av1/encoder/tune_vmaf.c

/*
 * Copyright (c) 2021, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 3-Clause Clear License
 * and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear
 * License was not distributed with this source code in the LICENSE file, you
 * can obtain it at aomedia.org/license/software-license/bsd-3-c-c/.  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
 * aomedia.org/license/patent-license/.
 */

#include "av1/encoder/tune_vmaf.h"

#include "aom_dsp/psnr.h"
#include "aom_ports/system_state.h"
#include "av1/encoder/extend.h"
#include "av1/encoder/rdopt.h"
#if CONFIG_USE_VMAF_RC
#include "config/aom_scale_rtcd.h"
#endif

static const double kBaselineVmaf = 97.42773;

static double get_layer_value(const double *array, int layer) {
  while (array[layer] < 0.0 && layer > 0) layer--;
  return AOMMAX(array[layer], 0.0);
}

static void motion_search(AV1_COMP *cpi, const YV12_BUFFER_CONFIG *src,
                          const YV12_BUFFER_CONFIG *ref,
                          const BLOCK_SIZE block_size, const int mb_row,
                          const int mb_col, FULLPEL_MV *ref_mv) {
  // Block information (ONLY Y-plane is used for motion search).
  const int mb_height = block_size_high[block_size];
  const int mb_width = block_size_wide[block_size];
  const int y_stride = src->y_stride;
  assert(y_stride == ref->y_stride);
  const int y_offset = mb_row * mb_height * y_stride + mb_col * mb_width;

  // Save input state.
  MACROBLOCK *const mb = &cpi->td.mb;
  MACROBLOCKD *const mbd = &mb->e_mbd;
  const struct buf_2d ori_src_buf = mb->plane[0].src;
  const struct buf_2d ori_pre_buf = mbd->plane[0].pre[0];

  // Parameters used for motion search.
  FULLPEL_MOTION_SEARCH_PARAMS full_ms_params;
  const SEARCH_METHODS search_method = NSTEP;
  const search_site_config *search_site_cfg =
      cpi->mv_search_params.search_site_cfg[SS_CFG_FPF];
  const int step_param =
      av1_init_search_range(AOMMAX(src->y_crop_width, src->y_crop_height));

  // Baseline position for motion search (used for rate distortion comparison).
  const MV baseline_mv = kZeroMv;

  // Setup.
  mb->plane[0].src.buf = src->y_buffer + y_offset;
  mb->plane[0].src.stride = y_stride;
  mbd->plane[0].pre[0].buf = ref->y_buffer + y_offset;
  mbd->plane[0].pre[0].stride = y_stride;

  // Unused intermediate results for motion search.
  int cost_list[5];

  // Do motion search.
  // Only do full search on the entire block.
  av1_make_default_fullpel_ms_params(&full_ms_params, cpi, mb, block_size,
                                     &baseline_mv, search_site_cfg,
                                     /*fine_search_interval=*/0);
  av1_set_mv_search_method(&full_ms_params, search_site_cfg, search_method);
  av1_full_pixel_search(*ref_mv, &full_ms_params, step_param,
                        cond_cost_list(cpi, cost_list), ref_mv, NULL);

  // Restore input state.
  mb->plane[0].src = ori_src_buf;
  mbd->plane[0].pre[0] = ori_pre_buf;
}

static unsigned int residual_variance(const AV1_COMP *cpi,
                                      const YV12_BUFFER_CONFIG *src,
                                      const YV12_BUFFER_CONFIG *ref,
                                      const BLOCK_SIZE block_size,
                                      const int mb_row, const int mb_col,
                                      FULLPEL_MV ref_mv, unsigned int *sse) {
  const int mb_height = block_size_high[block_size];
  const int mb_width = block_size_wide[block_size];
  const int y_stride = src->y_stride;
  assert(y_stride == ref->y_stride);
  const int y_offset = mb_row * mb_height * y_stride + mb_col * mb_width;
  const int mv_offset = ref_mv.row * y_stride + ref_mv.col;
  const unsigned int var =
      cpi->fn_ptr[block_size].vf(ref->y_buffer + y_offset + mv_offset, y_stride,
                                 src->y_buffer + y_offset, y_stride, sse);
  return var;
}

static double frame_average_variance(const AV1_COMP *const cpi,
                                     const YV12_BUFFER_CONFIG *const frame) {
  const uint16_t *const y_buffer = frame->y_buffer;
  const int y_stride = frame->y_stride;
  const BLOCK_SIZE block_size = BLOCK_64X64;

  const int block_w = mi_size_wide[block_size] * 4;
  const int block_h = mi_size_high[block_size] * 4;
  int row, col;
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  double var = 0.0, var_count = 0.0;

  // Loop through each block.
  for (row = 0; row < frame->y_height / block_h; ++row) {
    for (col = 0; col < frame->y_width / block_w; ++col) {
      struct buf_2d buf;
      const int row_offset_y = row * block_h;
      const int col_offset_y = col * block_w;

      buf.buf = y_buffer + row_offset_y * y_stride + col_offset_y;
      buf.stride = y_stride;

      var +=
          av1_high_get_sby_perpixel_variance(cpi, &buf, block_size, bit_depth);
      var_count += 1.0;
    }
  }
  var /= var_count;
  return var;
}

static double residual_frame_average_variance(AV1_COMP *cpi,
                                              const YV12_BUFFER_CONFIG *src,
                                              const YV12_BUFFER_CONFIG *ref,
                                              FULLPEL_MV *mvs) {
  if (ref == NULL) return frame_average_variance(cpi, src);
  const BLOCK_SIZE block_size = BLOCK_16X16;
  const int frame_height = src->y_height;
  const int frame_width = src->y_width;
  const int mb_height = block_size_high[block_size];
  const int mb_width = block_size_wide[block_size];
  const int mb_rows = (frame_height + mb_height - 1) / mb_height;
  const int mb_cols = (frame_width + mb_width - 1) / mb_width;
  const int num_planes = av1_num_planes(&cpi->common);
  const int mi_h = mi_size_high_log2[block_size];
  const int mi_w = mi_size_wide_log2[block_size];
  assert(num_planes >= 1 && num_planes <= MAX_MB_PLANE);

  // Save input state.
  MACROBLOCK *const mb = &cpi->td.mb;
  MACROBLOCKD *const mbd = &mb->e_mbd;
  uint16_t *input_buffer[MAX_MB_PLANE];
  for (int i = 0; i < num_planes; i++) {
    input_buffer[i] = mbd->plane[i].pre[0].buf;
  }
  MB_MODE_INFO **input_mb_mode_info = mbd->mi;

  bool do_motion_search = false;
  if (mvs == NULL) {
    do_motion_search = true;
    mvs = (FULLPEL_MV *)aom_malloc(sizeof(*mvs) * mb_rows * mb_cols);
    memset(mvs, 0, sizeof(*mvs) * mb_rows * mb_cols);
  }

  unsigned int variance = 0;
  // Perform temporal filtering block by block.
  for (int mb_row = 0; mb_row < mb_rows; mb_row++) {
    av1_set_mv_row_limits(&cpi->common.mi_params, &mb->mv_limits,
                          (mb_row << mi_h), (mb_height >> MI_SIZE_LOG2),
                          cpi->oxcf.border_in_pixels);
    for (int mb_col = 0; mb_col < mb_cols; mb_col++) {
      av1_set_mv_col_limits(&cpi->common.mi_params, &mb->mv_limits,
                            (mb_col << mi_w), (mb_width >> MI_SIZE_LOG2),
                            cpi->oxcf.border_in_pixels);
      FULLPEL_MV *ref_mv = &mvs[mb_col + mb_row * mb_cols];
      if (do_motion_search) {
        motion_search(cpi, src, ref, block_size, mb_row, mb_col, ref_mv);
      }
      unsigned int mv_sse;
      const unsigned int blk_var = residual_variance(
          cpi, src, ref, block_size, mb_row, mb_col, *ref_mv, &mv_sse);
      variance += blk_var;
    }
  }

  // Restore input state
  for (int i = 0; i < num_planes; i++) {
    mbd->plane[i].pre[0].buf = input_buffer[i];
  }
  mbd->mi = input_mb_mode_info;
  return (double)variance / (double)(mb_rows * mb_cols);
}

// TODO(sdeng): Add the SIMD implementation.
static AOM_INLINE void highbd_unsharp_rect(const uint16_t *source,
                                           int source_stride,
                                           const uint16_t *blurred,
                                           int blurred_stride, uint16_t *dst,
                                           int dst_stride, int w, int h,
                                           double amount, int bit_depth) {
  const int max_value = (1 << bit_depth) - 1;
  for (int i = 0; i < h; ++i) {
    for (int j = 0; j < w; ++j) {
      const double val =
          (double)source[j] + amount * ((double)source[j] - (double)blurred[j]);
      dst[j] = (uint16_t)clamp((int)(val + 0.5), 0, max_value);
    }
    source += source_stride;
    blurred += blurred_stride;
    dst += dst_stride;
  }
}

static AOM_INLINE void unsharp(const AV1_COMP *const cpi,
                               const YV12_BUFFER_CONFIG *source,
                               const YV12_BUFFER_CONFIG *blurred,
                               const YV12_BUFFER_CONFIG *dst, double amount) {
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  highbd_unsharp_rect(source->y_buffer, source->y_stride, blurred->y_buffer,
                      blurred->y_stride, dst->y_buffer, dst->y_stride,
                      source->y_width, source->y_height, amount, bit_depth);
}

// 8-tap Gaussian convolution filter with sigma = 1.0, sums to 128,
// all co-efficients must be even.
DECLARE_ALIGNED(16, static const int16_t, gauss_filter[8]) = { 0,  8, 30, 52,
                                                               30, 8, 0,  0 };
static AOM_INLINE void gaussian_blur(const int bit_depth,
                                     const YV12_BUFFER_CONFIG *source,
                                     const YV12_BUFFER_CONFIG *dst) {
  const int block_size = BLOCK_128X128;
  const int block_w = mi_size_wide[block_size] * 4;
  const int block_h = mi_size_high[block_size] * 4;
  const int num_cols = (source->y_width + block_w - 1) / block_w;
  const int num_rows = (source->y_height + block_h - 1) / block_h;
  int row, col;

  ConvolveParams conv_params = get_conv_params(0, 0, bit_depth);
  InterpFilterParams filter = { .filter_ptr = gauss_filter,
                                .taps = 8,
                                .interp_filter = EIGHTTAP_REGULAR };

  for (row = 0; row < num_rows; ++row) {
    for (col = 0; col < num_cols; ++col) {
      const int row_offset_y = row * block_h;
      const int col_offset_y = col * block_w;

      uint16_t *src_buf =
          source->y_buffer + row_offset_y * source->y_stride + col_offset_y;
      uint16_t *dst_buf =
          dst->y_buffer + row_offset_y * dst->y_stride + col_offset_y;

      av1_highbd_convolve_2d_sr(src_buf, source->y_stride, dst_buf,
                                dst->y_stride, block_w, block_h, &filter,
                                &filter, 0, 0, &conv_params, bit_depth);
    }
  }
}

static AOM_INLINE double cal_approx_vmaf(const AV1_COMP *const cpi,
#if CONFIG_USE_VMAF_RC
                                         VmafContext *vmaf_context,
                                         int *vmaf_cal_index,
#endif
                                         double source_variance,
                                         YV12_BUFFER_CONFIG *const source,
                                         YV12_BUFFER_CONFIG *const sharpened) {
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  double new_vmaf;

#if CONFIG_USE_VMAF_RC
  aom_calc_vmaf_at_index_rc(vmaf_context, cpi->vmaf_info.vmaf_model, source,
                            sharpened, bit_depth, *vmaf_cal_index, &new_vmaf);
  (*vmaf_cal_index)++;
#else
  aom_calc_vmaf(cpi->oxcf.tune_cfg.vmaf_model_path, source, sharpened,
                bit_depth, &new_vmaf);
#endif

  const double sharpened_var = frame_average_variance(cpi, sharpened);
  return source_variance / sharpened_var * (new_vmaf - kBaselineVmaf);
}

static double find_best_frame_unsharp_amount_loop(
    const AV1_COMP *const cpi,
#if CONFIG_USE_VMAF_RC
    VmafContext *vmaf_context, int *vmaf_cal_index,
#endif
    YV12_BUFFER_CONFIG *const source, YV12_BUFFER_CONFIG *const blurred,
    YV12_BUFFER_CONFIG *const sharpened, double best_vmaf,
    const double baseline_variance, const double unsharp_amount_start,
    const double step_size, const int max_loop_count, const double max_amount) {
  const double min_amount = 0.0;
  int loop_count = 0;
  double approx_vmaf = best_vmaf;
  double unsharp_amount = unsharp_amount_start;
  do {
    best_vmaf = approx_vmaf;
    unsharp_amount += step_size;
    if (unsharp_amount > max_amount || unsharp_amount < min_amount) break;
    unsharp(cpi, source, blurred, sharpened, unsharp_amount);
    approx_vmaf = cal_approx_vmaf(cpi,
#if CONFIG_USE_VMAF_RC
                                  vmaf_context, vmaf_cal_index,
#endif
                                  baseline_variance, source, sharpened);

    loop_count++;
  } while (approx_vmaf > best_vmaf && loop_count < max_loop_count);
  unsharp_amount =
      approx_vmaf > best_vmaf ? unsharp_amount : unsharp_amount - step_size;
  return AOMMIN(max_amount, AOMMAX(unsharp_amount, min_amount));
}

static double find_best_frame_unsharp_amount(const AV1_COMP *const cpi,
                                             YV12_BUFFER_CONFIG *const source,
                                             YV12_BUFFER_CONFIG *const blurred,
                                             const double unsharp_amount_start,
                                             const double step_size,
                                             const int max_loop_count,
                                             const double max_filter_amount) {
  const AV1_COMMON *const cm = &cpi->common;
  const int width = source->y_width;
  const int height = source->y_height;
#if CONFIG_USE_VMAF_RC
  VmafContext *vmaf_context;
  aom_init_vmaf_context_rc(
      &vmaf_context, cpi->vmaf_info.vmaf_model,
      cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_NEG_MAX_GAIN);
  int vmaf_cal_index = 0;
#endif
  YV12_BUFFER_CONFIG sharpened;
  memset(&sharpened, 0, sizeof(sharpened));
  aom_alloc_frame_buffer(&sharpened, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);

  const double baseline_variance = frame_average_variance(cpi, source);
  double unsharp_amount;
  if (unsharp_amount_start <= step_size) {
    unsharp_amount = find_best_frame_unsharp_amount_loop(
        cpi,
#if CONFIG_USE_VMAF_RC
        vmaf_context, &vmaf_cal_index,
#endif
        source, blurred, &sharpened, 0.0, baseline_variance, 0.0, step_size,
        max_loop_count, max_filter_amount);
  } else {
    double a0 = unsharp_amount_start - step_size, a1 = unsharp_amount_start;
    double v0, v1;
    unsharp(cpi, source, blurred, &sharpened, a0);
    v0 = cal_approx_vmaf(cpi,
#if CONFIG_USE_VMAF_RC
                         vmaf_context, &vmaf_cal_index,
#endif
                         baseline_variance, source, &sharpened);
    unsharp(cpi, source, blurred, &sharpened, a1);
    v1 = cal_approx_vmaf(cpi,
#if CONFIG_USE_VMAF_RC
                         vmaf_context, &vmaf_cal_index,
#endif
                         baseline_variance, source, &sharpened);
    if (fabs(v0 - v1) < 0.01) {
      unsharp_amount = a0;
    } else if (v0 > v1) {
      unsharp_amount = find_best_frame_unsharp_amount_loop(
          cpi,
#if CONFIG_USE_VMAF_RC
          vmaf_context, &vmaf_cal_index,
#endif
          source, blurred, &sharpened, v0, baseline_variance, a0, -step_size,
          max_loop_count, max_filter_amount);
    } else {
      unsharp_amount = find_best_frame_unsharp_amount_loop(
          cpi,
#if CONFIG_USE_VMAF_RC
          vmaf_context, &vmaf_cal_index,
#endif
          source, blurred, &sharpened, v1, baseline_variance, a1, step_size,
          max_loop_count, max_filter_amount);
    }
  }

  aom_free_frame_buffer(&sharpened);
#if CONFIG_USE_VMAF_RC
  aom_close_vmaf_context_rc(vmaf_context);
#endif
  return unsharp_amount;
}

#if CONFIG_USE_VMAF_RC
void av1_vmaf_neg_preprocessing(AV1_COMP *const cpi,
                                YV12_BUFFER_CONFIG *const source) {
  aom_clear_system_state();
  const AV1_COMMON *const cm = &cpi->common;
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  const int width = source->y_width;
  const int height = source->y_height;

  const GF_GROUP *const gf_group = &cpi->gf_group;
  const int layer_depth =
      AOMMIN(gf_group->layer_depth[gf_group->index], MAX_ARF_LAYERS - 1);
  const double best_frame_unsharp_amount =
      get_layer_value(cpi->vmaf_info.best_unsharp_amount, layer_depth);

  if (best_frame_unsharp_amount <= 0.0) return;

  YV12_BUFFER_CONFIG blurred;
  memset(&blurred, 0, sizeof(blurred));
  aom_alloc_frame_buffer(&blurred, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);

  gaussian_blur(bit_depth, source, &blurred);
  unsharp(cpi, source, &blurred, source, best_frame_unsharp_amount);
  aom_free_frame_buffer(&blurred);
  aom_clear_system_state();
}
#endif

void av1_vmaf_frame_preprocessing(AV1_COMP *const cpi,
                                  YV12_BUFFER_CONFIG *const source) {
  aom_clear_system_state();
  const AV1_COMMON *const cm = &cpi->common;
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  const int width = source->y_width;
  const int height = source->y_height;

  YV12_BUFFER_CONFIG source_extended, blurred;
  memset(&source_extended, 0, sizeof(source_extended));
  memset(&blurred, 0, sizeof(blurred));
  aom_alloc_frame_buffer(&source_extended, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&blurred, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);

  av1_copy_and_extend_frame(source, &source_extended);
  gaussian_blur(bit_depth, &source_extended, &blurred);
  aom_free_frame_buffer(&source_extended);

  const GF_GROUP *const gf_group = &cpi->gf_group;
  const int layer_depth =
      AOMMIN(gf_group->layer_depth[gf_group->index], MAX_ARF_LAYERS - 1);
  const double last_frame_unsharp_amount =
      get_layer_value(cpi->vmaf_info.last_frame_unsharp_amount, layer_depth);

  const double best_frame_unsharp_amount = find_best_frame_unsharp_amount(
      cpi, source, &blurred, last_frame_unsharp_amount, 0.05, 20, 1.01);

  cpi->vmaf_info.last_frame_unsharp_amount[layer_depth] =
      best_frame_unsharp_amount;

  unsharp(cpi, source, &blurred, source, best_frame_unsharp_amount);
  aom_free_frame_buffer(&blurred);
  aom_clear_system_state();
}

void av1_vmaf_blk_preprocessing(AV1_COMP *const cpi,
                                YV12_BUFFER_CONFIG *const source) {
  aom_clear_system_state();
  const AV1_COMMON *const cm = &cpi->common;
  const int width = source->y_width;
  const int height = source->y_height;
  const int bit_depth = cpi->td.mb.e_mbd.bd;

  YV12_BUFFER_CONFIG source_extended, blurred;
  memset(&blurred, 0, sizeof(blurred));
  memset(&source_extended, 0, sizeof(source_extended));
  aom_alloc_frame_buffer(&blurred, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&source_extended, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);

  av1_copy_and_extend_frame(source, &source_extended);
  gaussian_blur(bit_depth, &source_extended, &blurred);
  aom_free_frame_buffer(&source_extended);

  const GF_GROUP *const gf_group = &cpi->gf_group;
  const int layer_depth =
      AOMMIN(gf_group->layer_depth[gf_group->index], MAX_ARF_LAYERS - 1);
  const double last_frame_unsharp_amount =
      get_layer_value(cpi->vmaf_info.last_frame_unsharp_amount, layer_depth);

  const double best_frame_unsharp_amount = find_best_frame_unsharp_amount(
      cpi, source, &blurred, last_frame_unsharp_amount, 0.05, 20, 1.01);

  cpi->vmaf_info.last_frame_unsharp_amount[layer_depth] =
      best_frame_unsharp_amount;

  const int block_size = BLOCK_64X64;
  const int block_w = mi_size_wide[block_size] * 4;
  const int block_h = mi_size_high[block_size] * 4;
  const int num_cols = (source->y_width + block_w - 1) / block_w;
  const int num_rows = (source->y_height + block_h - 1) / block_h;
  double *best_unsharp_amounts =
      aom_malloc(sizeof(*best_unsharp_amounts) * num_cols * num_rows);
  memset(best_unsharp_amounts, 0,
         sizeof(*best_unsharp_amounts) * num_cols * num_rows);

  YV12_BUFFER_CONFIG source_block, blurred_block;
  memset(&source_block, 0, sizeof(source_block));
  memset(&blurred_block, 0, sizeof(blurred_block));
  aom_alloc_frame_buffer(&source_block, block_w, block_h, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&blurred_block, block_w, block_h, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);

  for (int row = 0; row < num_rows; ++row) {
    for (int col = 0; col < num_cols; ++col) {
      const int row_offset_y = row * block_h;
      const int col_offset_y = col * block_w;
      const int block_width = AOMMIN(width - col_offset_y, block_w);
      const int block_height = AOMMIN(height - row_offset_y, block_h);
      const int index = col + row * num_cols;

      uint16_t *frame_src_buf =
          source->y_buffer + row_offset_y * source->y_stride + col_offset_y;
      uint16_t *frame_blurred_buf =
          blurred.y_buffer + row_offset_y * blurred.y_stride + col_offset_y;
      uint16_t *blurred_dst = blurred_block.y_buffer;
      uint16_t *src_dst = source_block.y_buffer;

      // Copy block from source frame.
      for (int i = 0; i < block_h; ++i) {
        for (int j = 0; j < block_w; ++j) {
          if (i >= block_height || j >= block_width) {
            src_dst[j] = 0;
            blurred_dst[j] = 0;
          } else {
            src_dst[j] = frame_src_buf[j];
            blurred_dst[j] = frame_blurred_buf[j];
          }
        }
        frame_src_buf += source->y_stride;
        frame_blurred_buf += blurred.y_stride;
        src_dst += source_block.y_stride;
        blurred_dst += blurred_block.y_stride;
      }

      best_unsharp_amounts[index] = find_best_frame_unsharp_amount(
          cpi, &source_block, &blurred_block, best_frame_unsharp_amount, 0.1, 3,
          1.5);
    }
  }

  // Apply best blur amounts
  for (int row = 0; row < num_rows; ++row) {
    for (int col = 0; col < num_cols; ++col) {
      const int row_offset_y = row * block_h;
      const int col_offset_y = col * block_w;
      const int block_width = AOMMIN(source->y_width - col_offset_y, block_w);
      const int block_height = AOMMIN(source->y_height - row_offset_y, block_h);
      const int index = col + row * num_cols;

      uint16_t *src_buf =
          source->y_buffer + row_offset_y * source->y_stride + col_offset_y;
      uint16_t *blurred_buf =
          blurred.y_buffer + row_offset_y * blurred.y_stride + col_offset_y;
      highbd_unsharp_rect(src_buf, source->y_stride, blurred_buf,
                          blurred.y_stride, src_buf, source->y_stride,
                          block_width, block_height,
                          best_unsharp_amounts[index], bit_depth);
    }
  }

  aom_free_frame_buffer(&source_block);
  aom_free_frame_buffer(&blurred_block);
  aom_free_frame_buffer(&blurred);
  aom_free(best_unsharp_amounts);
  aom_clear_system_state();
}

#if !CONFIG_USE_VMAF_RC
typedef struct FrameData {
  const YV12_BUFFER_CONFIG *source, *blurred;
  int block_w, block_h, num_rows, num_cols, row, col, bit_depth;
} FrameData;

// A callback function used to pass data to VMAF.
// Returns 0 after reading a frame.
// Returns 2 when there is no more frame to read.
static int update_frame(float *ref_data, float *main_data, float *temp_data,
                        int stride, void *user_data) {
  FrameData *frames = (FrameData *)user_data;
  const int width = frames->source->y_width;
  const int height = frames->source->y_height;
  const int row = frames->row;
  const int col = frames->col;
  const int num_rows = frames->num_rows;
  const int num_cols = frames->num_cols;
  const int block_w = frames->block_w;
  const int block_h = frames->block_h;
  const YV12_BUFFER_CONFIG *source = frames->source;
  const YV12_BUFFER_CONFIG *blurred = frames->blurred;
  const int bit_depth = frames->bit_depth;
  const float scale_factor = 1.0f / (float)(1 << (bit_depth - 8));
  (void)temp_data;
  stride /= (int)sizeof(*ref_data);

  for (int i = 0; i < height; ++i) {
    float *ref, *main;
    ref = ref_data + i * stride;
    main = main_data + i * stride;
    uint16_t *src = source->y_buffer + i * source->y_stride;
    for (int j = 0; j < width; ++j) {
      ref[j] = main[j] = scale_factor * (float)src[j];
    }
  }
  if (row < num_rows && col < num_cols) {
    // Set current block
    const int row_offset = row * block_h;
    const int col_offset = col * block_w;
    const int block_width = AOMMIN(width - col_offset, block_w);
    const int block_height = AOMMIN(height - row_offset, block_h);

    float *main_buf = main_data + col_offset + row_offset * stride;
    uint16_t *blurred_buf =
        blurred->y_buffer + row_offset * blurred->y_stride + col_offset;
    for (int i = 0; i < block_height; ++i) {
      for (int j = 0; j < block_width; ++j) {
        main_buf[j] = scale_factor * (float)blurred_buf[j];
      }
      main_buf += stride;
      blurred_buf += blurred->y_stride;
    }

    frames->col++;
    if (frames->col >= num_cols) {
      frames->col = 0;
      frames->row++;
    }
    return 0;
  } else {
    return 2;
  }
}
#endif

void av1_set_mb_vmaf_rdmult_scaling(AV1_COMP *cpi) {
  AV1_COMMON *cm = &cpi->common;
  const int y_width = cpi->source->y_width;
  const int y_height = cpi->source->y_height;
  const int resized_block_size = BLOCK_32X32;
  const int resize_factor = 2;
  const int bit_depth = cpi->td.mb.e_mbd.bd;

  aom_clear_system_state();
  YV12_BUFFER_CONFIG resized_source;
  memset(&resized_source, 0, sizeof(resized_source));
  aom_alloc_frame_buffer(
      &resized_source, y_width / resize_factor, y_height / resize_factor, 1, 1,
      cpi->oxcf.border_in_pixels, cm->features.byte_alignment);
  av1_resize_and_extend_frame_nonnormative(cpi->source, &resized_source,
                                           bit_depth, av1_num_planes(cm));

  const int resized_y_width = resized_source.y_width;
  const int resized_y_height = resized_source.y_height;
  const int resized_block_w = mi_size_wide[resized_block_size] * 4;
  const int resized_block_h = mi_size_high[resized_block_size] * 4;
  const int num_cols =
      (resized_y_width + resized_block_w - 1) / resized_block_w;
  const int num_rows =
      (resized_y_height + resized_block_h - 1) / resized_block_h;

  YV12_BUFFER_CONFIG blurred;
  memset(&blurred, 0, sizeof(blurred));
  aom_alloc_frame_buffer(&blurred, resized_y_width, resized_y_height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  gaussian_blur(bit_depth, &resized_source, &blurred);

#if CONFIG_USE_VMAF_RC
  YV12_BUFFER_CONFIG recon;
  memset(&recon, 0, sizeof(recon));
  aom_alloc_frame_buffer(&recon, resized_y_width, resized_y_height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_yv12_copy_frame(&resized_source, &recon, 1);

  VmafContext *vmaf_context;
  aom_init_vmaf_context_rc(
      &vmaf_context, cpi->vmaf_info.vmaf_model,
      cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_NEG_MAX_GAIN);
#else
  double *scores = aom_malloc(sizeof(*scores) * (num_rows * num_cols));
  memset(scores, 0, sizeof(*scores) * (num_rows * num_cols));
  FrameData frame_data;
  frame_data.source = &resized_source;
  frame_data.blurred = &blurred;
  frame_data.block_w = resized_block_w;
  frame_data.block_h = resized_block_h;
  frame_data.num_rows = num_rows;
  frame_data.num_cols = num_cols;
  frame_data.row = 0;
  frame_data.col = 0;
  frame_data.bit_depth = bit_depth;
  aom_calc_vmaf_multi_frame(&frame_data, cpi->oxcf.tune_cfg.vmaf_model_path,
                            update_frame, resized_y_width, resized_y_height,
                            bit_depth, scores);
#endif

  // Loop through each 'block_size' block.
  for (int row = 0; row < num_rows; ++row) {
    for (int col = 0; col < num_cols; ++col) {
      const int index = row * num_cols + col;
      const int row_offset_y = row * resized_block_h;
      const int col_offset_y = col * resized_block_w;

      uint16_t *const orig_buf = resized_source.y_buffer +
                                 row_offset_y * resized_source.y_stride +
                                 col_offset_y;
      uint16_t *const blurred_buf =
          blurred.y_buffer + row_offset_y * blurred.y_stride + col_offset_y;

      unsigned int sse;
      cpi->fn_ptr[resized_block_size].vf(orig_buf, resized_source.y_stride,
                                         blurred_buf, blurred.y_stride, &sse);

#if CONFIG_USE_VMAF_RC
      uint16_t *const recon_buf =
          recon.y_buffer + row_offset_y * recon.y_stride + col_offset_y;
      // Set recon buf
      highbd_unsharp_rect(blurred_buf, blurred.y_stride, blurred_buf,
                          blurred.y_stride, recon_buf, recon.y_stride,
                          resized_block_w, resized_block_h, 0.0, bit_depth);

      double vmaf;
      aom_calc_vmaf_at_index_rc(vmaf_context, cpi->vmaf_info.vmaf_model,
                                &resized_source, &recon, bit_depth, index,
                                &vmaf);

      // Restore recon buf
      highbd_unsharp_rect(orig_buf, resized_source.y_stride, orig_buf,
                          resized_source.y_stride, recon_buf, recon.y_stride,
                          resized_block_w, resized_block_h, 0.0, bit_depth);
#else
      const double vmaf = scores[index];
#endif
      const double dvmaf = kBaselineVmaf - vmaf;

      const double mse =
          (double)sse / (double)(resized_y_width * resized_y_height);
      double weight;
      const double eps = 0.01 / (num_rows * num_cols);
      if (dvmaf < eps || mse < eps) {
        weight = 1.0;
      } else {
        weight = mse / dvmaf;
      }

      // Normalize it with a data fitted model.
      weight = 6.0 * (1.0 - exp(-0.05 * weight)) + 0.8;
      cpi->vmaf_info.rdmult_scaling_factors[index] = weight;
    }
  }

  aom_free_frame_buffer(&resized_source);
  aom_free_frame_buffer(&blurred);
#if CONFIG_USE_VMAF_RC
  aom_close_vmaf_context_rc(vmaf_context);
#else
  aom_free(scores);
#endif
  aom_clear_system_state();
}

void av1_set_vmaf_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
                         const BLOCK_SIZE bsize, const int mi_row,
                         const int mi_col, int *const rdmult) {
  const AV1_COMMON *const cm = &cpi->common;

  const int bsize_base = BLOCK_64X64;
  const int num_mi_w = mi_size_wide[bsize_base];
  const int num_mi_h = mi_size_high[bsize_base];
  const int num_cols = (cm->mi_params.mi_cols + num_mi_w - 1) / num_mi_w;
  const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h;
  const int num_bcols = (mi_size_wide[bsize] + num_mi_w - 1) / num_mi_w;
  const int num_brows = (mi_size_high[bsize] + num_mi_h - 1) / num_mi_h;
  int row, col;
  double num_of_mi = 0.0;
  double geom_mean_of_scale = 0.0;

  aom_clear_system_state();
  for (row = mi_row / num_mi_w;
       row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
    for (col = mi_col / num_mi_h;
         col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) {
      const int index = row * num_cols + col;
      geom_mean_of_scale += log(cpi->vmaf_info.rdmult_scaling_factors[index]);
      num_of_mi += 1.0;
    }
  }
  geom_mean_of_scale = exp(geom_mean_of_scale / num_of_mi);

  *rdmult = (int)((double)(*rdmult) * geom_mean_of_scale + 0.5);
  *rdmult = AOMMAX(*rdmult, 0);
  av1_set_error_per_bit(&x->mv_costs, *rdmult);
  aom_clear_system_state();
}

// TODO(sdeng): replace them with the SIMD versions.
static AOM_INLINE double highbd_image_sad_c(const uint16_t *src, int src_stride,
                                            const uint16_t *ref, int ref_stride,
                                            int w, int h) {
  double accum = 0.0;
  int i, j;

  for (i = 0; i < h; ++i) {
    for (j = 0; j < w; ++j) {
      double img1px = src[i * src_stride + j];
      double img2px = ref[i * ref_stride + j];

      accum += fabs(img1px - img2px);
    }
  }

  return accum / (double)(h * w);
}

static double calc_vmaf_motion_score(const AV1_COMP *const cpi,
                                     const AV1_COMMON *const cm,
                                     const YV12_BUFFER_CONFIG *const cur,
                                     const YV12_BUFFER_CONFIG *const last,
                                     const YV12_BUFFER_CONFIG *const next) {
  const int y_width = cur->y_width;
  const int y_height = cur->y_height;
  YV12_BUFFER_CONFIG blurred_cur, blurred_last, blurred_next;
  const int bit_depth = cpi->td.mb.e_mbd.bd;

  memset(&blurred_cur, 0, sizeof(blurred_cur));
  memset(&blurred_last, 0, sizeof(blurred_last));
  memset(&blurred_next, 0, sizeof(blurred_next));

  aom_alloc_frame_buffer(&blurred_cur, y_width, y_height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&blurred_last, y_width, y_height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&blurred_next, y_width, y_height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);

  gaussian_blur(bit_depth, cur, &blurred_cur);
  gaussian_blur(bit_depth, last, &blurred_last);
  if (next) gaussian_blur(bit_depth, next, &blurred_next);

  double motion1, motion2 = 65536.0;
  const float scale_factor = 1.0f / (float)(1 << (bit_depth - 8));
  motion1 = highbd_image_sad_c(blurred_cur.y_buffer, blurred_cur.y_stride,
                               blurred_last.y_buffer, blurred_last.y_stride,
                               y_width, y_height) *
            scale_factor;
  if (next) {
    motion2 = highbd_image_sad_c(blurred_cur.y_buffer, blurred_cur.y_stride,
                                 blurred_next.y_buffer, blurred_next.y_stride,
                                 y_width, y_height) *
              scale_factor;
  }

  aom_free_frame_buffer(&blurred_cur);
  aom_free_frame_buffer(&blurred_last);
  aom_free_frame_buffer(&blurred_next);

  return AOMMIN(motion1, motion2);
}

static AOM_INLINE void get_neighbor_frames(const AV1_COMP *const cpi,
                                           YV12_BUFFER_CONFIG **last,
                                           YV12_BUFFER_CONFIG **next) {
  const AV1_COMMON *const cm = &cpi->common;
  const GF_GROUP *gf_group = &cpi->gf_group;
  const int src_index =
      cm->show_frame != 0 ? 0 : gf_group->arf_src_offset[gf_group->index];
  struct lookahead_entry *last_entry =
      av1_lookahead_peek(cpi->lookahead, src_index - 1, cpi->compressor_stage);
  struct lookahead_entry *next_entry =
      av1_lookahead_peek(cpi->lookahead, src_index + 1, cpi->compressor_stage);
  *next = &next_entry->img;
  *last = cm->show_frame ? cpi->last_source : &last_entry->img;
}

// Calculates the new qindex from the VMAF motion score. This is based on the
// observation: when the motion score becomes higher, the VMAF score of the
// same source and distorted frames would become higher.
int av1_get_vmaf_base_qindex(const AV1_COMP *const cpi, int current_qindex) {
  const AV1_COMMON *const cm = &cpi->common;
  if (cm->current_frame.frame_number == 0 || cpi->oxcf.pass == 1) {
    return current_qindex;
  }
  aom_clear_system_state();
  const GF_GROUP *const gf_group = &cpi->gf_group;
  const int layer_depth =
      AOMMIN(gf_group->layer_depth[gf_group->index], MAX_ARF_LAYERS - 1);
  const double last_frame_ysse =
      get_layer_value(cpi->vmaf_info.last_frame_ysse, layer_depth);
  const double last_frame_vmaf =
      get_layer_value(cpi->vmaf_info.last_frame_vmaf, layer_depth);
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  const double approx_sse = last_frame_ysse / (double)((1 << (bit_depth - 8)) *
                                                       (1 << (bit_depth - 8)));
  const double approx_dvmaf = kBaselineVmaf - last_frame_vmaf;
  const double sse_threshold =
      0.01 * cpi->source->y_width * cpi->source->y_height;
  const double vmaf_threshold = 0.01;
  if (approx_sse < sse_threshold || approx_dvmaf < vmaf_threshold) {
    return current_qindex;
  }
  YV12_BUFFER_CONFIG *cur_buf = cpi->source;
  if (cm->show_frame == 0) {
    const int src_index = gf_group->arf_src_offset[gf_group->index];
    struct lookahead_entry *cur_entry =
        av1_lookahead_peek(cpi->lookahead, src_index, cpi->compressor_stage);
    cur_buf = &cur_entry->img;
  }
  assert(cur_buf);

  YV12_BUFFER_CONFIG *next_buf, *last_buf;
  get_neighbor_frames(cpi, &last_buf, &next_buf);
  assert(last_buf);

  const double motion =
      calc_vmaf_motion_score(cpi, cm, cur_buf, last_buf, next_buf);

  // Get dVMAF through a data fitted model.
  const double dvmaf = 26.11 * (1.0 - exp(-0.06 * motion));
  const double dsse = dvmaf * approx_sse / approx_dvmaf;

  const double beta = approx_sse / (dsse + approx_sse);
  const int offset = av1_get_deltaq_offset(cpi, current_qindex, beta);
  int qindex = current_qindex + offset;

  qindex = AOMMIN(qindex, MAXQ);
  qindex = AOMMAX(qindex, MINQ);

  aom_clear_system_state();
  return qindex;
}

#if CONFIG_USE_VMAF_RC
static AOM_INLINE double cal_approx_score(
    AV1_COMP *const cpi, VmafContext *vmaf_context, int vmaf_cal_index,
    double src_variance, double new_variance, double src_score,
    YV12_BUFFER_CONFIG *const src, YV12_BUFFER_CONFIG *const recon_sharpened) {
  double score;
  const uint32_t bit_depth = cpi->td.mb.e_mbd.bd;
  aom_calc_vmaf_at_index_rc(vmaf_context, cpi->vmaf_info.vmaf_model, src,
                            recon_sharpened, bit_depth, vmaf_cal_index, &score);
  return src_variance / new_variance * (score - src_score);
}

static double find_best_frame_unsharp_amount_loop_neg(
    AV1_COMP *const cpi, VmafContext *vmaf_context, double src_variance,
    double base_score, YV12_BUFFER_CONFIG *const src,
    YV12_BUFFER_CONFIG *const recon, YV12_BUFFER_CONFIG *const ref,
    FULLPEL_MV *mvs, double best_score, const double unsharp_amount_start,
    const double step_size, const int max_loop_count, const double max_amount) {
  const double min_amount = 0.0;
  int loop_count = 0;
  double approx_score = best_score;
  double unsharp_amount = unsharp_amount_start;
  int vmaf_cal_index = 2;

  const AV1_COMMON *const cm = &cpi->common;
  const int width = recon->y_width;
  const int height = recon->y_height;
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  YV12_BUFFER_CONFIG src_blurred, recon_blurred, src_sharpened, recon_sharpened;
  memset(&recon_sharpened, 0, sizeof(recon_sharpened));
  memset(&src_sharpened, 0, sizeof(src_sharpened));
  memset(&recon_blurred, 0, sizeof(recon_blurred));
  memset(&src_blurred, 0, sizeof(src_blurred));
  aom_alloc_frame_buffer(&recon_sharpened, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&src_sharpened, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&recon_blurred, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);
  aom_alloc_frame_buffer(&src_blurred, width, height, 1, 1,
                         cpi->oxcf.border_in_pixels,
                         cm->features.byte_alignment);

  gaussian_blur(bit_depth, recon, &recon_blurred);
  gaussian_blur(bit_depth, src, &src_blurred);
  do {
    best_score = approx_score;
    unsharp_amount += step_size;
    if (unsharp_amount > max_amount || unsharp_amount < min_amount) break;
    unsharp(cpi, recon, &recon_blurred, &recon_sharpened, unsharp_amount);
    unsharp(cpi, src, &src_blurred, &src_sharpened, unsharp_amount);
    const double new_variance =
        residual_frame_average_variance(cpi, &src_sharpened, ref, mvs);
    approx_score =
        cal_approx_score(cpi, vmaf_context, vmaf_cal_index++, src_variance,
                         new_variance, base_score, src, &recon_sharpened);

    loop_count++;
  } while (approx_score > best_score && loop_count < max_loop_count);
  unsharp_amount =
      approx_score > best_score ? unsharp_amount : unsharp_amount - step_size;

  aom_free_frame_buffer(&recon_sharpened);
  aom_free_frame_buffer(&src_sharpened);
  aom_free_frame_buffer(&recon_blurred);
  aom_free_frame_buffer(&src_blurred);
  return AOMMIN(max_amount, AOMMAX(unsharp_amount, min_amount));
}

static double find_best_frame_unsharp_amount_neg(
    AV1_COMP *const cpi, VmafContext *vmaf_context,
    YV12_BUFFER_CONFIG *const src, YV12_BUFFER_CONFIG *const recon,
    YV12_BUFFER_CONFIG *const ref, const double unsharp_amount_start,
    const double step_size, const int max_loop_count,
    const double max_filter_amount) {
  double base_score = 0.0;
  aom_calc_vmaf_at_index_rc(vmaf_context, cpi->vmaf_info.vmaf_model, src, recon,
                            cpi->td.mb.e_mbd.bd, 1, &base_score);

  FULLPEL_MV *mvs = NULL;
  const double src_variance =
      residual_frame_average_variance(cpi, src, ref, mvs);
  const double unsharp_amount = find_best_frame_unsharp_amount_loop_neg(
      cpi, vmaf_context, src_variance, base_score, src, recon, ref, mvs, 0.0,
      unsharp_amount_start, step_size, max_loop_count, max_filter_amount);

  aom_free(mvs);
  return unsharp_amount;
}
#endif  // CONFIG_USE_VMAF_RC

void av1_update_vmaf_curve(AV1_COMP *cpi) {
  YV12_BUFFER_CONFIG *source = cpi->source;
  YV12_BUFFER_CONFIG *recon = &cpi->common.cur_frame->buf;
  const int bit_depth = cpi->td.mb.e_mbd.bd;
  const GF_GROUP *const gf_group = &cpi->gf_group;
  const int layer_depth =
      AOMMIN(gf_group->layer_depth[gf_group->index], MAX_ARF_LAYERS - 1);
#if CONFIG_USE_VMAF_RC
  VmafContext *vmaf_context;
  aom_init_vmaf_context_rc(
      &vmaf_context, cpi->vmaf_info.vmaf_model,
      cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_NEG_MAX_GAIN);
  aom_calc_vmaf_at_index_rc(vmaf_context, cpi->vmaf_info.vmaf_model, source,
                            recon, bit_depth, 0,
                            &cpi->vmaf_info.last_frame_vmaf[layer_depth]);
#else
  aom_calc_vmaf(cpi->oxcf.tune_cfg.vmaf_model_path, source, recon, bit_depth,
                &cpi->vmaf_info.last_frame_vmaf[layer_depth]);
#endif  // CONFIG_USE_VMAF_RC
  cpi->vmaf_info.last_frame_ysse[layer_depth] =
      (double)aom_highbd_get_y_sse(source, recon);

#if CONFIG_USE_VMAF_RC
  if (cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_NEG_MAX_GAIN) {
    YV12_BUFFER_CONFIG *last, *next;
    get_neighbor_frames(cpi, &last, &next);
    cpi->vmaf_info.best_unsharp_amount[layer_depth] =
        find_best_frame_unsharp_amount_neg(cpi, vmaf_context, source, recon,
                                           last, 0.0, 0.025, 20, 1.01);
  }
  aom_close_vmaf_context_rc(vmaf_context);
#endif  // CONFIG_USE_VMAF_RC
}