#include <windows.h>
#include <filesystem>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include <gtest/gtest.h>
#include <stb/stb_image.h>

#include "../fixtures/RHIIntegrationFixture.h"
#include "XCEngine/Debug/ConsoleLogSink.h"
#include "XCEngine/Debug/Logger.h"
#include "XCEngine/RHI/RHIBuffer.h"
#include "XCEngine/RHI/RHIDescriptorPool.h"
#include "XCEngine/RHI/RHIDescriptorSet.h"
#include "XCEngine/RHI/RHIPipelineLayout.h"
#include "XCEngine/RHI/RHIPipelineState.h"
#include "XCEngine/RHI/RHIResourceView.h"
#include "XCEngine/RHI/RHISampler.h"
#include "XCEngine/RHI/RHITexture.h"

using namespace XCEngine::Debug;
using namespace XCEngine::RHI;
using namespace XCEngine::RHI::Integration;

namespace {

struct Vertex {
    float pos[4];
    float uv[2];
};

constexpr Vertex kQuadVertices[] = {
    { { -0.5f, -0.5f, 0.0f, 1.0f }, { 0.0f, 1.0f } },
    { { -0.5f,  0.5f, 0.0f, 1.0f }, { 0.0f, 0.0f } },
    { {  0.5f, -0.5f, 0.0f, 1.0f }, { 1.0f, 1.0f } },
    { {  0.5f,  0.5f, 0.0f, 1.0f }, { 1.0f, 0.0f } },
};

constexpr uint32_t kQuadIndices[] = { 0, 1, 2, 2, 1, 3 };

std::filesystem::path GetExecutableDirectory() {
    char exePath[MAX_PATH] = {};
    const DWORD length = GetModuleFileNameA(nullptr, exePath, MAX_PATH);
    if (length == 0 || length >= MAX_PATH) {
        return std::filesystem::current_path();
    }

    return std::filesystem::path(exePath).parent_path();
}

std::filesystem::path ResolveRuntimePath(const char* relativePath) {
    return GetExecutableDirectory() / relativePath;
}

const char kQuadHlsl[] = R"(
Texture2D gTexture : register(t0);
SamplerState gSampler : register(s0);

struct VSInput {
    float4 position : POSITION;
    float2 texcoord : TEXCOORD;
};

struct PSInput {
    float4 position : SV_POSITION;
    float2 texcoord : TEXCOORD;
};

PSInput MainVS(VSInput input) {
    PSInput output;
    output.position = input.position;
    output.texcoord = input.texcoord;
    return output;
}

float4 MainPS(PSInput input) : SV_TARGET {
    return gTexture.Sample(gSampler, input.texcoord);
}
)";

const char kQuadVertexShader[] = R"(#version 430
layout(location = 0) in vec4 aPosition;
layout(location = 1) in vec2 aTexCoord;

out vec2 vTexCoord;

void main() {
    gl_Position = aPosition;
    vTexCoord = aTexCoord;
}
)";

const char kQuadFragmentShader[] = R"(#version 430
layout(binding = 0) uniform sampler2D uTexture;

in vec2 vTexCoord;

layout(location = 0) out vec4 fragColor;

void main() {
    fragColor = texture(uTexture, vTexCoord);
}
)";

const char* GetScreenshotFilename(RHIType type) {
    return type == RHIType::D3D12 ? "quad_d3d12.ppm" : "quad_opengl.ppm";
}

RHITexture* LoadQuadTexture(RHIDevice* device) {
    const std::filesystem::path texturePath = ResolveRuntimePath("Res/Image/earth.png");
    const std::string texturePathString = texturePath.string();

    stbi_set_flip_vertically_on_load(0);

    int width = 0;
    int height = 0;
    int channels = 0;
    stbi_uc* pixels = stbi_load(texturePathString.c_str(), &width, &height, &channels, STBI_rgb_alpha);
    if (pixels == nullptr) {
        Log("[TEST] Failed to load quad texture: %s", texturePathString.c_str());
        return nullptr;
    }

    TextureDesc textureDesc = {};
    textureDesc.width = static_cast<uint32_t>(width);
    textureDesc.height = static_cast<uint32_t>(height);
    textureDesc.depth = 1;
    textureDesc.mipLevels = 1;
    textureDesc.arraySize = 1;
    textureDesc.format = static_cast<uint32_t>(Format::R8G8B8A8_UNorm);
    textureDesc.textureType = static_cast<uint32_t>(TextureType::Texture2D);
    textureDesc.sampleCount = 1;
    textureDesc.sampleQuality = 0;
    textureDesc.flags = 0;

    RHITexture* texture = device->CreateTexture(
        textureDesc,
        pixels,
        static_cast<size_t>(width) * static_cast<size_t>(height) * 4,
        static_cast<uint32_t>(width) * 4);
    stbi_image_free(pixels);
    return texture;
}

int GetComparisonThreshold(RHIType type) {
    return type == RHIType::OpenGL ? 5 : 0;
}

GraphicsPipelineDesc CreateQuadPipelineDesc(RHIType type, RHIPipelineLayout* pipelineLayout) {
    GraphicsPipelineDesc desc = {};
    desc.pipelineLayout = pipelineLayout;
    desc.topologyType = static_cast<uint32_t>(PrimitiveTopologyType::Triangle);
    desc.renderTargetFormats[0] = static_cast<uint32_t>(Format::R8G8B8A8_UNorm);
    desc.depthStencilFormat = static_cast<uint32_t>(Format::Unknown);
    desc.sampleCount = 1;

    desc.rasterizerState.fillMode = static_cast<uint32_t>(FillMode::Solid);
    desc.rasterizerState.cullMode = static_cast<uint32_t>(CullMode::None);
    desc.rasterizerState.frontFace = static_cast<uint32_t>(FrontFace::CounterClockwise);
    desc.rasterizerState.depthClipEnable = true;

    desc.depthStencilState.depthTestEnable = false;
    desc.depthStencilState.depthWriteEnable = false;
    desc.depthStencilState.stencilEnable = false;

    InputElementDesc position = {};
    position.semanticName = "POSITION";
    position.semanticIndex = 0;
    position.format = static_cast<uint32_t>(Format::R32G32B32A32_Float);
    position.inputSlot = 0;
    position.alignedByteOffset = 0;
    desc.inputLayout.elements.push_back(position);

    InputElementDesc texcoord = {};
    texcoord.semanticName = "TEXCOORD";
    texcoord.semanticIndex = 0;
    texcoord.format = static_cast<uint32_t>(Format::R32G32_Float);
    texcoord.inputSlot = 0;
    texcoord.alignedByteOffset = sizeof(float) * 4;
    desc.inputLayout.elements.push_back(texcoord);

    if (type == RHIType::D3D12) {
        desc.vertexShader.source.assign(kQuadHlsl, kQuadHlsl + strlen(kQuadHlsl));
        desc.vertexShader.sourceLanguage = ShaderLanguage::HLSL;
        desc.vertexShader.entryPoint = L"MainVS";
        desc.vertexShader.profile = L"vs_5_0";

        desc.fragmentShader.source.assign(kQuadHlsl, kQuadHlsl + strlen(kQuadHlsl));
        desc.fragmentShader.sourceLanguage = ShaderLanguage::HLSL;
        desc.fragmentShader.entryPoint = L"MainPS";
        desc.fragmentShader.profile = L"ps_5_0";
    } else {
        desc.vertexShader.source.assign(kQuadVertexShader, kQuadVertexShader + strlen(kQuadVertexShader));
        desc.vertexShader.sourceLanguage = ShaderLanguage::GLSL;
        desc.vertexShader.profile = L"vs_4_30";

        desc.fragmentShader.source.assign(kQuadFragmentShader, kQuadFragmentShader + strlen(kQuadFragmentShader));
        desc.fragmentShader.sourceLanguage = ShaderLanguage::GLSL;
        desc.fragmentShader.profile = L"fs_4_30";
    }

    return desc;
}

class QuadTest : public RHIIntegrationFixture {
protected:
    void SetUp() override;
    void TearDown() override;
    void RenderFrame() override;

private:
    void InitializeQuadResources();
    void ShutdownQuadResources();

    RHIBuffer* mVertexBuffer = nullptr;
    RHIResourceView* mVertexBufferView = nullptr;
    RHIBuffer* mIndexBuffer = nullptr;
    RHIResourceView* mIndexBufferView = nullptr;
    RHITexture* mTexture = nullptr;
    RHIResourceView* mTextureView = nullptr;
    RHISampler* mSampler = nullptr;
    RHIDescriptorPool* mTexturePool = nullptr;
    RHIDescriptorSet* mTextureSet = nullptr;
    RHIDescriptorPool* mSamplerPool = nullptr;
    RHIDescriptorSet* mSamplerSet = nullptr;
    RHIPipelineLayout* mPipelineLayout = nullptr;
    RHIPipelineState* mPipelineState = nullptr;
};

void QuadTest::SetUp() {
    RHIIntegrationFixture::SetUp();
    InitializeQuadResources();
}

void QuadTest::TearDown() {
    ShutdownQuadResources();
    RHIIntegrationFixture::TearDown();
}

void QuadTest::InitializeQuadResources() {
    BufferDesc vertexBufferDesc = {};
    vertexBufferDesc.size = sizeof(kQuadVertices);
    vertexBufferDesc.stride = sizeof(Vertex);
    vertexBufferDesc.bufferType = static_cast<uint32_t>(BufferType::Vertex);
    vertexBufferDesc.flags = 0;

    mVertexBuffer = GetDevice()->CreateBuffer(vertexBufferDesc);
    ASSERT_NE(mVertexBuffer, nullptr);
    mVertexBuffer->SetData(kQuadVertices, sizeof(kQuadVertices));
    mVertexBuffer->SetStride(sizeof(Vertex));
    mVertexBuffer->SetBufferType(BufferType::Vertex);

    ResourceViewDesc vertexViewDesc = {};
    vertexViewDesc.dimension = ResourceViewDimension::Buffer;
    vertexViewDesc.structureByteStride = sizeof(Vertex);
    mVertexBufferView = GetDevice()->CreateVertexBufferView(mVertexBuffer, vertexViewDesc);
    ASSERT_NE(mVertexBufferView, nullptr);

    BufferDesc indexBufferDesc = {};
    indexBufferDesc.size = sizeof(kQuadIndices);
    indexBufferDesc.stride = sizeof(uint32_t);
    indexBufferDesc.bufferType = static_cast<uint32_t>(BufferType::Index);
    indexBufferDesc.flags = 0;

    mIndexBuffer = GetDevice()->CreateBuffer(indexBufferDesc);
    ASSERT_NE(mIndexBuffer, nullptr);
    mIndexBuffer->SetData(kQuadIndices, sizeof(kQuadIndices));
    mIndexBuffer->SetStride(sizeof(uint32_t));
    mIndexBuffer->SetBufferType(BufferType::Index);

    ResourceViewDesc indexViewDesc = {};
    indexViewDesc.dimension = ResourceViewDimension::Buffer;
    indexViewDesc.format = static_cast<uint32_t>(Format::R32_UInt);
    mIndexBufferView = GetDevice()->CreateIndexBufferView(mIndexBuffer, indexViewDesc);
    ASSERT_NE(mIndexBufferView, nullptr);

    mTexture = LoadQuadTexture(GetDevice());
    ASSERT_NE(mTexture, nullptr);

    ResourceViewDesc textureViewDesc = {};
    textureViewDesc.format = static_cast<uint32_t>(Format::R8G8B8A8_UNorm);
    textureViewDesc.dimension = ResourceViewDimension::Texture2D;
    textureViewDesc.mipLevel = 0;
    mTextureView = GetDevice()->CreateShaderResourceView(mTexture, textureViewDesc);
    ASSERT_NE(mTextureView, nullptr);

    SamplerDesc samplerDesc = {};
    samplerDesc.filter = static_cast<uint32_t>(FilterMode::Linear);
    samplerDesc.addressU = static_cast<uint32_t>(TextureAddressMode::Clamp);
    samplerDesc.addressV = static_cast<uint32_t>(TextureAddressMode::Clamp);
    samplerDesc.addressW = static_cast<uint32_t>(TextureAddressMode::Clamp);
    samplerDesc.mipLodBias = 0.0f;
    samplerDesc.maxAnisotropy = 1;
    samplerDesc.comparisonFunc = static_cast<uint32_t>(ComparisonFunc::Always);
    samplerDesc.borderColorR = 0.0f;
    samplerDesc.borderColorG = 0.0f;
    samplerDesc.borderColorB = 0.0f;
    samplerDesc.borderColorA = 0.0f;
    samplerDesc.minLod = 0.0f;
    samplerDesc.maxLod = 1000.0f;
    mSampler = GetDevice()->CreateSampler(samplerDesc);
    ASSERT_NE(mSampler, nullptr);

    DescriptorPoolDesc texturePoolDesc = {};
    texturePoolDesc.type = DescriptorHeapType::CBV_SRV_UAV;
    texturePoolDesc.descriptorCount = 1;
    texturePoolDesc.shaderVisible = true;
    mTexturePool = GetDevice()->CreateDescriptorPool(texturePoolDesc);
    ASSERT_NE(mTexturePool, nullptr);

    DescriptorSetLayoutBinding textureBinding = {};
    textureBinding.binding = 0;
    textureBinding.type = static_cast<uint32_t>(DescriptorType::SRV);
    textureBinding.count = 1;

    DescriptorSetLayoutDesc textureLayoutDesc = {};
    textureLayoutDesc.bindings = &textureBinding;
    textureLayoutDesc.bindingCount = 1;

    mTextureSet = mTexturePool->AllocateSet(textureLayoutDesc);
    ASSERT_NE(mTextureSet, nullptr);
    mTextureSet->Update(0, mTextureView);

    DescriptorPoolDesc samplerPoolDesc = {};
    samplerPoolDesc.type = DescriptorHeapType::Sampler;
    samplerPoolDesc.descriptorCount = 1;
    samplerPoolDesc.shaderVisible = true;
    mSamplerPool = GetDevice()->CreateDescriptorPool(samplerPoolDesc);
    ASSERT_NE(mSamplerPool, nullptr);

    DescriptorSetLayoutBinding samplerBinding = {};
    samplerBinding.binding = 0;
    samplerBinding.type = static_cast<uint32_t>(DescriptorType::Sampler);
    samplerBinding.count = 1;

    DescriptorSetLayoutDesc samplerLayoutDesc = {};
    samplerLayoutDesc.bindings = &samplerBinding;
    samplerLayoutDesc.bindingCount = 1;

    mSamplerSet = mSamplerPool->AllocateSet(samplerLayoutDesc);
    ASSERT_NE(mSamplerSet, nullptr);
    mSamplerSet->UpdateSampler(0, mSampler);

    RHIPipelineLayoutDesc pipelineLayoutDesc = {};
    pipelineLayoutDesc.textureCount = 1;
    pipelineLayoutDesc.samplerCount = 1;
    mPipelineLayout = GetDevice()->CreatePipelineLayout(pipelineLayoutDesc);
    ASSERT_NE(mPipelineLayout, nullptr);

    GraphicsPipelineDesc pipelineDesc = CreateQuadPipelineDesc(GetBackendType(), mPipelineLayout);
    mPipelineState = GetDevice()->CreatePipelineState(pipelineDesc);
    ASSERT_NE(mPipelineState, nullptr);
    ASSERT_TRUE(mPipelineState->IsValid());

    Log("[TEST] Quad resources initialized for backend %d", static_cast<int>(GetBackendType()));
}

void QuadTest::ShutdownQuadResources() {
    if (mPipelineState != nullptr) {
        mPipelineState->Shutdown();
        delete mPipelineState;
        mPipelineState = nullptr;
    }

    if (mPipelineLayout != nullptr) {
        mPipelineLayout->Shutdown();
        delete mPipelineLayout;
        mPipelineLayout = nullptr;
    }

    if (mTextureSet != nullptr) {
        mTextureSet->Shutdown();
        delete mTextureSet;
        mTextureSet = nullptr;
    }

    if (mSamplerSet != nullptr) {
        mSamplerSet->Shutdown();
        delete mSamplerSet;
        mSamplerSet = nullptr;
    }

    if (mTexturePool != nullptr) {
        mTexturePool->Shutdown();
        delete mTexturePool;
        mTexturePool = nullptr;
    }

    if (mSamplerPool != nullptr) {
        mSamplerPool->Shutdown();
        delete mSamplerPool;
        mSamplerPool = nullptr;
    }

    if (mSampler != nullptr) {
        mSampler->Shutdown();
        delete mSampler;
        mSampler = nullptr;
    }

    if (mTextureView != nullptr) {
        mTextureView->Shutdown();
        delete mTextureView;
        mTextureView = nullptr;
    }

    if (mTexture != nullptr) {
        mTexture->Shutdown();
        delete mTexture;
        mTexture = nullptr;
    }

    if (mVertexBufferView != nullptr) {
        mVertexBufferView->Shutdown();
        delete mVertexBufferView;
        mVertexBufferView = nullptr;
    }

    if (mIndexBufferView != nullptr) {
        mIndexBufferView->Shutdown();
        delete mIndexBufferView;
        mIndexBufferView = nullptr;
    }

    if (mVertexBuffer != nullptr) {
        mVertexBuffer->Shutdown();
        delete mVertexBuffer;
        mVertexBuffer = nullptr;
    }

    if (mIndexBuffer != nullptr) {
        mIndexBuffer->Shutdown();
        delete mIndexBuffer;
        mIndexBuffer = nullptr;
    }
}

void QuadTest::RenderFrame() {
    ASSERT_NE(mPipelineState, nullptr);
    ASSERT_NE(mVertexBufferView, nullptr);
    ASSERT_NE(mIndexBufferView, nullptr);
    ASSERT_NE(mTextureSet, nullptr);
    ASSERT_NE(mSamplerSet, nullptr);
    ASSERT_NE(mPipelineLayout, nullptr);

    RHICommandList* cmdList = GetCommandList();
    RHICommandQueue* cmdQueue = GetCommandQueue();
    ASSERT_NE(cmdList, nullptr);
    ASSERT_NE(cmdQueue, nullptr);

    cmdList->Reset();
    SetRenderTargetForClear();

    Viewport viewport = { 0.0f, 0.0f, 1280.0f, 720.0f, 0.0f, 1.0f };
    Rect scissorRect = { 0, 0, 1280, 720 };
    cmdList->SetViewport(viewport);
    cmdList->SetScissorRect(scissorRect);
    cmdList->Clear(0.0f, 0.0f, 1.0f, 1.0f, 1);

    cmdList->SetPipelineState(mPipelineState);
    RHIDescriptorSet* descriptorSets[] = { mTextureSet, mSamplerSet };
    cmdList->SetGraphicsDescriptorSets(0, 2, descriptorSets, mPipelineLayout);
    cmdList->SetPrimitiveTopology(PrimitiveTopology::TriangleList);

    RHIResourceView* vertexBuffers[] = { mVertexBufferView };
    uint64_t offsets[] = { 0 };
    uint32_t strides[] = { sizeof(Vertex) };
    cmdList->SetVertexBuffers(0, 1, vertexBuffers, offsets, strides);
    cmdList->SetIndexBuffer(mIndexBufferView, 0);
    cmdList->DrawIndexed(static_cast<uint32_t>(sizeof(kQuadIndices) / sizeof(kQuadIndices[0])));

    EndRender();

    cmdList->Close();
    void* cmdLists[] = { cmdList };
    cmdQueue->ExecuteCommandLists(1, cmdLists);
}

TEST_P(QuadTest, RenderQuad) {
    RHICommandQueue* cmdQueue = GetCommandQueue();
    RHISwapChain* swapChain = GetSwapChain();
    const int targetFrameCount = 30;
    const char* screenshotFilename = GetScreenshotFilename(GetBackendType());
    const int comparisonThreshold = GetComparisonThreshold(GetBackendType());

    for (int frameCount = 0; frameCount <= targetFrameCount; ++frameCount) {
        if (frameCount > 0) {
            cmdQueue->WaitForPreviousFrame();
        }

        Log("[TEST] Quad MainLoop: frame %d", frameCount);
        BeginRender();
        RenderFrame();

        if (frameCount >= targetFrameCount) {
            cmdQueue->WaitForIdle();
            Log("[TEST] Quad MainLoop: frame %d reached, capturing %s", frameCount, screenshotFilename);
            ASSERT_TRUE(TakeScreenshot(screenshotFilename));
            ASSERT_TRUE(CompareWithGoldenTemplate(screenshotFilename, "GT.ppm", static_cast<float>(comparisonThreshold)));
            Log("[TEST] Quad MainLoop: frame %d compare passed", frameCount);
            break;
        }

        swapChain->Present(0, 0);
    }
}

} // namespace

INSTANTIATE_TEST_SUITE_P(D3D12, QuadTest, ::testing::Values(RHIType::D3D12));
INSTANTIATE_TEST_SUITE_P(OpenGL, QuadTest, ::testing::Values(RHIType::OpenGL));

GTEST_API_ int main(int argc, char** argv) {
    Logger::Get().Initialize();
    Logger::Get().AddSink(std::make_unique<ConsoleLogSink>());
    Logger::Get().SetMinimumLevel(LogLevel::Debug);

    testing::InitGoogleTest(&argc, argv);
    return RUN_ALL_TESTS();
}