path: root/av1/qmode_rc/ducky_encode.cc

/*
 * Copyright (c) 2022, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */
#include <stdlib.h>
#include <string.h>
#include <algorithm>
#include <memory>
#include <numeric>
#include <vector>

#include "av1/common/enums.h"
#include "av1/encoder/rd.h"
#include "config/aom_config.h"

#include "aom/aom_encoder.h"

#include "av1/av1_cx_iface.h"
#include "av1/av1_iface_common.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/ethread.h"
#include "av1/encoder/firstpass.h"
#include "av1/encoder/temporal_filter.h"
#include "av1/qmode_rc/ducky_encode.h"

#include "common/tools_common.h"

namespace aom {
struct EncoderResource {
  STATS_BUFFER_CTX *stats_buf_ctx;
  FIRSTPASS_STATS *stats_buffer;
  aom_image_t img;
  AV1_PRIMARY *ppi;
  int lookahead_push_count;
  int encode_frame_count;  // Use in second pass only
};

class DuckyEncode::EncodeImpl {
 public:
  VideoInfo video_info;
  int g_usage;
  int max_ref_frames;
  int speed;
  int base_qindex;
  BLOCK_SIZE sb_size;
  enum aom_rc_mode rc_end_usage;
  aom_rational64_t timestamp_ratio;
  std::vector<FIRSTPASS_STATS> stats_list;
  EncoderResource enc_resource;
  struct AvxInputContext input;
};

DuckyEncode::DuckyEncode(const VideoInfo &video_info, BLOCK_SIZE sb_size,
                         int max_ref_frames, int speed, int base_qindex) {
  impl_ptr_ = std::unique_ptr<EncodeImpl>(new EncodeImpl());
  impl_ptr_->video_info = video_info;
  impl_ptr_->g_usage = GOOD;
  impl_ptr_->max_ref_frames = max_ref_frames;
  impl_ptr_->speed = speed;
  impl_ptr_->base_qindex = base_qindex;
  impl_ptr_->sb_size = sb_size;
  impl_ptr_->rc_end_usage = AOM_Q;
  // TODO(angiebird): Set timestamp_ratio properly
  // timestamp_ratio.den = cfg->g_timebase.den;
  // timestamp_ratio.num = (int64_t)cfg->g_timebase.num * TICKS_PER_SEC;
  impl_ptr_->timestamp_ratio = { 1, 1 };
  // TODO(angiebird): How to set ptsvol and duration?
  impl_ptr_->input.filename = impl_ptr_->video_info.file_path.c_str();
}

DuckyEncode::~DuckyEncode() {}

static AV1EncoderConfig GetEncoderConfig(const VideoInfo &video_info,
                                         int g_usage, aom_enc_pass pass) {
  const aom_codec_iface *codec = aom_codec_av1_cx();
  aom_codec_enc_cfg_t cfg;
  aom_codec_enc_config_default(codec, &cfg, g_usage);
  cfg.g_w = video_info.frame_width;
  cfg.g_h = video_info.frame_height;
  cfg.g_pass = pass;
  // g_timebase is the inverse of frame_rate
  cfg.g_timebase.num = video_info.frame_rate.den;
  cfg.g_timebase.den = video_info.frame_rate.num;
  if (pass == AOM_RC_SECOND_PASS) {
    cfg.rc_twopass_stats_in.sz =
        (video_info.frame_count + 1) * sizeof(FIRSTPASS_STATS);
  }
  AV1EncoderConfig oxcf = av1_get_encoder_config(&cfg);
  // TODO(angiebird): Why didn't we init use_highbitdepth in
  // av1_get_encoder_config()?
  oxcf.use_highbitdepth = 0;

  // TODO(jingning): Change this to 35 when the baseline rate control
  // logic is in place.
  // Force maximum look ahead buffer to be 19. This will disable the use
  // of maximum 32 GOP length.
  oxcf.gf_cfg.lag_in_frames = 19;

  return oxcf;
}

static STATS_BUFFER_CTX *CreateStatsBufferCtx(int frame_count,
                                              FIRSTPASS_STATS **stats_buffer) {
  STATS_BUFFER_CTX *stats_buf_ctx = new STATS_BUFFER_CTX;
  // +2 is for total_stats and total_left_stats
  *stats_buffer = new FIRSTPASS_STATS[frame_count + 2];
  stats_buf_ctx->stats_in_start = *stats_buffer;
  stats_buf_ctx->stats_in_end = stats_buf_ctx->stats_in_start;
  stats_buf_ctx->stats_in_buf_end = stats_buf_ctx->stats_in_start + frame_count;
  stats_buf_ctx->total_stats = stats_buf_ctx->stats_in_buf_end;
  stats_buf_ctx->total_left_stats =
      stats_buf_ctx->stats_in_start + frame_count + 1;
  for (FIRSTPASS_STATS *buffer = stats_buf_ctx->stats_in_start;
       buffer <= stats_buf_ctx->total_left_stats; ++buffer) {
    av1_twopass_zero_stats(buffer);
  }
  return stats_buf_ctx;
}

static void DestroyStatsBufferCtx(STATS_BUFFER_CTX **stats_buf_context,
                                  FIRSTPASS_STATS **stats_buffer) {
  (*stats_buf_context)->stats_in_start = nullptr;
  (*stats_buf_context)->stats_in_end = nullptr;
  (*stats_buf_context)->stats_in_buf_end = nullptr;
  (*stats_buf_context)->total_stats = nullptr;
  (*stats_buf_context)->total_left_stats = nullptr;
  delete *stats_buf_context;
  *stats_buf_context = nullptr;
  delete[](*stats_buffer);
  *stats_buffer = nullptr;
}

static FIRSTPASS_STATS ComputeTotalStats(
    const std::vector<FIRSTPASS_STATS> &stats_list) {
  FIRSTPASS_STATS total_stats = {};
  for (size_t i = 0; i < stats_list.size(); ++i) {
    av1_accumulate_stats(&total_stats, &stats_list[i]);
  }
  return total_stats;
}

static bool FileIsY4m(const char detect[4]) {
  return memcmp(detect, "YUV4", 4) == 0;
}

static bool FourccIsIvf(const char detect[4]) {
  return memcmp(detect, "DKIF", 4) == 0;
}

static void OpenInputFile(struct AvxInputContext *input) {
  input->file = fopen(input->filename, "rb");
  /* For RAW input sources, these bytes will applied on the first frame
   *  in read_frame().
   */
  input->detect.buf_read = fread(input->detect.buf, 1, 4, input->file);
  input->detect.position = 0;
  aom_chroma_sample_position_t const csp = AOM_CSP_UNKNOWN;
  if (input->detect.buf_read == 4 && FileIsY4m(input->detect.buf)) {
    if (y4m_input_open(&input->y4m, input->file, input->detect.buf, 4, csp,
                       input->only_i420) >= 0) {
      input->file_type = FILE_TYPE_Y4M;
      input->width = input->y4m.pic_w;
      input->height = input->y4m.pic_h;
      input->pixel_aspect_ratio.numerator = input->y4m.par_n;
      input->pixel_aspect_ratio.denominator = input->y4m.par_d;
      input->framerate.numerator = input->y4m.fps_n;
      input->framerate.denominator = input->y4m.fps_d;
      input->fmt = input->y4m.aom_fmt;
      input->bit_depth = static_cast<aom_bit_depth_t>(input->y4m.bit_depth);
      input->color_range = input->y4m.color_range;
    } else
      fatal("Unsupported Y4M stream.");
  } else if (input->detect.buf_read == 4 && FourccIsIvf(input->detect.buf)) {
    fatal("IVF is not supported as input.");
  } else {
    input->file_type = FILE_TYPE_RAW;
  }
}

void DuckyEncode::InitEncoder(aom_enc_pass pass,
                              const std::vector<FIRSTPASS_STATS> *stats_list) {
  EncoderResource enc_resource = {};
  enc_resource.lookahead_push_count = 0;
  OpenInputFile(&impl_ptr_->input);
  if (impl_ptr_->input.file_type != FILE_TYPE_Y4M) {
    aom_img_alloc(&enc_resource.img, impl_ptr_->video_info.img_fmt,
                  impl_ptr_->video_info.frame_width,
                  impl_ptr_->video_info.frame_height, /*align=*/1);
  }
  AV1EncoderConfig oxcf =
      GetEncoderConfig(impl_ptr_->video_info, impl_ptr_->g_usage, pass);
  oxcf.dec_model_cfg.decoder_model_info_present_flag = 0;
  oxcf.dec_model_cfg.display_model_info_present_flag = 0;
  oxcf.ref_frm_cfg.max_reference_frames = impl_ptr_->max_ref_frames;
  oxcf.speed = impl_ptr_->speed;
  if (impl_ptr_->sb_size == BLOCK_64X64)
    oxcf.tool_cfg.superblock_size = AOM_SUPERBLOCK_SIZE_64X64;
  else
    oxcf.tool_cfg.superblock_size = AOM_SUPERBLOCK_SIZE_128X128;

  av1_initialize_enc(impl_ptr_->g_usage, impl_ptr_->rc_end_usage);
  AV1_PRIMARY *ppi =
      av1_create_primary_compressor(nullptr,
                                    /*num_lap_buffers=*/0, &oxcf);
  enc_resource.ppi = ppi;

  assert(ppi != nullptr);
  // Turn off ppi->b_calculate_psnr to avoid calling generate_psnr_packet() in
  // av1_post_encode_updates().
  // TODO(angiebird): Modify generate_psnr_packet() to handle the case that
  // cpi->ppi->output_pkt_list = nullptr.
  ppi->b_calculate_psnr = 0;

  aom_codec_err_t res = AOM_CODEC_OK;
  (void)res;
  enc_resource.stats_buf_ctx = CreateStatsBufferCtx(
      impl_ptr_->video_info.frame_count, &enc_resource.stats_buffer);
  if (pass == AOM_RC_SECOND_PASS) {
    assert(stats_list != nullptr);
    std::copy(stats_list->begin(), stats_list->end(),
              enc_resource.stats_buffer);
    *enc_resource.stats_buf_ctx->total_stats = ComputeTotalStats(*stats_list);
    oxcf.twopass_stats_in.buf = enc_resource.stats_buffer;
    // We need +1 here because av1 encoder assumes
    // oxcf.twopass_stats_in.buf[video_info.frame_count] has the total_stats
    oxcf.twopass_stats_in.sz = (impl_ptr_->video_info.frame_count + 1) *
                               sizeof(enc_resource.stats_buffer[0]);
  } else {
    assert(pass == AOM_RC_FIRST_PASS);
    // We don't use stats_list for AOM_RC_FIRST_PASS.
    assert(stats_list == nullptr);
  }
  ppi->twopass.stats_buf_ctx = enc_resource.stats_buf_ctx;
  BufferPool *buffer_pool = nullptr;
  res = av1_create_context_and_bufferpool(ppi, &ppi->cpi, &buffer_pool, &oxcf,
                                          ENCODE_STAGE, -1);
  // TODO(angiebird): Why didn't we set initial_dimensions in
  // av1_create_compressor()?
  ppi->cpi->initial_dimensions.width = oxcf.frm_dim_cfg.width;
  ppi->cpi->initial_dimensions.height = oxcf.frm_dim_cfg.height;
  // use_ducky_encode is the flag we use to change AV1 behavior
  // slightly based on DuckyEncode's need. We should minimize this kind of
  // change unless it's necessary.
  ppi->cpi->use_ducky_encode = 1;
  assert(res == AOM_CODEC_OK);
  assert(ppi->cpi != nullptr);
  assert(buffer_pool != nullptr);
  const AV1_COMP *cpi = ppi->cpi;
  SequenceHeader *seq_params = ppi->cpi->common.seq_params;
  set_sb_size(seq_params, impl_ptr_->sb_size);
  ppi->seq_params_locked = 1;
  assert(ppi->lookahead == nullptr);

  int lag_in_frames = cpi->oxcf.gf_cfg.lag_in_frames;
  ppi->lookahead = av1_lookahead_init(
      cpi->oxcf.frm_dim_cfg.width, cpi->oxcf.frm_dim_cfg.height,
      seq_params->subsampling_x, seq_params->subsampling_y,
      seq_params->use_highbitdepth, lag_in_frames, cpi->oxcf.border_in_pixels,
      cpi->common.features.byte_alignment,
      /*num_lap_buffers=*/0, /*is_all_intra=*/0,
      cpi->oxcf.tool_cfg.enable_global_motion);

  av1_tf_info_alloc(&cpi->ppi->tf_info, cpi);
  assert(ppi->lookahead != nullptr);

  impl_ptr_->enc_resource = enc_resource;
}

static void CloseInputFile(struct AvxInputContext *input) {
  fclose(input->file);
  if (input->file_type == FILE_TYPE_Y4M) y4m_input_close(&input->y4m);
}

void DuckyEncode::FreeEncoder() {
  EncoderResource *enc_resource = &impl_ptr_->enc_resource;
  CloseInputFile(&impl_ptr_->input);
  aom_img_free(&enc_resource->img);
  DestroyStatsBufferCtx(&enc_resource->stats_buf_ctx,
                        &enc_resource->stats_buffer);
  BufferPool *buffer_pool = enc_resource->ppi->cpi->common.buffer_pool;
  av1_destroy_context_and_bufferpool(enc_resource->ppi->cpi, &buffer_pool);
  av1_remove_primary_compressor(enc_resource->ppi);
  enc_resource->ppi = nullptr;
}

static int ReadFrame(struct AvxInputContext *input_ctx, aom_image_t *img) {
  FILE *f = input_ctx->file;
  y4m_input *y4m = &input_ctx->y4m;
  int shortread = 0;

  if (input_ctx->file_type == FILE_TYPE_Y4M) {
    if (y4m_input_fetch_frame(y4m, f, img) < 1) return 0;
  } else {
    shortread = read_yuv_frame(input_ctx, img);
  }

  return !shortread;
}

std::vector<FIRSTPASS_STATS> DuckyEncode::ComputeFirstPassStats() {
  aom_enc_pass pass = AOM_RC_FIRST_PASS;
  InitEncoder(pass, nullptr);
  AV1_PRIMARY *ppi = impl_ptr_->enc_resource.ppi;
  EncoderResource *enc_resource = &impl_ptr_->enc_resource;
  struct lookahead_ctx *lookahead = ppi->lookahead;
  int frame_count = impl_ptr_->video_info.frame_count;
  aom_rational64_t timestamp_ratio = impl_ptr_->timestamp_ratio;
  // TODO(angiebird): Ideally, ComputeFirstPassStats() doesn't output
  // bitstream. Do we need bitstream buffer here?
  std::vector<uint8_t> buf(1000);
  std::vector<FIRSTPASS_STATS> stats_list;
  for (int i = 0; i < frame_count; ++i) {
    if (ReadFrame(&impl_ptr_->input, &impl_ptr_->enc_resource.img)) {
      // TODO(angiebird): Set ts_start/ts_end properly
      int64_t ts_start = enc_resource->lookahead_push_count;
      int64_t ts_end = ts_start + 1;
      YV12_BUFFER_CONFIG sd;
      image2yuvconfig(&enc_resource->img, &sd);
      av1_lookahead_push(lookahead, &sd, ts_start, ts_end,
                         /*use_highbitdepth=*/0, /*flags=*/0);
      ++enc_resource->lookahead_push_count;
      AV1_COMP_DATA cpi_data = {};
      cpi_data.cx_data = buf.data();
      cpi_data.cx_data_sz = buf.size();
      cpi_data.frame_size = 0;
      cpi_data.flush = 1;  // Makes av1_get_compressed_data process a frame
      cpi_data.ts_frame_start = ts_start;
      cpi_data.ts_frame_end = ts_end;
      cpi_data.pop_lookahead = 1;
      cpi_data.timestamp_ratio = &timestamp_ratio;
      // av1_get_compressed_data only generates first pass stats not
      // compresses data
      int res = av1_get_compressed_data(ppi->cpi, &cpi_data);
      (void)res;
      assert(res == static_cast<int>(AOM_CODEC_OK));
      stats_list.push_back(*(ppi->twopass.stats_buf_ctx->stats_in_end - 1));
      av1_post_encode_updates(ppi->cpi, &cpi_data);
    }
  }
  av1_end_first_pass(ppi->cpi);

  FreeEncoder();
  return stats_list;
}

void DuckyEncode::StartEncode(const std::vector<FIRSTPASS_STATS> &stats_list) {
  aom_enc_pass pass = AOM_RC_SECOND_PASS;
  impl_ptr_->stats_list = stats_list;
  InitEncoder(pass, &stats_list);
  write_temp_delimiter_ = true;
}

static void DuckyEncodeInfoSetGopStruct(AV1_PRIMARY *ppi,
                                        const GopStruct &gop_struct,
                                        const GopEncodeInfo &gop_encode_info) {
  GF_GROUP *gf_group = &ppi->gf_group;
  ppi->p_rc.baseline_gf_interval = gop_struct.show_frame_count;
  ppi->internal_altref_allowed = 1;

  gf_group->size = static_cast<int>(gop_struct.gop_frame_list.size());
  gf_group->max_layer_depth = 0;

  int i = 0;
  for (const auto &frame : gop_struct.gop_frame_list) {
    gf_group->update_type[i] = (int)frame.update_type;
    if (frame.update_type == GopFrameType::kRegularArf) gf_group->arf_index = i;

    gf_group->frame_type[i] = !frame.is_key_frame;

    gf_group->q_val[i] = gop_encode_info.param_list[i].q_index;
    gf_group->rdmult_val[i] = gop_encode_info.param_list[i].rdmult;

    gf_group->cur_frame_idx[i] = 0;
    gf_group->arf_src_offset[i] = frame.order_idx - frame.display_idx;
    gf_group->cur_frame_idx[i] = frame.display_idx;
    gf_group->src_offset[i] = 0;

    // TODO(jingning): Placeholder - update the arf boost.
    gf_group->arf_boost[i] = 500;
    gf_group->layer_depth[i] = frame.layer_depth;
    gf_group->max_layer_depth =
        AOMMAX(frame.layer_depth, gf_group->max_layer_depth);
    gf_group->refbuf_state[i] =
        frame.is_key_frame ? REFBUF_RESET : REFBUF_UPDATE;

    std::fill_n(gf_group->ref_frame_list[i], REF_FRAMES, -1);
    gf_group->update_ref_idx[i] = -1;
    for (int ref_idx = 0;
         ref_idx < static_cast<int>(frame.ref_frame_list.size()); ++ref_idx) {
      int ref_frame = static_cast<int>(frame.ref_frame_list[ref_idx].name);
      gf_group->ref_frame_list[i][ref_frame] =
          static_cast<int8_t>(frame.ref_frame_list[ref_idx].index);
    }
    gf_group->update_ref_idx[i] = frame.update_ref_idx;
    gf_group->primary_ref_idx[i] = frame.primary_ref_frame.index;
    ++i;
  }
  ppi->cpi->gf_frame_index = 0;
}

static void DuckyEncodeInfoSetEncodeFrameDecision(
    DuckyEncodeInfo *ducky_encode_info, const EncodeFrameDecision &decision) {
  DuckyEncodeFrameInfo *frame_info = &ducky_encode_info->frame_info;
  *frame_info = {};
  frame_info->qp_mode = static_cast<DUCKY_ENCODE_FRAME_MODE>(decision.qp_mode);
  frame_info->gop_mode = static_cast<DUCKY_ENCODE_GOP_MODE>(decision.gop_mode);
  frame_info->q_index = decision.parameters.q_index;
  frame_info->rdmult = decision.parameters.rdmult;
  const size_t num_superblocks =
      decision.parameters.superblock_encode_params.size();
  frame_info->delta_q_enabled = 0;
  if (num_superblocks > 1) {
    frame_info->delta_q_enabled = 1;
    frame_info->superblock_encode_qindex = new int[num_superblocks];
    frame_info->superblock_encode_rdmult = new int[num_superblocks];
    for (size_t i = 0; i < num_superblocks; ++i) {
      frame_info->superblock_encode_qindex[i] =
          decision.parameters.superblock_encode_params[i].q_index;
      frame_info->superblock_encode_rdmult[i] =
          decision.parameters.superblock_encode_params[i].rdmult;
    }
  }
}

static void DuckyEncodeInfoGetEncodeFrameResult(
    const DuckyEncodeInfo *ducky_encode_info, EncodeFrameResult *result) {
  const DuckyEncodeFrameResult &frame_result = ducky_encode_info->frame_result;
  result->global_order_idx = frame_result.global_order_idx;
  result->q_index = frame_result.q_index;
  result->rdmult = frame_result.rdmult;
  result->rate = frame_result.rate;
  result->dist = frame_result.dist;
  result->psnr = frame_result.psnr;
}

static void WriteObu(AV1_PRIMARY *ppi, AV1_COMP_DATA *cpi_data) {
  AV1_COMP *const cpi = ppi->cpi;
  uint32_t obu_header_size = 1;
  const uint32_t obu_payload_size = 0;
  const size_t length_field_size = aom_uleb_size_in_bytes(obu_payload_size);

  const size_t move_offset = obu_header_size + length_field_size;
  memmove(cpi_data->cx_data + move_offset, cpi_data->cx_data,
          cpi_data->frame_size);
  obu_header_size =
      av1_write_obu_header(&ppi->level_params, &cpi->frame_header_count,
                           OBU_TEMPORAL_DELIMITER, 0, cpi_data->cx_data);

  // OBUs are preceded/succeeded by an unsigned leb128 coded integer.
  if (av1_write_uleb_obu_size(obu_header_size, obu_payload_size,
                              cpi_data->cx_data) != AOM_CODEC_OK) {
    aom_internal_error(&ppi->error, AOM_CODEC_ERROR, NULL);
  }

  cpi_data->frame_size +=
      obu_header_size + obu_payload_size + length_field_size;
}

TplGopStats DuckyEncode::ObtainTplStats(const GopStruct gop_struct,
                                        bool rate_dist_present) {
  TplGopStats tpl_gop_stats;

  AV1_PRIMARY *ppi = impl_ptr_->enc_resource.ppi;
  const uint8_t block_mis_log2 = ppi->tpl_data.tpl_stats_block_mis_log2;

  for (size_t idx = 0; idx < gop_struct.gop_frame_list.size(); ++idx) {
    TplFrameStats tpl_frame_stats = {};
    tpl_frame_stats.rate_dist_present = rate_dist_present;

    TplDepFrame *tpl_frame = &ppi->tpl_data.tpl_frame[idx];
    if (gop_struct.gop_frame_list[idx].update_type == GopFrameType::kOverlay ||
        gop_struct.gop_frame_list[idx].update_type ==
            GopFrameType::kIntermediateOverlay) {
      tpl_gop_stats.frame_stats_list.push_back(tpl_frame_stats);
      continue;
    }

    int ref_frame_index_mapping[REF_FRAMES] = { 0 };
    const GopFrame &gop_frame = gop_struct.gop_frame_list[idx];

    for (auto &rf : gop_frame.ref_frame_list) {
      ref_frame_index_mapping[static_cast<int>(rf.name)] = rf.index;
    }

    const int mi_rows = tpl_frame->mi_rows;
    const int mi_cols = tpl_frame->mi_cols;
    const int tpl_frame_stride = tpl_frame->stride;
    tpl_frame_stats.frame_height = mi_rows * MI_SIZE;
    tpl_frame_stats.frame_width = mi_cols * MI_SIZE;
    tpl_frame_stats.min_block_size = (1 << block_mis_log2) * MI_SIZE;

    const int mi_step = 1 << block_mis_log2;
    for (int mi_row = 0; mi_row < mi_rows; mi_row += mi_step) {
      for (int mi_col = 0; mi_col < mi_cols; mi_col += mi_step) {
        int tpl_blk_pos = (mi_row >> block_mis_log2) * tpl_frame_stride +
                          (mi_col >> block_mis_log2);
        TplDepStats *tpl_stats_ptr = &tpl_frame->tpl_stats_ptr[tpl_blk_pos];

        TplBlockStats block_stats;
        block_stats.row = mi_row * MI_SIZE;
        block_stats.col = mi_col * MI_SIZE;
        block_stats.height = (1 << block_mis_log2) * MI_SIZE;
        block_stats.width = (1 << block_mis_log2) * MI_SIZE;

        block_stats.inter_cost =
            RDCOST(tpl_frame->base_rdmult, tpl_stats_ptr->recrf_rate,
                   tpl_stats_ptr->recrf_dist);
        block_stats.intra_cost =
            RDCOST(tpl_frame->base_rdmult, tpl_stats_ptr->intra_rate,
                   tpl_stats_ptr->intra_dist);

        if (tpl_frame_stats.rate_dist_present) {
          block_stats.recrf_dist = tpl_stats_ptr->recrf_dist;
          block_stats.recrf_rate = tpl_stats_ptr->recrf_rate;
          block_stats.intra_pred_err = tpl_stats_ptr->intra_sse;
          block_stats.inter_pred_err = tpl_stats_ptr->recrf_sse;
        }

        block_stats.ref_frame_index = { -1, -1 };

        for (int i = 0; i < kBlockRefCount; ++i) {
          if (tpl_stats_ptr->ref_frame_index[i] >= 0) {
            block_stats.ref_frame_index[i] =
                ref_frame_index_mapping[tpl_stats_ptr->ref_frame_index[i] + 1];
            block_stats.mv[i] = {
              tpl_stats_ptr->mv[tpl_stats_ptr->ref_frame_index[i]].as_mv.row,
              tpl_stats_ptr->mv[tpl_stats_ptr->ref_frame_index[i]].as_mv.col, 3
            };
          }
        }
        tpl_frame_stats.block_stats_list.push_back(block_stats);
      }
    }

    tpl_gop_stats.frame_stats_list.push_back(tpl_frame_stats);
  }

  return tpl_gop_stats;
}

// Obtain TPL stats through ducky_encode.
// TODO(jianj): Populate rate_dist_present flag through qmode_rc_encoder
std::vector<TplGopStats> DuckyEncode::ComputeTplStats(
    const std::vector<FIRSTPASS_STATS> &stats_list,
    const GopStructList &gop_list,
    const GopEncodeInfoList &gop_encode_info_list) {
  StartEncode(stats_list);
  std::vector<TplGopStats> tpl_gop_stats_list;
  AV1_PRIMARY *ppi = impl_ptr_->enc_resource.ppi;
  const VideoInfo &video_info = impl_ptr_->video_info;
  write_temp_delimiter_ = true;
  AllocateBitstreamBuffer(video_info);

  // Go through each gop and encode each frame in the gop
  for (size_t i = 0; i < gop_list.size(); ++i) {
    const aom::GopStruct &gop_struct = gop_list[i];
    const aom::GopEncodeInfo &gop_encode_info = gop_encode_info_list[i];

    DuckyEncodeInfoSetGopStruct(ppi, gop_struct, gop_encode_info);

    aom::TplGopStats tpl_gop_stats;
    for (auto &frame_param : gop_encode_info.param_list) {
      // encoding frame frame_number
      aom::EncodeFrameDecision frame_decision = { aom::EncodeFrameMode::kQindex,
                                                  aom::EncodeGopMode::kGopRcl,
                                                  frame_param };
      EncodeFrame(frame_decision);
      if (ppi->cpi->common.show_frame) pending_ctx_size_ = 0;
      write_temp_delimiter_ = ppi->cpi->common.show_frame;
    }
    // The rate_dist_present needs to be populated.
    tpl_gop_stats = ObtainTplStats(gop_struct, 0);
    tpl_gop_stats_list.push_back(tpl_gop_stats);
  }
  EndEncode();
  return tpl_gop_stats_list;
}

std::vector<TplGopStats> DuckyEncode::ComputeTwoPassTplStats(
    const std::vector<FIRSTPASS_STATS> &stats_list,
    const GopStructList &gop_list,
    const GopEncodeInfoList &gop_encode_info_list,
    const GopEncodeInfoList &alt_gop_encode_info_list) {
  std::vector<TplGopStats> first_tpl_gop_stats_list =
      ComputeTplStats(stats_list, gop_list, gop_encode_info_list);
  const std::vector<TplGopStats> second_tpl_gop_stats_list =
      ComputeTplStats(stats_list, gop_list, alt_gop_encode_info_list);
  assert(first_tpl_gop_stats_list.size() == second_tpl_gop_stats_list.size());

  // Set alternate_block_stats_list in first_tpl_gop_stats_list
  // and return first_tpl_gop_stats_list
  for (size_t i = 0; i < first_tpl_gop_stats_list.size(); ++i) {
    for (size_t j = 0; j < first_tpl_gop_stats_list[i].frame_stats_list.size();
         ++j) {
      first_tpl_gop_stats_list[i]
          .frame_stats_list[j]
          .alternate_block_stats_list =
          second_tpl_gop_stats_list[i].frame_stats_list[j].block_stats_list;
    }
  }
  return first_tpl_gop_stats_list;
}

// Conduct final encoding process.
std::vector<EncodeFrameResult> DuckyEncode::EncodeVideo(
    const GopStructList &gop_list,
    const GopEncodeInfoList &gop_encode_info_list) {
  AV1_PRIMARY *ppi = impl_ptr_->enc_resource.ppi;
  std::vector<EncodeFrameResult> encoded_frame_list;
  const VideoInfo &video_info = impl_ptr_->video_info;

  write_temp_delimiter_ = true;
  AllocateBitstreamBuffer(video_info);

  // Go through each gop and encode each frame in the gop
  for (size_t i = 0; i < gop_list.size(); ++i) {
    const aom::GopStruct &gop_struct = gop_list[i];
    const aom::GopEncodeInfo &gop_encode_info = gop_encode_info_list[i];
    DuckyEncodeInfoSetGopStruct(ppi, gop_struct, gop_encode_info);

    for (auto &frame_param : gop_encode_info.param_list) {
      aom::EncodeFrameDecision frame_decision = { aom::EncodeFrameMode::kQindex,
                                                  aom::EncodeGopMode::kGopRcl,
                                                  frame_param };
      EncodeFrameResult temp_result = EncodeFrame(frame_decision);
      if (ppi->cpi->common.show_frame) {
        bitstream_buf_.resize(pending_ctx_size_);
        EncodeFrameResult encode_frame_result = temp_result;
        encode_frame_result.bitstream_buf = bitstream_buf_;
        encoded_frame_list.push_back(encode_frame_result);

        AllocateBitstreamBuffer(video_info);
      }
      write_temp_delimiter_ = ppi->cpi->common.show_frame;
    }
  }

  return encoded_frame_list;
}

EncodeFrameResult DuckyEncode::EncodeFrame(
    const EncodeFrameDecision &decision) {
  EncodeFrameResult encode_frame_result = {};
  encode_frame_result.bitstream_buf = bitstream_buf_;
  AV1_PRIMARY *ppi = impl_ptr_->enc_resource.ppi;
  aom_image_t *img = &impl_ptr_->enc_resource.img;
  AV1_COMP *const cpi = ppi->cpi;
  struct lookahead_ctx *lookahead = ppi->lookahead;

  while (!av1_lookahead_full(lookahead)) {
    if (ReadFrame(&impl_ptr_->input, img)) {
      YV12_BUFFER_CONFIG sd;
      image2yuvconfig(img, &sd);
      int64_t ts_start = impl_ptr_->enc_resource.lookahead_push_count;
      int64_t ts_end = ts_start + 1;
      av1_lookahead_push(lookahead, &sd, ts_start, ts_end,
                         /*use_highbitdepth=*/0, /*flags=*/0);
      ++impl_ptr_->enc_resource.lookahead_push_count;
    } else {
      break;
    }
  }

  AV1_COMP_DATA cpi_data = {};
  cpi_data.cx_data = bitstream_buf_.data() + pending_ctx_size_;
  cpi_data.cx_data_sz = bitstream_buf_.size() - pending_ctx_size_;
  cpi_data.frame_size = 0;
  cpi_data.flush = 1;
  // ts_frame_start and ts_frame_end are not as important since we are focusing
  // on q mode
  cpi_data.ts_frame_start = impl_ptr_->enc_resource.encode_frame_count;
  cpi_data.ts_frame_end = cpi_data.ts_frame_start + 1;
  cpi_data.pop_lookahead = 1;
  cpi_data.timestamp_ratio = &impl_ptr_->timestamp_ratio;
  ++impl_ptr_->enc_resource.encode_frame_count;

  av1_compute_num_workers_for_mt(cpi);
  av1_init_frame_mt(ppi, cpi);

  DuckyEncodeInfoSetEncodeFrameDecision(&cpi->ducky_encode_info, decision);
  const int status = av1_get_compressed_data(cpi, &cpi_data);

  if (write_temp_delimiter_) WriteObu(ppi, &cpi_data);
  (void)status;
  assert(status == static_cast<int>(AOM_CODEC_OK));
  DuckyEncodeInfoGetEncodeFrameResult(&cpi->ducky_encode_info,
                                      &encode_frame_result);
  av1_post_encode_updates(cpi, &cpi_data);
  if (cpi->common.show_frame) {
    // decrement frames_left counter
    ppi->frames_left = AOMMAX(0, ppi->frames_left - 1);
  }

  pending_ctx_size_ += cpi_data.frame_size;

  fprintf(stderr, "frame %d, qp = %d, size %d, PSNR %f\n",
          encode_frame_result.global_order_idx, encode_frame_result.q_index,
          encode_frame_result.rate, encode_frame_result.psnr);
  delete[] cpi->ducky_encode_info.frame_info.superblock_encode_qindex;
  delete[] cpi->ducky_encode_info.frame_info.superblock_encode_rdmult;
  return encode_frame_result;
}

void DuckyEncode::EndEncode() { FreeEncoder(); }

void DuckyEncode::AllocateBitstreamBuffer(const VideoInfo &video_info) {
  pending_ctx_size_ = 0;
  // TODO(angiebird): Set bitstream_buf size to a conservatve upperbound.
  bitstream_buf_.assign(
      video_info.frame_width * video_info.frame_height * 3 * 8, 0);
}
}  // namespace aom