// gpr_poc —— POC-B headless 度量 CLI。
//
// 串起整条地基：发现 14 通道 .iprb + .ord → assembleGprSurvey → buildGprVolume
// → ChunkedVolumeStore::write → buildPyramid → WholeVolumeSource(load)，
// 在真实/合成数据上输出可测的真实指标（耗时/维度/体积/压缩比/加载/峰值内存）。
//
// 子命令：
//   gpr_poc build <dir> [--line 001] [--cellXY 0.2] [--cellZ 0.05] [--out <storeDir>] [--levels 2]
//   gpr_poc load  <storeDir>
//   gpr_poc selftest
//   gpr_poc offscreen-smoke               —— 离屏 GL 闸门冒烟
//   gpr_poc renderB <storeDir> [--frames 120]  —— 离屏体绘制/切片 fps 基准

#include <algorithm>
#include <cmath>
#include <cstdint>
#include <cstdlib>
#include <filesystem>
#include <fstream>
#include <iostream>
#include <map>
#include <string>
#include <vector>

#include "Probe.hpp"

#include "core/algo/GprVolumeBuilder.hpp"
#include "core/algo/IInterpolator.hpp"
#include "core/model/GprSurvey.hpp"
#include "core/model/ColorScale.hpp"
#include "core/model/ScalarVolumeI16.hpp"
#include "data/store/ChunkedVolumeStore.hpp"
#include "io/gpr/GprSurveyAssembler.hpp"
#include "render/actors/VoxelActor.hpp"
#include "render/source/WholeVolumeSource.hpp"

// ---- VTK 离屏渲染 ----
#include <vtkActor.h>
#include <vtkCamera.h>
#include <vtkCubeSource.h>
#include <vtkImageActor.h>
#include <vtkImageData.h>
#include <vtkImageMapToColors.h>
#include <vtkImageMapper3D.h>
#include <vtkImageReslice.h>
#include <vtkLookupTable.h>
#include <vtkNew.h>
#include <vtkObjectFactory.h>
#include <vtkOutputWindow.h>
#include <vtkPolyDataMapper.h>
#include <vtkProperty.h>
#include <vtkRenderWindow.h>
#include <vtkRenderer.h>
#include <vtkSmartPointer.h>
#include <vtkSmartVolumeMapper.h>
#include <vtkUnsignedCharArray.h>
#include <vtkVolume.h>

namespace fs = std::filesystem;
using geopro::tools::Probe;
using geopro::tools::Stopwatch;

namespace {

constexpr int kChannels = 14;

// ---- 命令行参数解析（极简 --key value）----
struct Args {
  std::map<std::string, std::string> kv;
  std::vector<std::string> positional;

  std::string get(const std::string& key, const std::string& def) const {
    auto it = kv.find(key);
    return it != kv.end() ? it->second : def;
  }
};

Args parseArgs(int argc, char** argv, int start) {
  Args a;
  for (int i = start; i < argc; ++i) {
    std::string tok = argv[i];
    if (tok.rfind("--", 0) == 0 && i + 1 < argc) {
      a.kv[tok.substr(2)] = argv[i + 1];
      ++i;
    } else {
      a.positional.push_back(tok);
    }
  }
  return a;
}

// 把一行指标追加写入 last-metrics.txt（与可执行同目录的工具源目录无关，
// 写到当前工作目录便于汇总；CSV 风格一行）。
void writeMetricLine(const std::string& line) {
  std::ofstream f("last-metrics.txt", std::ios::app);
  if (f) f << line << "\n";
}

// 发现某线 14 通道 .iprb（按通道号 A01..A14 排序）+ 该线 .ord。
struct LineFiles {
  std::vector<std::string> iprb;  // 已按通道号排序
  std::string ord;
};

LineFiles discoverLine(const std::string& dir, const std::string& line) {
  LineFiles lf;
  std::map<int, std::string> byChannel;  // 通道号 → 路径（自动按号排序）
  std::string ordPath;

  for (const auto& e : fs::directory_iterator(dir)) {
    if (!e.is_regular_file()) continue;
    const std::string name = e.path().filename().string();
    const std::string ext = e.path().extension().string();

    // .ord：优先匹配本线（含 "_<line>" 且以 .ord 结尾），否则记下工区任一 .ord 作兜底。
    if (ext == ".ord") {
      if (name.find("_" + line + ".") != std::string::npos) {
        ordPath = e.path().string();
      } else if (ordPath.empty()) {
        ordPath = e.path().string();
      }
      continue;
    }

    // .iprb：匹配 "*_<line>_A<NN>.iprb"。
    if (ext != ".iprb") continue;
    const std::string tag = "_" + line + "_A";
    const std::size_t pos = name.find(tag);
    if (pos == std::string::npos) continue;
    const std::size_t numStart = pos + tag.size();
    std::size_t numEnd = numStart;
    while (numEnd < name.size() &&
           std::isdigit(static_cast<unsigned char>(name[numEnd]))) {
      ++numEnd;
    }
    if (numEnd == numStart) continue;
    const int ch = std::stoi(name.substr(numStart, numEnd - numStart));
    byChannel[ch] = e.path().string();
  }

  for (const auto& [ch, path] : byChannel) lf.iprb.push_back(path);
  lf.ord = ordPath;
  return lf;
}

// 由 survey 推 GridSpec：X 沿测线，Y 跨通道，Z 深度。
geopro::core::GridSpec specFromSurvey(const geopro::core::GprSurvey& s,
                                      double cellXY, double cellZ) {
  geopro::core::GridSpec spec{};

  const double rangeX =
      (s.ntraces > 1) ? (s.ntraces - 1) * s.dx : 0.0;
  const double y0 = s.channelY.empty() ? 0.0 : s.channelY.front();
  const double y1 = s.channelY.empty() ? 0.0 : s.channelY.back();
  const double rangeY = y1 - y0;
  const double rangeZ =
      (s.samples > 1) ? (s.samples - 1) * s.dz : 0.0;

  auto cells = [](double range, double cell) {
    if (cell <= 0.0) return 1;
    return static_cast<int>(std::ceil(range / cell)) + 1;
  };

  spec.ox = s.x0;
  spec.oy = y0;
  spec.oz = s.z0;
  spec.dx = cellXY;
  spec.dy = cellXY;
  spec.dz = cellZ;
  spec.nx = cells(rangeX, cellXY);
  spec.ny = cells(rangeY, cellXY);
  spec.nz = cells(rangeZ, cellZ);
  spec.power = 2.0;
  spec.maxDist = cellXY * 2.0;
  return spec;
}

// 落盘 data.bin 体积（所有 data*.bin 之和，含金字塔各级）。
std::int64_t storeDataBytes(const std::string& dir) {
  std::int64_t total = 0;
  for (const auto& e : fs::directory_iterator(dir)) {
    if (!e.is_regular_file()) continue;
    const std::string name = e.path().filename().string();
    if (name.rfind("data", 0) == 0 &&
        e.path().extension().string() == ".bin") {
      total += static_cast<std::int64_t>(e.file_size());
    }
  }
  return total;
}

int cmdBuild(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc build <dir> [--line 001] [--cellXY 0.2] "
                 "[--cellZ 0.05] [--out <storeDir>] [--levels 2]\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  const std::string line = a.get("line", "001");
  const double cellXY = std::stod(a.get("cellXY", "0.2"));
  const double cellZ = std::stod(a.get("cellZ", "0.05"));
  const int levels = std::stoi(a.get("levels", "2"));
  const std::string out =
      a.get("out", (fs::temp_directory_path() / ("gpr_store_" + line)).string());

  std::cout << "[build] dir=" << dir << " line=" << line
            << " cellXY=" << cellXY << " cellZ=" << cellZ
            << " levels=" << levels << " out=" << out << "\n";

  const LineFiles lf = discoverLine(dir, line);
  std::cout << "[build] 发现通道数=" << lf.iprb.size()
            << " ord=" << (lf.ord.empty() ? "(无)" : lf.ord) << "\n";
  if (lf.iprb.size() != static_cast<std::size_t>(kChannels)) {
    std::cerr << "[build] 警告: 通道数 != " << kChannels
              << "（仍按发现数继续）\n";
  }
  if (lf.iprb.empty() || lf.ord.empty()) {
    std::cerr << "[build] 错误: 未发现 .iprb 或 .ord\n";
    return 1;
  }

  // 1) 装配
  Stopwatch swAsm;
  geopro::core::GprSurvey survey =
      geopro::io::gpr::assembleGprSurvey(lf.iprb, lf.ord);
  const double asmMs = swAsm.elapsedMs();
  std::cout << "[build] 装配完成 ntraces=" << survey.ntraces
            << " samples=" << survey.samples
            << " channels=" << survey.channelY.size()
            << " dx=" << survey.dx << " dz=" << survey.dz << "\n";

  // 2) 建体
  const geopro::core::GridSpec spec = specFromSurvey(survey, cellXY, cellZ);
  std::cout << "[build] GridSpec nx=" << spec.nx << " ny=" << spec.ny
            << " nz=" << spec.nz << " dx=" << spec.dx << " dy=" << spec.dy
            << " dz=" << spec.dz << " maxDist=" << spec.maxDist << "\n";

  Stopwatch swBuild;
  geopro::core::BuiltI16 built = geopro::core::buildGprVolume(survey, spec);
  const double buildMs = swBuild.elapsedMs();

  const std::int64_t nx = built.vol.nx(), ny = built.vol.ny(), nz = built.vol.nz();
  const std::int64_t voxels = nx * ny * nz;
  const std::int64_t rawBytes = voxels * 2;  // int16

  // 3) 落盘 + 金字塔
  fs::create_directories(out);
  Stopwatch swWrite;
  geopro::data::ChunkedVolumeStore::write(out, built);
  const double writeMs = swWrite.elapsedMs();

  Stopwatch swPyr;
  {
    geopro::data::ChunkedVolumeStore store(out);
    store.buildPyramid(levels);
  }
  const double pyrMs = swPyr.elapsedMs();

  const std::int64_t dataBytes = storeDataBytes(out);
  const double ratio =
      dataBytes > 0 ? static_cast<double>(rawBytes) / dataBytes : 0.0;
  const double peak = Probe::peakMemMB();

  std::cout << "\n=== build 指标 ===\n";
  std::cout << "装配耗时(ms)   : " << asmMs << "\n";
  std::cout << "建体耗时(ms)   : " << buildMs << "\n";
  std::cout << "落盘耗时(ms)   : " << writeMs << "\n";
  std::cout << "金字塔耗时(ms) : " << pyrMs << "\n";
  std::cout << "体维度         : " << nx << " x " << ny << " x " << nz << "\n";
  std::cout << "体素数         : " << voxels << "\n";
  std::cout << "原始体积(B)    : " << rawBytes << "  ("
            << rawBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "data.bin(B)    : " << dataBytes << "  ("
            << dataBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "压缩比         : " << ratio << " x\n";
  std::cout << "峰值内存(MB)   : " << peak << "\n";

  writeMetricLine(
      "build,line=" + line + ",cellXY=" + std::to_string(cellXY) +
      ",cellZ=" + std::to_string(cellZ) + ",nx=" + std::to_string(nx) +
      ",ny=" + std::to_string(ny) + ",nz=" + std::to_string(nz) +
      ",voxels=" + std::to_string(voxels) +
      ",rawB=" + std::to_string(rawBytes) +
      ",dataB=" + std::to_string(dataBytes) +
      ",ratio=" + std::to_string(ratio) + ",asmMs=" + std::to_string(asmMs) +
      ",buildMs=" + std::to_string(buildMs) +
      ",writeMs=" + std::to_string(writeMs) +
      ",pyrMs=" + std::to_string(pyrMs) + ",peakMB=" + std::to_string(peak));
  return 0;
}

int cmdLoad(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc load <storeDir>\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  std::cout << "[load] storeDir=" << dir << "\n";

  Stopwatch sw;
  geopro::render::WholeVolumeSource src(dir);
  const double loadMs = sw.elapsedMs();

  const auto& m = src.meta();
  const std::int64_t voxels =
      static_cast<std::int64_t>(m.nx) * m.ny * m.nz;
  const std::int64_t wholeBytes = voxels * 2;  // VTK_SHORT
  const double peak = Probe::peakMemMB();

  std::cout << "\n=== load 指标 ===\n";
  std::cout << "加载耗时(ms) : " << loadMs << "\n";
  std::cout << "整卷维度     : " << m.nx << " x " << m.ny << " x " << m.nz
            << "\n";
  std::cout << "整卷字节(B)  : " << wholeBytes << "  ("
            << wholeBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "峰值内存(MB) : " << peak << "\n";

  writeMetricLine("load,dir=" + dir + ",nx=" + std::to_string(m.nx) +
                  ",ny=" + std::to_string(m.ny) +
                  ",nz=" + std::to_string(m.nz) +
                  ",wholeB=" + std::to_string(wholeBytes) +
                  ",loadMs=" + std::to_string(loadMs) +
                  ",peakMB=" + std::to_string(peak));
  return 0;
}

// ---- selftest：合成极小数据走完整 build→load 管线 ----

// 写一个极小通道的 .iprb + .iprh（samples 采样、traces 道，值 = base + t + s）。
void writeSyntheticChannel(const fs::path& iprbPath, int samples, int traces,
                           std::int16_t base) {
  const fs::path iprhPath =
      fs::path(iprbPath).replace_extension(".iprh");
  std::ofstream h(iprhPath);
  h << "SAMPLES: " << samples << "\n";
  h << "LAST TRACE: " << (traces - 1) << "\n";
  h << "CHANNELS: 2\n";
  h << "TIMEWINDOW: 100.0\n";
  h << "SOIL VELOCITY: 100.0\n";       // m/µs → ×1e6 → 1e8 m/s
  h << "DISTANCE INTERVAL: 0.05\n";
  h.close();

  std::ofstream b(iprbPath, std::ios::binary);
  // 布局 [trace*samples + s]，s 最快。
  for (int t = 0; t < traces; ++t) {
    for (int s = 0; s < samples; ++s) {
      const std::int16_t v =
          static_cast<std::int16_t>(base + t + s);
      b.write(reinterpret_cast<const char*>(&v), sizeof(v));
    }
  }
}

int cmdSelftest() {
  std::cout << "[selftest] 构造极小合成 survey（2 通道）...\n";
  const fs::path tmp =
      fs::temp_directory_path() / "gpr_poc_selftest";
  std::error_code ec;
  fs::remove_all(tmp, ec);
  fs::create_directories(tmp);

  const int samples = 8;
  const int traces = 12;

  // 2 通道 .iprb/.iprh + .ord（末列==1 标记有效通道，第 2 列为横偏 Y）。
  writeSyntheticChannel(tmp / "syn_001_A01.iprb", samples, traces,
                        /*base=*/100);
  writeSyntheticChannel(tmp / "syn_001_A02.iprb", samples, traces,
                        /*base=*/200);
  {
    std::ofstream ord(tmp / "syn_001.ord");
    ord << "0 0.000000 -1.5 1\n";
    ord << "1 1.000000 -1.5 1\n";
  }

  const std::vector<std::string> iprb = {
      (tmp / "syn_001_A01.iprb").string(),
      (tmp / "syn_001_A02.iprb").string()};
  const std::string ord = (tmp / "syn_001.ord").string();

  bool ok = true;
  auto check = [&](bool cond, const std::string& msg) {
    if (!cond) {
      std::cerr << "[selftest] FAIL: " << msg << "\n";
      ok = false;
    }
  };

  try {
    // 装配
    geopro::core::GprSurvey survey =
        geopro::io::gpr::assembleGprSurvey(iprb, ord);
    check(survey.ntraces == traces, "ntraces");
    check(survey.samples == samples, "samples");
    check(survey.channelY.size() == 2, "channels");
    // channelY 升序：A01 偏移 0.0 在前，A02 偏移 1.0 在后。
    check(survey.channelY.front() < survey.channelY.back(), "channelY 升序");

    // 建体：cellXY 取通道间距 1.0 → ny=2；cellZ 较细确保 nz>1。
    const double cellXY = 1.0;
    const double cellZ = std::max(survey.dz, 1e-12);
    const geopro::core::GridSpec spec =
        specFromSurvey(survey, cellXY, cellZ);
    std::cout << "[selftest] GridSpec " << spec.nx << "x" << spec.ny << "x"
              << spec.nz << " dz=" << spec.dz << "\n";
    check(spec.ny == 2, "ny==2");

    geopro::core::BuiltI16 built =
        geopro::core::buildGprVolume(survey, spec);
    check(built.vol.nx() == spec.nx, "built nx");
    check(built.vol.ny() == spec.ny, "built ny");
    check(built.vol.nz() == spec.nz, "built nz");

    // 落盘 + 金字塔
    const std::string store = (tmp / "store").string();
    fs::create_directories(store);
    geopro::data::ChunkedVolumeStore::write(store, built, /*brick=*/4);
    {
      geopro::data::ChunkedVolumeStore s(store);
      s.buildPyramid(1);
      check(s.levels() == 2, "金字塔层数==2");
    }

    // 加载整卷，校验维度一致
    geopro::render::WholeVolumeSource src(store);
    check(src.meta().nx == spec.nx, "load nx");
    check(src.meta().ny == spec.ny, "load ny");
    check(src.meta().nz == spec.nz, "load nz");

    // 某体素值合理性：x0/y0 角点应有非 blank 量化值（落格命中首道首通道）。
    const std::int16_t q = built.vol.at(0, 0, 0);
    check(q != geopro::core::ScalarVolumeI16::kBlank, "(0,0,0) 非 blank");
  } catch (const std::exception& e) {
    std::cerr << "[selftest] 异常: " << e.what() << "\n";
    ok = false;
  }

  fs::remove_all(tmp, ec);
  std::cout << "[selftest] " << (ok ? "PASS" : "FAIL") << "\n";
  return ok ? 0 : 1;
}

// ============================================================================
// 离屏 GPU 渲染基准（POC-B）
// ============================================================================

// 捕获 VTK 错误输出的 OutputWindow：用于侦测体绘制时 vtkVolumeTexture 报的
// "Invalid texture dimensions" / "MAX_3D_TEXTURE_SIZE" —— 一旦出现，说明整卷
// 单张 3D 纹理上传失败，体绘制 fps 无意义，必须如实标 INVALID（绝不当真上报）。
class CapturingOutputWindow : public vtkOutputWindow {
 public:
  static CapturingOutputWindow* New();
  vtkTypeMacro(CapturingOutputWindow, vtkOutputWindow);

  void DisplayText(const char* txt) override {
    if (txt) {
      const std::string s(txt);
      captured_ += s;
      if (s.find("texture dimensions") != std::string::npos ||
          s.find("MAX_3D_TEXTURE_SIZE") != std::string::npos) {
        textureError_ = true;
      }
    }
    // 仍透传到 stderr，便于人工查看。
    if (txt) std::cerr << txt;
  }

  bool textureError() const { return textureError_; }
  const std::string& captured() const { return captured_; }

 private:
  std::string captured_;
  bool textureError_ = false;
};
vtkStandardNewMacro(CapturingOutputWindow);

// 创建一个离屏 vtkRenderWindow（VTK9.6：SetShowWindow(false)+OffScreenRenderingOn）。
vtkSmartPointer<vtkRenderWindow> makeOffscreenWindow(int w, int h) {
  auto rw = vtkSmartPointer<vtkRenderWindow>::New();
  rw->SetOffScreenRendering(1);
  rw->SetShowWindow(false);
  rw->SetSize(w, h);
  return rw;
}

// 闸门：最小离屏渲染冒烟。返回 0=OK，非 0=离屏 GL 起不来（BLOCKED_OFFSCREEN）。
// 流程：离屏窗口 → 加一个 cube actor → Render() → 读回像素，确认非全黑/读得到。
int cmdOffscreenSmoke() {
  std::cout << "[offscreen-smoke] 创建离屏 vtkRenderWindow...\n";
  try {
    auto rw = makeOffscreenWindow(256, 256);

    vtkNew<vtkRenderer> ren;
    ren->SetBackground(0.1, 0.1, 0.2);
    rw->AddRenderer(ren);

    vtkNew<vtkCubeSource> cube;
    cube->SetXLength(1.0);
    cube->SetYLength(1.0);
    cube->SetZLength(1.0);

    vtkNew<vtkPolyDataMapper> mapper;
    mapper->SetInputConnection(cube->GetOutputPort());

    vtkNew<vtkActor> actor;
    actor->SetMapper(mapper);
    actor->GetProperty()->SetColor(1.0, 0.6, 0.2);
    ren->AddActor(actor);
    ren->ResetCamera();

    // Render()：若 GL 上下文创建失败，VTK 会输出错误（多数返回，少数抛）。
    rw->Render();

    // 读回像素验证：取整窗 RGB，确认能读到且非全 0。
    const int* sz = rw->GetSize();
    const int w = sz[0], h = sz[1];
    if (w <= 0 || h <= 0) {
      std::cout << "[offscreen-smoke] FAIL: 窗口尺寸为 0（上下文未建立）\n";
      std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
      return 1;
    }

    auto pixels = vtkSmartPointer<vtkUnsignedCharArray>::New();
    // GetRGBACharPixelData(x0,y0,x1,y1,front,arr)：front=1 读前缓冲。
    const int ok =
        rw->GetRGBACharPixelData(0, 0, w - 1, h - 1, /*front=*/1, pixels);
    if (ok == 0 || pixels->GetNumberOfTuples() == 0) {
      std::cout << "[offscreen-smoke] FAIL: 读不到像素\n";
      std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
      return 1;
    }

    // 统计非背景像素（cube 应渲出橙色，存在像素 R 通道明显高于背景）。
    vtkIdType nonBlack = 0;
    const vtkIdType n = pixels->GetNumberOfTuples();
    for (vtkIdType i = 0; i < n; ++i) {
      const double r = pixels->GetComponent(i, 0);
      const double g = pixels->GetComponent(i, 1);
      const double b = pixels->GetComponent(i, 2);
      if (r > 80 || g > 80 || b > 80) ++nonBlack;
    }

    const char* caps = rw->ReportCapabilities();
    std::cout << "[offscreen-smoke] 读回像素 " << n << " 个，非背景像素 "
              << nonBlack << "\n";
    std::cout << "[offscreen-smoke] GL 能力:\n"
              << (caps ? caps : "(无)") << "\n";

    if (nonBlack == 0) {
      std::cout << "[offscreen-smoke] FAIL: 渲染结果全为背景（actor 未画出）\n";
      std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
      return 1;
    }

    std::cout << "[offscreen-smoke] OK：离屏 GL 可用，可继续真实基准。\n";
    std::cout << "STATUS=OK\n";
    return 0;
  } catch (const std::exception& e) {
    std::cout << "[offscreen-smoke] FAIL: 异常 " << e.what() << "\n";
    std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
    return 1;
  }
}

// 体绘制 fps：每帧绕 azimuth 旋相机再 Render()，避免被驱动优化成空渲染。
double benchVolumeFps(vtkRenderWindow* rw, vtkRenderer* ren, int frames) {
  ren->ResetCamera();
  vtkCamera* cam = ren->GetActiveCamera();
  rw->Render();  // 预热一帧（首帧含上传显存/编译 shader，不计时）
  Stopwatch sw;
  for (int f = 0; f < frames; ++f) {
    cam->Azimuth(360.0 / frames);  // 每帧转一点，扫满一圈
    rw->Render();
  }
  const double ms = sw.elapsedMs();
  return ms > 0.0 ? frames * 1000.0 / ms : 0.0;
}

// 切片扫描 fps：沿 K 轴(深度)逐偏移 reslice 取轴向切面 + 纹理渲染，每帧推进偏移。
double benchSliceFps(vtkRenderWindow* rw, vtkRenderer* ren,
                     vtkImageData* full, vtkLookupTable* lut, int frames) {
  // reslice：固定轴向(XY 平面)，沿 Z 改变 ResliceAxesOrigin 扫过整卷。
  vtkNew<vtkImageReslice> reslice;
  reslice->SetInputData(full);
  reslice->SetOutputDimensionality(2);
  reslice->SetInterpolationModeToLinear();

  vtkNew<vtkImageMapToColors> colorize;
  colorize->SetLookupTable(lut);
  colorize->SetInputConnection(reslice->GetOutputPort());

  vtkNew<vtkImageActor> imgActor;
  imgActor->GetMapper()->SetInputConnection(colorize->GetOutputPort());
  ren->AddViewProp(imgActor);
  ren->ResetCamera();

  double bounds[6];
  full->GetBounds(bounds);
  const double zMin = bounds[4], zMax = bounds[5];
  const double ox = 0.5 * (bounds[0] + bounds[1]);
  const double oy = 0.5 * (bounds[2] + bounds[3]);

  rw->Render();  // 预热
  Stopwatch sw;
  for (int f = 0; f < frames; ++f) {
    const double t = static_cast<double>(f) / std::max(1, frames - 1);
    const double z = zMin + (zMax - zMin) * t;
    reslice->SetResliceAxesOrigin(ox, oy, z);
    reslice->Modified();
    rw->Render();
  }
  const double ms = sw.elapsedMs();
  return ms > 0.0 ? frames * 1000.0 / ms : 0.0;
}

// 由 ColorScale 物理区间建 256 级 VTK LUT（切片纹理着色用，与体绘制色阶同源）。
vtkSmartPointer<vtkLookupTable> makeLut(const geopro::core::ColorScale& cs,
                                        double vmin, double vmax) {
  auto lut = vtkSmartPointer<vtkLookupTable>::New();
  const int n = 256;
  lut->SetNumberOfTableValues(n);
  lut->SetRange(vmin, vmax);
  for (int i = 0; i < n; ++i) {
    const double v = vmin + (vmax - vmin) * i / (n - 1);
    const auto c = cs.colorAt(v);
    lut->SetTableValue(i, c.r / 255.0, c.g / 255.0, c.b / 255.0, 1.0);
  }
  lut->Build();
  return lut;
}

// 简单蓝-白-红色阶（与 test_color_scale 同款最简构造）。
geopro::core::ColorScale makeColorScale(double vmin, double vmax) {
  geopro::core::ColorScale cs;
  const double mid = 0.5 * (vmin + vmax);
  cs.addStop(vmin, geopro::core::Rgba{0, 0, 255, 255});
  cs.addStop(mid, geopro::core::Rgba{255, 255, 255, 255});
  cs.addStop(vmax, geopro::core::Rgba{255, 0, 0, 255});
  return cs;
}

int cmdRenderB(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc renderB <storeDir> [--frames 120]\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  const int frames = std::stoi(a.get("frames", "120"));
  std::cout << "[renderB] storeDir=" << dir << " frames=" << frames << "\n";

  // 闸门复检：renderB 前先确认离屏可用（避免在不可渲染机上跑出假数据）。
  std::cout << "[renderB] 离屏闸门复检...\n";
  if (cmdOffscreenSmoke() != 0) {
    std::cout << "[renderB] 闸门失败，中止，不产出 fps。\n";
    return 1;
  }

  // 1) 加载整卷（VTK_SHORT）。
  Stopwatch swLoad;
  geopro::render::WholeVolumeSource src(dir);
  const double loadMs = swLoad.elapsedMs();
  const auto& m = src.meta();

  const std::int64_t voxels =
      static_cast<std::int64_t>(m.nx) * m.ny * m.nz;
  const std::int64_t wholeBytes = voxels * 2;  // VTK_SHORT
  std::cout << "[renderB] 整卷 " << m.nx << "x" << m.ny << "x" << m.nz
            << " 体素=" << voxels << " 字节=" << wholeBytes << " ("
            << wholeBytes / (1024.0 * 1024.0) << " MB)，加载 " << loadMs
            << "ms\n";

  auto images = src.currentImages();
  if (images.empty() || !images.front()) {
    std::cerr << "[renderB] 错误: currentImages 为空\n";
    return 1;
  }
  vtkImageData* shortImg = images.front().Get();

  // 色阶用 meta 的物理区间。
  const double vmin = m.vminPhys, vmax = m.vmaxPhys;
  const geopro::core::ColorScale cs = makeColorScale(vmin, vmax);

  // 2) 体绘制（离屏）。
  auto rw = makeOffscreenWindow(1024, 768);
  vtkNew<vtkRenderer> ren;
  ren->SetBackground(0.0, 0.0, 0.0);
  rw->AddRenderer(ren);

  vtkSmartPointer<vtkVolume> volume =
      geopro::render::buildVoxelI16FromImage(shortImg, m.quant, cs, vmin, vmax);
  ren->AddVolume(volume);

  // 装上捕获式 OutputWindow：拦截体绘制时的 3D 纹理维度错误。
  auto capWin = vtkSmartPointer<CapturingOutputWindow>::New();
  vtkOutputWindow::SetInstance(capWin);

  std::cout << "[renderB] 体绘制基准（" << frames << " 帧旋转相机）...\n";
  const double volFpsRaw = benchVolumeFps(rw, ren, frames);

  const bool textureErr = capWin->textureError();
  vtkOutputWindow::SetInstance(nullptr);  // 还原默认输出窗口

  // 进显存判据：SmartVolumeMapper 实际用的渲染模式（2=GPURenderMode）。
  int renderMode = -1;
  bool lowResResample = false;
  if (auto* svm =
          vtkSmartVolumeMapper::SafeDownCast(volume->GetMapper())) {
    renderMode = svm->GetLastUsedRenderMode();
    // 大体可能触发降质重采样（GPU 显存不足时 SmartVolumeMapper 走低分辨率）。
    lowResResample = (svm->GetInteractiveAdjustSampleDistances() == 0 &&
                      renderMode != vtkSmartVolumeMapper::GPURenderMode);
  }
  const bool onGpu = (renderMode == vtkSmartVolumeMapper::GPURenderMode);

  // 任一维度超过 GL_MAX_3D_TEXTURE_SIZE（本机实测 16384）→ 整卷无法成单张 3D 纹理。
  constexpr int kMax3DTexObserved = 16384;
  const bool dimOversize =
      (m.nx > kMax3DTexObserved || m.ny > kMax3DTexObserved ||
       m.nz > kMax3DTexObserved);
  // 体绘制 fps 是否可信：上传成功（无纹理错误且未超限）才算真实整卷体绘制帧率。
  const bool volFpsValid = !textureErr && !dimOversize;
  const double volFps = volFpsValid ? volFpsRaw : -1.0;

  std::cout << "[renderB] 体绘制 raw_fps=" << volFpsRaw
            << "  渲染模式=" << renderMode << (onGpu ? "(GPU)" : "(非GPU)")
            << "  纹理维度错误=" << (textureErr ? "是" : "否")
            << "  超 16384=" << (dimOversize ? "是" : "否") << "\n";
  if (!volFpsValid) {
    std::cout << "[renderB] 警告: 整卷未能成单张 3D 纹理（X=" << m.nx
              << " > " << kMax3DTexObserved
              << "），体绘制 fps 无意义 → 标 INVALID。\n";
  }

  // 3) 切片扫描（离屏，沿 Z 扫整卷）。
  vtkNew<vtkRenderer> ren2;
  ren2->SetBackground(0.0, 0.0, 0.0);
  auto rw2 = makeOffscreenWindow(1024, 768);
  rw2->AddRenderer(ren2);
  vtkSmartPointer<vtkLookupTable> lut = makeLut(cs, vmin, vmax);

  std::cout << "[renderB] 切片扫描基准（" << frames << " 帧沿 Z 推进）...\n";
  const double sliceFps =
      benchSliceFps(rw2, ren2, src.sliceSource(), lut, frames);
  std::cout << "[renderB] 切片 fps=" << sliceFps << "\n";

  const double peak = Probe::peakMemMB();
  const std::string vram = "N/A";  // VTK 安装未带 GLEW 头，无法直查 NVX 显存

  // 4) 汇总打印。
  const std::string volFpsStr =
      volFpsValid ? std::to_string(volFps) : "INVALID(整卷超 3D 纹理上限)";

  std::cout << "\n=== renderB GPU 指标 ===\n";
  std::cout << "离屏闸门       : OK\n";
  std::cout << "体维度         : " << m.nx << " x " << m.ny << " x " << m.nz
            << "\n";
  std::cout << "体素数         : " << voxels << "\n";
  std::cout << "整卷字节(B)    : " << wholeBytes << "  ("
            << wholeBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "体绘制 fps     : " << volFpsStr << "\n";
  if (!volFpsValid) {
    std::cout << "  (raw_fps=" << volFpsRaw
              << " 为空纹理渲染，X=" << m.nx << " > 16384，不可信)\n";
  }
  std::cout << "切片扫描 fps   : " << sliceFps << "  (2D 纹理，无 3D 上限约束)\n";
  std::cout << "渲染模式       : " << renderMode
            << (onGpu ? " (GPU 路径)" : " (非 GPU)") << "\n";
  std::cout << "整卷进显存     : "
            << (volFpsValid && onGpu ? "是（单张 3D 纹理）"
                                     : "否（超 GL_MAX_3D_TEXTURE_SIZE 16384）")
            << "\n";
  std::cout << "降质重采样     : " << (lowResResample ? "是" : "否") << "\n";
  std::cout << "GPU 显存       : " << vram << "\n";
  std::cout << "进程峰值内存(MB): " << peak << "\n";

  writeMetricLine(
      "renderB,dir=" + dir + ",nx=" + std::to_string(m.nx) +
      ",ny=" + std::to_string(m.ny) + ",nz=" + std::to_string(m.nz) +
      ",voxels=" + std::to_string(voxels) +
      ",wholeB=" + std::to_string(wholeBytes) +
      ",volFps=" + volFpsStr +
      ",volFpsRaw=" + std::to_string(volFpsRaw) +
      ",volFpsValid=" + std::to_string(volFpsValid ? 1 : 0) +
      ",sliceFps=" + std::to_string(sliceFps) +
      ",renderMode=" + std::to_string(renderMode) +
      ",onGpu=" + std::to_string(onGpu ? 1 : 0) +
      ",loadMs=" + std::to_string(loadMs) +
      ",peakMB=" + std::to_string(peak));
  return 0;
}

void usage() {
  std::cerr << "gpr_poc —— POC-B headless 度量 CLI\n"
               "  gpr_poc build <dir> [--line 001] [--cellXY 0.2] "
               "[--cellZ 0.05] [--out <storeDir>] [--levels 2]\n"
               "  gpr_poc load  <storeDir>\n"
               "  gpr_poc selftest\n"
               "  gpr_poc offscreen-smoke\n"
               "  gpr_poc renderB <storeDir> [--frames 120]\n";
}

}  // namespace

int main(int argc, char** argv) {
  if (argc < 2) {
    usage();
    return 2;
  }
  const std::string cmd = argv[1];
  try {
    if (cmd == "build") return cmdBuild(argc, argv);
    if (cmd == "load") return cmdLoad(argc, argv);
    if (cmd == "selftest") return cmdSelftest();
    if (cmd == "offscreen-smoke") return cmdOffscreenSmoke();
    if (cmd == "renderB") return cmdRenderB(argc, argv);
  } catch (const std::exception& e) {
    std::cerr << "错误: " << e.what() << "\n";
    return 1;
  }
  usage();
  return 2;
}