uvg266/src/encoderstate.c

/*****************************************************************************
 * This file is part of Kvazaar HEVC encoder.
 *
 * Copyright (C) 2013-2015 Tampere University of Technology and others (see
 * COPYING file).
 *
 * Kvazaar is free software: you can redistribute it and/or modify it under
 * the terms of the GNU Lesser General Public License as published by the
 * Free Software Foundation; either version 2.1 of the License, or (at your
 * option) any later version.
 *
 * Kvazaar is distributed in the hope that it will be useful, but WITHOUT ANY
 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
 * FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
 * more details.
 *
 * You should have received a copy of the GNU General Public License along
 * with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
 ****************************************************************************/

#include "encoderstate.h"

#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include "cabac.h"
#include "context.h"
#include "encode_coding_tree.h"
#include "encoder_state-bitstream.h"
#include "filter.h"
#include "image.h"
#include "rate_control.h"
#include "sao.h"
#include "search.h"
#include "tables.h"
#include "threadqueue.h"


int kvz_encoder_state_match_children_of_previous_frame(encoder_state_t * const state) {
  int i;
  for (i = 0; state->children[i].encoder_control; ++i) {
    //Child should also exist for previous encoder
    assert(state->previous_encoder_state->children[i].encoder_control);
    state->children[i].previous_encoder_state = &state->previous_encoder_state->children[i];
    kvz_encoder_state_match_children_of_previous_frame(&state->children[i]);
  }
  return 1;
}

/**
 * \brief Save edge pixels before SAO to buffers.
 *
 * Copies pixels at the edges of the area that will be filtered with SAO to
 * the given buffers. If deblocking is enabled, the pixels must have been
 * deblocked before this.
 *
 * The saved pixels will be needed later when doing SAO for the neighboring
 * areas.
 */
static void encoder_state_recdata_before_sao_to_bufs(
    encoder_state_t * const state,
    const lcu_order_element_t * const lcu,
    yuv_t * const hor_buf,
    yuv_t * const ver_buf)
{
  videoframe_t* const frame = state->tile->frame;

  if (hor_buf && lcu->below) {
    // Copy the bottommost row that will be filtered with SAO to the
    // horizontal buffer.
    vector2d_t pos = {
      .x = lcu->position_px.x,
      .y = lcu->position_px.y + LCU_WIDTH - SAO_DELAY_PX - 1,
    };
    // Copy all pixels that have been deblocked.
    int length = lcu->size.x - DEBLOCK_DELAY_PX;

    if (!lcu->right) {
      // If there is no LCU to the right, the last pixels will be
      // filtered too.
      length += DEBLOCK_DELAY_PX;
    }

    if (lcu->left) {
      // The rightmost pixels of the CTU to the left will also be filtered.
      pos.x -= DEBLOCK_DELAY_PX;
      length += DEBLOCK_DELAY_PX;
    }

    const unsigned from_index = pos.x + pos.y * frame->rec->stride;
    // NOTE: The horizontal buffer is indexed by
    //    x_px + y_lcu * frame->width
    // where x_px is in pixels and y_lcu in number of LCUs.
    const unsigned to_index = pos.x + lcu->position.y * frame->width;

    kvz_pixels_blit(&frame->rec->y[from_index],
                    &hor_buf->y[to_index],
                    length, 1,
                    frame->rec->stride,
                    frame->width);

    if (state->encoder_control->chroma_format != KVZ_CSP_400) {
      const unsigned from_index_c = (pos.x / 2) + (pos.y / 2) * frame->rec->stride / 2;
      const unsigned to_index_c = (pos.x / 2) + lcu->position.y * frame->width / 2;

      kvz_pixels_blit(&frame->rec->u[from_index_c],
                      &hor_buf->u[to_index_c],
                      length / 2, 1,
                      frame->rec->stride / 2,
                      frame->width / 2);
      kvz_pixels_blit(&frame->rec->v[from_index_c],
                      &hor_buf->v[to_index_c],
                      length / 2, 1,
                      frame->rec->stride / 2,
                      frame->width / 2);
    }
  }

  if (ver_buf && lcu->right) {
    // Copy the rightmost column that will be filtered with SAO to the
    // vertical buffer.
    vector2d_t pos = {
      .x = lcu->position_px.x + LCU_WIDTH - SAO_DELAY_PX - 1,
      .y = lcu->position_px.y,
    };
    int length = lcu->size.y - DEBLOCK_DELAY_PX;

    if (!lcu->below) {
      // If there is no LCU below, the last pixels will be filtered too.
      length += DEBLOCK_DELAY_PX;
    }

    if (lcu->above) {
      // The bottommost pixels of the CTU above will also be filtered.
      pos.y -= DEBLOCK_DELAY_PX;
      length += DEBLOCK_DELAY_PX;
    }

    const unsigned from_index = pos.x + pos.y * frame->rec->stride;
    // NOTE: The vertical buffer is indexed by
    //    x_lcu * frame->height + y_px
    // where x_lcu is in number of LCUs and y_px in pixels.
    const unsigned to_index = lcu->position.x * frame->height + pos.y;

    kvz_pixels_blit(&frame->rec->y[from_index],
                    &ver_buf->y[to_index],
                    1, length,
                    frame->rec->stride, 1);

    if (state->encoder_control->chroma_format != KVZ_CSP_400) {
      const unsigned from_index_c = (pos.x / 2) + (pos.y / 2) * frame->rec->stride / 2;
      const unsigned to_index_c = lcu->position.x * frame->height / 2 + pos.y / 2;

      kvz_pixels_blit(&frame->rec->u[from_index_c],
                      &ver_buf->u[to_index_c],
                      1, length / 2,
                      frame->rec->stride / 2, 1);
      kvz_pixels_blit(&frame->rec->v[from_index_c],
                      &ver_buf->v[to_index_c],
                      1, length / 2,
                      frame->rec->stride / 2, 1);
    }
  }
}

static void encoder_state_recdata_to_bufs(encoder_state_t * const state,
                                          const lcu_order_element_t * const lcu,
                                          yuv_t * const hor_buf,
                                          yuv_t * const ver_buf)
{
  videoframe_t* const frame = state->tile->frame;
  
  if (hor_buf) {
    //Copy the bottom row of this LCU to the horizontal buffer
    vector2d_t bottom = { lcu->position_px.x, lcu->position_px.y + lcu->size.y - 1 };
    const int lcu_row = lcu->position.y;

    unsigned from_index = bottom.y * frame->rec->stride + bottom.x;
    unsigned to_index = lcu->position_px.x + lcu_row * frame->width;
    
    kvz_pixels_blit(&frame->rec->y[from_index],
                    &hor_buf->y[to_index],
                    lcu->size.x, 1,
                    frame->rec->stride, frame->width);

    if (state->encoder_control->chroma_format != KVZ_CSP_400) {
      unsigned from_index_c = (bottom.y / 2) * frame->rec->stride / 2 + (bottom.x / 2);
      unsigned to_index_c = lcu->position_px.x / 2 + lcu_row * frame->width / 2;

      kvz_pixels_blit(&frame->rec->u[from_index_c],
                      &hor_buf->u[to_index_c],
                      lcu->size.x / 2, 1, 
                      frame->rec->stride / 2, frame->width / 2);
      kvz_pixels_blit(&frame->rec->v[from_index_c],
                      &hor_buf->v[to_index_c],
                      lcu->size.x / 2, 1,
                      frame->rec->stride / 2, frame->width / 2);
    }
  }
  
  if (ver_buf) {
    //Copy the right row of this LCU to the vertical buffer.
    
    const int lcu_col = lcu->position.x;
    vector2d_t left = { lcu->position_px.x + lcu->size.x - 1, lcu->position_px.y };
    
    kvz_pixels_blit(&frame->rec->y[left.y * frame->rec->stride + left.x],
                    &ver_buf->y[lcu->position_px.y + lcu_col * frame->height],
                    1, lcu->size.y,
                    frame->rec->stride, 1);

    if (state->encoder_control->chroma_format != KVZ_CSP_400) {
      unsigned from_index = (left.y / 2) * frame->rec->stride / 2 + (left.x / 2);
      unsigned to_index = lcu->position_px.y / 2 + lcu_col * frame->height / 2;

      kvz_pixels_blit(&frame->rec->u[from_index],
                      &ver_buf->u[to_index],
                      1, lcu->size.y / 2,
                      frame->rec->stride / 2, 1);
      kvz_pixels_blit(&frame->rec->v[from_index],
                      &ver_buf->v[to_index],
                      1, lcu->size.y / 2,
                      frame->rec->stride / 2, 1);
    }
  }
  
}

/**
 * \brief Do SAO reconstuction for all available pixels.
 *
 * Does SAO reconstruction for all pixels that are available after the
 * given LCU has been deblocked. This means the following pixels:
 *  - bottom-right block of SAO_DELAY_PX times SAO_DELAY_PX in the lcu to
 *    the left and up
 *  - the rightmost SAO_DELAY_PX pixels of the LCU to the left (excluding
 *    the bottommost pixel)
 *  - the bottommost SAO_DELAY_PX pixels of the LCU above (excluding the
 *    rightmost pixels)
 *  - all pixels inside the LCU, excluding the rightmost SAO_DELAY_PX and
 *    bottommost SAO_DELAY_PX
 */
static void encoder_sao_reconstruct(const encoder_state_t *const state,
                                    const lcu_order_element_t *const lcu)
{
  videoframe_t *const frame = state->tile->frame;


  // Temporary buffers for SAO input pixels. The buffers cover the pixels
  // inside the LCU (LCU_WIDTH x LCU_WIDTH), SAO_DELAY_PX wide bands to the
  // left and above the LCU, and one pixel border on the left and top
  // sides. We add two extra pixels to the buffers because the AVX2 SAO
  // reconstruction reads up to two extra bytes when using edge SAO in the
  // horizontal direction.
#define SAO_BUF_WIDTH   (1 + SAO_DELAY_PX   + LCU_WIDTH)
#define SAO_BUF_WIDTH_C (1 + SAO_DELAY_PX/2 + LCU_WIDTH_C)
  kvz_pixel sao_buf_y_array[SAO_BUF_WIDTH   * SAO_BUF_WIDTH   + 2];
  kvz_pixel sao_buf_u_array[SAO_BUF_WIDTH_C * SAO_BUF_WIDTH_C + 2];
  kvz_pixel sao_buf_v_array[SAO_BUF_WIDTH_C * SAO_BUF_WIDTH_C + 2];

  // Pointers to the top-left pixel of the LCU in the buffers.
  kvz_pixel *const sao_buf_y = &sao_buf_y_array[(SAO_DELAY_PX + 1) * (SAO_BUF_WIDTH + 1)];
  kvz_pixel *const sao_buf_u = &sao_buf_u_array[(SAO_DELAY_PX/2 + 1) * (SAO_BUF_WIDTH_C + 1)];
  kvz_pixel *const sao_buf_v = &sao_buf_v_array[(SAO_DELAY_PX/2 + 1) * (SAO_BUF_WIDTH_C + 1)];

  const int x_offsets[3] = {
    // If there is an lcu to the left, we need to filter its rightmost
    // pixels.
    lcu->left ? -SAO_DELAY_PX : 0,
    0,
    // If there is an lcu to the right, the rightmost pixels of this LCU
    // are filtered when filtering that LCU. Otherwise we filter them now.
    lcu->size.x - (lcu->right ? SAO_DELAY_PX : 0),
  };

  const int y_offsets[3] = {
    // If there is an lcu above, we need to filter its bottommost pixels.
    lcu->above ? -SAO_DELAY_PX : 0,
    0,
    // If there is an lcu below, the bottommost pixels of this LCU are
    // filtered when filtering that LCU. Otherwise we filter them now.
    lcu->size.y - (lcu->below ? SAO_DELAY_PX : 0),
  };

  // Number of pixels around the block that need to be copied to the
  // buffers.
  const int border_left  = lcu->left  ? 1 : 0;
  const int border_right = lcu->right ? 1 : 0;
  const int border_above = lcu->above ? 1 : 0;
  const int border_below = lcu->below ? 1 : 0;

  // Index of the pixel at the intersection of the top and left borders.
  const int border_index = (x_offsets[0] - border_left) +
                           (y_offsets[0] - border_above) * SAO_BUF_WIDTH;
  const int border_index_c = (x_offsets[0]/2 - border_left) +
                             (y_offsets[0]/2 - border_above) * SAO_BUF_WIDTH_C;
  // Width and height of the whole area to filter.
  const int width  = x_offsets[2] - x_offsets[0];
  const int height = y_offsets[2] - y_offsets[0];

  // Copy bordering pixels from above and left to buffers.
  if (lcu->above) {
    const int from_index = (lcu->position_px.x + x_offsets[0] - border_left) +
                           (lcu->position.y - 1) * frame->width;
    kvz_pixels_blit(&state->tile->hor_buf_before_sao->y[from_index],
                    &sao_buf_y[border_index],
                    width + border_left + border_right,
                    1,
                    frame->width,
                    SAO_BUF_WIDTH);
    if (state->encoder_control->chroma_format != KVZ_CSP_400) {
      const int from_index_c = (lcu->position_px.x + x_offsets[0])/2 - border_left +
                               (lcu->position.y - 1) * frame->width/2;
      kvz_pixels_blit(&state->tile->hor_buf_before_sao->u[from_index_c],
                      &sao_buf_u[border_index_c],
                      width/2 + border_left + border_right,
                      1,
                      frame->width/2,
                      SAO_BUF_WIDTH_C);
      kvz_pixels_blit(&state->tile->hor_buf_before_sao->v[from_index_c],
                      &sao_buf_v[border_index_c],
                      width/2 + border_left + border_right,
                      1,
                      frame->width/2,
                      SAO_BUF_WIDTH_C);
    }
  }
  if (lcu->left) {
    const int from_index = (lcu->position.x - 1) * frame->height +
                           (lcu->position_px.y + y_offsets[0] - border_above);
    kvz_pixels_blit(&state->tile->ver_buf_before_sao->y[from_index],
                    &sao_buf_y[border_index],
                    1,
                    height + border_above + border_below,
                    1,
                    SAO_BUF_WIDTH);
    if (state->encoder_control->chroma_format != KVZ_CSP_400) {
      const int from_index_c = (lcu->position.x - 1) * frame->height/2 +
                               (lcu->position_px.y + y_offsets[0])/2 - border_above;
      kvz_pixels_blit(&state->tile->ver_buf_before_sao->u[from_index_c],
                      &sao_buf_u[border_index_c],
                      1,
                      height/2 + border_above + border_below,
                      1,
                      SAO_BUF_WIDTH_C);
      kvz_pixels_blit(&state->tile->ver_buf_before_sao->v[from_index_c],
                      &sao_buf_v[border_index_c],
                      1,
                      height/2 + border_above + border_below,
                      1,
                      SAO_BUF_WIDTH_C);
    }
  }
  // Copy pixels that will be filtered and bordering pixels from right and
  // below.
  const int from_index = (lcu->position_px.x + x_offsets[0]) +
                         (lcu->position_px.y + y_offsets[0]) * frame->rec->stride;
  const int to_index = x_offsets[0] + y_offsets[0] * SAO_BUF_WIDTH;
  kvz_pixels_blit(&frame->rec->y[from_index],
                  &sao_buf_y[to_index],
                  width + border_right,
                  height + border_below,
                  frame->rec->stride,
                  SAO_BUF_WIDTH);
  if (state->encoder_control->chroma_format != KVZ_CSP_400) {
    const int from_index_c = (lcu->position_px.x + x_offsets[0])/2 +
                             (lcu->position_px.y + y_offsets[0])/2 * frame->rec->stride/2;
    const int to_index_c = x_offsets[0]/2 + y_offsets[0]/2 * SAO_BUF_WIDTH_C;
    kvz_pixels_blit(&frame->rec->u[from_index_c],
                    &sao_buf_u[to_index_c],
                    width/2 + border_right,
                    height/2 + border_below,
                    frame->rec->stride/2,
                    SAO_BUF_WIDTH_C);
    kvz_pixels_blit(&frame->rec->v[from_index_c],
                    &sao_buf_v[to_index_c],
                    width/2 + border_right,
                    height/2 + border_below,
                    frame->rec->stride/2,
                    SAO_BUF_WIDTH_C);
  }

  // We filter the pixels in four parts:
  //  1. Pixels that belong to the LCU above and to the left
  //  2. Pixels that belong to the LCU above
  //  3. Pixels that belong to the LCU to the left
  //  4. Pixels that belong to the current LCU
  for (int y_offset_index = 0; y_offset_index < 2; y_offset_index++) {
    for (int x_offset_index = 0; x_offset_index < 2; x_offset_index++) {
      const int x = x_offsets[x_offset_index];
      const int y = y_offsets[y_offset_index];
      const int width = x_offsets[x_offset_index + 1] - x;
      const int height = y_offsets[y_offset_index + 1] - y;

      if (width == 0 || height == 0) continue;

      const int lcu_x = (lcu->position_px.x + x) >> LOG2_LCU_WIDTH;
      const int lcu_y = (lcu->position_px.y + y) >> LOG2_LCU_WIDTH;
      const int lcu_index = lcu_x + lcu_y * frame->width_in_lcu;
      const sao_info_t *sao_luma   = &frame->sao_luma[lcu_index];
      const sao_info_t *sao_chroma = &frame->sao_chroma[lcu_index];

      kvz_sao_reconstruct(state,
                          &sao_buf_y[x + y * SAO_BUF_WIDTH],
                          SAO_BUF_WIDTH,
                          lcu->position_px.x + x,
                          lcu->position_px.y + y,
                          width,
                          height,
                          sao_luma,
                          COLOR_Y);

      if (state->encoder_control->chroma_format != KVZ_CSP_400) {
        // Coordinates in chroma pixels.
        int x_c = x >> 1;
        int y_c = y >> 1;

        kvz_sao_reconstruct(state,
                            &sao_buf_u[x_c + y_c * SAO_BUF_WIDTH_C],
                            SAO_BUF_WIDTH_C,
                            lcu->position_px.x / 2 + x_c,
                            lcu->position_px.y / 2 + y_c,
                            width / 2,
                            height / 2,
                            sao_chroma,
                            COLOR_U);
        kvz_sao_reconstruct(state,
                            &sao_buf_v[x_c + y_c * SAO_BUF_WIDTH_C],
                            SAO_BUF_WIDTH_C,
                            lcu->position_px.x / 2 + x_c,
                            lcu->position_px.y / 2 + y_c,
                            width / 2,
                            height / 2,
                            sao_chroma,
                            COLOR_V);
      }
    }
  }
}

static void encode_sao_color(encoder_state_t * const state, sao_info_t *sao,
                             color_t color_i)
{
  cabac_data_t * const cabac = &state->cabac;
  sao_eo_cat i;
  int offset_index = (color_i == COLOR_V) ? 5 : 0;

  // Skip colors with no SAO.
  //FIXME: for now, we always have SAO for all channels
  if (color_i == COLOR_Y && 0) return;
  if (color_i != COLOR_Y && 0) return;

  /// sao_type_idx_luma:   TR, cMax = 2, cRiceParam = 0, bins = {0, bypass}
  /// sao_type_idx_chroma: TR, cMax = 2, cRiceParam = 0, bins = {0, bypass}
  // Encode sao_type_idx for Y and U+V.
  if (color_i != COLOR_V) {
    cabac->cur_ctx = &(cabac->ctx.sao_type_idx_model);
    CABAC_BIN(cabac, sao->type != SAO_TYPE_NONE, "sao_type_idx");
    if (sao->type == SAO_TYPE_BAND) {
      CABAC_BIN_EP(cabac, 0, "sao_type_idx_ep");
    } else if (sao->type == SAO_TYPE_EDGE) {
      CABAC_BIN_EP(cabac, 1, "sao_type_idx_ep");
    }
  }

  if (sao->type == SAO_TYPE_NONE) return;

  /// sao_offset_abs[][][][]: TR, cMax = (1 << (Min(bitDepth, 10) - 5)) - 1,
  ///                         cRiceParam = 0, bins = {bypass x N}
  for (i = SAO_EO_CAT1; i <= SAO_EO_CAT4; ++i) {
    kvz_cabac_write_unary_max_symbol_ep(cabac, abs(sao->offsets[i + offset_index]), SAO_ABS_OFFSET_MAX);
  }

  /// sao_offset_sign[][][][]: FL, cMax = 1, bins = {bypass}
  /// sao_band_position[][][]: FL, cMax = 31, bins = {bypass x N}
  /// sao_eo_class_luma:       FL, cMax = 3, bins = {bypass x 3}
  /// sao_eo_class_chroma:     FL, cMax = 3, bins = {bypass x 3}
  if (sao->type == SAO_TYPE_BAND) {
    for (i = SAO_EO_CAT1; i <= SAO_EO_CAT4; ++i) {
      // Positive sign is coded as 0.
      if (sao->offsets[i + offset_index] != 0) {
        CABAC_BIN_EP(cabac, sao->offsets[i + offset_index] < 0 ? 1 : 0, "sao_offset_sign");
      }
    }
    // TODO: sao_band_position
    // FL cMax=31 (5 bits)
    CABAC_BINS_EP(cabac, sao->band_position[color_i == COLOR_V ? 1:0], 5, "sao_band_position");
  } else if (color_i != COLOR_V) {
    CABAC_BINS_EP(cabac, sao->eo_class, 2, "sao_eo_class");
  }
}

static void encode_sao_merge_flags(encoder_state_t * const state, sao_info_t *sao, unsigned x_ctb, unsigned y_ctb)
{
  cabac_data_t * const cabac = &state->cabac;
  // SAO merge flags are not present for the first row and column.
  if (x_ctb > 0) {
    cabac->cur_ctx = &(cabac->ctx.sao_merge_flag_model);
    CABAC_BIN(cabac, sao->merge_left_flag, "sao_merge_left_flag");
  }
  if (y_ctb > 0 && !sao->merge_left_flag) {
    cabac->cur_ctx = &(cabac->ctx.sao_merge_flag_model);
    CABAC_BIN(cabac, sao->merge_up_flag, "sao_merge_up_flag");
  }
}


/**
 * \brief Encode SAO information.
 */
static void encode_sao(encoder_state_t * const state,
                       unsigned x_lcu, uint16_t y_lcu,
                       sao_info_t *sao_luma, sao_info_t *sao_chroma)
{
  // TODO: transmit merge flags outside sao_info
  encode_sao_merge_flags(state, sao_luma, x_lcu, y_lcu);

  // If SAO is merged, nothing else needs to be coded.
  if (!sao_luma->merge_left_flag && !sao_luma->merge_up_flag) {
    encode_sao_color(state, sao_luma, COLOR_Y);
    if (state->encoder_control->chroma_format != KVZ_CSP_400) {
      encode_sao_color(state, sao_chroma, COLOR_U);
      encode_sao_color(state, sao_chroma, COLOR_V);
    }
  }
}


/**
 * \brief Sets the QP for each CU in state->tile->frame->cu_array.
 *
 * The QPs are used in deblocking and QP prediction.
 *
 * The QP delta for a quantization group is coded when the first CU with
 * coded block flag set is encountered. Hence, for the purposes of
 * deblocking and QP prediction, all CUs in before the first one that has
 * cbf set use the QP predictor and all CUs after that use (QP predictor
 * + QP delta).
 *
 * \param state           encoder state
 * \param x               x-coordinate of the left edge of the root CU
 * \param y               y-coordinate of the top edge of the root CU
 * \param depth           depth in the CU quadtree
 * \param last_qp         QP of the last CU in the last quantization group
 * \param prev_qp         -1 if QP delta has not been coded in current QG,
 *                        otherwise the QP of the current QG
 */
static void set_cu_qps(encoder_state_t *state, int x, int y, int depth, int *last_qp, int *prev_qp)
{

  // Stop recursion if the CU is completely outside the frame.
  if (x >= state->tile->frame->width || y >= state->tile->frame->height) return;

  cu_info_t *cu = kvz_cu_array_at(state->tile->frame->cu_array, x, y);
  const int cu_width = LCU_WIDTH >> depth;

  if (depth <= state->encoder_control->max_qp_delta_depth) {
    *prev_qp = -1;
  }

  if (cu->depth > depth) {
    // Recursively process sub-CUs.
    const int d = cu_width >> 1;
    set_cu_qps(state, x,     y,     depth + 1, last_qp, prev_qp);
    set_cu_qps(state, x + d, y,     depth + 1, last_qp, prev_qp);
    set_cu_qps(state, x,     y + d, depth + 1, last_qp, prev_qp);
    set_cu_qps(state, x + d, y + d, depth + 1, last_qp, prev_qp);

  } else {
    bool cbf_found = *prev_qp >= 0;

    if (cu->tr_depth > depth) {
      // The CU is split into smaller transform units. Check whether coded
      // block flag is set for any of the TUs.
      const int tu_width = LCU_WIDTH >> cu->tr_depth;
      for (int y_scu = y; !cbf_found && y_scu < y + cu_width; y_scu += tu_width) {
        for (int x_scu = x; !cbf_found && x_scu < x + cu_width; x_scu += tu_width) {
          cu_info_t *tu = kvz_cu_array_at(state->tile->frame->cu_array, x_scu, y_scu);
          if (cbf_is_set_any(tu->cbf, cu->depth)) {
            cbf_found = true;
          }
        }
      }
    } else if (cbf_is_set_any(cu->cbf, cu->depth)) {
      cbf_found = true;
    }

    int8_t qp;
    if (cbf_found) {
      *prev_qp = qp = cu->qp;
    } else {
      qp = kvz_get_cu_ref_qp(state, x, y, *last_qp);
    }

    // Set the correct QP for all state->tile->frame->cu_array elements in
    // the area covered by the CU.
    for (int y_scu = y; y_scu < y + cu_width; y_scu += SCU_WIDTH) {
      for (int x_scu = x; x_scu < x + cu_width; x_scu += SCU_WIDTH) {
        kvz_cu_array_at(state->tile->frame->cu_array, x_scu, y_scu)->qp = qp;
      }
    }

    if (is_last_cu_in_qg(state, x, y, depth)) {
      *last_qp = cu->qp;
    }
  }
}


static void encoder_state_worker_encode_lcu(void * opaque)
{
  const lcu_order_element_t * const lcu = opaque;
  encoder_state_t *state = lcu->encoder_state;
  const encoder_control_t * const encoder = state->encoder_control;
  videoframe_t* const frame = state->tile->frame;

  kvz_set_lcu_lambda_and_qp(state, lcu->position);

  lcu_coeff_t coeff;
  state->coeff = &coeff;

  //This part doesn't write to bitstream, it's only search, deblock and sao
  kvz_search_lcu(state, lcu->position_px.x, lcu->position_px.y, state->tile->hor_buf_search, state->tile->ver_buf_search);

  encoder_state_recdata_to_bufs(state, lcu, state->tile->hor_buf_search, state->tile->ver_buf_search);

  if (encoder->max_qp_delta_depth >= 0) {
    int last_qp = state->last_qp;
    int prev_qp = -1;
    set_cu_qps(state, lcu->position_px.x, lcu->position_px.y, 0, &last_qp, &prev_qp);
  }

  if (encoder->cfg.deblock_enable) {
    kvz_filter_deblock_lcu(state, lcu->position_px.x, lcu->position_px.y);
  }

  if (encoder->cfg.sao_type) {
    // Save the post-deblocking but pre-SAO pixels of the LCU to a buffer
    // so that they can be used in SAO reconstruction later.
    encoder_state_recdata_before_sao_to_bufs(state,
                                             lcu,
                                             state->tile->hor_buf_before_sao,
                                             state->tile->ver_buf_before_sao);
    kvz_sao_search_lcu(state, lcu->position.x, lcu->position.y);
    encoder_sao_reconstruct(state, lcu);
  }

  //Now write data to bitstream (required to have a correct CABAC state)
  const uint64_t existing_bits = kvz_bitstream_tell(&state->stream);

  //Encode SAO
  if (encoder->cfg.sao_type) {
    encode_sao(state, lcu->position.x, lcu->position.y, &frame->sao_luma[lcu->position.y * frame->width_in_lcu + lcu->position.x], &frame->sao_chroma[lcu->position.y * frame->width_in_lcu + lcu->position.x]);
  }

  //Encode coding tree
  kvz_encode_coding_tree(state, lcu->position.x * LCU_WIDTH, lcu->position.y * LCU_WIDTH, 0);

  // Coeffs are not needed anymore.
  state->coeff = NULL;

  bool end_of_slice_segment_flag;
  if (state->encoder_control->cfg.slices & KVZ_SLICES_WPP) {
    // Slice segments end after each WPP row.
    end_of_slice_segment_flag = lcu->last_column;
  } else if (state->encoder_control->cfg.slices & KVZ_SLICES_TILES) {
    // Slices end after each tile.
    end_of_slice_segment_flag = lcu->last_column && lcu->last_row;
  } else {
    // Slice ends after the last row of the last tile.
    int last_tile_id = -1 + encoder->cfg.tiles_width_count * encoder->cfg.tiles_height_count;
    bool is_last_tile = state->tile->id == last_tile_id;
    end_of_slice_segment_flag = is_last_tile && lcu->last_column && lcu->last_row;
  }
  kvz_cabac_encode_bin_trm(&state->cabac, end_of_slice_segment_flag);

  {
    const bool end_of_tile = lcu->last_column && lcu->last_row;
    const bool end_of_wpp_row = encoder->cfg.wpp && lcu->last_column;


    if (end_of_tile || end_of_wpp_row) {
      if (!end_of_slice_segment_flag) {
        // end_of_sub_stream_one_bit
        kvz_cabac_encode_bin_trm(&state->cabac, 1);
      }

      // Finish the substream by writing out remaining state.
      kvz_cabac_finish(&state->cabac);

      // Write a rbsp_trailing_bits or a byte_alignment. The first one is used
      // for ending a slice_segment_layer_rbsp and the second one for ending
      // a substream. They are identical and align the byte stream.
      kvz_bitstream_put(state->cabac.stream, 1, 1);
      kvz_bitstream_align_zero(state->cabac.stream);

      kvz_cabac_start(&state->cabac);

      kvz_crypto_delete(&state->crypto_hdl);
    }
  }

  const uint32_t bits = kvz_bitstream_tell(&state->stream) - existing_bits;
  kvz_get_lcu_stats(state, lcu->position.x, lcu->position.y)->bits = bits;

  //Wavefronts need the context to be copied to the next row
  if (state->type == ENCODER_STATE_TYPE_WAVEFRONT_ROW && lcu->index == 1) {
    int j;
    //Find next encoder (next row)
    for (j=0; state->parent->children[j].encoder_control; ++j) {
      if (state->parent->children[j].wfrow->lcu_offset_y == state->wfrow->lcu_offset_y + 1) {
        //And copy context
        kvz_context_copy(&state->parent->children[j], state);
      }
    }
  }
}

static void encoder_state_encode_leaf(encoder_state_t * const state)
{
  assert(state->is_leaf);
  assert(state->lcu_order_count > 0);

  const encoder_control_t *ctrl = state->encoder_control;
  const kvz_config *cfg = &ctrl->cfg;

  // Signaled slice QP may be different to frame QP with set-qp-in-cu enabled.
  state->last_qp = ctrl->cfg.set_qp_in_cu ? 26 : state->frame->QP;

  if (cfg->crypto_features) {
    state->crypto_hdl = kvz_crypto_create(cfg);
    state->crypto_prev_pos = 0;
  }

  // Select whether to encode the frame/tile in current thread or to define
  // wavefront jobs for other threads to handle.
  bool wavefront = state->type == ENCODER_STATE_TYPE_WAVEFRONT_ROW;
  bool use_parallel_encoding = (wavefront && state->parent->children[1].encoder_control);
  if (!use_parallel_encoding) {
    // Encode every LCU in order and perform SAO reconstruction after every
    // frame is encoded. Deblocking and SAO search is done during LCU encoding.

    for (int i = 0; i < state->lcu_order_count; ++i) {
      encoder_state_worker_encode_lcu(&state->lcu_order[i]);
    }
  } else {
    // Add each LCU in the wavefront row as it's own job to the queue.

    // Select which frame dependancies should be set to.
    const encoder_state_t * ref_state = NULL;

    if (state->frame->slicetype == KVZ_SLICE_I) {
      // I-frames have no references.
      ref_state = NULL;
    } else if (cfg->gop_lowdelay &&
               cfg->gop_len > 0 &&
               state->previous_encoder_state != state)
    {
      // For LP-gop, depend on the state of the first reference.
      int ref_neg = cfg->gop[state->frame->gop_offset].ref_neg[0];
      if (ref_neg > cfg->owf) {
        // If frame is not within OWF range, it's already done.
        ref_state = NULL;
      } else {
        ref_state = state->previous_encoder_state;
        while (ref_neg > 1) {
          ref_neg -= 1;
          ref_state = ref_state->previous_encoder_state;
        }
      }
    } else {
      // Otherwise, depend on the previous frame.
      ref_state = state->previous_encoder_state;
    }

    for (int i = 0; i < state->lcu_order_count; ++i) {
      const lcu_order_element_t * const lcu = &state->lcu_order[i];

      kvz_threadqueue_free_job(&state->tile->wf_jobs[lcu->id]);
      state->tile->wf_jobs[lcu->id] = kvz_threadqueue_job_create(encoder_state_worker_encode_lcu, (void*)lcu);
      threadqueue_job_t **job = &state->tile->wf_jobs[lcu->id];

      // If job object was returned, add dependancies and allow it to run.
      if (job[0]) {
        // Add inter frame dependancies when ecoding more than one frame at
        // once. The added dependancy is for the first LCU of each wavefront
        // row to depend on the reconstruction status of the row below in the
        // previous frame.
        if (ref_state != NULL &&
            state->previous_encoder_state->tqj_recon_done &&
            state->frame->slicetype != KVZ_SLICE_I)
        {
          // We need to wait until the CTUs whose pixels we refer to are
          // done before we can start this CTU.
          const lcu_order_element_t *dep_lcu = lcu;
          for (int i = 0; dep_lcu->below && i < ctrl->max_inter_ref_lcu.down; i++) {
            dep_lcu = dep_lcu->below;
          }
          for (int i = 0; dep_lcu->right && i < ctrl->max_inter_ref_lcu.right; i++) {
            dep_lcu = dep_lcu->right;
          }
          kvz_threadqueue_job_dep_add(job[0], ref_state->tile->wf_jobs[dep_lcu->id]);

          // Very spesific bug that happens when owf length is longer than the
          // gop length. Takes care of that.
          if(!state->encoder_control->cfg.gop_lowdelay &&
             state->encoder_control->cfg.open_gop &&
             state->encoder_control->cfg.gop_len != 0 &&
             state->encoder_control->cfg.owf > state->encoder_control->cfg.gop_len &&
             ref_state->frame->slicetype == KVZ_SLICE_I &&
             ref_state->frame->num != 0){

            while (ref_state->frame->poc != state->frame->poc - state->encoder_control->cfg.gop_len){
              ref_state = ref_state->previous_encoder_state;
            }
            kvz_threadqueue_job_dep_add(job[0], ref_state->tile->wf_jobs[dep_lcu->id]);
          }
        }

        // Add local WPP dependancy to the LCU on the left.
        if (lcu->left) {
          kvz_threadqueue_job_dep_add(job[0], job[-1]);
        }
        // Add local WPP dependancy to the LCU on the top right.
        if (lcu->above) {
          if (lcu->above->right) {
            kvz_threadqueue_job_dep_add(job[0], job[-state->tile->frame->width_in_lcu + 1]);
          } else {
            kvz_threadqueue_job_dep_add(job[0], job[-state->tile->frame->width_in_lcu]);
          }
        }

        kvz_threadqueue_submit(state->encoder_control->threadqueue, state->tile->wf_jobs[lcu->id]);

        // The wavefront row is done when the last LCU in the row is done.
        if (i + 1 == state->lcu_order_count) {
          assert(!state->tqj_recon_done);
          state->tqj_recon_done =
            kvz_threadqueue_copy_ref(state->tile->wf_jobs[lcu->id]);
        }
      }
    }
  }
}

static void encoder_state_encode(encoder_state_t * const main_state);

static void encoder_state_worker_encode_children(void * opaque)
{
  encoder_state_t *sub_state = opaque;
  encoder_state_encode(sub_state);

  if (sub_state->is_leaf && sub_state->type == ENCODER_STATE_TYPE_WAVEFRONT_ROW) {
    // Set the last wavefront job of this row as the job that completes
    // the bitstream for this wavefront row state.

    int wpp_row = sub_state->wfrow->lcu_offset_y;
    int tile_width = sub_state->tile->frame->width_in_lcu;
    int end_of_row = (wpp_row + 1) * tile_width - 1;
    assert(!sub_state->tqj_bitstream_written);
    if (sub_state->tile->wf_jobs[end_of_row]) {
      sub_state->tqj_bitstream_written =
        kvz_threadqueue_copy_ref(sub_state->tile->wf_jobs[end_of_row]);
    }
  }
}

static int encoder_state_tree_is_a_chain(const encoder_state_t * const state) {
  if (!state->children[0].encoder_control) return 1;
  if (state->children[1].encoder_control) return 0;
  return encoder_state_tree_is_a_chain(&state->children[0]);
}

static void encoder_state_encode(encoder_state_t * const main_state) {
  //If we have children, encode at child level
  if (main_state->children[0].encoder_control) {
    //If we have only one child, than it cannot be the last split in tree
    int node_is_the_last_split_in_tree = (main_state->children[1].encoder_control != 0);

    for (int i = 0; main_state->children[i].encoder_control; ++i) {
      encoder_state_t *sub_state = &(main_state->children[i]);

      if (sub_state->tile != main_state->tile) {
        const int offset_x = sub_state->tile->offset_x;
        const int offset_y = sub_state->tile->offset_y;
        const int width = MIN(sub_state->tile->frame->width_in_lcu * LCU_WIDTH, main_state->tile->frame->width - offset_x);
        const int height = MIN(sub_state->tile->frame->height_in_lcu * LCU_WIDTH, main_state->tile->frame->height - offset_y);

        kvz_image_free(sub_state->tile->frame->source);
        sub_state->tile->frame->source = NULL;

        kvz_image_free(sub_state->tile->frame->rec);
        sub_state->tile->frame->rec = NULL;

        kvz_cu_array_free(&sub_state->tile->frame->cu_array);

        sub_state->tile->frame->source = kvz_image_make_subimage(
            main_state->tile->frame->source,
            offset_x,
            offset_y,
            width,
            height
        );
        sub_state->tile->frame->rec = kvz_image_make_subimage(
            main_state->tile->frame->rec,
            offset_x,
            offset_y,
            width,
            height
        );
        sub_state->tile->frame->cu_array = kvz_cu_subarray(
            main_state->tile->frame->cu_array,
            offset_x,
            offset_y,
            sub_state->tile->frame->width_in_lcu * LCU_WIDTH,
            sub_state->tile->frame->height_in_lcu * LCU_WIDTH
        );
      }

      //To be the last split, we require that every child is a chain
      node_is_the_last_split_in_tree =
        node_is_the_last_split_in_tree &&
        encoder_state_tree_is_a_chain(&main_state->children[i]);
    }
    //If it's the latest split point
    if (node_is_the_last_split_in_tree) {
      for (int i = 0; main_state->children[i].encoder_control; ++i) {
        //If we don't have wavefronts, parallelize encoding of children.
        if (main_state->children[i].type != ENCODER_STATE_TYPE_WAVEFRONT_ROW) {
          kvz_threadqueue_free_job(&main_state->children[i].tqj_recon_done);
          main_state->children[i].tqj_recon_done =
            kvz_threadqueue_job_create(encoder_state_worker_encode_children, &main_state->children[i]);
          if (main_state->children[i].previous_encoder_state != &main_state->children[i] &&
              main_state->children[i].previous_encoder_state->tqj_recon_done &&
              !main_state->children[i].frame->is_irap)
          {
#if 0
            // Disabled due to non-determinism.
            if (main_state->encoder_control->cfg->mv_constraint == KVZ_MV_CONSTRAIN_FRAME_AND_TILE_MARGIN)
            {
              // When MV's don't cross tile boundaries, add dependancy only to the same tile.
              kvz_threadqueue_job_dep_add(main_state->children[i].tqj_recon_done, main_state->children[i].previous_encoder_state->tqj_recon_done);
            } else 
#endif      
            {
              // Add dependancy to each child in the previous frame.
              for (int child_id = 0; main_state->children[child_id].encoder_control; ++child_id) {
                kvz_threadqueue_job_dep_add(main_state->children[i].tqj_recon_done, main_state->children[child_id].previous_encoder_state->tqj_recon_done);
              }
            }
          }
          kvz_threadqueue_submit(main_state->encoder_control->threadqueue, main_state->children[i].tqj_recon_done);
        } else {
          //Wavefront rows have parallelism at LCU level, so we should not launch multiple threads here!
          //FIXME: add an assert: we can only have wavefront children
          encoder_state_worker_encode_children(&(main_state->children[i]));
        }
      }
    } else {
      for (int i = 0; main_state->children[i].encoder_control; ++i) {
        encoder_state_worker_encode_children(&(main_state->children[i]));
      }
    }
  } else {
    switch (main_state->type) {
      case ENCODER_STATE_TYPE_TILE:
      case ENCODER_STATE_TYPE_SLICE:
      case ENCODER_STATE_TYPE_WAVEFRONT_ROW:
        encoder_state_encode_leaf(main_state);
        break;
      default:
        fprintf(stderr, "Unsupported leaf type %c!\n", main_state->type);
        assert(0);
    }
  }
}


static void encoder_ref_insertion_sort(const encoder_state_t *const state,
                                       uint8_t reflist[16],
                                       uint8_t length,
                                       bool reverse)
{

  for (uint8_t i = 1; i < length; ++i) {
    const uint8_t cur_idx = reflist[i];
    const int32_t cur_poc = state->frame->ref->pocs[cur_idx];
    int8_t j = i;
    while ((j > 0 && !reverse && cur_poc > state->frame->ref->pocs[reflist[j - 1]]) ||
           (j > 0 &&  reverse && cur_poc < state->frame->ref->pocs[reflist[j - 1]]))
    {
      reflist[j] = reflist[j - 1];
      --j;
    }
    reflist[j] = cur_idx;
  }
}

/**
 * \brief Generate reference picture lists.
 *
 * \param state             main encoder state
 */
void kvz_encoder_create_ref_lists(const encoder_state_t *const state)
{
  const kvz_config *cfg = &state->encoder_control->cfg;

  FILL_ARRAY(state->frame->ref_LX_size, 0, 2);

  int num_negative = 0;
  int num_positive = 0;

  // Add positive references to L1 list
  for (int i = 0; i < state->frame->ref->used_size; i++) {
    if (state->frame->ref->pocs[i] > state->frame->poc) {
      state->frame->ref_LX[1][state->frame->ref_LX_size[1]] = i;
      state->frame->ref_LX_size[1] += 1;
      num_positive++;
    }
  }

  // Add negative references to L1 list when bipred is enabled and GOP is
  // either disabled or does not use picture reordering.
  bool l1_negative_refs =
    (cfg->bipred && (cfg->gop_len == 0 || cfg->gop_lowdelay));

  // Add negative references to L0 and L1 lists.
  for (int i = 0; i < state->frame->ref->used_size; i++) {
    if (state->frame->ref->pocs[i] < state->frame->poc) {
      state->frame->ref_LX[0][state->frame->ref_LX_size[0]] = i;
      state->frame->ref_LX_size[0] += 1;
      if (l1_negative_refs) {
        state->frame->ref_LX[1][state->frame->ref_LX_size[1]] = i;
        state->frame->ref_LX_size[1] += 1;
      }
      num_negative++;
    }
  }

  // Fill the rest with -1.
  for (int i = state->frame->ref_LX_size[0]; i < 16; i++) {
    state->frame->ref_LX[0][i] = 0xff;
  }
  for (int i = state->frame->ref_LX_size[1]; i < 16; i++) {
    state->frame->ref_LX[1][i] = 0xff;
  }

  // Sort reference lists.
  encoder_ref_insertion_sort(state, state->frame->ref_LX[0], num_negative, false);
  encoder_ref_insertion_sort(state, state->frame->ref_LX[1], num_positive, true);
  if (l1_negative_refs) {
    encoder_ref_insertion_sort(state, state->frame->ref_LX[1] + num_positive, num_negative, false);
  }
}

/**
 * \brief Remove any references that should no longer be used.
 */
static void encoder_state_remove_refs(encoder_state_t *state) {
  const encoder_control_t * const encoder = state->encoder_control;

  int neg_refs = encoder->cfg.gop[state->frame->gop_offset].ref_neg_count;
  int pos_refs = encoder->cfg.gop[state->frame->gop_offset].ref_pos_count;

  unsigned target_ref_num;
  if (encoder->cfg.gop_len) {
    target_ref_num = neg_refs + pos_refs;
  } else {
    target_ref_num = encoder->cfg.ref_frames;
  }

  if (state->frame->pictype == KVZ_NAL_IDR_W_RADL ||
      state->frame->pictype == KVZ_NAL_IDR_N_LP)
  {
    target_ref_num = 0;
  }

  if (encoder->cfg.gop_len && target_ref_num > 0) {
    // With GOP in use, go through all the existing reference pictures and
    // remove any picture that is not referenced by the current picture.

    for (int ref = state->frame->ref->used_size - 1; ref >= 0; --ref) {
      bool is_referenced = false;

      int ref_poc = state->frame->ref->pocs[ref];

      for (int i = 0; i < neg_refs; i++) {
        int ref_relative_poc = -encoder->cfg.gop[state->frame->gop_offset].ref_neg[i];
        if (ref_poc == state->frame->poc + ref_relative_poc) {
          is_referenced = true;
          break;
        }
      }

      for (int i = 0; i < pos_refs; i++) {
        int ref_relative_poc = encoder->cfg.gop[state->frame->gop_offset].ref_pos[i];
        if (ref_poc == state->frame->poc + ref_relative_poc) {
          is_referenced = true;
          break;
        }
      }

      if (ref_poc < state->frame->irap_poc &&
          state->frame->irap_poc < state->frame->poc)
      {
        // Trailing frames cannot refer to leading frames.
        is_referenced = false;
      }

      if (encoder->cfg.intra_period > 0 &&
          ref_poc < state->frame->irap_poc - encoder->cfg.intra_period)
      {
        // No frame can refer past the two preceding IRAP frames.
        is_referenced = false;
      }

      if (!is_referenced) {
        // This reference is not referred to by this frame, it must be removed.
        kvz_image_list_rem(state->frame->ref, ref);
      }
    }
  } else {
    // Without GOP, remove the oldest picture.
    while (state->frame->ref->used_size > target_ref_num) {
      int8_t oldest_ref = state->frame->ref->used_size - 1;
      kvz_image_list_rem(state->frame->ref, oldest_ref);
    }
  }

  assert(state->frame->ref->used_size <= target_ref_num);
}

static void encoder_set_source_picture(encoder_state_t * const state, kvz_picture* frame)
{
  assert(!state->tile->frame->source);
  assert(!state->tile->frame->rec);

  state->tile->frame->source = frame;
  if (state->encoder_control->cfg.lossless) {
    // In lossless mode, the reconstruction is equal to the source frame.
    state->tile->frame->rec = kvz_image_copy_ref(frame);
  } else {
    state->tile->frame->rec = kvz_image_alloc(state->encoder_control->chroma_format, frame->width, frame->height);
    state->tile->frame->rec->dts = frame->dts;
    state->tile->frame->rec->pts = frame->pts;
  }

  kvz_videoframe_set_poc(state->tile->frame, state->frame->poc);
}

static void encoder_state_init_children(encoder_state_t * const state) {
  kvz_bitstream_clear(&state->stream);

  if (state->is_leaf) {
    //Leaf states have cabac and context
    kvz_cabac_start(&state->cabac);
    kvz_init_contexts(state, state->encoder_control->cfg.set_qp_in_cu ? 26 : state->frame->QP, state->frame->slicetype);
  }

  //Clear the jobs
  kvz_threadqueue_free_job(&state->tqj_bitstream_written);
  kvz_threadqueue_free_job(&state->tqj_recon_done);

  //Copy the constraint pointer
  // TODO: Try to do it in the if (state->is_leaf)
  //if (state->parent != NULL) {
    // state->constraint = state->parent->constraint;
  //}

  for (int i = 0; state->children[i].encoder_control; ++i) {
    encoder_state_init_children(&state->children[i]);
  }
}

static void normalize_lcu_weights(encoder_state_t * const state)
{
  if (state->frame->num == 0) return;

  const uint32_t num_lcus = state->encoder_control->in.width_in_lcu *
                            state->encoder_control->in.height_in_lcu;
  double sum = 0.0;
  for (uint32_t i = 0; i < num_lcus; i++) {
    sum += state->frame->lcu_stats[i].weight;
  }

  for (uint32_t i = 0; i < num_lcus; i++) {
    state->frame->lcu_stats[i].weight /= sum;
  }
}

// Calculate pixel value variance. Takes in arrays of kvz_pixel
static double pixel_var(kvz_pixel * const arr, const uint32_t len) {
  double var = 0;
  double arr_mean = 0;

  // Calculate array mean
  int i = 0;
  double sum = 0;

  for (; i < len; ++i) {
    sum += arr[i];
  }
  arr_mean = sum / (double)len;

  // Calculate array variance  
  for (i = 0; i < len; ++i) {
    double tmp = (double)arr[i] - arr_mean;
    var += tmp*tmp;
  }

  var /= len;

  return var;
}

static void encoder_state_init_new_frame(encoder_state_t * const state, kvz_picture* frame) {
  assert(state->type == ENCODER_STATE_TYPE_MAIN);

  const kvz_config * const cfg = &state->encoder_control->cfg;

  encoder_set_source_picture(state, frame);

  assert(!state->tile->frame->cu_array);
  state->tile->frame->cu_array = kvz_cu_array_alloc(
      state->tile->frame->width,
      state->tile->frame->height
  );

  // Variance adaptive quantization
  if (cfg->vaq) {
    double d = 1.5; // Empirically decided constant. Affects delta-QP strength
    
    // Calculate frame pixel variance
    uint32_t len = state->tile->frame->width * state->tile->frame->height;
    double frame_var = pixel_var(state->tile->frame->source->y, len);

    // Loop through LCUs
    // For each LCU calculate: D * (log(LCU pixel variance) - log(frame pixel variance))
    int x = 0;
    int y = 0;
    unsigned x_lim = state->tile->frame->width_in_lcu;
    unsigned y_lim = state->tile->frame->height_in_lcu;
    
    unsigned id = 0;
    for (; y < y_lim; ++y) {
      for (; x < x_lim; ++x) {
        kvz_pixel tmp[LCU_LUMA_SIZE];
        int x_max = MIN(x + LCU_WIDTH, frame->width) - x;
        int y_max = MIN(y + LCU_WIDTH, frame->height) - y;
        // blit pixel array
        kvz_pixels_blit(&state->tile->frame->source->y[x + y * state->tile->frame->source->stride], tmp,
          x_max, y_max, state->tile->frame->source->stride, LCU_WIDTH);
        
        double lcu_var = pixel_var(tmp, LCU_LUMA_SIZE);
        state->frame->aq_offsets[id] = d * (log(lcu_var) - log(frame_var));
        id++; 
      }
    }
  }
  // Variance adaptive quantization - END

  // Use this flag to handle closed gop irap picture selection.
  // If set to true, irap is already set and we avoid
  // setting it based on the intra period
  bool is_closed_normal_gop = false;

  // Set POC.
  if (state->frame->num == 0) {
    state->frame->poc = 0;
  } else if (cfg->gop_len && !cfg->gop_lowdelay) {

    int32_t framenum = state->frame->num - 1;
    // Handle closed GOP
    // Closed GOP structure has an extra IDR between the GOPs
    if (cfg->intra_period > 0 && !cfg->open_gop) {
      is_closed_normal_gop = true;
      if (framenum % (cfg->intra_period + 1) == cfg->intra_period) {
        // Insert IDR before each new GOP after intra period in closed GOP configuration
        state->frame->poc = 0;
      } else {
        // Calculate frame number again and use that for the POC
        framenum = framenum % (cfg->intra_period + 1);
        int32_t poc_offset = cfg->gop[state->frame->gop_offset].poc_offset;
        state->frame->poc = framenum - framenum % cfg->gop_len + poc_offset;
        // This should not be an irap picture in closed GOP
        state->frame->is_irap = false;
      }
    } else { // Open GOP
      // Calculate POC according to the global frame counter and GOP structure
      int32_t poc_offset = cfg->gop[state->frame->gop_offset].poc_offset;
      state->frame->poc = framenum - framenum % cfg->gop_len + poc_offset;
    }
    
    kvz_videoframe_set_poc(state->tile->frame, state->frame->poc);
  } else if (cfg->intra_period > 0) {
    state->frame->poc = state->frame->num % cfg->intra_period;
  } else {
    state->frame->poc = state->frame->num;
  }

  // Check whether the frame is a keyframe or not.
  if (state->frame->num == 0 || state->frame->poc == 0) {
    state->frame->is_irap = true;
  } else if(!is_closed_normal_gop) { // In closed-GOP IDR frames are poc==0 so skip this check
    state->frame->is_irap =
      cfg->intra_period > 0 &&
      (state->frame->poc % cfg->intra_period) == 0;
  }
  if (state->frame->is_irap) {
    state->frame->irap_poc = state->frame->poc;
  }

  // Set pictype.
  if (state->frame->is_irap) {
    if (state->frame->num == 0 ||
        cfg->intra_period == 1 ||
        cfg->gop_len == 0 ||
        cfg->gop_lowdelay ||
        !cfg->open_gop) // Closed GOP uses IDR pictures
    {
      state->frame->pictype = KVZ_NAL_IDR_W_RADL;
    } else {
      state->frame->pictype = KVZ_NAL_CRA_NUT;
    }
  } else if (state->frame->poc < state->frame->irap_poc) {
    state->frame->pictype = KVZ_NAL_RASL_R;
  } else {
    state->frame->pictype = KVZ_NAL_TRAIL_R;
  }

  encoder_state_remove_refs(state);
  kvz_encoder_create_ref_lists(state);

  // Set slicetype.
  if (state->frame->is_irap) {
    state->frame->slicetype = KVZ_SLICE_I;
  } else if (state->frame->ref_LX_size[1] > 0) {
    state->frame->slicetype = KVZ_SLICE_B;
  } else {
    state->frame->slicetype = KVZ_SLICE_P;
  }

  if (cfg->target_bitrate > 0 && state->frame->num > cfg->owf) {
    normalize_lcu_weights(state);
  }
  kvz_set_picture_lambda_and_qp(state);

  encoder_state_init_children(state);
}

static void _encode_one_frame_add_bitstream_deps(const encoder_state_t * const state, threadqueue_job_t * const job) {
  int i;
  for (i = 0; state->children[i].encoder_control; ++i) {
    _encode_one_frame_add_bitstream_deps(&state->children[i], job);
  }
  if (state->tqj_bitstream_written) {
    kvz_threadqueue_job_dep_add(job, state->tqj_bitstream_written);
  }
  if (state->tqj_recon_done) {
    kvz_threadqueue_job_dep_add(job, state->tqj_recon_done);
  }
}


void kvz_encode_one_frame(encoder_state_t * const state, kvz_picture* frame)
{
  encoder_state_init_new_frame(state, frame);
  encoder_state_encode(state);

  threadqueue_job_t *job =
    kvz_threadqueue_job_create(kvz_encoder_state_worker_write_bitstream, state);

  _encode_one_frame_add_bitstream_deps(state, job);
  if (state->previous_encoder_state != state && state->previous_encoder_state->tqj_bitstream_written) {
    //We need to depend on previous bitstream generation
    kvz_threadqueue_job_dep_add(job, state->previous_encoder_state->tqj_bitstream_written);
  }
  kvz_threadqueue_submit(state->encoder_control->threadqueue, job);
  assert(!state->tqj_bitstream_written);
  state->tqj_bitstream_written = job;

  state->frame->done = 0;
}


/**
 * Prepare the encoder state for encoding the next frame.
 *
 * - Add the previous reconstructed picture as a reference, if needed.
 * - Free the previous reconstructed and source pictures.
 * - Create a new cu array, if needed.
 * - Update frame count and POC.
 */
void kvz_encoder_prepare(encoder_state_t *state)
{
  const encoder_control_t * const encoder = state->encoder_control;

  // The previous frame must be done before the next one is started.
  assert(state->frame->done);

  if (state->frame->num == -1) {
    // We're at the first frame, so don't care about all this stuff.
    state->frame->num = 0;
    state->frame->poc = 0;
    state->frame->irap_poc = 0;
    assert(!state->tile->frame->source);
    assert(!state->tile->frame->rec);
    assert(!state->tile->frame->cu_array);
    state->frame->prepared = 1;

    return;
  }

  // NOTE: prev_state is equal to state when OWF is zero
  encoder_state_t *prev_state = state->previous_encoder_state;

  if (state->previous_encoder_state != state) {
    kvz_cu_array_free(&state->tile->frame->cu_array);
    unsigned width  = state->tile->frame->width_in_lcu  * LCU_WIDTH;
    unsigned height = state->tile->frame->height_in_lcu * LCU_WIDTH;
    state->tile->frame->cu_array = kvz_cu_array_alloc(width, height);

    kvz_image_list_copy_contents(state->frame->ref, prev_state->frame->ref);
    kvz_encoder_create_ref_lists(state);
  }

  if (!encoder->cfg.gop_len ||
      !prev_state->frame->poc ||
      encoder->cfg.gop[prev_state->frame->gop_offset].is_ref) {

    // Store current list of POCs for use in TMVP derivation
    memcpy(prev_state->tile->frame->rec->ref_pocs, state->frame->ref->pocs, sizeof(int32_t)*state->frame->ref->used_size);

    // Add previous reconstructed picture as a reference
    kvz_image_list_add(state->frame->ref,
                   prev_state->tile->frame->rec,
                   prev_state->tile->frame->cu_array,
                   prev_state->frame->poc,
                   prev_state->frame->ref_LX);
    kvz_cu_array_free(&state->tile->frame->cu_array);
    unsigned height = state->tile->frame->height_in_lcu * LCU_WIDTH;
    unsigned width  = state->tile->frame->width_in_lcu  * LCU_WIDTH;
    state->tile->frame->cu_array = kvz_cu_array_alloc(width, height);
  }

  // Remove source and reconstructed picture.
  kvz_image_free(state->tile->frame->source);
  state->tile->frame->source = NULL;

  kvz_image_free(state->tile->frame->rec);
  state->tile->frame->rec = NULL;

  kvz_cu_array_free(&state->tile->frame->cu_array);

  // Update POC and frame count.
  state->frame->num = prev_state->frame->num + 1;
  state->frame->poc = prev_state->frame->poc + 1;
  state->frame->irap_poc = prev_state->frame->irap_poc;

  state->frame->prepared = 1;


}

coeff_scan_order_t kvz_get_scan_order(int8_t cu_type, int intra_mode, int depth)
{
  // Scan mode is diagonal, except for 4x4+8x8 luma and 4x4 chroma, where:
  // - angular 6-14 = vertical
  // - angular 22-30 = horizontal
  if (cu_type == CU_INTRA && depth >= 3) {
    if (intra_mode >= 6 && intra_mode <= 14) {
      return SCAN_VER;
    } else if (intra_mode >= 22 && intra_mode <= 30) {
      return SCAN_HOR;
    }
  }

  return SCAN_DIAG;
}

lcu_stats_t* kvz_get_lcu_stats(encoder_state_t *state, int lcu_x, int lcu_y)
{
  const int index = lcu_x + state->tile->lcu_offset_x +
                    (lcu_y + state->tile->lcu_offset_y) *
                    state->encoder_control->in.width_in_lcu;
  return &state->frame->lcu_stats[index];
}

int kvz_get_cu_ref_qp(const encoder_state_t *state, int x, int y, int last_qp)
{
  const encoder_control_t *ctrl = state->encoder_control;
  const cu_array_t *cua = state->tile->frame->cu_array;
  // Quantization group width
  const int qg_width = LCU_WIDTH >> MIN(ctrl->max_qp_delta_depth, kvz_cu_array_at_const(cua, x, y)->depth);

  // Coordinates of the top-left corner of the quantization group
  const int x_qg = x & ~(qg_width - 1);
  const int y_qg = y & ~(qg_width - 1);

  int qp_pred_a = last_qp;
  if (x_qg % LCU_WIDTH > 0) {
    qp_pred_a = kvz_cu_array_at_const(cua, x_qg - 1, y_qg)->qp;
  }

  int qp_pred_b = last_qp;
  if (y_qg % LCU_WIDTH > 0) {
    qp_pred_b = kvz_cu_array_at_const(cua, x_qg, y_qg - 1)->qp;
  }

  return ((qp_pred_a + qp_pred_b + 1) >> 1);
}