path: root/third_party/jpeg-xl/lib/jxl/enc_frame.cc

// Copyright (c) the JPEG XL Project Authors. All rights reserved.
//
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

#include "lib/jxl/enc_frame.h"

#include <stddef.h>
#include <stdint.h>

#include <algorithm>
#include <array>
#include <atomic>
#include <cmath>
#include <limits>
#include <numeric>
#include <vector>

#include "lib/jxl/ac_context.h"
#include "lib/jxl/ac_strategy.h"
#include "lib/jxl/ans_params.h"
#include "lib/jxl/base/bits.h"
#include "lib/jxl/base/compiler_specific.h"
#include "lib/jxl/base/data_parallel.h"
#include "lib/jxl/base/override.h"
#include "lib/jxl/base/padded_bytes.h"
#include "lib/jxl/base/profiler.h"
#include "lib/jxl/base/status.h"
#include "lib/jxl/chroma_from_luma.h"
#include "lib/jxl/coeff_order.h"
#include "lib/jxl/coeff_order_fwd.h"
#include "lib/jxl/color_encoding_internal.h"
#include "lib/jxl/color_management.h"
#include "lib/jxl/common.h"
#include "lib/jxl/compressed_dc.h"
#include "lib/jxl/dct_util.h"
#include "lib/jxl/enc_adaptive_quantization.h"
#include "lib/jxl/enc_ans.h"
#include "lib/jxl/enc_aux_out.h"
#include "lib/jxl/enc_bit_writer.h"
#include "lib/jxl/enc_cache.h"
#include "lib/jxl/enc_chroma_from_luma.h"
#include "lib/jxl/enc_coeff_order.h"
#include "lib/jxl/enc_context_map.h"
#include "lib/jxl/enc_entropy_coder.h"
#include "lib/jxl/enc_fields.h"
#include "lib/jxl/enc_gaborish.h"
#include "lib/jxl/enc_group.h"
#include "lib/jxl/enc_modular.h"
#include "lib/jxl/enc_noise.h"
#include "lib/jxl/enc_params.h"
#include "lib/jxl/enc_patch_dictionary.h"
#include "lib/jxl/enc_photon_noise.h"
#include "lib/jxl/enc_quant_weights.h"
#include "lib/jxl/enc_splines.h"
#include "lib/jxl/enc_toc.h"
#include "lib/jxl/enc_xyb.h"
#include "lib/jxl/fields.h"
#include "lib/jxl/frame_header.h"
#include "lib/jxl/image.h"
#include "lib/jxl/image_bundle.h"
#include "lib/jxl/image_ops.h"
#include "lib/jxl/loop_filter.h"
#include "lib/jxl/modular/options.h"
#include "lib/jxl/quant_weights.h"
#include "lib/jxl/quantizer.h"
#include "lib/jxl/splines.h"
#include "lib/jxl/toc.h"

namespace jxl {
namespace {

PassDefinition progressive_passes_dc_vlf_lf_full_ac[] = {
    {/*num_coefficients=*/2, /*shift=*/0,
     /*suitable_for_downsampling_of_at_least=*/4},
    {/*num_coefficients=*/3, /*shift=*/0,
     /*suitable_for_downsampling_of_at_least=*/2},
    {/*num_coefficients=*/8, /*shift=*/0,
     /*suitable_for_downsampling_of_at_least=*/0},
};

PassDefinition progressive_passes_dc_quant_ac_full_ac[] = {
    {/*num_coefficients=*/8, /*shift=*/1,
     /*suitable_for_downsampling_of_at_least=*/2},
    {/*num_coefficients=*/8, /*shift=*/0,
     /*suitable_for_downsampling_of_at_least=*/0},
};

void ClusterGroups(PassesEncoderState* enc_state) {
  if (enc_state->shared.frame_header.passes.num_passes > 1) {
    // TODO(veluca): implement this for progressive modes.
    return;
  }
  // This only considers pass 0 for now.
  std::vector<uint8_t> context_map;
  EntropyEncodingData codes;
  auto& ac = enc_state->passes[0].ac_tokens;
  size_t limit = std::ceil(std::sqrt(ac.size()));
  if (limit == 1) return;
  size_t num_contexts = enc_state->shared.block_ctx_map.NumACContexts();
  std::vector<float> costs(ac.size());
  HistogramParams params;
  params.uint_method = HistogramParams::HybridUintMethod::kNone;
  params.lz77_method = HistogramParams::LZ77Method::kNone;
  params.ans_histogram_strategy =
      HistogramParams::ANSHistogramStrategy::kApproximate;
  size_t max = 0;
  auto token_cost = [&](std::vector<std::vector<Token>>& tokens, size_t num_ctx,
                        bool estimate = true) {
    // TODO(veluca): not estimating is very expensive.
    BitWriter writer;
    size_t c = BuildAndEncodeHistograms(
        params, num_ctx, tokens, &codes, &context_map,
        estimate ? nullptr : &writer, 0, /*aux_out=*/0);
    if (estimate) return c;
    for (size_t i = 0; i < tokens.size(); i++) {
      WriteTokens(tokens[i], codes, context_map, &writer, 0, nullptr);
    }
    return writer.BitsWritten();
  };
  for (size_t i = 0; i < ac.size(); i++) {
    std::vector<std::vector<Token>> tokens{ac[i]};
    costs[i] =
        token_cost(tokens, enc_state->shared.block_ctx_map.NumACContexts());
    if (costs[i] > costs[max]) {
      max = i;
    }
  }
  auto dist = [&](int i, int j) {
    std::vector<std::vector<Token>> tokens{ac[i], ac[j]};
    return token_cost(tokens, num_contexts) - costs[i] - costs[j];
  };
  std::vector<size_t> out{max};
  std::vector<float> dists(ac.size());
  size_t farthest = 0;
  for (size_t i = 0; i < ac.size(); i++) {
    if (i == max) continue;
    dists[i] = dist(max, i);
    if (dists[i] > dists[farthest]) {
      farthest = i;
    }
  }

  while (dists[farthest] > 0 && out.size() < limit) {
    out.push_back(farthest);
    dists[farthest] = 0;
    enc_state->histogram_idx[farthest] = out.size() - 1;
    for (size_t i = 0; i < ac.size(); i++) {
      float d = dist(out.back(), i);
      if (d < dists[i]) {
        dists[i] = d;
        enc_state->histogram_idx[i] = out.size() - 1;
      }
      if (dists[i] > dists[farthest]) {
        farthest = i;
      }
    }
  }

  std::vector<size_t> remap(out.size());
  std::iota(remap.begin(), remap.end(), 0);
  for (size_t i = 0; i < enc_state->histogram_idx.size(); i++) {
    enc_state->histogram_idx[i] = remap[enc_state->histogram_idx[i]];
  }
  auto remap_cost = [&](std::vector<size_t> remap) {
    std::vector<size_t> re_remap(remap.size(), remap.size());
    size_t r = 0;
    for (size_t i = 0; i < remap.size(); i++) {
      if (re_remap[remap[i]] == remap.size()) {
        re_remap[remap[i]] = r++;
      }
      remap[i] = re_remap[remap[i]];
    }
    auto tokens = ac;
    size_t max_hist = 0;
    for (size_t i = 0; i < tokens.size(); i++) {
      for (size_t j = 0; j < tokens[i].size(); j++) {
        size_t hist = remap[enc_state->histogram_idx[i]];
        tokens[i][j].context += hist * num_contexts;
        max_hist = std::max(hist + 1, max_hist);
      }
    }
    return token_cost(tokens, max_hist * num_contexts, /*estimate=*/false);
  };

  for (size_t src = 0; src < out.size(); src++) {
    float cost = remap_cost(remap);
    size_t best = src;
    for (size_t j = src + 1; j < out.size(); j++) {
      if (remap[src] == remap[j]) continue;
      auto remap_c = remap;
      std::replace(remap_c.begin(), remap_c.end(), remap[src], remap[j]);
      float c = remap_cost(remap_c);
      if (c < cost) {
        best = j;
        cost = c;
      }
    }
    if (src != best) {
      std::replace(remap.begin(), remap.end(), remap[src], remap[best]);
    }
  }
  std::vector<size_t> re_remap(remap.size(), remap.size());
  size_t r = 0;
  for (size_t i = 0; i < remap.size(); i++) {
    if (re_remap[remap[i]] == remap.size()) {
      re_remap[remap[i]] = r++;
    }
    remap[i] = re_remap[remap[i]];
  }

  enc_state->shared.num_histograms =
      *std::max_element(remap.begin(), remap.end()) + 1;
  for (size_t i = 0; i < enc_state->histogram_idx.size(); i++) {
    enc_state->histogram_idx[i] = remap[enc_state->histogram_idx[i]];
  }
  for (size_t i = 0; i < ac.size(); i++) {
    for (size_t j = 0; j < ac[i].size(); j++) {
      ac[i][j].context += enc_state->histogram_idx[i] * num_contexts;
    }
  }
}

uint64_t FrameFlagsFromParams(const CompressParams& cparams) {
  uint64_t flags = 0;

  const float dist = cparams.butteraugli_distance;

  // We don't add noise at low butteraugli distances because the original
  // noise is stored within the compressed image and adding noise makes things
  // worse.
  if (ApplyOverride(cparams.noise, dist >= kMinButteraugliForNoise) ||
      cparams.photon_noise_iso > 0 ||
      cparams.manual_noise.size() == NoiseParams::kNumNoisePoints) {
    flags |= FrameHeader::kNoise;
  }

  if (cparams.progressive_dc > 0 && cparams.modular_mode == false) {
    flags |= FrameHeader::kUseDcFrame;
  }

  return flags;
}

Status LoopFilterFromParams(const CompressParams& cparams,
                            FrameHeader* JXL_RESTRICT frame_header) {
  LoopFilter* loop_filter = &frame_header->loop_filter;

  // Gaborish defaults to enabled in Hare or slower.
  loop_filter->gab = ApplyOverride(
      cparams.gaborish, cparams.speed_tier <= SpeedTier::kHare &&
                            frame_header->encoding == FrameEncoding::kVarDCT &&
                            cparams.decoding_speed_tier < 4);

  if (cparams.epf != -1) {
    loop_filter->epf_iters = cparams.epf;
  } else {
    if (frame_header->encoding == FrameEncoding::kModular) {
      loop_filter->epf_iters = 0;
    } else {
      constexpr float kThresholds[3] = {0.7, 1.5, 4.0};
      loop_filter->epf_iters = 0;
      if (cparams.decoding_speed_tier < 3) {
        for (size_t i = cparams.decoding_speed_tier == 2 ? 1 : 0; i < 3; i++) {
          if (cparams.butteraugli_distance >= kThresholds[i]) {
            loop_filter->epf_iters++;
          }
        }
      }
    }
  }
  // Strength of EPF in modular mode.
  if (frame_header->encoding == FrameEncoding::kModular &&
      !cparams.IsLossless()) {
    // TODO(veluca): this formula is nonsense.
    loop_filter->epf_sigma_for_modular = cparams.butteraugli_distance;
  }
  if (frame_header->encoding == FrameEncoding::kModular &&
      cparams.lossy_palette) {
    loop_filter->epf_sigma_for_modular = 1.0f;
  }

  return true;
}

Status MakeFrameHeader(const CompressParams& cparams,
                       const ProgressiveSplitter& progressive_splitter,
                       const FrameInfo& frame_info, const ImageBundle& ib,
                       FrameHeader* JXL_RESTRICT frame_header) {
  frame_header->nonserialized_is_preview = frame_info.is_preview;
  frame_header->is_last = frame_info.is_last;
  frame_header->save_before_color_transform =
      frame_info.save_before_color_transform;
  frame_header->frame_type = frame_info.frame_type;
  frame_header->name = ib.name;

  progressive_splitter.InitPasses(&frame_header->passes);

  if (cparams.modular_mode) {
    frame_header->encoding = FrameEncoding::kModular;
    frame_header->group_size_shift = cparams.modular_group_size_shift;
  }

  frame_header->chroma_subsampling = ib.chroma_subsampling;
  if (ib.IsJPEG()) {
    // we are transcoding a JPEG, so we don't get to choose
    frame_header->encoding = FrameEncoding::kVarDCT;
    frame_header->color_transform = ib.color_transform;
  } else {
    frame_header->color_transform = cparams.color_transform;
    if (!cparams.modular_mode &&
        (frame_header->chroma_subsampling.MaxHShift() != 0 ||
         frame_header->chroma_subsampling.MaxVShift() != 0)) {
      return JXL_FAILURE(
          "Chroma subsampling is not supported in VarDCT mode when not "
          "recompressing JPEGs");
    }
  }
  if (frame_header->color_transform != ColorTransform::kYCbCr &&
      (frame_header->chroma_subsampling.MaxHShift() != 0 ||
       frame_header->chroma_subsampling.MaxVShift() != 0)) {
    return JXL_FAILURE(
        "Chroma subsampling is not supported when color transform is not "
        "YCbCr");
  }

  frame_header->flags = FrameFlagsFromParams(cparams);
  // Non-photon noise is not supported in the Modular encoder for now.
  if (frame_header->encoding != FrameEncoding::kVarDCT &&
      cparams.photon_noise_iso == 0 && cparams.manual_noise.empty()) {
    frame_header->UpdateFlag(false, FrameHeader::Flags::kNoise);
  }

  JXL_RETURN_IF_ERROR(LoopFilterFromParams(cparams, frame_header));

  frame_header->dc_level = frame_info.dc_level;
  if (frame_header->dc_level > 2) {
    // With 3 or more progressive_dc frames, the implementation does not yet
    // work, see enc_cache.cc.
    return JXL_FAILURE("progressive_dc > 2 is not yet supported");
  }
  if (cparams.progressive_dc > 0 &&
      (cparams.ec_resampling != 1 || cparams.resampling != 1)) {
    return JXL_FAILURE("Resampling not supported with DC frames");
  }
  if (cparams.resampling != 1 && cparams.resampling != 2 &&
      cparams.resampling != 4 && cparams.resampling != 8) {
    return JXL_FAILURE("Invalid resampling factor");
  }
  if (cparams.ec_resampling != 1 && cparams.ec_resampling != 2 &&
      cparams.ec_resampling != 4 && cparams.ec_resampling != 8) {
    return JXL_FAILURE("Invalid ec_resampling factor");
  }
  // Resized frames.
  if (frame_info.frame_type != FrameType::kDCFrame) {
    frame_header->frame_origin = ib.origin;
    size_t ups = 1;
    if (cparams.already_downsampled) ups = cparams.resampling;

    // TODO(lode): this is not correct in case of odd original image sizes in
    // combination with cparams.already_downsampled. Likely these values should
    // be set to respectively frame_header->default_xsize() and
    // frame_header->default_ysize() instead, the original (non downsampled)
    // intended decoded image dimensions. But it may be more subtle than that
    // if combined with crop. This issue causes custom_size_or_origin to be
    // incorrectly set to true in case of already_downsampled with odd output
    // image size when no cropping is used.
    frame_header->frame_size.xsize = ib.xsize() * ups;
    frame_header->frame_size.ysize = ib.ysize() * ups;
    if (ib.origin.x0 != 0 || ib.origin.y0 != 0 ||
        frame_header->frame_size.xsize != frame_header->default_xsize() ||
        frame_header->frame_size.ysize != frame_header->default_ysize()) {
      frame_header->custom_size_or_origin = true;
    }
  }
  // Upsampling.
  frame_header->upsampling = cparams.resampling;
  const std::vector<ExtraChannelInfo>& extra_channels =
      frame_header->nonserialized_metadata->m.extra_channel_info;
  frame_header->extra_channel_upsampling.clear();
  frame_header->extra_channel_upsampling.resize(extra_channels.size(),
                                                cparams.ec_resampling);
  frame_header->save_as_reference = frame_info.save_as_reference;

  // Set blending-related information.
  if (ib.blend || frame_header->custom_size_or_origin) {
    // Set blend_channel to the first alpha channel. These values are only
    // encoded in case a blend mode involving alpha is used and there are more
    // than one extra channels.
    size_t index = 0;
    if (frame_info.alpha_channel == -1) {
      if (extra_channels.size() > 1) {
        for (size_t i = 0; i < extra_channels.size(); i++) {
          if (extra_channels[i].type == ExtraChannel::kAlpha) {
            index = i;
            break;
          }
        }
      }
    } else {
      index = static_cast<size_t>(frame_info.alpha_channel);
      JXL_ASSERT(index == 0 || index < extra_channels.size());
    }
    frame_header->blending_info.alpha_channel = index;
    frame_header->blending_info.mode =
        ib.blend ? ib.blendmode : BlendMode::kReplace;
    frame_header->blending_info.source = frame_info.source;
    frame_header->blending_info.clamp = frame_info.clamp;
    const auto& extra_channel_info = frame_info.extra_channel_blending_info;
    for (size_t i = 0; i < extra_channels.size(); i++) {
      if (i < extra_channel_info.size()) {
        frame_header->extra_channel_blending_info[i] = extra_channel_info[i];
      } else {
        frame_header->extra_channel_blending_info[i].alpha_channel = index;
        BlendMode default_blend = ib.blendmode;
        if (extra_channels[i].type != ExtraChannel::kBlack && i != index) {
          // K needs to be blended, spot colors and other stuff gets added
          default_blend = BlendMode::kAdd;
        }
        frame_header->extra_channel_blending_info[i].mode =
            ib.blend ? default_blend : BlendMode::kReplace;
        frame_header->extra_channel_blending_info[i].source = 1;
      }
    }
  }

  frame_header->animation_frame.duration = ib.duration;
  frame_header->animation_frame.timecode = ib.timecode;

  return true;
}

// Invisible (alpha = 0) pixels tend to be a mess in optimized PNGs.
// Since they have no visual impact whatsoever, we can replace them with
// something that compresses better and reduces artifacts near the edges. This
// does some kind of smooth stuff that seems to work.
// Replace invisible pixels with a weighted average of the pixel to the left,
// the pixel to the topright, and non-invisible neighbours.
// Produces downward-blurry smears, with in the upwards direction only a 1px
// edge duplication but not more. It would probably be better to smear in all
// directions. That requires an alpha-weighed convolution with a large enough
// kernel though, which might be overkill...
void SimplifyInvisible(Image3F* image, const ImageF& alpha, bool lossless) {
  for (size_t c = 0; c < 3; ++c) {
    for (size_t y = 0; y < image->ysize(); ++y) {
      float* JXL_RESTRICT row = image->PlaneRow(c, y);
      const float* JXL_RESTRICT prow =
          (y > 0 ? image->PlaneRow(c, y - 1) : nullptr);
      const float* JXL_RESTRICT nrow =
          (y + 1 < image->ysize() ? image->PlaneRow(c, y + 1) : nullptr);
      const float* JXL_RESTRICT a = alpha.Row(y);
      const float* JXL_RESTRICT pa = (y > 0 ? alpha.Row(y - 1) : nullptr);
      const float* JXL_RESTRICT na =
          (y + 1 < image->ysize() ? alpha.Row(y + 1) : nullptr);
      for (size_t x = 0; x < image->xsize(); ++x) {
        if (a[x] == 0) {
          if (lossless) {
            row[x] = 0;
            continue;
          }
          float d = 0.f;
          row[x] = 0;
          if (x > 0) {
            row[x] += row[x - 1];
            d++;
            if (a[x - 1] > 0.f) {
              row[x] += row[x - 1];
              d++;
            }
          }
          if (x + 1 < image->xsize()) {
            if (y > 0) {
              row[x] += prow[x + 1];
              d++;
            }
            if (a[x + 1] > 0.f) {
              row[x] += 2.f * row[x + 1];
              d += 2.f;
            }
            if (y > 0 && pa[x + 1] > 0.f) {
              row[x] += 2.f * prow[x + 1];
              d += 2.f;
            }
            if (y + 1 < image->ysize() && na[x + 1] > 0.f) {
              row[x] += 2.f * nrow[x + 1];
              d += 2.f;
            }
          }
          if (y > 0 && pa[x] > 0.f) {
            row[x] += 2.f * prow[x];
            d += 2.f;
          }
          if (y + 1 < image->ysize() && na[x] > 0.f) {
            row[x] += 2.f * nrow[x];
            d += 2.f;
          }
          if (d > 1.f) row[x] /= d;
        }
      }
    }
  }
}

struct PixelStatsForChromacityAdjustment {
  float dx = 0;
  float db = 0;
  float exposed_blue = 0;
  float CalcPlane(const ImageF* JXL_RESTRICT plane) const {
    float xmax = 0;
    float ymax = 0;
    for (size_t ty = 1; ty < plane->ysize(); ++ty) {
      for (size_t tx = 1; tx < plane->xsize(); ++tx) {
        float cur = plane->Row(ty)[tx];
        float prev_row = plane->Row(ty - 1)[tx];
        float prev = plane->Row(ty)[tx - 1];
        xmax = std::max(xmax, std::abs(cur - prev));
        ymax = std::max(ymax, std::abs(cur - prev_row));
      }
    }
    return std::max(xmax, ymax);
  }
  void CalcExposedBlue(const ImageF* JXL_RESTRICT plane_y,
                       const ImageF* JXL_RESTRICT plane_b) {
    float eb = 0;
    float xmax = 0;
    float ymax = 0;
    for (size_t ty = 1; ty < plane_y->ysize(); ++ty) {
      for (size_t tx = 1; tx < plane_y->xsize(); ++tx) {
        float cur_y = plane_y->Row(ty)[tx];
        float cur_b = plane_b->Row(ty)[tx];
        float exposed_b = cur_b - cur_y * 1.2;
        float diff_b = cur_b - cur_y;
        float prev_row = plane_b->Row(ty - 1)[tx];
        float prev = plane_b->Row(ty)[tx - 1];
        float diff_prev_row = prev_row - plane_y->Row(ty - 1)[tx];
        float diff_prev = prev - plane_y->Row(ty)[tx - 1];
        xmax = std::max(xmax, std::abs(diff_b - diff_prev));
        ymax = std::max(ymax, std::abs(diff_b - diff_prev_row));
        if (exposed_b >= 0) {
          exposed_b *= fabs(cur_b - prev) + fabs(cur_b - prev_row);
          eb = std::max(eb, exposed_b);
        }
      }
    }
    exposed_blue = eb;
    db = std::max(xmax, ymax);
  }
  void Calc(const Image3F* JXL_RESTRICT opsin) {
    dx = CalcPlane(&opsin->Plane(0));
    CalcExposedBlue(&opsin->Plane(1), &opsin->Plane(2));
  }
  int HowMuchIsXChannelPixelized() {
    if (dx >= 0.03) {
      return 2;
    }
    if (dx >= 0.017) {
      return 1;
    }
    return 0;
  }
  int HowMuchIsBChannelPixelized() {
    int add = exposed_blue >= 0.13 ? 1 : 0;
    if (db > 0.38) {
      return 2 + add;
    }
    if (db > 0.33) {
      return 1 + add;
    }
    if (db > 0.28) {
      return add;
    }
    return 0;
  }
};

}  // namespace

class LossyFrameEncoder {
 public:
  LossyFrameEncoder(const CompressParams& cparams,
                    const FrameHeader& frame_header,
                    PassesEncoderState* JXL_RESTRICT enc_state,
                    const JxlCmsInterface& cms, ThreadPool* pool,
                    AuxOut* aux_out)
      : enc_state_(enc_state), cms_(cms), pool_(pool), aux_out_(aux_out) {
    JXL_CHECK(InitializePassesSharedState(frame_header, &enc_state_->shared,
                                          /*encoder=*/true));
    enc_state_->cparams = cparams;
    enc_state_->passes.clear();
  }

  Status ComputeEncodingData(const ImageBundle* linear,
                             Image3F* JXL_RESTRICT opsin,
                             const JxlCmsInterface& cms, ThreadPool* pool,
                             ModularFrameEncoder* modular_frame_encoder,
                             FrameHeader* frame_header) {
    PROFILER_ZONE("ComputeEncodingData uninstrumented");
    JXL_ASSERT((opsin->xsize() % kBlockDim) == 0 &&
               (opsin->ysize() % kBlockDim) == 0);
    PassesSharedState& shared = enc_state_->shared;

    if (!enc_state_->cparams.max_error_mode) {
      // Compute chromacity adjustments using two approaches.
      // 1) Distance based approach for chromacity adjustment:
      float x_qm_scale_steps[4] = {1.25f, 7.0f, 15.0f, 24.0f};
      shared.frame_header.x_qm_scale = 2;
      for (float x_qm_scale_step : x_qm_scale_steps) {
        if (enc_state_->cparams.original_butteraugli_distance >
            x_qm_scale_step) {
          shared.frame_header.x_qm_scale++;
        }
      }
      if (enc_state_->cparams.butteraugli_distance < 0.299f) {
        // Favor chromacity preservation for making images appear more
        // faithful to original even with extreme (5-10x) zooming.
        shared.frame_header.x_qm_scale++;
      }
      // 2) Pixel-based approach for chromacity adjustment:
      // look at the individual pixels and make a guess how difficult
      // the image would be based on the worst case pixel.
      PixelStatsForChromacityAdjustment pixel_stats;
      if (enc_state_->cparams.speed_tier <= SpeedTier::kWombat) {
        pixel_stats.Calc(opsin);
      }
      // For X take the most severe adjustment.
      shared.frame_header.x_qm_scale =
          std::max<int>(shared.frame_header.x_qm_scale,
                        2 + pixel_stats.HowMuchIsXChannelPixelized());
      // B only ajudsted by pixel-based approach.
      shared.frame_header.b_qm_scale =
          2 + pixel_stats.HowMuchIsBChannelPixelized();
    }

    JXL_RETURN_IF_ERROR(enc_state_->heuristics->LossyFrameHeuristics(
        enc_state_, modular_frame_encoder, linear, opsin, cms_, pool_,
        aux_out_));

    JXL_RETURN_IF_ERROR(InitializePassesEncoder(
        *opsin, cms, pool_, enc_state_, modular_frame_encoder, aux_out_));

    enc_state_->passes.resize(enc_state_->progressive_splitter.GetNumPasses());
    for (PassesEncoderState::PassData& pass : enc_state_->passes) {
      pass.ac_tokens.resize(shared.frame_dim.num_groups);
    }

    ComputeAllCoeffOrders(shared.frame_dim);
    shared.num_histograms = 1;

    const auto tokenize_group_init = [&](const size_t num_threads) {
      group_caches_.resize(num_threads);
      return true;
    };
    const auto tokenize_group = [&](const uint32_t group_index,
                                    const size_t thread) {
      // Tokenize coefficients.
      const Rect rect = shared.BlockGroupRect(group_index);
      for (size_t idx_pass = 0; idx_pass < enc_state_->passes.size();
           idx_pass++) {
        JXL_ASSERT(enc_state_->coeffs[idx_pass]->Type() == ACType::k32);
        const int32_t* JXL_RESTRICT ac_rows[3] = {
            enc_state_->coeffs[idx_pass]->PlaneRow(0, group_index, 0).ptr32,
            enc_state_->coeffs[idx_pass]->PlaneRow(1, group_index, 0).ptr32,
            enc_state_->coeffs[idx_pass]->PlaneRow(2, group_index, 0).ptr32,
        };
        // Ensure group cache is initialized.
        group_caches_[thread].InitOnce();
        TokenizeCoefficients(
            &shared.coeff_orders[idx_pass * shared.coeff_order_size], rect,
            ac_rows, shared.ac_strategy, frame_header->chroma_subsampling,
            &group_caches_[thread].num_nzeroes,
            &enc_state_->passes[idx_pass].ac_tokens[group_index],
            enc_state_->shared.quant_dc, enc_state_->shared.raw_quant_field,
            enc_state_->shared.block_ctx_map);
      }
    };
    JXL_RETURN_IF_ERROR(RunOnPool(pool_, 0, shared.frame_dim.num_groups,
                                  tokenize_group_init, tokenize_group,
                                  "TokenizeGroup"));

    *frame_header = shared.frame_header;
    return true;
  }

  Status ComputeJPEGTranscodingData(const jpeg::JPEGData& jpeg_data,
                                    ModularFrameEncoder* modular_frame_encoder,
                                    FrameHeader* frame_header) {
    PROFILER_ZONE("ComputeJPEGTranscodingData uninstrumented");
    PassesSharedState& shared = enc_state_->shared;

    frame_header->x_qm_scale = 2;
    frame_header->b_qm_scale = 2;

    FrameDimensions frame_dim = frame_header->ToFrameDimensions();

    const size_t xsize = frame_dim.xsize_padded;
    const size_t ysize = frame_dim.ysize_padded;
    const size_t xsize_blocks = frame_dim.xsize_blocks;
    const size_t ysize_blocks = frame_dim.ysize_blocks;

    // no-op chroma from luma
    shared.cmap = ColorCorrelationMap(xsize, ysize, false);
    shared.ac_strategy.FillDCT8();
    FillImage(uint8_t(0), &shared.epf_sharpness);

    enc_state_->passes.resize(enc_state_->progressive_splitter.GetNumPasses());
    for (PassesEncoderState::PassData& pass : enc_state_->passes) {
      pass.ac_tokens.resize(shared.frame_dim.num_groups);
    }

    enc_state_->coeffs.clear();
    while (enc_state_->coeffs.size() < enc_state_->passes.size()) {
      enc_state_->coeffs.emplace_back(make_unique<ACImageT<int32_t>>(
          kGroupDim * kGroupDim, frame_dim.num_groups));
    }

    // convert JPEG quantization table to a Quantizer object
    float dcquantization[3];
    std::vector<QuantEncoding> qe(DequantMatrices::kNum,
                                  QuantEncoding::Library(0));

    auto jpeg_c_map = JpegOrder(frame_header->color_transform,
                                jpeg_data.components.size() == 1);

    std::vector<int> qt(192);
    for (size_t c = 0; c < 3; c++) {
      size_t jpeg_c = jpeg_c_map[c];
      const int32_t* quant =
          jpeg_data.quant[jpeg_data.components[jpeg_c].quant_idx].values.data();

      dcquantization[c] = 255 * 8.0f / quant[0];
      for (size_t y = 0; y < 8; y++) {
        for (size_t x = 0; x < 8; x++) {
          // JPEG XL transposes the DCT, JPEG doesn't.
          qt[c * 64 + 8 * x + y] = quant[8 * y + x];
        }
      }
    }
    DequantMatricesSetCustomDC(&shared.matrices, dcquantization);
    float dcquantization_r[3] = {1.0f / dcquantization[0],
                                 1.0f / dcquantization[1],
                                 1.0f / dcquantization[2]};

    qe[AcStrategy::Type::DCT] = QuantEncoding::RAW(qt);
    DequantMatricesSetCustom(&shared.matrices, qe, modular_frame_encoder);

    // Ensure that InvGlobalScale() is 1.
    shared.quantizer = Quantizer(&shared.matrices, 1, kGlobalScaleDenom);
    // Recompute MulDC() and InvMulDC().
    shared.quantizer.RecomputeFromGlobalScale();

    // Per-block dequant scaling should be 1.
    FillImage(static_cast<int32_t>(shared.quantizer.InvGlobalScale()),
              &shared.raw_quant_field);

    std::vector<int32_t> scaled_qtable(192);
    for (size_t c = 0; c < 3; c++) {
      for (size_t i = 0; i < 64; i++) {
        scaled_qtable[64 * c + i] =
            (1 << kCFLFixedPointPrecision) * qt[64 + i] / qt[64 * c + i];
      }
    }

    auto jpeg_row = [&](size_t c, size_t y) {
      return jpeg_data.components[jpeg_c_map[c]].coeffs.data() +
             jpeg_data.components[jpeg_c_map[c]].width_in_blocks *
                 kDCTBlockSize * y;
    };

    Image3F dc = Image3F(xsize_blocks, ysize_blocks);
    bool DCzero =
        (shared.frame_header.color_transform == ColorTransform::kYCbCr);
    // Compute chroma-from-luma for AC (doesn't seem to be useful for DC)
    if (frame_header->chroma_subsampling.Is444() &&
        enc_state_->cparams.force_cfl_jpeg_recompression &&
        jpeg_data.components.size() == 3) {
      for (size_t c : {0, 2}) {
        ImageSB* map = (c == 0 ? &shared.cmap.ytox_map : &shared.cmap.ytob_map);
        const float kScale = kDefaultColorFactor;
        const int kOffset = 127;
        const float kBase =
            c == 0 ? shared.cmap.YtoXRatio(0) : shared.cmap.YtoBRatio(0);
        const float kZeroThresh =
            kScale * kZeroBiasDefault[c] *
            0.9999f;  // just epsilon less for better rounding

        auto process_row = [&](const uint32_t task, const size_t thread) {
          size_t ty = task;
          int8_t* JXL_RESTRICT row_out = map->Row(ty);
          for (size_t tx = 0; tx < map->xsize(); ++tx) {
            const size_t y0 = ty * kColorTileDimInBlocks;
            const size_t x0 = tx * kColorTileDimInBlocks;
            const size_t y1 = std::min(frame_dim.ysize_blocks,
                                       (ty + 1) * kColorTileDimInBlocks);
            const size_t x1 = std::min(frame_dim.xsize_blocks,
                                       (tx + 1) * kColorTileDimInBlocks);
            int32_t d_num_zeros[257] = {0};
            // TODO(veluca): this needs SIMD + fixed point adaptation, and/or
            // conversion to the new CfL algorithm.
            for (size_t y = y0; y < y1; ++y) {
              const int16_t* JXL_RESTRICT row_m = jpeg_row(1, y);
              const int16_t* JXL_RESTRICT row_s = jpeg_row(c, y);
              for (size_t x = x0; x < x1; ++x) {
                for (size_t coeffpos = 1; coeffpos < kDCTBlockSize;
                     coeffpos++) {
                  const float scaled_m =
                      row_m[x * kDCTBlockSize + coeffpos] *
                      scaled_qtable[64 * c + coeffpos] *
                      (1.0f / (1 << kCFLFixedPointPrecision));
                  const float scaled_s =
                      kScale * row_s[x * kDCTBlockSize + coeffpos] +
                      (kOffset - kBase * kScale) * scaled_m;
                  if (std::abs(scaled_m) > 1e-8f) {
                    float from, to;
                    if (scaled_m > 0) {
                      from = (scaled_s - kZeroThresh) / scaled_m;
                      to = (scaled_s + kZeroThresh) / scaled_m;
                    } else {
                      from = (scaled_s + kZeroThresh) / scaled_m;
                      to = (scaled_s - kZeroThresh) / scaled_m;
                    }
                    if (from < 0.0f) {
                      from = 0.0f;
                    }
                    if (to > 255.0f) {
                      to = 255.0f;
                    }
                    // Instead of clamping the both values
                    // we just check that range is sane.
                    if (from <= to) {
                      d_num_zeros[static_cast<int>(std::ceil(from))]++;
                      d_num_zeros[static_cast<int>(std::floor(to + 1))]--;
                    }
                  }
                }
              }
            }
            int best = 0;
            int32_t best_sum = 0;
            FindIndexOfSumMaximum(d_num_zeros, 256, &best, &best_sum);
            int32_t offset_sum = 0;
            for (int i = 0; i < 256; ++i) {
              if (i <= kOffset) {
                offset_sum += d_num_zeros[i];
              }
            }
            row_out[tx] = 0;
            if (best_sum > offset_sum + 1) {
              row_out[tx] = best - kOffset;
            }
          }
        };

        JXL_RETURN_IF_ERROR(RunOnPool(pool_, 0, map->ysize(),
                                      ThreadPool::NoInit, process_row,
                                      "FindCorrelation"));
      }
    }

    if (!frame_header->chroma_subsampling.Is444()) {
      ZeroFillImage(&dc);
      for (auto& coeff : enc_state_->coeffs) {
        coeff->ZeroFill();
      }
    }
    // JPEG DC is from -1024 to 1023.
    std::vector<size_t> dc_counts[3] = {};
    dc_counts[0].resize(2048);
    dc_counts[1].resize(2048);
    dc_counts[2].resize(2048);
    size_t total_dc[3] = {};
    for (size_t c : {1, 0, 2}) {
      if (jpeg_data.components.size() == 1 && c != 1) {
        for (auto& coeff : enc_state_->coeffs) {
          coeff->ZeroFillPlane(c);
        }
        ZeroFillImage(&dc.Plane(c));
        // Ensure no division by 0.
        dc_counts[c][1024] = 1;
        total_dc[c] = 1;
        continue;
      }
      size_t hshift = frame_header->chroma_subsampling.HShift(c);
      size_t vshift = frame_header->chroma_subsampling.VShift(c);
      ImageSB& map = (c == 0 ? shared.cmap.ytox_map : shared.cmap.ytob_map);
      for (size_t group_index = 0; group_index < frame_dim.num_groups;
           group_index++) {
        const size_t gx = group_index % frame_dim.xsize_groups;
        const size_t gy = group_index / frame_dim.xsize_groups;
        int32_t* coeffs[kMaxNumPasses];
        for (size_t i = 0; i < enc_state_->coeffs.size(); i++) {
          coeffs[i] = enc_state_->coeffs[i]->PlaneRow(c, group_index, 0).ptr32;
        }
        int32_t block[64];
        for (size_t by = gy * kGroupDimInBlocks;
             by < ysize_blocks && by < (gy + 1) * kGroupDimInBlocks; ++by) {
          if ((by >> vshift) << vshift != by) continue;
          const int16_t* JXL_RESTRICT inputjpeg = jpeg_row(c, by >> vshift);
          const int16_t* JXL_RESTRICT inputjpegY = jpeg_row(1, by);
          float* JXL_RESTRICT fdc = dc.PlaneRow(c, by >> vshift);
          const int8_t* JXL_RESTRICT cm =
              map.ConstRow(by / kColorTileDimInBlocks);
          for (size_t bx = gx * kGroupDimInBlocks;
               bx < xsize_blocks && bx < (gx + 1) * kGroupDimInBlocks; ++bx) {
            if ((bx >> hshift) << hshift != bx) continue;
            size_t base = (bx >> hshift) * kDCTBlockSize;
            int idc;
            if (DCzero) {
              idc = inputjpeg[base];
            } else {
              idc = inputjpeg[base] + 1024 / qt[c * 64];
            }
            dc_counts[c][std::min(static_cast<uint32_t>(idc + 1024),
                                  uint32_t(2047))]++;
            total_dc[c]++;
            fdc[bx >> hshift] = idc * dcquantization_r[c];
            if (c == 1 || !enc_state_->cparams.force_cfl_jpeg_recompression ||
                !frame_header->chroma_subsampling.Is444()) {
              for (size_t y = 0; y < 8; y++) {
                for (size_t x = 0; x < 8; x++) {
                  block[y * 8 + x] = inputjpeg[base + x * 8 + y];
                }
              }
            } else {
              const int32_t scale =
                  shared.cmap.RatioJPEG(cm[bx / kColorTileDimInBlocks]);

              for (size_t y = 0; y < 8; y++) {
                for (size_t x = 0; x < 8; x++) {
                  int Y = inputjpegY[kDCTBlockSize * bx + x * 8 + y];
                  int QChroma = inputjpeg[kDCTBlockSize * bx + x * 8 + y];
                  // Fixed-point multiply of CfL scale with quant table ratio
                  // first, and Y value second.
                  int coeff_scale = (scale * scaled_qtable[64 * c + y * 8 + x] +
                                     (1 << (kCFLFixedPointPrecision - 1))) >>
                                    kCFLFixedPointPrecision;
                  int cfl_factor = (Y * coeff_scale +
                                    (1 << (kCFLFixedPointPrecision - 1))) >>
                                   kCFLFixedPointPrecision;
                  int QCR = QChroma - cfl_factor;
                  block[y * 8 + x] = QCR;
                }
              }
            }
            enc_state_->progressive_splitter.SplitACCoefficients(
                block, AcStrategy::FromRawStrategy(AcStrategy::Type::DCT), bx,
                by, coeffs);
            for (size_t i = 0; i < enc_state_->coeffs.size(); i++) {
              coeffs[i] += kDCTBlockSize;
            }
          }
        }
      }
    }

    auto& dct = enc_state_->shared.block_ctx_map.dc_thresholds;
    auto& num_dc_ctxs = enc_state_->shared.block_ctx_map.num_dc_ctxs;
    num_dc_ctxs = 1;
    for (size_t i = 0; i < 3; i++) {
      dct[i].clear();
      int num_thresholds = (CeilLog2Nonzero(total_dc[i]) - 12) / 2;
      // up to 3 buckets per channel:
      // dark/medium/bright, yellow/unsat/blue, green/unsat/red
      num_thresholds = std::min(std::max(num_thresholds, 0), 2);
      size_t cumsum = 0;
      size_t cut = total_dc[i] / (num_thresholds + 1);
      for (int j = 0; j < 2048; j++) {
        cumsum += dc_counts[i][j];
        if (cumsum > cut) {
          dct[i].push_back(j - 1025);
          cut = total_dc[i] * (dct[i].size() + 1) / (num_thresholds + 1);
        }
      }
      num_dc_ctxs *= dct[i].size() + 1;
    }

    auto& ctx_map = enc_state_->shared.block_ctx_map.ctx_map;
    ctx_map.clear();
    ctx_map.resize(3 * kNumOrders * num_dc_ctxs, 0);

    int lbuckets = (dct[1].size() + 1);
    for (size_t i = 0; i < num_dc_ctxs; i++) {
      // up to 9 contexts for luma
      ctx_map[i] = i / lbuckets;
      // up to 3 contexts for chroma
      ctx_map[kNumOrders * num_dc_ctxs + i] =
          ctx_map[2 * kNumOrders * num_dc_ctxs + i] =
              num_dc_ctxs / lbuckets + (i % lbuckets);
    }
    enc_state_->shared.block_ctx_map.num_ctxs =
        *std::max_element(ctx_map.begin(), ctx_map.end()) + 1;

    enc_state_->histogram_idx.resize(shared.frame_dim.num_groups);

    // disable DC frame for now
    shared.frame_header.UpdateFlag(false, FrameHeader::kUseDcFrame);
    auto compute_dc_coeffs = [&](const uint32_t group_index,
                                 size_t /* thread */) {
      modular_frame_encoder->AddVarDCTDC(dc, group_index, /*nl_dc=*/false,
                                         enc_state_, /*jpeg_transcode=*/true);
      modular_frame_encoder->AddACMetadata(group_index, /*jpeg_transcode=*/true,
                                           enc_state_);
    };
    JXL_RETURN_IF_ERROR(RunOnPool(pool_, 0, shared.frame_dim.num_dc_groups,
                                  ThreadPool::NoInit, compute_dc_coeffs,
                                  "Compute DC coeffs"));

    // Must happen before WriteFrameHeader!
    shared.frame_header.UpdateFlag(true, FrameHeader::kSkipAdaptiveDCSmoothing);

    ComputeAllCoeffOrders(frame_dim);
    shared.num_histograms = 1;

    const auto tokenize_group_init = [&](const size_t num_threads) {
      group_caches_.resize(num_threads);
      return true;
    };
    const auto tokenize_group = [&](const uint32_t group_index,
                                    const size_t thread) {
      // Tokenize coefficients.
      const Rect rect = shared.BlockGroupRect(group_index);
      for (size_t idx_pass = 0; idx_pass < enc_state_->passes.size();
           idx_pass++) {
        JXL_ASSERT(enc_state_->coeffs[idx_pass]->Type() == ACType::k32);
        const int32_t* JXL_RESTRICT ac_rows[3] = {
            enc_state_->coeffs[idx_pass]->PlaneRow(0, group_index, 0).ptr32,
            enc_state_->coeffs[idx_pass]->PlaneRow(1, group_index, 0).ptr32,
            enc_state_->coeffs[idx_pass]->PlaneRow(2, group_index, 0).ptr32,
        };
        // Ensure group cache is initialized.
        group_caches_[thread].InitOnce();
        TokenizeCoefficients(
            &shared.coeff_orders[idx_pass * shared.coeff_order_size], rect,
            ac_rows, shared.ac_strategy, frame_header->chroma_subsampling,
            &group_caches_[thread].num_nzeroes,
            &enc_state_->passes[idx_pass].ac_tokens[group_index],
            enc_state_->shared.quant_dc, enc_state_->shared.raw_quant_field,
            enc_state_->shared.block_ctx_map);
      }
    };
    JXL_RETURN_IF_ERROR(RunOnPool(pool_, 0, shared.frame_dim.num_groups,
                                  tokenize_group_init, tokenize_group,
                                  "TokenizeGroup"));
    *frame_header = shared.frame_header;
    doing_jpeg_recompression = true;
    return true;
  }

  Status EncodeGlobalDCInfo(const FrameHeader& frame_header,
                            BitWriter* writer) const {
    // Encode quantizer DC and global scale.
    QuantizerParams params = enc_state_->shared.quantizer.GetParams();
    JXL_RETURN_IF_ERROR(
        WriteQuantizerParams(params, writer, kLayerQuant, aux_out_));
    EncodeBlockCtxMap(enc_state_->shared.block_ctx_map, writer, aux_out_);
    ColorCorrelationMapEncodeDC(&enc_state_->shared.cmap, writer, kLayerDC,
                                aux_out_);
    return true;
  }

  Status EncodeGlobalACInfo(BitWriter* writer,
                            ModularFrameEncoder* modular_frame_encoder) {
    JXL_RETURN_IF_ERROR(DequantMatricesEncode(&enc_state_->shared.matrices,
                                              writer, kLayerQuant, aux_out_,
                                              modular_frame_encoder));
    if (enc_state_->cparams.speed_tier <= SpeedTier::kTortoise) {
      if (!doing_jpeg_recompression) ClusterGroups(enc_state_);
    }
    size_t num_histo_bits =
        CeilLog2Nonzero(enc_state_->shared.frame_dim.num_groups);
    if (num_histo_bits != 0) {
      BitWriter::Allotment allotment(writer, num_histo_bits);
      writer->Write(num_histo_bits, enc_state_->shared.num_histograms - 1);
      allotment.ReclaimAndCharge(writer, kLayerAC, aux_out_);
    }

    for (size_t i = 0; i < enc_state_->progressive_splitter.GetNumPasses();
         i++) {
      // Encode coefficient orders.
      size_t order_bits = 0;
      JXL_RETURN_IF_ERROR(U32Coder::CanEncode(
          kOrderEnc, enc_state_->used_orders[i], &order_bits));
      BitWriter::Allotment allotment(writer, order_bits);
      JXL_CHECK(U32Coder::Write(kOrderEnc, enc_state_->used_orders[i], writer));
      allotment.ReclaimAndCharge(writer, kLayerOrder, aux_out_);
      EncodeCoeffOrders(
          enc_state_->used_orders[i],
          &enc_state_->shared
               .coeff_orders[i * enc_state_->shared.coeff_order_size],
          writer, kLayerOrder, aux_out_);

      // Encode histograms.
      HistogramParams hist_params(
          enc_state_->cparams.speed_tier,
          enc_state_->shared.block_ctx_map.NumACContexts());
      if (enc_state_->cparams.speed_tier > SpeedTier::kTortoise) {
        hist_params.lz77_method = HistogramParams::LZ77Method::kNone;
      }
      if (enc_state_->cparams.decoding_speed_tier >= 1) {
        hist_params.max_histograms = 6;
      }
      BuildAndEncodeHistograms(
          hist_params,
          enc_state_->shared.num_histograms *
              enc_state_->shared.block_ctx_map.NumACContexts(),
          enc_state_->passes[i].ac_tokens, &enc_state_->passes[i].codes,
          &enc_state_->passes[i].context_map, writer, kLayerAC, aux_out_);
    }

    return true;
  }

  Status EncodeACGroup(size_t pass, size_t group_index, BitWriter* group_code,
                       AuxOut* local_aux_out) {
    return EncodeGroupTokenizedCoefficients(
        group_index, pass, enc_state_->histogram_idx[group_index], *enc_state_,
        group_code, local_aux_out);
  }

  PassesEncoderState* State() { return enc_state_; }

 private:
  void ComputeAllCoeffOrders(const FrameDimensions& frame_dim) {
    PROFILER_FUNC;
    // No coefficient reordering in Falcon or faster.
    auto used_orders_info = ComputeUsedOrders(
        enc_state_->cparams.speed_tier, enc_state_->shared.ac_strategy,
        Rect(enc_state_->shared.raw_quant_field));
    enc_state_->used_orders.clear();
    enc_state_->used_orders.resize(
        enc_state_->progressive_splitter.GetNumPasses(),
        used_orders_info.second);
    for (size_t i = 0; i < enc_state_->progressive_splitter.GetNumPasses();
         i++) {
      ComputeCoeffOrder(
          enc_state_->cparams.speed_tier, *enc_state_->coeffs[i],
          enc_state_->shared.ac_strategy, frame_dim, enc_state_->used_orders[i],
          used_orders_info.first,
          &enc_state_->shared
               .coeff_orders[i * enc_state_->shared.coeff_order_size]);
    }
  }

  template <typename V, typename R>
  static inline void FindIndexOfSumMaximum(const V* array, const size_t len,
                                           R* idx, V* sum) {
    JXL_ASSERT(len > 0);
    V maxval = 0;
    V val = 0;
    R maxidx = 0;
    for (size_t i = 0; i < len; ++i) {
      val += array[i];
      if (val > maxval) {
        maxval = val;
        maxidx = i;
      }
    }
    *idx = maxidx;
    *sum = maxval;
  }

  PassesEncoderState* JXL_RESTRICT enc_state_;
  JxlCmsInterface cms_;
  ThreadPool* pool_;
  AuxOut* aux_out_;
  std::vector<EncCache> group_caches_;
  bool doing_jpeg_recompression = false;
};

Status ParamsPostInit(CompressParams* p) {
  if (!p->manual_noise.empty() &&
      p->manual_noise.size() != NoiseParams::kNumNoisePoints) {
    return JXL_FAILURE("Invalid number of noise lut entries");
  }
  if (!p->manual_xyb_factors.empty() && p->manual_xyb_factors.size() != 3) {
    return JXL_FAILURE("Invalid number of XYB quantization factors");
  }
  if (!p->modular_mode && p->butteraugli_distance == 0.0) {
    p->butteraugli_distance = kMinButteraugliDistance;
  }
  if (p->original_butteraugli_distance == -1.0) {
    p->original_butteraugli_distance = p->butteraugli_distance;
  }
  if (p->resampling <= 0) {
    p->resampling = 1;
    // For very low bit rates, using 2x2 resampling gives better results on
    // most photographic images, with an adjusted butteraugli score chosen to
    // give roughly the same amount of bits per pixel.
    if (!p->already_downsampled && p->butteraugli_distance >= 20) {
      p->resampling = 2;
      p->butteraugli_distance = 6 + ((p->butteraugli_distance - 20) * 0.25);
    }
  }
  if (p->ec_resampling <= 0) {
    p->ec_resampling = p->resampling;
  }
  return true;
}

Status EncodeFrame(const CompressParams& cparams_orig,
                   const FrameInfo& frame_info, const CodecMetadata* metadata,
                   const ImageBundle& ib, PassesEncoderState* passes_enc_state,
                   const JxlCmsInterface& cms, ThreadPool* pool,
                   BitWriter* writer, AuxOut* aux_out) {
  CompressParams cparams = cparams_orig;
  if (cparams.speed_tier == SpeedTier::kGlacier && !cparams.IsLossless()) {
    cparams.speed_tier = SpeedTier::kTortoise;
  }
  if (cparams.speed_tier == SpeedTier::kGlacier) {
    std::vector<CompressParams> all_params;
    std::vector<size_t> size;

    CompressParams cparams_attempt = cparams_orig;
    cparams_attempt.speed_tier = SpeedTier::kTortoise;
    cparams_attempt.options.max_properties = 4;

    for (float x : {0.0f, 80.f}) {
      cparams_attempt.channel_colors_percent = x;
      for (float y : {0.0f, 95.0f}) {
        cparams_attempt.channel_colors_pre_transform_percent = y;
        // 70000 ensures that the number of palette colors is representable in
        // modular headers.
        for (int K : {0, 1 << 10, 70000}) {
          cparams_attempt.palette_colors = K;
          for (int tree_mode : {-1, (int)ModularOptions::TreeMode::kNoWP,
                                (int)ModularOptions::TreeMode::kDefault}) {
            if (tree_mode == -1) {
              // LZ77 only
              cparams_attempt.options.nb_repeats = 0;
            } else {
              cparams_attempt.options.nb_repeats = 1;
              cparams_attempt.options.wp_tree_mode =
                  static_cast<ModularOptions::TreeMode>(tree_mode);
            }
            for (Predictor pred : {Predictor::Zero, Predictor::Variable}) {
              cparams_attempt.options.predictor = pred;
              for (int g : {0, 1, 3}) {
                cparams_attempt.modular_group_size_shift = g;
                for (Override patches : {Override::kDefault, Override::kOff}) {
                  cparams_attempt.patches = patches;
                  all_params.push_back(cparams_attempt);
                }
              }
            }
          }
        }
      }
    }

    size.resize(all_params.size());

    std::atomic<int> num_errors{0};

    JXL_RETURN_IF_ERROR(RunOnPool(
        pool, 0, all_params.size(), ThreadPool::NoInit,
        [&](size_t task, size_t) {
          BitWriter w;
          PassesEncoderState state;
          if (!EncodeFrame(all_params[task], frame_info, metadata, ib, &state,
                           cms, nullptr, &w, aux_out)) {
            num_errors.fetch_add(1, std::memory_order_relaxed);
            return;
          }
          size[task] = w.BitsWritten();
        },
        "Compress kGlacier"));
    JXL_RETURN_IF_ERROR(num_errors.load(std::memory_order_relaxed) == 0);

    size_t best_idx = 0;
    for (size_t i = 1; i < all_params.size(); i++) {
      if (size[best_idx] > size[i]) {
        best_idx = i;
      }
    }
    cparams = all_params[best_idx];
  }

  if (cparams_orig.target_bitrate > 0.0f &&
      frame_info.frame_type == FrameType::kRegularFrame) {
    cparams.target_bitrate = 0.0f;
    const float target_bitrate = cparams_orig.target_bitrate;
    float bitrate = 0.0f;
    float prev_bitrate = 0.0f;
    float rescale = 1.0f;
    size_t prev_bits = 0;
    float error = 0.0f;
    float best_error = 100.0f;
    float best_rescale = 1.0f;
    for (size_t i = 0; i < 10; ++i) {
      std::unique_ptr<PassesEncoderState> state =
          jxl::make_unique<PassesEncoderState>();
      BitWriter bw;
      JXL_CHECK(EncodeFrame(cparams, frame_info, metadata, ib, state.get(), cms,
                            pool, &bw, nullptr));
      bitrate = bw.BitsWritten() * 1.0 / (ib.xsize() * ib.ysize());
      error = target_bitrate / bitrate - 1.0f;
      if (std::abs(error) < std::abs(best_error)) {
        best_error = error;
        best_rescale = cparams.quant_ac_rescale;
      }
      if (bw.BitsWritten() == prev_bits || std::abs(error) < 0.0005f) {
        break;
      }
      float lambda = 1.0f;
      if (i > 0) {
        lambda = (((bitrate / prev_bitrate) - 1.0f) / (rescale - 1.0f));
      }
      rescale = (1.0f + ((target_bitrate / bitrate) - 1.0f) / lambda);
      if (rescale < 0.0f) {
        break;
      }
      cparams.quant_ac_rescale *= rescale;
      prev_bitrate = bitrate;
      prev_bits = bw.BitsWritten();
    }
    if (aux_out) {
      aux_out->max_quant_rescale = best_rescale;
      aux_out->min_quant_rescale = best_rescale;
      aux_out->min_bitrate_error = best_error;
      aux_out->max_bitrate_error = best_error;
    }
    cparams.quant_ac_rescale = best_rescale;
  }
  ib.VerifyMetadata();

  passes_enc_state->special_frames.clear();

  if (cparams.qprogressive_mode) {
    passes_enc_state->progressive_splitter.SetProgressiveMode(
        ProgressiveMode{progressive_passes_dc_quant_ac_full_ac});
  } else if (cparams.progressive_mode) {
    passes_enc_state->progressive_splitter.SetProgressiveMode(
        ProgressiveMode{progressive_passes_dc_vlf_lf_full_ac});
  }

  JXL_RETURN_IF_ERROR(ParamsPostInit(&cparams));

  if (cparams.progressive_dc < 0) {
    if (cparams.progressive_dc != -1) {
      return JXL_FAILURE("Invalid progressive DC setting value (%d)",
                         cparams.progressive_dc);
    }
    cparams.progressive_dc = 0;
  }
  if (cparams.ec_resampling < cparams.resampling) {
    cparams.ec_resampling = cparams.resampling;
  }
  if (cparams.resampling > 1 || frame_info.is_preview) {
    cparams.progressive_dc = 0;
  }

  if (frame_info.dc_level + cparams.progressive_dc > 4) {
    return JXL_FAILURE("Too many levels of progressive DC");
  }

  if (cparams.butteraugli_distance != 0 &&
      cparams.butteraugli_distance < kMinButteraugliDistance) {
    return JXL_FAILURE("Butteraugli distance is too low (%f)",
                       cparams.butteraugli_distance);
  }

  if (ib.IsJPEG()) {
    cparams.gaborish = Override::kOff;
    cparams.epf = 0;
    cparams.modular_mode = false;
  }

  if (ib.xsize() == 0 || ib.ysize() == 0) return JXL_FAILURE("Empty image");

  // Assert that this metadata is correctly set up for the compression params,
  // this should have been done by enc_file.cc
  JXL_ASSERT(metadata->m.xyb_encoded ==
             (cparams.color_transform == ColorTransform::kXYB));
  std::unique_ptr<FrameHeader> frame_header =
      jxl::make_unique<FrameHeader>(metadata);
  JXL_RETURN_IF_ERROR(MakeFrameHeader(cparams,
                                      passes_enc_state->progressive_splitter,
                                      frame_info, ib, frame_header.get()));
  // Check that if the codestream header says xyb_encoded, the color_transform
  // matches the requirement. This is checked from the cparams here, even though
  // optimally we'd be able to check this against what has actually been written
  // in the main codestream header, but since ib is a const object and the data
  // written to the main codestream header is (in modified form) in ib, the
  // encoder cannot indicate this fact in the ib's metadata.
  if (cparams_orig.color_transform == ColorTransform::kXYB) {
    if (frame_header->color_transform != ColorTransform::kXYB) {
      return JXL_FAILURE(
          "The color transform of frames must be xyb if the codestream is xyb "
          "encoded");
    }
  } else {
    if (frame_header->color_transform == ColorTransform::kXYB) {
      return JXL_FAILURE(
          "The color transform of frames cannot be xyb if the codestream is "
          "not xyb encoded");
    }
  }

  FrameDimensions frame_dim = frame_header->ToFrameDimensions();

  const size_t num_groups = frame_dim.num_groups;

  Image3F opsin;
  const ColorEncoding& c_linear = ColorEncoding::LinearSRGB(ib.IsGray());
  std::unique_ptr<ImageMetadata> metadata_linear =
      jxl::make_unique<ImageMetadata>();
  metadata_linear->xyb_encoded =
      (cparams.color_transform == ColorTransform::kXYB);
  metadata_linear->color_encoding = c_linear;
  ImageBundle linear_storage(metadata_linear.get());

  std::vector<AuxOut> aux_outs;
  // LossyFrameEncoder stores a reference to a std::function<Status(size_t)>
  // so we need to keep the std::function<Status(size_t)> being referenced
  // alive while lossy_frame_encoder is used. We could make resize_aux_outs a
  // lambda type by making LossyFrameEncoder a template instead, but this is
  // simpler.
  const std::function<Status(size_t)> resize_aux_outs =
      [&aux_outs, aux_out](const size_t num_threads) -> Status {
    if (aux_out != nullptr) {
      size_t old_size = aux_outs.size();
      for (size_t i = num_threads; i < old_size; i++) {
        aux_out->Assimilate(aux_outs[i]);
      }
      aux_outs.resize(num_threads);
      // Each thread needs these INPUTS. Don't copy the entire AuxOut
      // because it may contain stats which would be Assimilated multiple
      // times below.
      for (size_t i = old_size; i < aux_outs.size(); i++) {
        aux_outs[i].dump_image = aux_out->dump_image;
        aux_outs[i].debug_prefix = aux_out->debug_prefix;
      }
    }
    return true;
  };

  LossyFrameEncoder lossy_frame_encoder(cparams, *frame_header,
                                        passes_enc_state, cms, pool, aux_out);
  std::unique_ptr<ModularFrameEncoder> modular_frame_encoder =
      jxl::make_unique<ModularFrameEncoder>(*frame_header, cparams);

  const std::vector<ImageF>* extra_channels = &ib.extra_channels();
  std::vector<ImageF> extra_channels_storage;
  // Clear patches
  passes_enc_state->shared.image_features.patches = PatchDictionary();
  passes_enc_state->shared.image_features.patches.SetPassesSharedState(
      &passes_enc_state->shared);

  if (ib.IsJPEG()) {
    JXL_RETURN_IF_ERROR(lossy_frame_encoder.ComputeJPEGTranscodingData(
        *ib.jpeg_data, modular_frame_encoder.get(), frame_header.get()));
  } else if (!lossy_frame_encoder.State()->heuristics->HandlesColorConversion(
                 cparams, ib) ||
             frame_header->encoding != FrameEncoding::kVarDCT) {
    // Allocating a large enough image avoids a copy when padding.
    opsin =
        Image3F(RoundUpToBlockDim(ib.xsize()), RoundUpToBlockDim(ib.ysize()));
    opsin.ShrinkTo(ib.xsize(), ib.ysize());

    const bool want_linear = frame_header->encoding == FrameEncoding::kVarDCT &&
                             cparams.speed_tier <= SpeedTier::kKitten;
    const ImageBundle* JXL_RESTRICT ib_or_linear = &ib;

    if (frame_header->color_transform == ColorTransform::kXYB &&
        frame_info.ib_needs_color_transform) {
      // linear_storage would only be used by the Butteraugli loop (passing
      // linear sRGB avoids a color conversion there). Otherwise, don't
      // fill it to reduce memory usage.
      ib_or_linear =
          ToXYB(ib, pool, &opsin, cms, want_linear ? &linear_storage : nullptr);
    } else {  // RGB or YCbCr: don't do anything (forward YCbCr is not
              // implemented, this is only used when the input is already in
              // YCbCr)
              // If encoding a special DC or reference frame, don't do anything:
              // input is already in XYB.
      CopyImageTo(ib.color(), &opsin);
    }
    bool lossless = cparams.IsLossless();
    if (ib.HasAlpha() && !ib.AlphaIsPremultiplied() &&
        frame_header->frame_type == FrameType::kRegularFrame &&
        !ApplyOverride(cparams.keep_invisible, lossless) &&
        cparams.ec_resampling == cparams.resampling) {
      // simplify invisible pixels
      SimplifyInvisible(&opsin, ib.alpha(), lossless);
      if (want_linear) {
        SimplifyInvisible(const_cast<Image3F*>(&ib_or_linear->color()),
                          ib.alpha(), lossless);
      }
    }
    if (aux_out != nullptr) {
      JXL_RETURN_IF_ERROR(
          aux_out->InspectImage3F("enc_frame:OpsinDynamicsImage", opsin));
    }
    if (frame_header->encoding == FrameEncoding::kVarDCT) {
      PadImageToBlockMultipleInPlace(&opsin);
      JXL_RETURN_IF_ERROR(lossy_frame_encoder.ComputeEncodingData(
          ib_or_linear, &opsin, cms, pool, modular_frame_encoder.get(),
          frame_header.get()));
    } else if (frame_header->upsampling != 1 && !cparams.already_downsampled) {
      // In VarDCT mode, LossyFrameHeuristics takes care of running downsampling
      // after noise, if necessary.
      DownsampleImage(&opsin, frame_header->upsampling);
    }
  } else {
    JXL_RETURN_IF_ERROR(lossy_frame_encoder.ComputeEncodingData(
        &ib, &opsin, cms, pool, modular_frame_encoder.get(),
        frame_header.get()));
  }
  if (cparams.ec_resampling != 1 && !cparams.already_downsampled) {
    extra_channels = &extra_channels_storage;
    for (size_t i = 0; i < ib.extra_channels().size(); i++) {
      extra_channels_storage.emplace_back(CopyImage(ib.extra_channels()[i]));
      DownsampleImage(&extra_channels_storage.back(), cparams.ec_resampling);
    }
  }
  // needs to happen *AFTER* VarDCT-ComputeEncodingData.
  JXL_RETURN_IF_ERROR(modular_frame_encoder->ComputeEncodingData(
      *frame_header, *ib.metadata(), &opsin, *extra_channels,
      lossy_frame_encoder.State(), cms, pool, aux_out,
      /* do_color=*/frame_header->encoding == FrameEncoding::kModular));

  writer->AppendByteAligned(lossy_frame_encoder.State()->special_frames);
  frame_header->UpdateFlag(
      lossy_frame_encoder.State()->shared.image_features.patches.HasAny(),
      FrameHeader::kPatches);
  frame_header->UpdateFlag(
      lossy_frame_encoder.State()->shared.image_features.splines.HasAny(),
      FrameHeader::kSplines);
  JXL_RETURN_IF_ERROR(WriteFrameHeader(*frame_header, writer, aux_out));

  const size_t num_passes =
      passes_enc_state->progressive_splitter.GetNumPasses();

  // DC global info + DC groups + AC global info + AC groups *
  // num_passes.
  const bool has_ac_global = true;
  std::vector<BitWriter> group_codes(NumTocEntries(frame_dim.num_groups,
                                                   frame_dim.num_dc_groups,
                                                   num_passes, has_ac_global));
  const size_t global_ac_index = frame_dim.num_dc_groups + 1;
  const bool is_small_image = frame_dim.num_groups == 1 && num_passes == 1;
  const auto get_output = [&](const size_t index) {
    return &group_codes[is_small_image ? 0 : index];
  };
  auto ac_group_code = [&](size_t pass, size_t group) {
    return get_output(AcGroupIndex(pass, group, frame_dim.num_groups,
                                   frame_dim.num_dc_groups, has_ac_global));
  };

  if (frame_header->flags & FrameHeader::kPatches) {
    PatchDictionaryEncoder::Encode(
        lossy_frame_encoder.State()->shared.image_features.patches,
        get_output(0), kLayerDictionary, aux_out);
  }

  if (frame_header->flags & FrameHeader::kSplines) {
    EncodeSplines(lossy_frame_encoder.State()->shared.image_features.splines,
                  get_output(0), kLayerSplines, HistogramParams(), aux_out);
  }

  if (cparams.photon_noise_iso > 0) {
    lossy_frame_encoder.State()->shared.image_features.noise_params =
        SimulatePhotonNoise(ib.xsize(), ib.ysize(), cparams.photon_noise_iso);
  }
  if (cparams.manual_noise.size() == NoiseParams::kNumNoisePoints) {
    for (size_t i = 0; i < NoiseParams::kNumNoisePoints; i++) {
      lossy_frame_encoder.State()->shared.image_features.noise_params.lut[i] =
          cparams.manual_noise[i];
    }
  }
  if (frame_header->flags & FrameHeader::kNoise) {
    EncodeNoise(lossy_frame_encoder.State()->shared.image_features.noise_params,
                get_output(0), kLayerNoise, aux_out);
  }

  JXL_RETURN_IF_ERROR(
      DequantMatricesEncodeDC(&lossy_frame_encoder.State()->shared.matrices,
                              get_output(0), kLayerQuant, aux_out));
  if (frame_header->encoding == FrameEncoding::kVarDCT) {
    JXL_RETURN_IF_ERROR(
        lossy_frame_encoder.EncodeGlobalDCInfo(*frame_header, get_output(0)));
  }
  JXL_RETURN_IF_ERROR(
      modular_frame_encoder->EncodeGlobalInfo(get_output(0), aux_out));
  JXL_RETURN_IF_ERROR(modular_frame_encoder->EncodeStream(
      get_output(0), aux_out, kLayerModularGlobal, ModularStreamId::Global()));

  const auto process_dc_group = [&](const uint32_t group_index,
                                    const size_t thread) {
    AuxOut* my_aux_out = aux_out ? &aux_outs[thread] : nullptr;
    BitWriter* output = get_output(group_index + 1);
    if (frame_header->encoding == FrameEncoding::kVarDCT &&
        !(frame_header->flags & FrameHeader::kUseDcFrame)) {
      BitWriter::Allotment allotment(output, 2);
      output->Write(2, modular_frame_encoder->extra_dc_precision[group_index]);
      allotment.ReclaimAndCharge(output, kLayerDC, my_aux_out);
      JXL_CHECK(modular_frame_encoder->EncodeStream(
          output, my_aux_out, kLayerDC,
          ModularStreamId::VarDCTDC(group_index)));
    }
    JXL_CHECK(modular_frame_encoder->EncodeStream(
        output, my_aux_out, kLayerModularDcGroup,
        ModularStreamId::ModularDC(group_index)));
    if (frame_header->encoding == FrameEncoding::kVarDCT) {
      const Rect& rect =
          lossy_frame_encoder.State()->shared.DCGroupRect(group_index);
      size_t nb_bits = CeilLog2Nonzero(rect.xsize() * rect.ysize());
      if (nb_bits != 0) {
        BitWriter::Allotment allotment(output, nb_bits);
        output->Write(nb_bits,
                      modular_frame_encoder->ac_metadata_size[group_index] - 1);
        allotment.ReclaimAndCharge(output, kLayerControlFields, my_aux_out);
      }
      JXL_CHECK(modular_frame_encoder->EncodeStream(
          output, my_aux_out, kLayerControlFields,
          ModularStreamId::ACMetadata(group_index)));
    }
  };
  JXL_RETURN_IF_ERROR(RunOnPool(pool, 0, frame_dim.num_dc_groups,
                                resize_aux_outs, process_dc_group,
                                "EncodeDCGroup"));

  if (frame_header->encoding == FrameEncoding::kVarDCT) {
    JXL_RETURN_IF_ERROR(lossy_frame_encoder.EncodeGlobalACInfo(
        get_output(global_ac_index), modular_frame_encoder.get()));
  }

  std::atomic<int> num_errors{0};
  const auto process_group = [&](const uint32_t group_index,
                                 const size_t thread) {
    AuxOut* my_aux_out = aux_out ? &aux_outs[thread] : nullptr;

    for (size_t i = 0; i < num_passes; i++) {
      if (frame_header->encoding == FrameEncoding::kVarDCT) {
        if (!lossy_frame_encoder.EncodeACGroup(
                i, group_index, ac_group_code(i, group_index), my_aux_out)) {
          num_errors.fetch_add(1, std::memory_order_relaxed);
          return;
        }
      }
      // Write all modular encoded data (color?, alpha, depth, extra channels)
      if (!modular_frame_encoder->EncodeStream(
              ac_group_code(i, group_index), my_aux_out, kLayerModularAcGroup,
              ModularStreamId::ModularAC(group_index, i))) {
        num_errors.fetch_add(1, std::memory_order_relaxed);
        return;
      }
    }
  };
  JXL_RETURN_IF_ERROR(RunOnPool(pool, 0, num_groups, resize_aux_outs,
                                process_group, "EncodeGroupCoefficients"));

  // Resizing aux_outs to 0 also Assimilates the array.
  static_cast<void>(resize_aux_outs(0));
  JXL_RETURN_IF_ERROR(num_errors.load(std::memory_order_relaxed) == 0);

  for (BitWriter& bw : group_codes) {
    BitWriter::Allotment allotment(&bw, 8);
    bw.ZeroPadToByte();  // end of group.
    allotment.ReclaimAndCharge(&bw, kLayerAC, aux_out);
  }

  std::vector<coeff_order_t>* permutation_ptr = nullptr;
  std::vector<coeff_order_t> permutation;
  if (cparams.centerfirst && !(num_passes == 1 && num_groups == 1)) {
    permutation_ptr = &permutation;
    // Don't permute global DC/AC or DC.
    permutation.resize(global_ac_index + 1);
    std::iota(permutation.begin(), permutation.end(), 0);
    std::vector<coeff_order_t> ac_group_order(num_groups);
    std::iota(ac_group_order.begin(), ac_group_order.end(), 0);
    size_t group_dim = frame_dim.group_dim;

    // The center of the image is either given by parameters or chosen
    // to be the middle of the image by default if center_x, center_y resp.
    // are not provided.

    int64_t imag_cx;
    if (cparams.center_x != static_cast<size_t>(-1)) {
      JXL_RETURN_IF_ERROR(cparams.center_x < ib.xsize());
      imag_cx = cparams.center_x;
    } else {
      imag_cx = ib.xsize() / 2;
    }

    int64_t imag_cy;
    if (cparams.center_y != static_cast<size_t>(-1)) {
      JXL_RETURN_IF_ERROR(cparams.center_y < ib.ysize());
      imag_cy = cparams.center_y;
    } else {
      imag_cy = ib.ysize() / 2;
    }

    // The center of the group containing the center of the image.
    int64_t cx = (imag_cx / group_dim) * group_dim + group_dim / 2;
    int64_t cy = (imag_cy / group_dim) * group_dim + group_dim / 2;
    // This identifies in what area of the central group the center of the image
    // lies in.
    double direction = -std::atan2(imag_cy - cy, imag_cx - cx);
    // This identifies the side of the central group the center of the image
    // lies closest to. This can take values 0, 1, 2, 3 corresponding to left,
    // bottom, right, top.
    int64_t side = std::fmod((direction + 5 * kPi / 4), 2 * kPi) * 2 / kPi;
    auto get_distance_from_center = [&](size_t gid) {
      Rect r = passes_enc_state->shared.GroupRect(gid);
      int64_t gcx = r.x0() + group_dim / 2;
      int64_t gcy = r.y0() + group_dim / 2;
      int64_t dx = gcx - cx;
      int64_t dy = gcy - cy;
      // The angle is determined by taking atan2 and adding an appropriate
      // starting point depending on the side we want to start on.
      double angle = std::remainder(
          std::atan2(dy, dx) + kPi / 4 + side * (kPi / 2), 2 * kPi);
      // Concentric squares in clockwise order.
      return std::make_pair(std::max(std::abs(dx), std::abs(dy)), angle);
    };
    std::sort(ac_group_order.begin(), ac_group_order.end(),
              [&](coeff_order_t a, coeff_order_t b) {
                return get_distance_from_center(a) <
                       get_distance_from_center(b);
              });
    std::vector<coeff_order_t> inv_ac_group_order(ac_group_order.size(), 0);
    for (size_t i = 0; i < ac_group_order.size(); i++) {
      inv_ac_group_order[ac_group_order[i]] = i;
    }
    for (size_t i = 0; i < num_passes; i++) {
      size_t pass_start = permutation.size();
      for (coeff_order_t v : inv_ac_group_order) {
        permutation.push_back(pass_start + v);
      }
    }
    std::vector<BitWriter> new_group_codes(group_codes.size());
    for (size_t i = 0; i < permutation.size(); i++) {
      new_group_codes[permutation[i]] = std::move(group_codes[i]);
    }
    group_codes = std::move(new_group_codes);
  }

  JXL_RETURN_IF_ERROR(
      WriteGroupOffsets(group_codes, permutation_ptr, writer, aux_out));
  writer->AppendByteAligned(group_codes);

  return true;
}

}  // namespace jxl