// CNN v2 Effect Implementation

#include "gpu/effects/cnn_v2_effect.h"

#if defined(USE_TEST_ASSETS)
#include "test_assets.h"
#else
#include "generated/assets.h"
#endif

#include "util/asset_manager.h"
#include "util/fatal_error.h"
#include <cstring>

CNNv2Effect::CNNv2Effect(const GpuContext& ctx)
    : PostProcessEffect(ctx),
      static_pipeline_(nullptr),
      static_bind_group_(nullptr),
      static_features_tex_(nullptr),
      static_features_view_(nullptr),
      layer_pipeline_(nullptr),
      weights_buffer_(nullptr),
      layer_params_buffer_(nullptr),
      input_mip_tex_(nullptr),
      current_input_view_(nullptr),
      initialized_(false) {
  std::memset(input_mip_view_, 0, sizeof(input_mip_view_));
}

CNNv2Effect::~CNNv2Effect() {
  cleanup();
}

void CNNv2Effect::init(MainSequence* demo) {
  (void)demo;
  if (initialized_) return;

  load_weights();
  create_textures();
  create_pipelines();

  initialized_ = true;
}

void CNNv2Effect::resize(int width, int height) {
  PostProcessEffect::resize(width, height);
  cleanup();
  create_textures();
  create_pipelines();
}

void CNNv2Effect::load_weights() {
  // Load binary weights asset
  size_t weights_size = 0;
  const uint8_t* weights_data = (const uint8_t*)GetAsset(AssetId::ASSET_WEIGHTS_CNN_V2, &weights_size);

  if (!weights_data || weights_size < 16) {
    // Weights not available - effect will skip
    return;
  }

  // Parse header (16 bytes)
  const uint32_t* header = (const uint32_t*)weights_data;
  uint32_t magic = header[0];
  uint32_t version = header[1];
  uint32_t num_layers = header[2];
  uint32_t total_weights = header[3];

  FATAL_CHECK(magic == 0x324e4e43, "Invalid CNN v2 weights magic\n");  // 'CNN2'
  FATAL_CHECK(version == 1, "Unsupported CNN v2 weights version\n");

  // Parse layer info (20 bytes per layer)
  const uint32_t* layer_data = header + 4;
  for (uint32_t i = 0; i < num_layers; ++i) {
    LayerInfo info;
    info.kernel_size = layer_data[i * 5 + 0];
    info.in_channels = layer_data[i * 5 + 1];
    info.out_channels = layer_data[i * 5 + 2];
    info.weight_offset = layer_data[i * 5 + 3];
    info.weight_count = layer_data[i * 5 + 4];
    layer_info_.push_back(info);
  }

  // Create GPU storage buffer for weights
  // Buffer contains: header + layer info + packed f16 weights (as u32)
  WGPUBufferDescriptor buffer_desc = {};
  buffer_desc.size = weights_size;
  buffer_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst;
  buffer_desc.mappedAtCreation = false;

  weights_buffer_ = wgpuDeviceCreateBuffer(ctx_.device, &buffer_desc);

  // Upload weights data
  wgpuQueueWriteBuffer(ctx_.queue, weights_buffer_, 0, weights_data, weights_size);

  // Create uniform buffer for layer params
  WGPUBufferDescriptor params_desc = {};
  params_desc.size = sizeof(LayerParams);
  params_desc.usage = WGPUBufferUsage_Uniform | WGPUBufferUsage_CopyDst;
  params_desc.mappedAtCreation = false;

  layer_params_buffer_ = wgpuDeviceCreateBuffer(ctx_.device, &params_desc);
}

void CNNv2Effect::create_textures() {
  const WGPUExtent3D size = {
    static_cast<uint32_t>(width_),
    static_cast<uint32_t>(height_),
    1
  };

  // Static features texture (8×f16 packed as 4×u32)
  WGPUTextureDescriptor static_desc = {};
  static_desc.usage = WGPUTextureUsage_StorageBinding | WGPUTextureUsage_TextureBinding;
  static_desc.dimension = WGPUTextureDimension_2D;
  static_desc.size = size;
  static_desc.format = WGPUTextureFormat_RGBA32Uint;
  static_desc.mipLevelCount = 1;
  static_desc.sampleCount = 1;
  static_features_tex_ = wgpuDeviceCreateTexture(ctx_.device, &static_desc);

  WGPUTextureViewDescriptor view_desc = {};
  view_desc.format = WGPUTextureFormat_RGBA32Uint;
  view_desc.dimension = WGPUTextureViewDimension_2D;
  view_desc.baseMipLevel = 0;
  view_desc.mipLevelCount = 1;
  view_desc.baseArrayLayer = 0;
  view_desc.arrayLayerCount = 1;
  static_features_view_ = wgpuTextureCreateView(static_features_tex_, &view_desc);

  // Input texture with mips (for multi-scale features)
  WGPUTextureDescriptor input_mip_desc = {};
  input_mip_desc.usage = WGPUTextureUsage_TextureBinding | WGPUTextureUsage_CopyDst;
  input_mip_desc.dimension = WGPUTextureDimension_2D;
  input_mip_desc.size = size;
  input_mip_desc.format = WGPUTextureFormat_RGBA8Unorm;
  input_mip_desc.mipLevelCount = 3;  // Levels 0, 1, 2
  input_mip_desc.sampleCount = 1;
  input_mip_tex_ = wgpuDeviceCreateTexture(ctx_.device, &input_mip_desc);

  for (int i = 0; i < 3; ++i) {
    WGPUTextureViewDescriptor mip_view_desc = {};
    mip_view_desc.format = WGPUTextureFormat_RGBA8Unorm;
    mip_view_desc.dimension = WGPUTextureViewDimension_2D;
    mip_view_desc.baseMipLevel = i;
    mip_view_desc.mipLevelCount = 1;
    mip_view_desc.baseArrayLayer = 0;
    mip_view_desc.arrayLayerCount = 1;
    input_mip_view_[i] = wgpuTextureCreateView(input_mip_tex_, &mip_view_desc);
  }

  // Create 2 layer textures (ping-pong buffers for intermediate results)
  // Each stores 8×f16 channels packed as 4×u32
  for (int i = 0; i < 2; ++i) {
    WGPUTextureDescriptor layer_desc = {};
    layer_desc.usage = WGPUTextureUsage_StorageBinding | WGPUTextureUsage_TextureBinding;
    layer_desc.dimension = WGPUTextureDimension_2D;
    layer_desc.size = size;
    layer_desc.format = WGPUTextureFormat_RGBA32Uint;
    layer_desc.mipLevelCount = 1;
    layer_desc.sampleCount = 1;

    WGPUTexture tex = wgpuDeviceCreateTexture(ctx_.device, &layer_desc);
    layer_textures_.push_back(tex);

    WGPUTextureViewDescriptor view_desc = {};
    view_desc.format = WGPUTextureFormat_RGBA32Uint;
    view_desc.dimension = WGPUTextureViewDimension_2D;
    view_desc.baseMipLevel = 0;
    view_desc.mipLevelCount = 1;
    view_desc.baseArrayLayer = 0;
    view_desc.arrayLayerCount = 1;

    WGPUTextureView view = wgpuTextureCreateView(tex, &view_desc);
    layer_views_.push_back(view);
  }
}

void CNNv2Effect::create_pipelines() {
  // Static features compute pipeline
  size_t shader_size = 0;
  const char* static_code = (const char*)GetAsset(AssetId::ASSET_SHADER_CNN_V2_STATIC, &shader_size);

  if (!static_code || shader_size == 0) {
    // Shader not available (e.g., in test mode) - skip pipeline creation
    return;
  }

  WGPUShaderSourceWGSL wgsl_src = {};
  wgsl_src.chain.sType = WGPUSType_ShaderSourceWGSL;
  wgsl_src.code = str_view(static_code);

  WGPUShaderModuleDescriptor shader_desc = {};
  shader_desc.nextInChain = &wgsl_src.chain;

  WGPUShaderModule static_module = wgpuDeviceCreateShaderModule(ctx_.device, &shader_desc);
  if (!static_module) {
    return;
  }

  WGPUComputePipelineDescriptor pipeline_desc = {};
  pipeline_desc.compute.module = static_module;
  pipeline_desc.compute.entryPoint = str_view("main");

  static_pipeline_ = wgpuDeviceCreateComputePipeline(ctx_.device, &pipeline_desc);
  wgpuShaderModuleRelease(static_module);

  // Create bind group layout for static features compute
  // Bindings: 0=input_tex, 1=input_mip1, 2=input_mip2, 3=depth_tex, 4=output
  WGPUBindGroupLayoutEntry bgl_entries[5] = {};

  // Binding 0: Input texture (mip 0)
  bgl_entries[0].binding = 0;
  bgl_entries[0].visibility = WGPUShaderStage_Compute;
  bgl_entries[0].texture.sampleType = WGPUTextureSampleType_Float;
  bgl_entries[0].texture.viewDimension = WGPUTextureViewDimension_2D;

  // Binding 1: Input texture (mip 1)
  bgl_entries[1].binding = 1;
  bgl_entries[1].visibility = WGPUShaderStage_Compute;
  bgl_entries[1].texture.sampleType = WGPUTextureSampleType_Float;
  bgl_entries[1].texture.viewDimension = WGPUTextureViewDimension_2D;

  // Binding 2: Input texture (mip 2)
  bgl_entries[2].binding = 2;
  bgl_entries[2].visibility = WGPUShaderStage_Compute;
  bgl_entries[2].texture.sampleType = WGPUTextureSampleType_Float;
  bgl_entries[2].texture.viewDimension = WGPUTextureViewDimension_2D;

  // Binding 3: Depth texture
  bgl_entries[3].binding = 3;
  bgl_entries[3].visibility = WGPUShaderStage_Compute;
  bgl_entries[3].texture.sampleType = WGPUTextureSampleType_Float;
  bgl_entries[3].texture.viewDimension = WGPUTextureViewDimension_2D;

  // Binding 4: Output (static features)
  bgl_entries[4].binding = 4;
  bgl_entries[4].visibility = WGPUShaderStage_Compute;
  bgl_entries[4].storageTexture.access = WGPUStorageTextureAccess_WriteOnly;
  bgl_entries[4].storageTexture.format = WGPUTextureFormat_RGBA32Uint;
  bgl_entries[4].storageTexture.viewDimension = WGPUTextureViewDimension_2D;

  WGPUBindGroupLayoutDescriptor bgl_desc = {};
  bgl_desc.entryCount = 5;
  bgl_desc.entries = bgl_entries;

  WGPUBindGroupLayout static_bgl = wgpuDeviceCreateBindGroupLayout(ctx_.device, &bgl_desc);

  // Update pipeline layout
  WGPUPipelineLayoutDescriptor pl_desc = {};
  pl_desc.bindGroupLayoutCount = 1;
  pl_desc.bindGroupLayouts = &static_bgl;
  WGPUPipelineLayout pipeline_layout = wgpuDeviceCreatePipelineLayout(ctx_.device, &pl_desc);

  // Recreate pipeline with proper layout
  WGPUComputePipelineDescriptor pipeline_desc2 = {};
  pipeline_desc2.compute.module = wgpuDeviceCreateShaderModule(ctx_.device, &shader_desc);
  pipeline_desc2.compute.entryPoint = str_view("main");
  pipeline_desc2.layout = pipeline_layout;

  if (static_pipeline_) wgpuComputePipelineRelease(static_pipeline_);
  static_pipeline_ = wgpuDeviceCreateComputePipeline(ctx_.device, &pipeline_desc2);

  wgpuShaderModuleRelease(pipeline_desc2.compute.module);
  wgpuPipelineLayoutRelease(pipeline_layout);
  wgpuBindGroupLayoutRelease(static_bgl);

  // CNN layer compute pipeline (storage buffer version)
  if (layer_info_.empty()) return;  // No weights loaded

  size_t layer_shader_size = 0;
  const char* layer_code = (const char*)GetAsset(AssetId::ASSET_SHADER_CNN_V2_COMPUTE, &layer_shader_size);

  if (!layer_code || layer_shader_size == 0) return;

  WGPUShaderSourceWGSL layer_wgsl = {};
  layer_wgsl.chain.sType = WGPUSType_ShaderSourceWGSL;
  layer_wgsl.code = str_view(layer_code);

  WGPUShaderModuleDescriptor layer_shader_desc = {};
  layer_shader_desc.nextInChain = &layer_wgsl.chain;

  WGPUShaderModule layer_module = wgpuDeviceCreateShaderModule(ctx_.device, &layer_shader_desc);
  if (!layer_module) return;

  // Create bind group layout for layer compute
  // 0=static_features, 1=layer_input, 2=output, 3=weights, 4=params
  WGPUBindGroupLayoutEntry layer_bgl_entries[5] = {};

  // Binding 0: Static features (texture)
  layer_bgl_entries[0].binding = 0;
  layer_bgl_entries[0].visibility = WGPUShaderStage_Compute;
  layer_bgl_entries[0].texture.sampleType = WGPUTextureSampleType_Uint;
  layer_bgl_entries[0].texture.viewDimension = WGPUTextureViewDimension_2D;

  // Binding 1: Layer input (texture)
  layer_bgl_entries[1].binding = 1;
  layer_bgl_entries[1].visibility = WGPUShaderStage_Compute;
  layer_bgl_entries[1].texture.sampleType = WGPUTextureSampleType_Uint;
  layer_bgl_entries[1].texture.viewDimension = WGPUTextureViewDimension_2D;

  // Binding 2: Output (storage texture)
  layer_bgl_entries[2].binding = 2;
  layer_bgl_entries[2].visibility = WGPUShaderStage_Compute;
  layer_bgl_entries[2].storageTexture.access = WGPUStorageTextureAccess_WriteOnly;
  layer_bgl_entries[2].storageTexture.format = WGPUTextureFormat_RGBA32Uint;
  layer_bgl_entries[2].storageTexture.viewDimension = WGPUTextureViewDimension_2D;

  // Binding 3: Weights (storage buffer)
  layer_bgl_entries[3].binding = 3;
  layer_bgl_entries[3].visibility = WGPUShaderStage_Compute;
  layer_bgl_entries[3].buffer.type = WGPUBufferBindingType_ReadOnlyStorage;

  // Binding 4: Layer params (uniform buffer)
  layer_bgl_entries[4].binding = 4;
  layer_bgl_entries[4].visibility = WGPUShaderStage_Compute;
  layer_bgl_entries[4].buffer.type = WGPUBufferBindingType_Uniform;
  layer_bgl_entries[4].buffer.minBindingSize = sizeof(LayerParams);

  WGPUBindGroupLayoutDescriptor layer_bgl_desc = {};
  layer_bgl_desc.entryCount = 5;
  layer_bgl_desc.entries = layer_bgl_entries;

  WGPUBindGroupLayout layer_bgl = wgpuDeviceCreateBindGroupLayout(ctx_.device, &layer_bgl_desc);

  WGPUPipelineLayoutDescriptor layer_pl_desc = {};
  layer_pl_desc.bindGroupLayoutCount = 1;
  layer_pl_desc.bindGroupLayouts = &layer_bgl;

  WGPUPipelineLayout layer_pipeline_layout = wgpuDeviceCreatePipelineLayout(ctx_.device, &layer_pl_desc);

  WGPUComputePipelineDescriptor layer_pipeline_desc = {};
  layer_pipeline_desc.compute.module = layer_module;
  layer_pipeline_desc.compute.entryPoint = str_view("main");
  layer_pipeline_desc.layout = layer_pipeline_layout;

  layer_pipeline_ = wgpuDeviceCreateComputePipeline(ctx_.device, &layer_pipeline_desc);

  wgpuShaderModuleRelease(layer_module);
  wgpuPipelineLayoutRelease(layer_pipeline_layout);
  wgpuBindGroupLayoutRelease(layer_bgl);
}

void CNNv2Effect::update_bind_group(WGPUTextureView input_view) {
  if (!static_pipeline_) return;

  // Cache input view
  current_input_view_ = input_view;

  // Release old bind group
  if (static_bind_group_) {
    wgpuBindGroupRelease(static_bind_group_);
    static_bind_group_ = nullptr;
  }

  // Create bind group for static features compute
  WGPUBindGroupEntry bg_entries[5] = {};

  // Binding 0: Input (mip 0)
  bg_entries[0].binding = 0;
  bg_entries[0].textureView = input_view;

  // Binding 1: Input (mip 1)
  bg_entries[1].binding = 1;
  bg_entries[1].textureView = input_mip_view_[0];  // Use mip 0 for now

  // Binding 2: Input (mip 2)
  bg_entries[2].binding = 2;
  bg_entries[2].textureView = input_mip_view_[0];  // Use mip 0 for now

  // Binding 3: Depth (use input for now, no depth available)
  bg_entries[3].binding = 3;
  bg_entries[3].textureView = input_view;

  // Binding 4: Output (static features)
  bg_entries[4].binding = 4;
  bg_entries[4].textureView = static_features_view_;

  WGPUBindGroupDescriptor bg_desc = {};
  bg_desc.layout = wgpuComputePipelineGetBindGroupLayout(static_pipeline_, 0);
  bg_desc.entryCount = 5;
  bg_desc.entries = bg_entries;

  static_bind_group_ = wgpuDeviceCreateBindGroup(ctx_.device, &bg_desc);

  wgpuBindGroupLayoutRelease(bg_desc.layout);

  // Create layer bind groups
  if (!layer_pipeline_ || layer_info_.empty()) return;

  // Release old layer bind groups
  for (auto bg : layer_bind_groups_) {
    wgpuBindGroupRelease(bg);
  }
  layer_bind_groups_.clear();

  // Get bind group layout from layer pipeline
  WGPUBindGroupLayout layer_bgl = wgpuComputePipelineGetBindGroupLayout(layer_pipeline_, 0);

  // Create bind group for each layer
  for (size_t i = 0; i < layer_info_.size(); ++i) {
    WGPUBindGroupEntry layer_entries[5] = {};

    // Binding 0: Static features (constant)
    layer_entries[0].binding = 0;
    layer_entries[0].textureView = static_features_view_;

    // Binding 1: Layer input (ping-pong: use previous layer's output)
    // First layer uses static features as input, others use ping-pong buffers
    layer_entries[1].binding = 1;
    layer_entries[1].textureView = (i == 0) ? static_features_view_ : layer_views_[i % 2];

    // Binding 2: Output texture (ping-pong)
    layer_entries[2].binding = 2;
    layer_entries[2].textureView = layer_views_[(i + 1) % 2];

    // Binding 3: Weights buffer (constant)
    layer_entries[3].binding = 3;
    layer_entries[3].buffer = weights_buffer_;
    layer_entries[3].size = wgpuBufferGetSize(weights_buffer_);

    // Binding 4: Layer params (will be updated per dispatch)
    layer_entries[4].binding = 4;
    layer_entries[4].buffer = layer_params_buffer_;
    layer_entries[4].size = sizeof(LayerParams);

    WGPUBindGroupDescriptor layer_bg_desc = {};
    layer_bg_desc.layout = layer_bgl;
    layer_bg_desc.entryCount = 5;
    layer_bg_desc.entries = layer_entries;

    WGPUBindGroup layer_bg = wgpuDeviceCreateBindGroup(ctx_.device, &layer_bg_desc);
    layer_bind_groups_.push_back(layer_bg);
  }

  wgpuBindGroupLayoutRelease(layer_bgl);
}

void CNNv2Effect::compute(WGPUCommandEncoder encoder,
                          const CommonPostProcessUniforms& uniforms) {
  (void)uniforms;
  if (!initialized_ || !static_pipeline_ || !static_bind_group_) return;

  // Pass 1: Compute static features
  WGPUComputePassEncoder pass = wgpuCommandEncoderBeginComputePass(encoder, nullptr);

  wgpuComputePassEncoderSetPipeline(pass, static_pipeline_);
  wgpuComputePassEncoderSetBindGroup(pass, 0, static_bind_group_, 0, nullptr);

  // Dispatch workgroups (8×8 threads per group)
  uint32_t workgroups_x = (width_ + 7) / 8;
  uint32_t workgroups_y = (height_ + 7) / 8;
  wgpuComputePassEncoderDispatchWorkgroups(pass, workgroups_x, workgroups_y, 1);

  wgpuComputePassEncoderEnd(pass);
  wgpuComputePassEncoderRelease(pass);

  // Execute CNN layer passes
  if (!layer_pipeline_ || layer_bind_groups_.empty()) return;

  for (size_t i = 0; i < layer_info_.size(); ++i) {
    const LayerInfo& info = layer_info_[i];

    // Update layer params uniform buffer
    LayerParams params;
    params.kernel_size = info.kernel_size;
    params.in_channels = info.in_channels;
    params.out_channels = info.out_channels;
    params.weight_offset = info.weight_offset;
    params.is_output_layer = (i == layer_info_.size() - 1) ? 1 : 0;

    wgpuQueueWriteBuffer(ctx_.queue, layer_params_buffer_, 0, &params, sizeof(params));

    // Execute layer compute pass
    WGPUComputePassEncoder layer_pass = wgpuCommandEncoderBeginComputePass(encoder, nullptr);

    wgpuComputePassEncoderSetPipeline(layer_pass, layer_pipeline_);
    wgpuComputePassEncoderSetBindGroup(layer_pass, 0, layer_bind_groups_[i], 0, nullptr);

    wgpuComputePassEncoderDispatchWorkgroups(layer_pass, workgroups_x, workgroups_y, 1);

    wgpuComputePassEncoderEnd(layer_pass);
    wgpuComputePassEncoderRelease(layer_pass);
  }
}

void CNNv2Effect::render(WGPURenderPassEncoder pass,
                         const CommonPostProcessUniforms& uniforms) {
  (void)pass;
  (void)uniforms;
  // Compute-only effect, rendering is done by default composite pass
}

void CNNv2Effect::cleanup() {
  if (static_features_view_) wgpuTextureViewRelease(static_features_view_);
  if (static_features_tex_) wgpuTextureRelease(static_features_tex_);
  if (static_bind_group_) wgpuBindGroupRelease(static_bind_group_);
  if (static_pipeline_) wgpuComputePipelineRelease(static_pipeline_);

  if (layer_pipeline_) wgpuComputePipelineRelease(layer_pipeline_);
  if (weights_buffer_) wgpuBufferRelease(weights_buffer_);
  if (layer_params_buffer_) wgpuBufferRelease(layer_params_buffer_);

  for (int i = 0; i < 3; ++i) {
    if (input_mip_view_[i]) wgpuTextureViewRelease(input_mip_view_[i]);
  }
  if (input_mip_tex_) wgpuTextureRelease(input_mip_tex_);

  for (auto view : layer_views_) wgpuTextureViewRelease(view);
  for (auto tex : layer_textures_) wgpuTextureRelease(tex);
  for (auto bg : layer_bind_groups_) wgpuBindGroupRelease(bg);

  layer_views_.clear();
  layer_textures_.clear();
  layer_bind_groups_.clear();
  layer_info_.clear();

  initialized_ = false;
}