraphicsQueueIndex = other.m_GraphicsQueueIndex;
    return *this;
}