#include "VoxelRenderer.h"
#include "wiJobSystem.h"
#include "wiPrimitive.h"
#include <algorithm>
#include <chrono>
#include <cmath>
#include <cstring>

using namespace wi::graphics;

namespace voxel {

// ── VoxelRenderer Implementation ────────────────────────────────

VoxelRenderer::VoxelRenderer() = default;
VoxelRenderer::~VoxelRenderer() { shutdown(); }

void VoxelRenderer::initialize(GraphicsDevice* dev) {
    device_ = dev;
    if (!device_) return;

    createPipeline();
    if (!pso_.IsValid()) {
        wi::backlog::post("VoxelRenderer: pipeline creation failed", wi::backlog::LogLevel::Error);
        initialized_ = false;
        return;
    }
    generateTextures();

    // Create mega quad buffer (SRV for vertex pulling)
    GPUBufferDesc megaDesc;
    megaDesc.size = MEGA_BUFFER_CAPACITY * sizeof(PackedQuad);
    megaDesc.bind_flags = BindFlag::SHADER_RESOURCE;
    megaDesc.misc_flags = ResourceMiscFlag::BUFFER_STRUCTURED;
    megaDesc.stride = sizeof(PackedQuad);
    megaDesc.usage = Usage::DEFAULT;
    device_->CreateBuffer(&megaDesc, nullptr, &megaQuadBuffer_);

    // Create chunk info buffer (SRV for VS chunk lookup)
    GPUBufferDesc infoDesc;
    infoDesc.size = MAX_CHUNKS * sizeof(GPUChunkInfo);
    infoDesc.bind_flags = BindFlag::SHADER_RESOURCE;
    infoDesc.misc_flags = ResourceMiscFlag::BUFFER_STRUCTURED;
    infoDesc.stride = sizeof(GPUChunkInfo);
    infoDesc.usage = Usage::DEFAULT;
    device_->CreateBuffer(&infoDesc, nullptr, &chunkInfoBuffer_);

    // Create indirect args buffer (for DrawInstancedIndirectCount, up to 6 draws per chunk)
    // UAV bind flag needed for GPU cull compute shader to write args
    GPUBufferDesc argsDesc;
    argsDesc.size = MAX_DRAWS * sizeof(IndirectDrawArgs);
    argsDesc.bind_flags = BindFlag::UNORDERED_ACCESS;
    argsDesc.misc_flags = ResourceMiscFlag::BUFFER_STRUCTURED | ResourceMiscFlag::INDIRECT_ARGS;
    argsDesc.stride = sizeof(IndirectDrawArgs);
    argsDesc.usage = Usage::DEFAULT;
    device_->CreateBuffer(&argsDesc, nullptr, &indirectArgsBuffer_);

    // Create draw count buffer (single uint32, raw for RWByteAddressBuffer)
    // UAV bind flag needed for GPU cull compute shader atomic counter
    GPUBufferDesc countDesc;
    countDesc.size = sizeof(uint32_t);
    countDesc.bind_flags = BindFlag::UNORDERED_ACCESS;
    countDesc.misc_flags = ResourceMiscFlag::BUFFER_RAW | ResourceMiscFlag::INDIRECT_ARGS;
    countDesc.usage = Usage::DEFAULT;
    device_->CreateBuffer(&countDesc, nullptr, &drawCountBuffer_);

    // ── GPU Timestamp Queries ──────────────────────────────────────
    GPUQueryHeapDesc queryDesc;
    queryDesc.type = GpuQueryType::TIMESTAMP;
    queryDesc.query_count = TS_COUNT;
    device_->CreateQueryHeap(&queryDesc, &timestampHeap_);

    GPUBufferDesc readbackDesc;
    readbackDesc.size = TS_COUNT * sizeof(uint64_t);
    readbackDesc.usage = Usage::READBACK;
    device_->CreateBuffer(&readbackDesc, nullptr, &timestampReadback_);

    // ── GPU Compute Mesher resources ─────────────────────────────
    wi::renderer::LoadShader(ShaderStage::CS, meshShader_, "voxel/voxelMeshCS.cso");
    gpuMesherAvailable_ = meshShader_.IsValid();
    if (gpuMesherAvailable_) {
        // Voxel data buffer: 1 chunk's worth (32^3 voxels / 2 per uint = 16384 uint)
        GPUBufferDesc voxDesc;
        voxDesc.size = (CHUNK_VOLUME / 2) * sizeof(uint32_t);
        voxDesc.bind_flags = BindFlag::SHADER_RESOURCE;
        voxDesc.misc_flags = ResourceMiscFlag::BUFFER_STRUCTURED;
        voxDesc.stride = sizeof(uint32_t);
        voxDesc.usage = Usage::DEFAULT;
        device_->CreateBuffer(&voxDesc, nullptr, &voxelDataBuffer_);

        // GPU quad output: same capacity as mega-buffer
        GPUBufferDesc gpuQDesc;
        gpuQDesc.size = MEGA_BUFFER_CAPACITY * sizeof(uint64_t); // PackedQuad = 8 bytes
        gpuQDesc.bind_flags = BindFlag::UNORDERED_ACCESS | BindFlag::SHADER_RESOURCE;
        gpuQDesc.misc_flags = ResourceMiscFlag::BUFFER_STRUCTURED;
        gpuQDesc.stride = sizeof(uint64_t); // uint2 = 8 bytes
        gpuQDesc.usage = Usage::DEFAULT;
        device_->CreateBuffer(&gpuQDesc, nullptr, &gpuQuadBuffer_);

        // Quad counter
        GPUBufferDesc cntDesc;
        cntDesc.size = sizeof(uint32_t);
        cntDesc.bind_flags = BindFlag::UNORDERED_ACCESS;
        cntDesc.misc_flags = ResourceMiscFlag::BUFFER_RAW;
        cntDesc.usage = Usage::DEFAULT;
        device_->CreateBuffer(&cntDesc, nullptr, &gpuQuadCounter_);

        // Readback buffer for quad counter (GPU → CPU)
        GPUBufferDesc rbDesc;
        rbDesc.size = sizeof(uint32_t);
        rbDesc.usage = Usage::READBACK;
        device_->CreateBuffer(&rbDesc, nullptr, &meshCounterReadback_);

        wi::backlog::post("VoxelRenderer: GPU compute mesher available");
    } else {
        wi::backlog::post("VoxelRenderer: GPU compute mesher not available", wi::backlog::LogLevel::Warning);
    }

    cpuMegaQuads_.reserve(MEGA_BUFFER_CAPACITY);
    cpuChunkInfo_.reserve(MAX_CHUNKS);
    chunkSlots_.reserve(MAX_CHUNKS);
    cpuIndirectArgs_.reserve(MAX_CHUNKS);

    initialized_ = true;
    wi::backlog::post("VoxelRenderer: initialized (mega-buffer: "
        + std::to_string(MEGA_BUFFER_CAPACITY) + " quads capacity)");
}

void VoxelRenderer::shutdown() {
    chunkSlots_.clear();
    cpuChunkInfo_.clear();
    cpuMegaQuads_.clear();
    initialized_ = false;
}

void VoxelRenderer::createPipeline() {
    // Constant buffer for per-frame data
    GPUBufferDesc cbDesc;
    cbDesc.size = sizeof(VoxelConstants);
    cbDesc.bind_flags = BindFlag::CONSTANT_BUFFER;
    cbDesc.usage = Usage::DEFAULT;
    device_->CreateBuffer(&cbDesc, nullptr, &constantBuffer_);

    // Anisotropic wrap sampler
    SamplerDesc samplerDesc;
    samplerDesc.filter = Filter::ANISOTROPIC;
    samplerDesc.address_u = TextureAddressMode::WRAP;
    samplerDesc.address_v = TextureAddressMode::WRAP;
    samplerDesc.address_w = TextureAddressMode::WRAP;
    samplerDesc.max_anisotropy = 16;
    device_->CreateSampler(&samplerDesc, &sampler_);

    // Load shaders
    wi::renderer::LoadShader(ShaderStage::VS, vertexShader_, "voxel/voxelVS.cso");
    wi::renderer::LoadShader(ShaderStage::PS, pixelShader_, "voxel/voxelPS.cso");
    wi::renderer::LoadShader(ShaderStage::CS, cullShader_, "voxel/voxelCullCS.cso");

    if (!vertexShader_.IsValid() || !pixelShader_.IsValid()) {
        wi::backlog::post("VoxelRenderer: shader loading failed", wi::backlog::LogLevel::Error);
        return;
    }
    if (cullShader_.IsValid()) {
        gpuCullingEnabled_ = true;
        wi::backlog::post("VoxelRenderer: GPU cull compute shader enabled");
    } else {
        gpuCullingEnabled_ = false;
        wi::backlog::post("VoxelRenderer: cull compute shader not available, using CPU fallback", wi::backlog::LogLevel::Warning);
    }

    // Pipeline: backface cull, depth test, opaque blend, triangle list
    PipelineStateDesc psoDesc;
    psoDesc.vs = &vertexShader_;
    psoDesc.ps = &pixelShader_;
    psoDesc.rs = wi::renderer::GetRasterizerState(wi::enums::RSTYPE_FRONT);
    psoDesc.dss = wi::renderer::GetDepthStencilState(wi::enums::DSSTYPE_DEFAULT);
    psoDesc.bs = wi::renderer::GetBlendState(wi::enums::BSTYPE_OPAQUE);
    psoDesc.pt = PrimitiveTopology::TRIANGLELIST;

    device_->CreatePipelineState(&psoDesc, &pso_);
}

// ── Procedural texture generation ───────────────────────────────

static void generateNoiseTexture(uint8_t* pixels, int w, int h,
    uint8_t r0, uint8_t g0, uint8_t b0,
    uint8_t r1, uint8_t g1, uint8_t b1,
    uint32_t seed, float heightFreq = 1.0f, float heightContrast = 1.0f)
{
    uint32_t s = seed;
    uint32_t s2 = seed * 7919u + 104729u; // separate seed for heightmap
    for (int y = 0; y < h; y++) {
        for (int x = 0; x < w; x++) {
            s = s * 1664525u + 1013904223u;
            float noise = (float)(s & 0xFFFF) / 65535.0f;
            float fx = (float)x / w;
            float fy = (float)y / h;
            float pattern = 0.5f + 0.5f * std::sin(fx * 20.0f + noise * 3.0f) *
                            std::cos(fy * 20.0f + noise * 3.0f);
            float t = noise * 0.6f + pattern * 0.4f;

            int idx = (y * w + x) * 4;
            pixels[idx + 0] = (uint8_t)(r0 + (r1 - r0) * t);
            pixels[idx + 1] = (uint8_t)(g0 + (g1 - g0) * t);
            pixels[idx + 2] = (uint8_t)(b0 + (b1 - b0) * t);

            // Heightmap in alpha: separate noise for height-based material blending
            s2 = s2 * 1664525u + 1013904223u;
            float hn = (float)(s2 & 0xFFFF) / 65535.0f;
            float hPattern = 0.5f + 0.5f * std::sin(fx * 12.0f * heightFreq + hn * 2.0f) *
                             std::cos(fy * 12.0f * heightFreq + hn * 2.0f);
            float heightVal = hn * 0.5f + hPattern * 0.5f;
            heightVal = std::clamp(heightVal * heightContrast, 0.0f, 1.0f);
            pixels[idx + 3] = (uint8_t)(heightVal * 255.0f);
        }
    }
}

void VoxelRenderer::generateTextures() {
    const int TEX_SIZE = 256;
    const int NUM_MATERIALS = 5;

    std::vector<uint8_t> allPixels(TEX_SIZE * TEX_SIZE * 4 * NUM_MATERIALS);

    struct MatColor {
        uint8_t r0,g0,b0, r1,g1,b1;
        uint32_t seed;
        float heightFreq;     // heightmap noise frequency
        float heightContrast; // heightmap contrast (higher = more defined peaks)
    };
    MatColor colors[NUM_MATERIALS] = {
        { 60, 140, 40,   80, 180, 60,  101, 1.5f, 0.8f },  // Grass: medium bumps
        { 100, 70, 40,  140, 100, 60,  202, 0.8f, 0.6f },  // Dirt: smooth mounds
        { 80, 80, 90,   120, 120, 130, 303, 2.5f, 0.5f },  // Stone: darker blue-gray, moderate height (was 1.2, lowered so neighbors bleed onto it more)
        { 220, 200, 130, 245, 230, 160, 404, 3.0f, 0.4f },  // Sand: warmer yellow, fine
        { 220, 225, 230, 245, 248, 252, 505, 1.0f, 0.5f },  // Snow: smooth, soft
    };

    for (int i = 0; i < NUM_MATERIALS; i++) {
        auto& c = colors[i];
        generateNoiseTexture(
            allPixels.data() + i * TEX_SIZE * TEX_SIZE * 4,
            TEX_SIZE, TEX_SIZE,
            c.r0, c.g0, c.b0, c.r1, c.g1, c.b1, c.seed,
            c.heightFreq, c.heightContrast
        );
    }

    TextureDesc texDesc;
    texDesc.type = TextureDesc::Type::TEXTURE_2D;
    texDesc.width = TEX_SIZE;
    texDesc.height = TEX_SIZE;
    texDesc.array_size = NUM_MATERIALS;
    texDesc.mip_levels = 1;
    texDesc.format = Format::R8G8B8A8_UNORM;
    texDesc.bind_flags = BindFlag::SHADER_RESOURCE;
    texDesc.usage = Usage::DEFAULT;

    std::vector<SubresourceData> subData(NUM_MATERIALS);
    for (int i = 0; i < NUM_MATERIALS; i++) {
        subData[i].data_ptr = allPixels.data() + i * TEX_SIZE * TEX_SIZE * 4;
        subData[i].row_pitch = TEX_SIZE * 4;
        subData[i].slice_pitch = TEX_SIZE * TEX_SIZE * 4;
    }

    device_->CreateTexture(&texDesc, subData.data(), &textureArray_);
}

// ── Mega-buffer rebuild ─────────────────────────────────────────
// Packs all chunk quads contiguously into a single buffer.
// Simple strategy: full rebuild whenever any chunk is dirty.

void VoxelRenderer::rebuildMegaBuffer(VoxelWorld& world) {
    cpuMegaQuads_.clear();
    chunkSlots_.clear();
    cpuChunkInfo_.clear();

    uint32_t offset = 0;
    float debugFlag = debugFaceColors_ ? 1.0f : 0.0f;

    world.forEachChunk([&](const ChunkPos& pos, Chunk& chunk) {
        if (chunk.quadCount == 0) return;
        if (offset + chunk.quadCount > MEGA_BUFFER_CAPACITY) return; // overflow guard

        ChunkSlot slot;
        slot.pos = pos;
        slot.quadOffset = offset;
        slot.quadCount = chunk.quadCount;
        chunkSlots_.push_back(slot);

        GPUChunkInfo info = {};
        info.worldPos = XMFLOAT4(
            (float)(pos.x * CHUNK_SIZE),
            (float)(pos.y * CHUNK_SIZE),
            (float)(pos.z * CHUNK_SIZE),
            debugFlag
        );
        info.quadOffset = offset;
        info.quadCount = chunk.quadCount;
        for (int f = 0; f < 6; f++) {
            info.faceOffsets[f] = chunk.faceOffsets[f];
            info.faceCounts[f] = chunk.faceCounts[f];
        }
        cpuChunkInfo_.push_back(info);

        cpuMegaQuads_.insert(cpuMegaQuads_.end(), chunk.quads.begin(), chunk.quads.end());
        offset += chunk.quadCount;
    });

    chunkCount_ = (uint32_t)chunkSlots_.size();
    totalQuads_ = offset;
}

// Build chunkInfoBuffer without CPU meshing (for GPU mesh path)
void VoxelRenderer::rebuildChunkInfoOnly(VoxelWorld& world) {
    chunkSlots_.clear();
    cpuChunkInfo_.clear();

    uint32_t idx = 0;
    float debugFlag = debugFaceColors_ ? 1.0f : 0.0f;

    world.forEachChunk([&](const ChunkPos& pos, Chunk& chunk) {
        ChunkSlot slot;
        slot.pos = pos;
        slot.quadOffset = 0; // not used in GPU mesh path
        slot.quadCount = 0;
        chunkSlots_.push_back(slot);

        GPUChunkInfo info = {};
        info.worldPos = XMFLOAT4(
            (float)(pos.x * CHUNK_SIZE),
            (float)(pos.y * CHUNK_SIZE),
            (float)(pos.z * CHUNK_SIZE),
            debugFlag
        );
        info.quadOffset = 0;
        info.quadCount = 0;
        cpuChunkInfo_.push_back(info);
        idx++;
    });

    chunkCount_ = (uint32_t)chunkSlots_.size();
}

void VoxelRenderer::updateMeshes(VoxelWorld& world) {
    if (!device_) return;

    // GPU mesh path: skip CPU meshing entirely, just rebuild chunk info
    if (gpuMeshEnabled_ && gpuMesherAvailable_) {
        bool anyDirty = false;
        world.forEachChunk([&](const ChunkPos& pos, Chunk& chunk) {
            if (chunk.dirty) { anyDirty = true; chunk.dirty = false; }
        });
        if (anyDirty || megaBufferDirty_) {
            rebuildChunkInfoOnly(world);
            // If cache wasn't already filled by fused regen+pack, mark for repack
            if (!gpuMeshDirty_) {
                // Non-fused dirty (e.g. initial load): need both repack and GPU update
                voxelCacheDirty_ = true;
                gpuMeshDirty_ = true;
            }
            // else: fused path already set gpuMeshDirty_=true, cache is clean
            chunkInfoDirty_ = true;
            megaBufferDirty_ = false;
        }
        return;
    }

    // CPU meshing path (fallback)
    // Collect dirty chunks for parallel meshing
    std::vector<Chunk*> dirtyChunks;
    world.forEachChunk([&](const ChunkPos& pos, Chunk& chunk) {
        if (chunk.dirty) dirtyChunks.push_back(&chunk);
    });
    bool anyDirty = !dirtyChunks.empty();

    // Parallel CPU greedy meshing via wi::jobsystem
    auto cpuStart = std::chrono::high_resolution_clock::now();
    if (anyDirty) {
        wi::jobsystem::context ctx;
        wi::jobsystem::Dispatch(ctx, (uint32_t)dirtyChunks.size(), 1,
            [&dirtyChunks, &world](wi::jobsystem::JobArgs args) {
                VoxelMesher::meshChunk(*dirtyChunks[args.jobIndex], world);
            });
        wi::jobsystem::Wait(ctx);
    }
    auto cpuEnd = std::chrono::high_resolution_clock::now();

    if (anyDirty) {
        cpuMeshTimeMs_ = std::chrono::duration<float, std::milli>(cpuEnd - cpuStart).count();
        // Trigger GPU benchmark on next render frame
        if (gpuMesherAvailable_ && benchState_ == BenchState::IDLE) {
            benchState_ = BenchState::DISPATCH;
        }
    }

    if (anyDirty || megaBufferDirty_) {
        rebuildMegaBuffer(world);
        megaBufferDirty_ = false;
    }
}

// ── GPU Mesh Benchmark (Phase 2.4) ──────────────────────────────
// Dispatches the baseline 1x1 GPU mesher for ALL chunks and measures timing.
// State machine: DISPATCH (frame N) → READBACK (frame N+1) → DONE.

void VoxelRenderer::dispatchGpuMeshBenchmark(CommandList cmd, const VoxelWorld& world) const {
    auto* dev = device_;

    // Zero the quad counter
    uint32_t zero = 0;
    dev->UpdateBuffer(&gpuQuadCounter_, &zero, cmd, sizeof(uint32_t));

    // Barrier: COPY_DST → UAV for counter, UNDEFINED → UAV for output buffer
    GPUBarrier preBarriers[] = {
        GPUBarrier::Buffer(&gpuQuadCounter_, ResourceState::COPY_DST, ResourceState::UNORDERED_ACCESS),
        GPUBarrier::Buffer(&gpuQuadBuffer_, ResourceState::UNDEFINED, ResourceState::UNORDERED_ACCESS),
    };
    dev->Barrier(preBarriers, 2, cmd);

    dev->BindComputeShader(&meshShader_, cmd);

    // GPU timestamp: mesh begin
    dev->QueryEnd(&timestampHeap_, TS_MESH_BEGIN, cmd);

    // Dispatch for each chunk
    uint32_t chunkIdx = 0;
    world.forEachChunk([&](const ChunkPos& pos, Chunk& chunk) {
        // Pack voxel data: 32^3 voxels → 16384 uint32s (2 voxels per uint)
        std::vector<uint32_t> packed(CHUNK_VOLUME / 2, 0);
        for (int i = 0; i < CHUNK_VOLUME; i++) {
            uint32_t v = chunk.voxels[i].packed;
            if (i & 1)
                packed[i >> 1] |= (v << 16);
            else
                packed[i >> 1] = v;
        }

        // Upload voxel data (re-uses the single-chunk buffer)
        dev->UpdateBuffer(&voxelDataBuffer_, packed.data(), cmd,
            packed.size() * sizeof(uint32_t));

        // Bind resources (after BindComputeShader, so PushConstants targets compute)
        dev->BindResource(&voxelDataBuffer_, 0, cmd);
        dev->BindUAV(&gpuQuadBuffer_, 0, cmd);
        dev->BindUAV(&gpuQuadCounter_, 1, cmd);

        // Push constants for this chunk
        struct MeshPush {
            uint32_t chunkIndex;
            uint32_t voxelBufferOffset;
            uint32_t quadBufferOffset;
            uint32_t maxOutputQuads;
            uint32_t pad[8];
        };
        MeshPush pushData = {};
        pushData.chunkIndex = chunkIdx;
        pushData.voxelBufferOffset = 0; // single-chunk buffer, always at offset 0
        pushData.quadBufferOffset = 0;  // all chunks share global atomic counter
        pushData.maxOutputQuads = MEGA_BUFFER_CAPACITY;
        dev->PushConstants(&pushData, sizeof(pushData), cmd);

        // Dispatch: 32/8 = 4 groups per axis → 64 groups total
        dev->Dispatch(4, 4, 4, cmd);

        chunkIdx++;
    });

    // GPU timestamp: mesh end
    dev->QueryEnd(&timestampHeap_, TS_MESH_END, cmd);

    // Copy quad counter to readback buffer
    GPUBarrier postBarrier = GPUBarrier::Buffer(
        &gpuQuadCounter_, ResourceState::UNORDERED_ACCESS, ResourceState::COPY_SRC);
    dev->Barrier(&postBarrier, 1, cmd);
    dev->CopyBuffer(&meshCounterReadback_, 0, &gpuQuadCounter_, 0, sizeof(uint32_t), cmd);

    // Resolve timestamps
    dev->QueryResolve(&timestampHeap_, TS_MESH_BEGIN, 2, &timestampReadback_,
        TS_MESH_BEGIN * sizeof(uint64_t), cmd);

    benchState_ = BenchState::READBACK;
}

void VoxelRenderer::readbackGpuMeshBenchmark() const {
    // Read quad count from readback buffer
    uint32_t* countData = (uint32_t*)meshCounterReadback_.mapped_data;
    if (countData) {
        gpuBaselineQuads_ = *countData;
    }

    // Read GPU mesh timestamps
    uint64_t* tsData = (uint64_t*)timestampReadback_.mapped_data;
    if (tsData) {
        double freq = (double)device_->GetTimestampFrequency();
        if (freq > 0.0 && tsData[TS_MESH_END] > tsData[TS_MESH_BEGIN]) {
            gpuMeshTimeMs_ = (float)((double)(tsData[TS_MESH_END] - tsData[TS_MESH_BEGIN]) / freq * 1000.0);
        }
    }

    // Log benchmark results
    char msg[256];
    snprintf(msg, sizeof(msg),
        "=== MESH BENCHMARK ===\n"
        "  CPU greedy:    %.2f ms, %u quads (%u chunks)\n"
        "  GPU baseline:  %.3f ms, %u quads (1x1, no merge)\n"
        "  Ratio quads: %.1fx more (GPU baseline vs CPU greedy)",
        cpuMeshTimeMs_, totalQuads_, chunkCount_,
        gpuMeshTimeMs_, gpuBaselineQuads_,
        totalQuads_ > 0 ? (float)gpuBaselineQuads_ / totalQuads_ : 0.0f);
    wi::backlog::post(msg);

    benchState_ = BenchState::DONE;
}

// ── GPU Mesh Dispatch (production path) ─────────────────────────
// Dispatches GPU mesher for ALL chunks every frame. Replaces CPU greedy meshing.
// Uses the atomic quad counter for 1-frame-delayed readback of total quad count.

void VoxelRenderer::dispatchGpuMesh(CommandList cmd, const VoxelWorld& world,
    ProfileAccum* profPack, ProfileAccum* profUpload, ProfileAccum* profDispatch) const {
    auto* dev = device_;

    // Zero the quad counter
    uint32_t zero = 0;
    dev->UpdateBuffer(&gpuQuadCounter_, &zero, cmd, sizeof(uint32_t));

    // Barrier: COPY_DST → UAV for counter, UNDEFINED → UAV for output buffer
    GPUBarrier preBarriers[] = {
        GPUBarrier::Buffer(&gpuQuadCounter_, ResourceState::COPY_DST, ResourceState::UNORDERED_ACCESS),
        GPUBarrier::Buffer(&gpuQuadBuffer_, ResourceState::UNDEFINED, ResourceState::UNORDERED_ACCESS),
    };
    dev->Barrier(preBarriers, 2, cmd);

    dev->BindComputeShader(&meshShader_, cmd);

    // Pack and upload all chunks' voxel data
    // Each chunk = 32^3/2 = 16384 uint32 (two voxels per uint)
    const uint32_t wordsPerChunk = CHUNK_VOLUME / 2;
    uint32_t totalWords = chunkCount_ * wordsPerChunk;

    // Resize voxel data buffer if needed
    if (totalWords > voxelDataCapacity_) {
        voxelDataCapacity_ = totalWords;
        GPUBufferDesc voxDesc;
        voxDesc.size = totalWords * sizeof(uint32_t);
        voxDesc.bind_flags = BindFlag::SHADER_RESOURCE;
        voxDesc.misc_flags = ResourceMiscFlag::BUFFER_STRUCTURED;
        voxDesc.stride = sizeof(uint32_t);
        voxDesc.usage = Usage::DEFAULT;
        dev->CreateBuffer(&voxDesc, nullptr, const_cast<GPUBuffer*>(&voxelDataBuffer_));
    }

    // Pack voxel data — use cached copy, only update when dirty.
    // VoxelData is exactly uint16_t, so voxels[] is a packed uint16 array.
    // Two consecutive uint16 = one uint32 → direct memcpy, no bit manipulation.
    static_assert(sizeof(VoxelData) == sizeof(uint16_t),
        "VoxelData must be 2 bytes for direct memcpy to GPU buffer");

    auto tPack0 = std::chrono::high_resolution_clock::now();
    if (voxelCacheDirty_) {
        packedVoxelCache_.resize(totalWords);
        uint32_t chunkI = 0;
        world.forEachChunk([&](const ChunkPos& pos, const Chunk& chunk) {
            std::memcpy(
                packedVoxelCache_.data() + chunkI * wordsPerChunk,
                chunk.voxels,
                wordsPerChunk * sizeof(uint32_t) // = CHUNK_VOLUME * 2 bytes
            );
            chunkI++;
        });
        voxelCacheDirty_ = false;
    }
    auto tPack1 = std::chrono::high_resolution_clock::now();
    if (profPack) profPack->add(std::chrono::duration<float, std::milli>(tPack1 - tPack0).count());

    // Upload all voxel data at once
    auto tUpload0 = std::chrono::high_resolution_clock::now();
    dev->UpdateBuffer(&voxelDataBuffer_, packedVoxelCache_.data(), cmd,
        totalWords * sizeof(uint32_t));
    auto tUpload1 = std::chrono::high_resolution_clock::now();
    if (profUpload) profUpload->add(std::chrono::duration<float, std::milli>(tUpload1 - tUpload0).count());

    // Bind resources (shared across all chunk dispatches)
    dev->BindResource(&voxelDataBuffer_, 0, cmd);
    dev->BindUAV(&gpuQuadBuffer_, 0, cmd);
    dev->BindUAV(&gpuQuadCounter_, 1, cmd);

    // Dispatch for each chunk
    struct MeshPush {
        uint32_t chunkIndex;
        uint32_t voxelBufferOffset;
        uint32_t quadBufferOffset;
        uint32_t maxOutputQuads;
        uint32_t pad[8];
    };

    auto tDisp0 = std::chrono::high_resolution_clock::now();
    uint32_t chunkIdx = 0;
    world.forEachChunk([&](const ChunkPos& pos, const Chunk& chunk) {
        MeshPush pushData = {};
        pushData.chunkIndex = chunkIdx;
        pushData.voxelBufferOffset = chunkIdx * wordsPerChunk;
        pushData.quadBufferOffset = 0;  // global atomic counter handles offsets
        pushData.maxOutputQuads = MEGA_BUFFER_CAPACITY;
        dev->PushConstants(&pushData, sizeof(pushData), cmd);

        // Dispatch: 32/8 = 4 groups per axis → 64 groups per chunk
        dev->Dispatch(4, 4, 4, cmd);
        chunkIdx++;
    });
    auto tDisp1 = std::chrono::high_resolution_clock::now();
    if (profDispatch) profDispatch->add(std::chrono::duration<float, std::milli>(tDisp1 - tDisp0).count());

    // Barriers: UAV → COPY_SRC for counter readback, UAV → SRV for quad buffer (rendering)
    GPUBarrier postBarriers[] = {
        GPUBarrier::Buffer(&gpuQuadCounter_, ResourceState::UNORDERED_ACCESS, ResourceState::COPY_SRC),
        GPUBarrier::Buffer(&gpuQuadBuffer_, ResourceState::UNORDERED_ACCESS, ResourceState::SHADER_RESOURCE),
    };
    dev->Barrier(postBarriers, 2, cmd);

    // Copy quad counter to readback buffer (result available next frame)
    dev->CopyBuffer(&meshCounterReadback_, 0, &gpuQuadCounter_, 0, sizeof(uint32_t), cmd);

    totalQuads_ = gpuMeshQuadCount_; // display previous frame's count in HUD
    gpuMeshDirty_ = false;
}

// ── Frustum plane extraction (Gribb-Hartmann method) ────────────
static void extractFrustumPlanes(const XMMATRIX& vp, XMFLOAT4 planes[6]) {
    XMFLOAT4X4 m;
    XMStoreFloat4x4(&m, vp);

    // Left
    planes[0] = XMFLOAT4(m._14 + m._11, m._24 + m._21, m._34 + m._31, m._44 + m._41);
    // Right
    planes[1] = XMFLOAT4(m._14 - m._11, m._24 - m._21, m._34 - m._31, m._44 - m._41);
    // Bottom
    planes[2] = XMFLOAT4(m._14 + m._12, m._24 + m._22, m._34 + m._32, m._44 + m._42);
    // Top
    planes[3] = XMFLOAT4(m._14 - m._12, m._24 - m._22, m._34 - m._32, m._44 - m._42);
    // Near
    planes[4] = XMFLOAT4(m._13, m._23, m._33, m._43);
    // Far
    planes[5] = XMFLOAT4(m._14 - m._13, m._24 - m._23, m._34 - m._33, m._44 - m._43);

    // Normalize each plane
    for (int i = 0; i < 6; i++) {
        float len = std::sqrt(planes[i].x * planes[i].x +
                              planes[i].y * planes[i].y +
                              planes[i].z * planes[i].z);
        if (len > 0.0001f) {
            planes[i].x /= len;
            planes[i].y /= len;
            planes[i].z /= len;
            planes[i].w /= len;
        }
    }
}

// ── Render pass ─────────────────────────────────────────────────

void VoxelRenderer::render(
    CommandList cmd,
    const wi::scene::CameraComponent& camera,
    const Texture& depthBuffer,
    const Texture& renderTarget
) const {
    if (!initialized_ || chunkCount_ == 0 || !pso_.IsValid()) return;

    auto* dev = device_;

    // ── GPU Mesh path: quads already dispatched in Render(), just draw ──
    if (gpuMeshEnabled_ && gpuMesherAvailable_) {
        // Upload chunk info only when chunks changed
        if (!cpuChunkInfo_.empty() && chunkInfoDirty_) {
            dev->UpdateBuffer(&chunkInfoBuffer_, cpuChunkInfo_.data(), cmd,
                cpuChunkInfo_.size() * sizeof(GPUChunkInfo));
            chunkInfoDirty_ = false;
        }

        // Per-frame constants
        VoxelConstants cb = {};
        XMMATRIX vpMatrix = camera.GetViewProjection();
        XMStoreFloat4x4(&cb.viewProjection, vpMatrix);
        cb.cameraPosition = XMFLOAT4(camera.Eye.x, camera.Eye.y, camera.Eye.z, 1.0f);
        cb.sunDirection = XMFLOAT4(-0.5f, -0.8f, -0.3f, 0.0f);
        cb.sunColor = XMFLOAT4(1.2f, 1.1f, 0.9f, 1.0f);
        cb.chunkSize = (float)CHUNK_SIZE;
        cb.textureTiling = 0.25f;
        cb.blendEnabled = 1.0f; // Phase 3: PS-based blending enabled in GPU mesh path
        cb.debugBlend = debugBlend_ ? 1.0f : 0.0f;
        cb.chunkCount = chunkCount_;
        // Per-material blend flags (bit N = material N):
        // canBleed: material can overflow visually onto adjacent voxels
        // resistBleed: adjacent materials cannot overflow onto this material
        // Material IDs: 1=Grass, 2=Dirt, 3=Stone, 4=Sand, 5=Snow
        cb.bleedMask = (1u << 1) | (1u << 2) | (1u << 4) | (1u << 5); // Grass, Dirt, Sand, Snow can bleed (NOT Stone)
        cb.resistBleedMask = (1u << 1); // Grass resists bleed (she bleeds onto others, not the reverse)
        cb._cullPad2 = 0;
        dev->UpdateBuffer(&constantBuffer_, &cb, cmd, sizeof(cb));

        // Render pass
        RenderPassImage rp[] = {
            RenderPassImage::RenderTarget(
                &renderTarget,
                RenderPassImage::LoadOp::CLEAR,
                RenderPassImage::StoreOp::STORE,
                ResourceState::SHADER_RESOURCE,
                ResourceState::SHADER_RESOURCE
            ),
            RenderPassImage::DepthStencil(
                &depthBuffer,
                RenderPassImage::LoadOp::CLEAR,
                RenderPassImage::StoreOp::STORE,
                ResourceState::DEPTHSTENCIL,
                ResourceState::DEPTHSTENCIL,
                ResourceState::DEPTHSTENCIL
            ),
        };
        dev->RenderPassBegin(rp, 2, cmd);

        Viewport vp;
        vp.width = (float)renderTarget.GetDesc().width;
        vp.height = (float)renderTarget.GetDesc().height;
        vp.min_depth = 0.0f;
        vp.max_depth = 1.0f;
        dev->BindViewports(1, &vp, cmd);

        Rect scissor = { 0, 0, (int)vp.width, (int)vp.height };
        dev->BindScissorRects(1, &scissor, cmd);

        dev->BindPipelineState(&pso_, cmd);
        dev->BindConstantBuffer(&constantBuffer_, 0, cmd);
        dev->BindResource(&gpuQuadBuffer_, 0, cmd);  // GPU quads, not mega-buffer
        dev->BindResource(&textureArray_, 1, cmd);
        dev->BindResource(&chunkInfoBuffer_, 2, cmd);
        dev->BindResource(&voxelDataBuffer_, 3, cmd); // Phase 3: voxel data for PS neighbor lookups
        dev->BindSampler(&sampler_, 0, cmd);

        // GPU mesh mode: flags=2, MUST be after BindPipelineState
        struct VoxelPush {
            uint32_t chunkIndex;
            uint32_t quadOffset;
            uint32_t flags;
            uint32_t pad[9];
        };
        VoxelPush pushData = {};
        pushData.flags = 2; // GPU mesh mode
        pushData.quadOffset = 0;
        dev->PushConstants(&pushData, sizeof(pushData), cmd);

        // Draw using previous frame's quad count (1-frame delay)
        if (gpuMeshQuadCount_ > 0) {
            dev->DrawInstanced(gpuMeshQuadCount_ * 6, 1, 0, 0, cmd);
            drawCalls_ = 1;
        }

        dev->RenderPassEnd(cmd);
        visibleChunks_ = chunkCount_;
        return;
    }

    // Upload mega-buffer and chunk info to GPU
    if (!cpuMegaQuads_.empty()) {
        dev->UpdateBuffer(&megaQuadBuffer_, cpuMegaQuads_.data(), cmd,
            cpuMegaQuads_.size() * sizeof(PackedQuad));
    }
    if (!cpuChunkInfo_.empty()) {
        dev->UpdateBuffer(&chunkInfoBuffer_, cpuChunkInfo_.data(), cmd,
            cpuChunkInfo_.size() * sizeof(GPUChunkInfo));
    }

    // Per-frame constants (with frustum planes for GPU cull shader)
    VoxelConstants cb = {};
    XMMATRIX vpMatrix = camera.GetViewProjection();
    XMStoreFloat4x4(&cb.viewProjection, vpMatrix);
    cb.cameraPosition = XMFLOAT4(camera.Eye.x, camera.Eye.y, camera.Eye.z, 1.0f);
    cb.sunDirection = XMFLOAT4(-0.5f, -0.8f, -0.3f, 0.0f);
    cb.sunColor = XMFLOAT4(1.2f, 1.1f, 0.9f, 1.0f);
    cb.chunkSize = (float)CHUNK_SIZE;
    cb.textureTiling = 0.25f;
    cb.blendEnabled = 0.0f; // Phase 3: blending disabled in CPU/MDI paths (no voxel data SRV)
    cb.debugBlend = 0.0f;
    cb.bleedMask = 0;
    cb.resistBleedMask = 0;
    cb._cullPad2 = 0;
    cb.chunkCount = chunkCount_;
    extractFrustumPlanes(vpMatrix, cb.frustumPlanes);
    dev->UpdateBuffer(&constantBuffer_, &cb, cmd, sizeof(cb));

    // Push constant structure (must be 48 bytes = 12 x uint32, matches b999)
    struct VoxelPush {
        uint32_t chunkIndex;
        uint32_t quadOffset;
        uint32_t flags;       // bit 0: 1=MDI mode, 0=CPU mode
        uint32_t pad[9];
    };

    visibleChunks_ = 0;
    drawCalls_ = 0;

    // ── GPU Cull + MDI path ────────────────────────────────────────
    if (gpuCullingEnabled_) {
        // DX12 buffer decay: all buffers return to COMMON after ExecuteCommandLists.
        // So every frame starts clean — no cross-frame state tracking needed.

        // Zero the draw count via UpdateBuffer (COMMON → COPY_DST implicit promotion)
        uint32_t zero = 0;
        dev->UpdateBuffer(&drawCountBuffer_, &zero, cmd, sizeof(uint32_t));

        // Barriers to UAV for compute shader writes:
        // - drawCountBuffer_: COPY_DST → UAV (was promoted to COPY_DST by UpdateBuffer)
        // - indirectArgsBuffer_: COMMON → UAV (explicit, required because COMMON can't
        //   be implicitly promoted to UAV)
        GPUBarrier preBarriers[] = {
            GPUBarrier::Buffer(&drawCountBuffer_, ResourceState::COPY_DST, ResourceState::UNORDERED_ACCESS),
            GPUBarrier::Buffer(&indirectArgsBuffer_, ResourceState::UNDEFINED, ResourceState::UNORDERED_ACCESS),
        };
        dev->Barrier(preBarriers, 2, cmd);

        // Timestamp: cull begin
        dev->QueryEnd(&timestampHeap_, TS_CULL_BEGIN, cmd);

        // Dispatch GPU frustum + backface cull compute shader
        dev->BindComputeShader(&cullShader_, cmd);
        dev->BindConstantBuffer(&constantBuffer_, 0, cmd);
        dev->BindResource(&chunkInfoBuffer_, 2, cmd);
        dev->BindUAV(&indirectArgsBuffer_, 0, cmd);
        dev->BindUAV(&drawCountBuffer_, 1, cmd);
        dev->Dispatch((chunkCount_ + 63) / 64, 1, 1, cmd);

        // Timestamp: cull end
        dev->QueryEnd(&timestampHeap_, TS_CULL_END, cmd);

        // Barriers: UAV → INDIRECT_ARGUMENT for DrawInstancedIndirectCount
        GPUBarrier postBarriers[] = {
            GPUBarrier::Buffer(&indirectArgsBuffer_, ResourceState::UNORDERED_ACCESS, ResourceState::INDIRECT_ARGUMENT),
            GPUBarrier::Buffer(&drawCountBuffer_, ResourceState::UNORDERED_ACCESS, ResourceState::INDIRECT_ARGUMENT),
        };
        dev->Barrier(postBarriers, 2, cmd);

        // ── Render pass ────────────────────────────────────────────
        RenderPassImage rp[] = {
            RenderPassImage::RenderTarget(
                &renderTarget,
                RenderPassImage::LoadOp::CLEAR,
                RenderPassImage::StoreOp::STORE,
                ResourceState::SHADER_RESOURCE,
                ResourceState::SHADER_RESOURCE
            ),
            RenderPassImage::DepthStencil(
                &depthBuffer,
                RenderPassImage::LoadOp::CLEAR,
                RenderPassImage::StoreOp::STORE,
                ResourceState::DEPTHSTENCIL,
                ResourceState::DEPTHSTENCIL,
                ResourceState::DEPTHSTENCIL
            ),
        };
        dev->RenderPassBegin(rp, 2, cmd);

        Viewport vp;
        vp.width = (float)renderTarget.GetDesc().width;
        vp.height = (float)renderTarget.GetDesc().height;
        vp.min_depth = 0.0f;
        vp.max_depth = 1.0f;
        dev->BindViewports(1, &vp, cmd);

        Rect scissor = { 0, 0, (int)vp.width, (int)vp.height };
        dev->BindScissorRects(1, &scissor, cmd);

        dev->BindPipelineState(&pso_, cmd);
        dev->BindConstantBuffer(&constantBuffer_, 0, cmd);
        dev->BindResource(&megaQuadBuffer_, 0, cmd);
        dev->BindResource(&textureArray_, 1, cmd);
        dev->BindResource(&chunkInfoBuffer_, 2, cmd);
        dev->BindSampler(&sampler_, 0, cmd);

        // IMPORTANT: PushConstants must be called AFTER BindPipelineState.
        // Wicked Engine's PushConstants uses SetGraphicsRoot32BitConstants only
        // when active_pso is set. If called before (with active_cs from compute),
        // it would set COMPUTE push constants instead of GRAPHICS ones.
        VoxelPush pushData = {};
        pushData.flags = 1; // MDI mode
        dev->PushConstants(&pushData, sizeof(pushData), cmd);

        // Timestamp: draw begin
        dev->QueryEnd(&timestampHeap_, TS_DRAW_BEGIN, cmd);

        // Single MDI call: GPU cull shader filled the indirect args
        dev->DrawInstancedIndirectCount(
            &indirectArgsBuffer_, 0,
            &drawCountBuffer_, 0,
            MAX_DRAWS, cmd
        );
        drawCalls_ = 1;

        // Timestamp: draw end
        dev->QueryEnd(&timestampHeap_, TS_DRAW_END, cmd);

        dev->RenderPassEnd(cmd);

        // Resolve timestamps for readback (results available next frame)
        dev->QueryResolve(&timestampHeap_, 0, TS_COUNT, &timestampReadback_, 0, cmd);

        // Read back previous frame's timestamps (persistently mapped READBACK buffer)
        uint64_t* tsData = (uint64_t*)timestampReadback_.mapped_data;
        if (tsData) {
            double freq = (double)dev->GetTimestampFrequency();
            if (freq > 0.0 && tsData[TS_CULL_END] > tsData[TS_CULL_BEGIN]) {
                gpuCullTimeMs_ = (float)((double)(tsData[TS_CULL_END] - tsData[TS_CULL_BEGIN]) / freq * 1000.0);
            }
            if (freq > 0.0 && tsData[TS_DRAW_END] > tsData[TS_DRAW_BEGIN]) {
                gpuDrawTimeMs_ = (float)((double)(tsData[TS_DRAW_END] - tsData[TS_DRAW_BEGIN]) / freq * 1000.0);
            }
        }

        // GPU cull handles visibility counting — approximate from chunkCount
        visibleChunks_ = chunkCount_; // exact count would require readback of drawCount

        return;
    }

    // ── CPU frustum + backface cull (shared by MDI and per-face paths) ──
    wi::primitive::Frustum frustum;
    frustum.Create(camera.GetViewProjection());

    // ── Phase 2.2: CPU-filled indirect args + MDI draw ──────────────
    if (mdiEnabled_) {
        // CPU cull: fill indirect args with visible face groups
        cpuIndirectArgs_.clear();
        uint32_t cpuDrawCount = 0;

        for (uint32_t i = 0; i < chunkCount_; i++) {
            const auto& slot = chunkSlots_[i];
            if (slot.quadCount == 0) continue;

            XMFLOAT3 aabbMin(
                (float)(slot.pos.x * CHUNK_SIZE),
                (float)(slot.pos.y * CHUNK_SIZE),
                (float)(slot.pos.z * CHUNK_SIZE)
            );
            XMFLOAT3 aabbMax(
                aabbMin.x + CHUNK_SIZE,
                aabbMin.y + CHUNK_SIZE,
                aabbMin.z + CHUNK_SIZE
            );
            wi::primitive::AABB aabb(aabbMin, aabbMax);
            if (!frustum.CheckBoxFast(aabb)) continue;

            visibleChunks_++;
            const auto& info = cpuChunkInfo_[i];

            for (uint32_t f = 0; f < 6; f++) {
                if (info.faceCounts[f] == 0) continue;

                bool backFacing = false;
                switch (f) {
                case 0: backFacing = (camera.Eye.x < aabbMin.x); break;
                case 1: backFacing = (camera.Eye.x > aabbMax.x); break;
                case 2: backFacing = (camera.Eye.y < aabbMin.y); break;
                case 3: backFacing = (camera.Eye.y > aabbMax.y); break;
                case 4: backFacing = (camera.Eye.z < aabbMin.z); break;
                case 5: backFacing = (camera.Eye.z > aabbMax.z); break;
                }
                if (backFacing) continue;

                IndirectDrawArgs args = {};
                // Pack chunkIndex (low 16 bits) + faceIndex (high 16 bits) into push constant.
                // The shader unpacks this to look up quadOffset from GPUChunkInfo.
                // We do NOT use startVertexLocation because SV_VertexID may not include it
                // reliably in ExecuteIndirect context.
                args.pushConstant = i | (f << 16);
                args.vertexCountPerInstance = info.faceCounts[f] * 6;
                args.instanceCount = 1;
                args.startVertexLocation = 0;
                args.startInstanceLocation = 0;
                cpuIndirectArgs_.push_back(args);
                cpuDrawCount++;
            }
        }

        // Upload indirect args and draw count to GPU
        // Note: no explicit barriers needed here. Buffers start in COMMON each frame
        // (DX12 buffer decay after command list execution). COMMON is implicitly
        // promoted to COPY_DST by UpdateBuffer, then to INDIRECT_ARGUMENT by
        // DrawInstancedIndirectCount. This matches Phase 2.1 pattern (no barriers
        // between UpdateBuffer and SRV usage for megaQuadBuffer_/chunkInfoBuffer_).
        if (!cpuIndirectArgs_.empty()) {
            dev->UpdateBuffer(&indirectArgsBuffer_, cpuIndirectArgs_.data(), cmd,
                cpuIndirectArgs_.size() * sizeof(IndirectDrawArgs));
        }
        dev->UpdateBuffer(&drawCountBuffer_, &cpuDrawCount, cmd, sizeof(uint32_t));

        // ── Render pass ────────────────────────────────────────────
        RenderPassImage rp[] = {
            RenderPassImage::RenderTarget(
                &renderTarget,
                RenderPassImage::LoadOp::CLEAR,
                RenderPassImage::StoreOp::STORE,
                ResourceState::SHADER_RESOURCE,
                ResourceState::SHADER_RESOURCE
            ),
            RenderPassImage::DepthStencil(
                &depthBuffer,
                RenderPassImage::LoadOp::CLEAR,
                RenderPassImage::StoreOp::STORE,
                ResourceState::DEPTHSTENCIL,
                ResourceState::DEPTHSTENCIL,
                ResourceState::DEPTHSTENCIL
            ),
        };
        dev->RenderPassBegin(rp, 2, cmd);

        Viewport vp;
        vp.width = (float)renderTarget.GetDesc().width;
        vp.height = (float)renderTarget.GetDesc().height;
        vp.min_depth = 0.0f;
        vp.max_depth = 1.0f;
        dev->BindViewports(1, &vp, cmd);

        Rect scissor = { 0, 0, (int)vp.width, (int)vp.height };
        dev->BindScissorRects(1, &scissor, cmd);

        dev->BindPipelineState(&pso_, cmd);
        dev->BindConstantBuffer(&constantBuffer_, 0, cmd);
        dev->BindResource(&megaQuadBuffer_, 0, cmd);
        dev->BindResource(&textureArray_, 1, cmd);
        dev->BindResource(&chunkInfoBuffer_, 2, cmd);
        dev->BindSampler(&sampler_, 0, cmd);

        // MDI mode: VS uses binary search to find chunk from SV_VertexID
        VoxelPush pushData = {};
        pushData.flags = 1; // MDI mode
        dev->PushConstants(&pushData, sizeof(pushData), cmd);

        dev->DrawInstancedIndirectCount(
            &indirectArgsBuffer_, 0,
            &drawCountBuffer_, 0,
            MAX_DRAWS, cmd
        );
        drawCalls_ = 1;

        dev->RenderPassEnd(cmd);
        return;
    }

    // ── Phase 2.1 Fallback: per-face-group DrawInstanced ────────────
    RenderPassImage rp[] = {
        RenderPassImage::RenderTarget(
            &renderTarget,
            RenderPassImage::LoadOp::CLEAR,
            RenderPassImage::StoreOp::STORE,
            ResourceState::SHADER_RESOURCE,
            ResourceState::SHADER_RESOURCE
        ),
        RenderPassImage::DepthStencil(
            &depthBuffer,
            RenderPassImage::LoadOp::CLEAR,
            RenderPassImage::StoreOp::STORE,
            ResourceState::DEPTHSTENCIL,
            ResourceState::DEPTHSTENCIL,
            ResourceState::DEPTHSTENCIL
        ),
    };
    dev->RenderPassBegin(rp, 2, cmd);

    Viewport vp;
    vp.width = (float)renderTarget.GetDesc().width;
    vp.height = (float)renderTarget.GetDesc().height;
    vp.min_depth = 0.0f;
    vp.max_depth = 1.0f;
    dev->BindViewports(1, &vp, cmd);

    Rect scissor = { 0, 0, (int)vp.width, (int)vp.height };
    dev->BindScissorRects(1, &scissor, cmd);

    dev->BindPipelineState(&pso_, cmd);
    dev->BindConstantBuffer(&constantBuffer_, 0, cmd);
    dev->BindResource(&megaQuadBuffer_, 0, cmd);
    dev->BindResource(&textureArray_, 1, cmd);
    dev->BindResource(&chunkInfoBuffer_, 2, cmd);
    dev->BindSampler(&sampler_, 0, cmd);

    for (uint32_t i = 0; i < chunkCount_; i++) {
        const auto& slot = chunkSlots_[i];
        if (slot.quadCount == 0) continue;

        XMFLOAT3 aabbMin(
            (float)(slot.pos.x * CHUNK_SIZE),
            (float)(slot.pos.y * CHUNK_SIZE),
            (float)(slot.pos.z * CHUNK_SIZE)
        );
        XMFLOAT3 aabbMax(
            aabbMin.x + CHUNK_SIZE,
            aabbMin.y + CHUNK_SIZE,
            aabbMin.z + CHUNK_SIZE
        );
        wi::primitive::AABB aabb(aabbMin, aabbMax);
        if (!frustum.CheckBoxFast(aabb)) continue;

        visibleChunks_++;
        const auto& info = cpuChunkInfo_[i];

        for (uint32_t f = 0; f < 6; f++) {
            if (info.faceCounts[f] == 0) continue;

            bool backFacing = false;
            switch (f) {
            case 0: backFacing = (camera.Eye.x < aabbMin.x); break;
            case 1: backFacing = (camera.Eye.x > aabbMax.x); break;
            case 2: backFacing = (camera.Eye.y < aabbMin.y); break;
            case 3: backFacing = (camera.Eye.y > aabbMax.y); break;
            case 4: backFacing = (camera.Eye.z < aabbMin.z); break;
            case 5: backFacing = (camera.Eye.z > aabbMax.z); break;
            }
            if (backFacing) continue;

            VoxelPush pushData = {};
            pushData.chunkIndex = i;
            pushData.quadOffset = slot.quadOffset + info.faceOffsets[f];
            pushData.flags = 0; // CPU mode
            dev->PushConstants(&pushData, sizeof(pushData), cmd);

            dev->DrawInstanced(info.faceCounts[f] * 6, 1, 0, 0, cmd);
            drawCalls_++;
        }
    }

    dev->RenderPassEnd(cmd);
}

// ── VoxelRenderPath (custom RenderPath3D) ───────────────────────

void VoxelRenderPath::Start() {
    RenderPath3D::Start();

    auto* device = wi::graphics::GetDevice();
    renderer.initialize(device);
    renderer.debugFaceColors_ = debugMode;

    // Generate world
    if (debugMode) {
        world.generateDebug();
        cameraPos = { 10.0f, 10.0f, 0.0f };
        cameraPitch = -0.4f;
        cameraYaw = 0.5f;
    } else {
        world.generateAround(cameraPos.x, cameraPos.y, cameraPos.z, 4);
    }
    if (renderer.isInitialized()) {
        renderer.updateMeshes(world);
    }

    // Phase 4: Initialize toping system and collect instances
    topingSystem.initialize();
    topingSystem.collectInstances(world);
    {
        char msg[256];
        snprintf(msg, sizeof(msg),
            "TopingSystem: %zu defs, %zu vertices, %zu instances",
            topingSystem.getDefCount(),
            topingSystem.getVertexCount(),
            topingSystem.getInstanceCount());
        wi::backlog::post(msg);
    }

    worldGenerated_ = true;

    setAO(AO_DISABLED);
    setFXAAEnabled(true);
    setBloomEnabled(false);

    createRenderTargets();
}

void VoxelRenderPath::createRenderTargets() {
    auto* device = wi::graphics::GetDevice();
    if (!device) return;

    uint32_t w = GetPhysicalWidth();
    uint32_t h = GetPhysicalHeight();
    if (w == 0 || h == 0) { w = 1920; h = 1080; }

    wi::graphics::TextureDesc rtDesc;
    rtDesc.type = wi::graphics::TextureDesc::Type::TEXTURE_2D;
    rtDesc.width = w;
    rtDesc.height = h;
    rtDesc.format = wi::graphics::Format::R8G8B8A8_UNORM;
    rtDesc.bind_flags = wi::graphics::BindFlag::RENDER_TARGET | wi::graphics::BindFlag::SHADER_RESOURCE;
    rtDesc.mip_levels = 1;
    rtDesc.sample_count = 1;
    rtDesc.layout = wi::graphics::ResourceState::SHADER_RESOURCE;
    device->CreateTexture(&rtDesc, nullptr, &voxelRT_);

    wi::graphics::TextureDesc depthDesc;
    depthDesc.type = wi::graphics::TextureDesc::Type::TEXTURE_2D;
    depthDesc.width = w;
    depthDesc.height = h;
    depthDesc.format = wi::graphics::Format::D32_FLOAT;
    depthDesc.bind_flags = wi::graphics::BindFlag::DEPTH_STENCIL | wi::graphics::BindFlag::SHADER_RESOURCE;
    depthDesc.mip_levels = 1;
    depthDesc.sample_count = 1;
    depthDesc.layout = wi::graphics::ResourceState::DEPTHSTENCIL;
    device->CreateTexture(&depthDesc, nullptr, &voxelDepth_);

    rtCreated_ = voxelRT_.IsValid() && voxelDepth_.IsValid();
    wi::backlog::post("VoxelRenderPath: render targets " + std::string(rtCreated_ ? "OK" : "FAILED")
        + " (" + std::to_string(w) + "x" + std::to_string(h) + ")");
}

// ── WASD camera input ───────────────────────────────────────────

static constexpr wi::input::BUTTON KEY_W = (wi::input::BUTTON)(wi::input::CHARACTER_RANGE_START + ('W' - 'A'));
static constexpr wi::input::BUTTON KEY_A = (wi::input::BUTTON)(wi::input::CHARACTER_RANGE_START + ('A' - 'A'));
static constexpr wi::input::BUTTON KEY_S = (wi::input::BUTTON)(wi::input::CHARACTER_RANGE_START + ('S' - 'A'));
static constexpr wi::input::BUTTON KEY_D = (wi::input::BUTTON)(wi::input::CHARACTER_RANGE_START + ('D' - 'A'));

void VoxelRenderPath::handleInput(float dt) {
    // F2: toggle backlog console
    if (wi::input::Press(wi::input::KEYBOARD_BUTTON_F2)) {
        wi::backlog::Toggle();
    }
    // F3: toggle animated terrain
    if (wi::input::Press(wi::input::KEYBOARD_BUTTON_F3)) {
        animatedTerrain_ = !animatedTerrain_;
        wi::backlog::post(animatedTerrain_ ? "Animation: ON (60 Hz)" : "Animation: OFF");
    }
    // F4: toggle blend debug visualization
    if (wi::input::Press(wi::input::KEYBOARD_BUTTON_F4)) {
        renderer.debugBlend_ = !renderer.debugBlend_;
        wi::backlog::post(renderer.debugBlend_ ? "Blend debug: ON" : "Blend debug: OFF");
    }
    if (wi::input::Press(wi::input::MOUSE_BUTTON_RIGHT)) {
        mouseCaptured = !mouseCaptured;
        wi::input::HidePointer(mouseCaptured);
    }

    if (mouseCaptured) {
        auto mouseState = wi::input::GetMouseState();
        cameraYaw += mouseState.delta_position.x * cameraSensitivity;
        cameraPitch += mouseState.delta_position.y * cameraSensitivity;
        cameraPitch = std::clamp(cameraPitch, -1.5f, 1.5f);
    }

    float cosPitch = std::cos(cameraPitch);
    XMFLOAT3 forward(
        std::sin(cameraYaw) * cosPitch,
        -std::sin(cameraPitch),
        std::cos(cameraYaw) * cosPitch
    );
    XMFLOAT3 right(std::cos(cameraYaw), 0.0f, -std::sin(cameraYaw));

    float speed = cameraSpeed * dt;
    if (wi::input::Down(wi::input::KEYBOARD_BUTTON_LSHIFT)) speed *= 3.0f;

    if (wi::input::Down(KEY_W)) { cameraPos.x += forward.x * speed; cameraPos.y += forward.y * speed; cameraPos.z += forward.z * speed; }
    if (wi::input::Down(KEY_S)) { cameraPos.x -= forward.x * speed; cameraPos.y -= forward.y * speed; cameraPos.z -= forward.z * speed; }
    if (wi::input::Down(KEY_A)) { cameraPos.x -= right.x * speed; cameraPos.z -= right.z * speed; }
    if (wi::input::Down(KEY_D)) { cameraPos.x += right.x * speed; cameraPos.z += right.z * speed; }
    if (wi::input::Down(wi::input::KEYBOARD_BUTTON_SPACE))    cameraPos.y += speed;
    if (wi::input::Down(wi::input::KEYBOARD_BUTTON_LCONTROL)) cameraPos.y -= speed;

    camera->Eye = cameraPos;
    camera->At = forward;
    camera->Up = XMFLOAT3(0, 1, 0);
    camera->UpdateCamera();
}

void VoxelRenderPath::Update(float dt) {
    auto frameStart = std::chrono::high_resolution_clock::now();
    lastDt_ = dt;
    float instantFps = (dt > 0.0f) ? (1.0f / dt) : 0.0f;
    smoothFps_ = smoothFps_ * 0.95f + instantFps * 0.05f;
    if (camera) handleInput(dt);

    // Animated terrain: regenerate at 60 Hz with time-shifted noise
    // Fused: regenerate + pack voxel data in the same parallel pass
    if (animatedTerrain_ && renderer.isInitialized()) {
        animAccum_ += dt;
        if (animAccum_ >= ANIM_INTERVAL) {
            animAccum_ -= ANIM_INTERVAL;
            animTime_ += ANIM_INTERVAL;

            // Prepare pack cache for fused regenerate+pack
            const uint32_t wordsPerChunk = CHUNK_VOLUME / 2;
            uint32_t totalWords = (uint32_t)world.chunkCount() * wordsPerChunk;
            renderer.packedVoxelCache_.resize(totalWords);

            auto t0 = std::chrono::high_resolution_clock::now();
            world.regenerateAnimated(animTime_,
                renderer.packedVoxelCache_.data(), totalWords);
            auto t1 = std::chrono::high_resolution_clock::now();
            profRegenerate_.add(std::chrono::duration<float, std::milli>(t1 - t0).count());

            renderer.voxelCacheDirty_ = false;    // cache already filled by fused pack
            renderer.gpuMeshDirty_ = true;        // GPU still needs upload + dispatch
        }
    }

    if (renderer.isInitialized()) {
        auto t0 = std::chrono::high_resolution_clock::now();
        renderer.updateMeshes(world);
        auto t1 = std::chrono::high_resolution_clock::now();
        profUpdateMeshes_.add(std::chrono::duration<float, std::milli>(t1 - t0).count());
    }
    RenderPath3D::Update(dt);

    // Profiling: accumulate frame time (will be completed in Compose)
    auto frameEnd = std::chrono::high_resolution_clock::now();
    profFrame_.add(std::chrono::duration<float, std::milli>(frameEnd - frameStart).count());

    // Log averages every 5 seconds
    profTimer_ += dt;
    if (profTimer_ >= PROF_INTERVAL) {
        logProfilingAverages();
        profTimer_ -= PROF_INTERVAL;
    }
}

void VoxelRenderPath::Render() const {
    RenderPath3D::Render();

    if (renderer.isInitialized() && camera && rtCreated_) {
        auto* device = wi::graphics::GetDevice();
        CommandList cmd = device->BeginCommandList();

        // GPU mesh path: only re-dispatch when voxel data changed
        if (renderer.gpuMeshEnabled_ && renderer.gpuMesherAvailable_) {
            // Always readback previous frame's quad count
            uint32_t* countData = (uint32_t*)renderer.meshCounterReadback_.mapped_data;
            if (countData) {
                renderer.gpuMeshQuadCount_ = *countData;
                renderer.totalQuads_ = renderer.gpuMeshQuadCount_;
            }
            // Only re-dispatch compute mesher when data changed
            if (renderer.gpuMeshDirty_) {
                renderer.dispatchGpuMesh(cmd, world,
                    &profVoxelPack_, &profGpuUpload_, &profGpuDispatch_);
            }
        }

        // GPU mesh benchmark state machine (runs once after world gen, CPU path only)
        if (!renderer.gpuMeshEnabled_) {
            if (renderer.benchState_ == VoxelRenderer::BenchState::DISPATCH) {
                renderer.dispatchGpuMeshBenchmark(cmd, world);
            } else if (renderer.benchState_ == VoxelRenderer::BenchState::READBACK) {
                renderer.readbackGpuMeshBenchmark();
            }
        }

        auto tRender0 = std::chrono::high_resolution_clock::now();
        renderer.render(cmd, *camera, voxelDepth_, voxelRT_);
        auto tRender1 = std::chrono::high_resolution_clock::now();
        profRender_.add(std::chrono::duration<float, std::milli>(tRender1 - tRender0).count());
    }
}

void VoxelRenderPath::logProfilingAverages() const {
    char msg[512];
    snprintf(msg, sizeof(msg),
        "=== PERF PROFILE (avg over %.0fs) ===\n"
        "  Regenerate:    %7.2f ms  (%u calls)\n"
        "  UpdateMeshes:  %7.2f ms  (%u calls)\n"
        "  VoxelPack:     %7.2f ms  (%u calls)\n"
        "  GPU Upload:    %7.2f ms  (%u calls)\n"
        "  GPU Dispatch:  %7.2f ms  (%u calls)\n"
        "  Render:        %7.2f ms  (%u calls)\n"
        "  Frame (Upd):   %7.2f ms  (%u calls, %.1f FPS)",
        PROF_INTERVAL,
        profRegenerate_.avg(), profRegenerate_.count,
        profUpdateMeshes_.avg(), profUpdateMeshes_.count,
        profVoxelPack_.avg(), profVoxelPack_.count,
        profGpuUpload_.avg(), profGpuUpload_.count,
        profGpuDispatch_.avg(), profGpuDispatch_.count,
        profRender_.avg(), profRender_.count,
        profFrame_.avg(), profFrame_.count,
        profFrame_.count > 0 ? (1000.0f / profFrame_.avg()) : 0.0f);
    wi::backlog::post(msg);

    profRegenerate_.reset();
    profUpdateMeshes_.reset();
    profVoxelPack_.reset();
    profGpuUpload_.reset();
    profGpuDispatch_.reset();
    profRender_.reset();
    profFrame_.reset();
}

void VoxelRenderPath::Compose(CommandList cmd) const {
    frameCount_++;

    RenderPath3D::Compose(cmd);

    if (rtCreated_ && voxelRT_.IsValid()) {
        wi::image::Params fx;
        fx.enableFullScreen();
        fx.blendFlag = wi::enums::BLENDMODE_OPAQUE;
        wi::image::Draw(&voxelRT_, fx, cmd);
    }

    // HUD overlay
    wi::font::Params fp;
    fp.posX = 10; fp.posY = 10; fp.size = 20;
    fp.color = wi::Color(255, 255, 255, 230);
    fp.shadowColor = wi::Color(0, 0, 0, 180);

    char fpsStr[16];
    snprintf(fpsStr, sizeof(fpsStr), "%.1f", smoothFps_);
    char dtStr[16];
    snprintf(dtStr, sizeof(dtStr), "%.2f", lastDt_ * 1000.0f);

    std::string stats = "BVLE Voxel Engine (Phase 4 — Toping)\n";
    stats += "FPS: " + std::string(fpsStr) + " (" + std::string(dtStr) + " ms)\n";
    if (debugMode) {
        stats += "=== DEBUG FACE MODE ===\n";
        stats += "+X=Red  -X=DkRed  +Y=Green  -Y=DkGreen  +Z=Blue  -Z=DkBlue\n";
    }
    stats += "Chunks: " + std::to_string(renderer.getVisibleChunks())
           + "/" + std::to_string(renderer.getChunkCount()) + "\n";
    stats += "Quads: " + std::to_string(renderer.getTotalQuads()) + "\n";
    std::string renderMode;
    if (renderer.isGpuMeshEnabled())
        renderMode = "GPU mesh (1x1) + DrawInstanced";
    else if (renderer.isGpuCulling())
        renderMode = "CPU greedy + MDI + GPU cull";
    else if (renderer.isMdiEnabled())
        renderMode = "CPU greedy + MDI + CPU cull";
    else
        renderMode = "CPU greedy + DrawInstanced + CPU cull";
    stats += "Draw Calls: " + std::to_string(renderer.getDrawCalls())
           + " (" + renderMode + ")\n";

    if (renderer.isGpuMeshEnabled()) {
        stats += "GPU Mesh Quads: " + std::to_string(renderer.getGpuMeshQuadCount()) + "\n";
    } else {
        char cullStr[16], drawStr[16];
        snprintf(cullStr, sizeof(cullStr), "%.3f", renderer.getGpuCullTimeMs());
        snprintf(drawStr, sizeof(drawStr), "%.3f", renderer.getGpuDrawTimeMs());
        stats += "GPU Cull: " + std::string(cullStr) + " ms | Draw: " + std::string(drawStr) + " ms\n";
    }
    stats += "Topings: " + std::to_string(topingSystem.getInstanceCount())
           + " instances (" + std::to_string(topingSystem.getDefCount()) + " types, "
           + std::to_string(topingSystem.getVertexCount()) + " verts)\n";
    stats += "WASD+Space/Ctrl: move | Shift: fast | Right-click: capture mouse\n";
    stats += "F2: console | F3: anim [" + std::string(animatedTerrain_ ? "ON" : "OFF")
        + "] | F4: dbg [" + std::string(renderer.debugBlend_ ? "ON" : "OFF") + "]";

    wi::font::Draw(stats, fp, cmd);
}

} // namespace voxel