Quelle  dec_group.cc   Sprache: C

// 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/dec_group.h"

#include <algorithm>
#include <cstdint>
#include <cstring>
#include <memory>
#include <utility>

#include "lib/jxl/chroma_from_luma.h"
#include "lib/jxl/frame_header.h"

#undef HWY_TARGET_INCLUDE
#define HWY_TARGET_INCLUDE "lib/jxl/dec_group.cc"
#include <hwy/foreach_target.h>
#include <hwy/highway.h>

#include "lib/jxl/ac_context.h"
#include "lib/jxl/ac_strategy.h"
#include "lib/jxl/base/bits.h"
#include "lib/jxl/base/common.h"
#include "lib/jxl/base/printf_macros.h"
#include "lib/jxl/base/rect.h"
#include "lib/jxl/base/status.h"
#include "lib/jxl/coeff_order.h"
#include "lib/jxl/common.h"  // kMaxNumPasses
#include "lib/jxl/dec_cache.h"
#include "lib/jxl/dec_transforms-inl.h"
#include "lib/jxl/dec_xyb.h"
#include "lib/jxl/entropy_coder.h"
#include "lib/jxl/quant_weights.h"
#include "lib/jxl/quantizer-inl.h"
#include "lib/jxl/quantizer.h"

#ifndef LIB_JXL_DEC_GROUP_CC
#define LIB_JXL_DEC_GROUP_CC
namespace jxl {

struct AuxOut;

// Interface for reading groups for DecodeGroupImpl.
class GetBlock {
 public:
  virtual void StartRow(size_t by) = 0;
  virtual Status LoadBlock(size_t bx, size_t by, const AcStrategy& acs,
                           size_t size, size_t log2_covered_blocks,
                           ACPtr block[3], ACType ac_type) = 0;
  virtual ~GetBlock() {}
};

// Controls whether DecodeGroupImpl renders to pixels or not.
enum DrawMode {
  // Render to pixels.
  kDraw = 0,
  // Don't render to pixels.
  kDontDraw = 1,
};

}  // namespace jxl
#endif  // LIB_JXL_DEC_GROUP_CC

HWY_BEFORE_NAMESPACE();
namespace jxl {
namespace HWY_NAMESPACE {

// These templates are not found via ADL.
using hwy::HWY_NAMESPACE::AllFalse;
using hwy::HWY_NAMESPACE::Gt;
using hwy::HWY_NAMESPACE::Le;
using hwy::HWY_NAMESPACE::MaskFromVec;
using hwy::HWY_NAMESPACE::Or;
using hwy::HWY_NAMESPACE::Rebind;
using hwy::HWY_NAMESPACE::ShiftRight;

using D = HWY_FULL(float);
using DU = HWY_FULL(uint32_t);
using DI = HWY_FULL(int32_t);
using DI16 = Rebind<int16_t, DI>;
using DI16_FULL = HWY_CAPPED(int16_t, kDCTBlockSize);
constexpr D d;
constexpr DI di;
constexpr DI16 di16;
constexpr DI16_FULL di16_full;

// TODO(veluca): consider SIMDfying.
void Transpose8x8InPlace(int32_t* JXL_RESTRICT block) {
  for (size_t x = 0; x < 8; x++) {
    for (size_t y = x + 1; y < 8; y++) {
      std::swap(block[y * 8 + x], block[x * 8 + y]);
    }
  }
}

template <ACType ac_type>
void DequantLane(Vec<D> scaled_dequant_x, Vec<D> scaled_dequant_y,
                 Vec<D> scaled_dequant_b,
                 const float* JXL_RESTRICT dequant_matrices, size_t size,
                 size_t k, Vec<D> x_cc_mul, Vec<D> b_cc_mul,
                 const float* JXL_RESTRICT biases, ACPtr qblock[3],
                 float* JXL_RESTRICT block) {
  const auto x_mul = Mul(Load(d, dequant_matrices + k), scaled_dequant_x);
  const auto y_mul =
      Mul(Load(d, dequant_matrices + size + k), scaled_dequant_y);
  const auto b_mul =
      Mul(Load(d, dequant_matrices + 2 * size + k), scaled_dequant_b);

  Vec<DI> quantized_x_int;
  Vec<DI> quantized_y_int;
  Vec<DI> quantized_b_int;
  if (ac_type == ACType::k16) {
    Rebind<int16_t, DI> di16;
    quantized_x_int = PromoteTo(di, Load(di16, qblock[0].ptr16 + k));
    quantized_y_int = PromoteTo(di, Load(di16, qblock[1].ptr16 + k));
    quantized_b_int = PromoteTo(di, Load(di16, qblock[2].ptr16 + k));
  } else {
    quantized_x_int = Load(di, qblock[0].ptr32 + k);
    quantized_y_int = Load(di, qblock[1].ptr32 + k);
    quantized_b_int = Load(di, qblock[2].ptr32 + k);
  }

  const auto dequant_x_cc =
      Mul(AdjustQuantBias(di, 0, quantized_x_int, biases), x_mul);
  const auto dequant_y =
      Mul(AdjustQuantBias(di, 1, quantized_y_int, biases), y_mul);
  const auto dequant_b_cc =
      Mul(AdjustQuantBias(di, 2, quantized_b_int, biases), b_mul);

  const auto dequant_x = MulAdd(x_cc_mul, dequant_y, dequant_x_cc);
  const auto dequant_b = MulAdd(b_cc_mul, dequant_y, dequant_b_cc);
  Store(dequant_x, d, block + k);
  Store(dequant_y, d, block + size + k);
  Store(dequant_b, d, block + 2 * size + k);
}

template <ACType ac_type>
void DequantBlock(const AcStrategy& acs, float inv_global_scale, int quant,
                  float x_dm_multiplier, float b_dm_multiplier, Vec<D> x_cc_mul,
                  Vec<D> b_cc_mul, AcStrategyType kind, size_t size,
                  const Quantizer& quantizer, size_t covered_blocks,
                  const size_t* sbx,
                  const float* JXL_RESTRICT* JXL_RESTRICT dc_row,
                  size_t dc_stride, const float* JXL_RESTRICT biases,
                  ACPtr qblock[3], float* JXL_RESTRICT block,
                  float* JXL_RESTRICT scratch) {
  const auto scaled_dequant_s = inv_global_scale / quant;

  const auto scaled_dequant_x = Set(d, scaled_dequant_s * x_dm_multiplier);
  const auto scaled_dequant_y = Set(d, scaled_dequant_s);
  const auto scaled_dequant_b = Set(d, scaled_dequant_s * b_dm_multiplier);

  const float* dequant_matrices = quantizer.DequantMatrix(kind, 0);

  for (size_t k = 0; k < covered_blocks * kDCTBlockSize; k += Lanes(d)) {
    DequantLane<ac_type>(scaled_dequant_x, scaled_dequant_y, scaled_dequant_b,
                         dequant_matrices, size, k, x_cc_mul, b_cc_mul, biases,
                         qblock, block);
  }
  for (size_t c = 0; c < 3; c++) {
    LowestFrequenciesFromDC(acs.Strategy(), dc_row[c] + sbx[c], dc_stride,
                            block + c * size, scratch);
  }
}

Status DecodeGroupImpl(const FrameHeader& frame_header,
                       GetBlock* JXL_RESTRICT get_block,
                       GroupDecCache* JXL_RESTRICT group_dec_cache,
                       PassesDecoderState* JXL_RESTRICT dec_state,
                       size_t thread, size_t group_idx,
                       RenderPipelineInput& render_pipeline_input,
                       jpeg::JPEGData* jpeg_data, DrawMode draw) {
  // TODO(veluca): investigate cache usage in this function.
  const Rect block_rect =
      dec_state->shared->frame_dim.BlockGroupRect(group_idx);
  const AcStrategyImage& ac_strategy = dec_state->shared->ac_strategy;

  const size_t xsize_blocks = block_rect.xsize();
  const size_t ysize_blocks = block_rect.ysize();

  const size_t dc_stride = dec_state->shared->dc->PixelsPerRow();

  const float inv_global_scale = dec_state->shared->quantizer.InvGlobalScale();

  const YCbCrChromaSubsampling& cs = frame_header.chroma_subsampling;

  const auto kJpegDctMin = Set(di16_full, -4095);
  const auto kJpegDctMax = Set(di16_full, 4095);

  size_t idct_stride[3];
  for (size_t c = 0; c < 3; c++) {
    idct_stride[c] = render_pipeline_input.GetBuffer(c).first->PixelsPerRow();
  }

  HWY_ALIGN int32_t scaled_qtable[64 * 3];

  ACType ac_type = dec_state->coefficients->Type();
  auto dequant_block = ac_type == ACType::k16 ? DequantBlock<ACType::k16>
                                              : DequantBlock<ACType::k32>;
  // Whether or not coefficients should be stored for future usage, and/or read
  // from past usage.
  bool accumulate = !dec_state->coefficients->IsEmpty();
  // Offset of the current block in the group.
  size_t offset = 0;

  std::array<int, 3> jpeg_c_map;
  bool jpeg_is_gray = false;
  std::array<int, 3> dcoff = {};

  // TODO(veluca): all of this should be done only once per image.
  const ColorCorrelation& color_correlation = dec_state->shared->cmap.base();
  if (jpeg_data) {
    if (!color_correlation.IsJPEGCompatible()) {
      return JXL_FAILURE("The CfL map is not JPEG-compatible");
    }
    jpeg_is_gray = (jpeg_data->components.size() == 1);
    JXL_ENSURE(frame_header.color_transform != ColorTransform::kXYB);
    jpeg_c_map = JpegOrder(frame_header.color_transform, jpeg_is_gray);
    const std::vector<QuantEncoding>& qe =
        dec_state->shared->matrices.encodings();
    if (qe.empty() || qe[0].mode != QuantEncoding::Mode::kQuantModeRAW ||
        std::abs(qe[0].qraw.qtable_den - 1.f / (8 * 255)) > 1e-8f) {
      return JXL_FAILURE(
          "Quantization table is not a JPEG quantization table.");
    }
    JXL_ENSURE(qe[0].qraw.qtable->size() == 3 * 8 * 8);
    int* qtable = qe[0].qraw.qtable->data();
    for (size_t c = 0; c < 3; c++) {
      if (frame_header.color_transform == ColorTransform::kNone) {
        dcoff[c] = 1024 / qtable[64 * c];
      }
      for (size_t i = 0; i < 64; i++) {
        // Transpose the matrix, as it will be used on the transposed block.
        int n = qtable[64 + i];
        int d = qtable[64 * c + i];
        if (n <= 0 || d <= 0 || n >= 65536 || d >= 65536) {
          return JXL_FAILURE("Invalid JPEG quantization table");
        }
        scaled_qtable[64 * c + (i % 8) * 8 + (i / 8)] =
            (1 << kCFLFixedPointPrecision) * n / d;
      }
    }
  }

  size_t hshift[3] = {cs.HShift(0), cs.HShift(1), cs.HShift(2)};
  size_t vshift[3] = {cs.VShift(0), cs.VShift(1), cs.VShift(2)};
  Rect r[3];
  for (size_t i = 0; i < 3; i++) {
    r[i] =
        Rect(block_rect.x0() >> hshift[i], block_rect.y0() >> vshift[i],
             block_rect.xsize() >> hshift[i], block_rect.ysize() >> vshift[i]);
    if (!r[i].IsInside({0, 0, dec_state->shared->dc->Plane(i).xsize(),
                        dec_state->shared->dc->Plane(i).ysize()})) {
      return JXL_FAILURE("Frame dimensions are too big for the image.");
    }
  }

  for (size_t by = 0; by < ysize_blocks; ++by) {
    get_block->StartRow(by);
    size_t sby[3] = {by >> vshift[0], by >> vshift[1], by >> vshift[2]};

    const int32_t* JXL_RESTRICT row_quant =
        block_rect.ConstRow(dec_state->shared->raw_quant_field, by);

    const float* JXL_RESTRICT dc_rows[3] = {
        r[0].ConstPlaneRow(*dec_state->shared->dc, 0, sby[0]),
        r[1].ConstPlaneRow(*dec_state->shared->dc, 1, sby[1]),
        r[2].ConstPlaneRow(*dec_state->shared->dc, 2, sby[2]),
    };

    const size_t ty = (block_rect.y0() + by) / kColorTileDimInBlocks;
    AcStrategyRow acs_row = ac_strategy.ConstRow(block_rect, by);

    const int8_t* JXL_RESTRICT row_cmap[3] = {
        dec_state->shared->cmap.ytox_map.ConstRow(ty),
        nullptr,
        dec_state->shared->cmap.ytob_map.ConstRow(ty),
    };

    float* JXL_RESTRICT idct_row[3];
    int16_t* JXL_RESTRICT jpeg_row[3];
    for (size_t c = 0; c < 3; c++) {
      const auto& buffer = render_pipeline_input.GetBuffer(c);
      idct_row[c] = buffer.second.Row(buffer.first, sby[c] * kBlockDim);
      if (jpeg_data) {
        auto& component = jpeg_data->components[jpeg_c_map[c]];
        jpeg_row[c] =
            component.coeffs.data() +
            (component.width_in_blocks * (r[c].y0() + sby[c]) + r[c].x0()) *
                kDCTBlockSize;
      }
    }

    size_t bx = 0;
    for (size_t tx = 0; tx < DivCeil(xsize_blocks, kColorTileDimInBlocks);
         tx++) {
      size_t abs_tx = tx + block_rect.x0() / kColorTileDimInBlocks;
      auto x_cc_mul = Set(d, color_correlation.YtoXRatio(row_cmap[0][abs_tx]));
      auto b_cc_mul = Set(d, color_correlation.YtoBRatio(row_cmap[2][abs_tx]));
      // Increment bx by llf_x because those iterations would otherwise
      // immediately continue (!IsFirstBlock). Reduces mispredictions.
      for (; bx < xsize_blocks && bx < (tx + 1) * kColorTileDimInBlocks;) {
        size_t sbx[3] = {bx >> hshift[0], bx >> hshift[1], bx >> hshift[2]};
        AcStrategy acs = acs_row[bx];
        const size_t llf_x = acs.covered_blocks_x();

        // Can only happen in the second or lower rows of a varblock.
        if (JXL_UNLIKELY(!acs.IsFirstBlock())) {
          bx += llf_x;
          continue;
        }
        const size_t log2_covered_blocks = acs.log2_covered_blocks();

        const size_t covered_blocks = 1 << log2_covered_blocks;
        const size_t size = covered_blocks * kDCTBlockSize;

        ACPtr qblock[3];
        if (accumulate) {
          for (size_t c = 0; c < 3; c++) {
            qblock[c] = dec_state->coefficients->PlaneRow(c, group_idx, offset);
          }
        } else {
          // No point in reading from bitstream without accumulating and not
          // drawing.
          JXL_ENSURE(draw == kDraw);
          if (ac_type == ACType::k16) {
            memset(group_dec_cache->dec_group_qblock16, 0,
                   size * 3 * sizeof(int16_t));
            for (size_t c = 0; c < 3; c++) {
              qblock[c].ptr16 = group_dec_cache->dec_group_qblock16 + c * size;
            }
          } else {
            memset(group_dec_cache->dec_group_qblock, 0,
                   size * 3 * sizeof(int32_t));
            for (size_t c = 0; c < 3; c++) {
              qblock[c].ptr32 = group_dec_cache->dec_group_qblock + c * size;
            }
          }
        }
        JXL_RETURN_IF_ERROR(get_block->LoadBlock(
            bx, by, acs, size, log2_covered_blocks, qblock, ac_type));
        offset += size;
        if (draw == kDontDraw) {
          bx += llf_x;
          continue;
        }

        if (JXL_UNLIKELY(jpeg_data)) {
          if (acs.Strategy() != AcStrategyType::DCT) {
            return JXL_FAILURE(
                "Can only decode to JPEG if only DCT-8 is used.");
          }

          HWY_ALIGN int32_t transposed_dct_y[64];
          for (size_t c : {1, 0, 2}) {
            // Propagate only Y for grayscale.
            if (jpeg_is_gray && c != 1) {
              continue;
            }
            if ((sbx[c] << hshift[c] != bx) || (sby[c] << vshift[c] != by)) {
              continue;
            }
            int16_t* JXL_RESTRICT jpeg_pos =
                jpeg_row[c] + sbx[c] * kDCTBlockSize;
            // JPEG XL is transposed, JPEG is not.
            auto* transposed_dct = qblock[c].ptr32;
            Transpose8x8InPlace(transposed_dct);
            // No CfL - no need to store the y block converted to integers.
            if (!cs.Is444() ||
                (row_cmap[0][abs_tx] == 0 && row_cmap[2][abs_tx] == 0)) {
              for (size_t i = 0; i < 64; i += Lanes(d)) {
                const auto ini = Load(di, transposed_dct + i);
                const auto ini16 = DemoteTo(di16, ini);
                StoreU(ini16, di16, jpeg_pos + i);
              }
            } else if (c == 1) {
              // Y channel: save for restoring X/B, but nothing else to do.
              for (size_t i = 0; i < 64; i += Lanes(d)) {
                const auto ini = Load(di, transposed_dct + i);
                Store(ini, di, transposed_dct_y + i);
                const auto ini16 = DemoteTo(di16, ini);
                StoreU(ini16, di16, jpeg_pos + i);
              }
            } else {
              // transposed_dct_y contains the y channel block, transposed.
              const auto scale =
                  Set(di, ColorCorrelation::RatioJPEG(row_cmap[c][abs_tx]));
              const auto round = Set(di, 1 << (kCFLFixedPointPrecision - 1));
              for (int i = 0; i < 64; i += Lanes(d)) {
                auto in = Load(di, transposed_dct + i);
                auto in_y = Load(di, transposed_dct_y + i);
                auto qt = Load(di, scaled_qtable + c * size + i);
                auto coeff_scale = ShiftRight<kCFLFixedPointPrecision>(
                    Add(Mul(qt, scale), round));
                auto cfl_factor = ShiftRight<kCFLFixedPointPrecision>(
                    Add(Mul(in_y, coeff_scale), round));
                StoreU(DemoteTo(di16, Add(in, cfl_factor)), di16, jpeg_pos + i);
              }
            }
            jpeg_pos[0] =
                Clamp1<float>(dc_rows[c][sbx[c]] - dcoff[c], -2047, 2047);
            auto overflow = MaskFromVec(Set(di16_full, 0));
            auto underflow = MaskFromVec(Set(di16_full, 0));
            for (int i = 0; i < 64; i += Lanes(di16_full)) {
              auto in = LoadU(di16_full, jpeg_pos + i);
              overflow = Or(overflow, Gt(in, kJpegDctMax));
              underflow = Or(underflow, Lt(in, kJpegDctMin));
            }
            if (!AllFalse(di16_full, Or(overflow, underflow))) {
              return JXL_FAILURE("JPEG DCT coefficients out of range");
            }
          }
        } else {
          HWY_ALIGN float* const block = group_dec_cache->dec_group_block;
          // Dequantize and add predictions.
          dequant_block(
              acs, inv_global_scale, row_quant[bx], dec_state->x_dm_multiplier,
              dec_state->b_dm_multiplier, x_cc_mul, b_cc_mul, acs.Strategy(),
              size, dec_state->shared->quantizer,
              acs.covered_blocks_y() * acs.covered_blocks_x(), sbx, dc_rows,
              dc_stride,
              dec_state->output_encoding_info.opsin_params.quant_biases, qblock,
              block, group_dec_cache->scratch_space);

          for (size_t c : {1, 0, 2}) {
            if ((sbx[c] << hshift[c] != bx) || (sby[c] << vshift[c] != by)) {
              continue;
            }
            // IDCT
            float* JXL_RESTRICT idct_pos = idct_row[c] + sbx[c] * kBlockDim;
            TransformToPixels(acs.Strategy(), block + c * size, idct_pos,
                              idct_stride[c], group_dec_cache->scratch_space);
          }
        }
        bx += llf_x;
      }
    }
  }
  return true;
}

// NOLINTNEXTLINE(google-readability-namespace-comments)
}  // namespace HWY_NAMESPACE
}  // namespace jxl
HWY_AFTER_NAMESPACE();

#if HWY_ONCE
namespace jxl {
namespace {
// Decode quantized AC coefficients of DCT blocks.
// LLF components in the output block will not be modified.
template <ACType ac_type, bool uses_lz77>
Status DecodeACVarBlock(size_t ctx_offset, size_t log2_covered_blocks,
                        int32_t* JXL_RESTRICT row_nzeros,
                        const int32_t* JXL_RESTRICT row_nzeros_top,
                        size_t nzeros_stride, size_t c, size_t bx, size_t by,
                        size_t lbx, AcStrategy acs,
                        const coeff_order_t* JXL_RESTRICT coeff_order,
                        BitReader* JXL_RESTRICT br,
                        ANSSymbolReader* JXL_RESTRICT decoder,
                        const std::vector<uint8_t>& context_map,
                        const uint8_t* qdc_row, const int32_t* qf_row,
                        const BlockCtxMap& block_ctx_map, ACPtr block,
                        size_t shift = 0) {
  // Equal to number of LLF coefficients.
  const size_t covered_blocks = 1 << log2_covered_blocks;
  const size_t size = covered_blocks * kDCTBlockSize;
  int32_t predicted_nzeros =
      PredictFromTopAndLeft(row_nzeros_top, row_nzeros, bx, 32);

  size_t ord = kStrategyOrder[acs.RawStrategy()];
  const coeff_order_t* JXL_RESTRICT order =
      &coeff_order[CoeffOrderOffset(ord, c)];

  size_t block_ctx = block_ctx_map.Context(qdc_row[lbx], qf_row[bx], ord, c);
  const int32_t nzero_ctx =
      block_ctx_map.NonZeroContext(predicted_nzeros, block_ctx) + ctx_offset;

  size_t nzeros =
      decoder->ReadHybridUintInlined<uses_lz77>(nzero_ctx, br, context_map);
  if (nzeros > size - covered_blocks) {
    return JXL_FAILURE("Invalid AC: nzeros %" PRIuS " too large for %" PRIuS
                       " 8x8 blocks",
                       nzeros, covered_blocks);
  }
  for (size_t y = 0; y < acs.covered_blocks_y(); y++) {
    for (size_t x = 0; x < acs.covered_blocks_x(); x++) {
      row_nzeros[bx + x + y * nzeros_stride] =
          (nzeros + covered_blocks - 1) >> log2_covered_blocks;
    }
  }

  const size_t histo_offset =
      ctx_offset + block_ctx_map.ZeroDensityContextsOffset(block_ctx);

  size_t prev = (nzeros > size / 16 ? 0 : 1);
  for (size_t k = covered_blocks; k < size && nzeros != 0; ++k) {
    const size_t ctx =
        histo_offset + ZeroDensityContext(nzeros, k, covered_blocks,
                                          log2_covered_blocks, prev);
    const size_t u_coeff =
        decoder->ReadHybridUintInlined<uses_lz77>(ctx, br, context_map);
    // Hand-rolled version of UnpackSigned, shifting before the conversion to
    // signed integer to avoid undefined behavior of shifting negative numbers.
    const size_t magnitude = u_coeff >> 1;
    const size_t neg_sign = (~u_coeff) & 1;
    const intptr_t coeff =
        static_cast<intptr_t>((magnitude ^ (neg_sign - 1)) << shift);
    if (ac_type == ACType::k16) {
      block.ptr16[order[k]] += coeff;
    } else {
      block.ptr32[order[k]] += coeff;
    }
    prev = static_cast<size_t>(u_coeff != 0);
    nzeros -= prev;
  }
  if (JXL_UNLIKELY(nzeros != 0)) {
    return JXL_FAILURE("Invalid AC: nzeros at end of block is %" PRIuS
                       ", should be 0. Block (%" PRIuS ", %" PRIuS
                       "), channel %" PRIuS,
                       nzeros, bx, by, c);
  }

  return true;
}

// Structs used by DecodeGroupImpl to get a quantized block.
// GetBlockFromBitstream uses ANS decoding (and thus keeps track of row
// pointers in row_nzeros), GetBlockFromEncoder simply reads the coefficient
// image provided by the encoder.

struct GetBlockFromBitstream : public GetBlock {
  void StartRow(size_t by) override {
    qf_row = rect.ConstRow(*qf, by);
    for (size_t c = 0; c < 3; c++) {
      size_t sby = by >> vshift[c];
      quant_dc_row = quant_dc->ConstRow(rect.y0() + by) + rect.x0();
      for (size_t i = 0; i < num_passes; i++) {
        row_nzeros[i][c] = group_dec_cache->num_nzeroes[i].PlaneRow(c, sby);
        row_nzeros_top[i][c] =
            sby == 0
                ? nullptr
                : group_dec_cache->num_nzeroes[i].ConstPlaneRow(c, sby - 1);
      }
    }
  }

  Status LoadBlock(size_t bx, size_t by, const AcStrategy& acs, size_t size,
                   size_t log2_covered_blocks, ACPtr block[3],
                   ACType ac_type) override {
    ;
    for (size_t c : {1, 0, 2}) {
      size_t sbx = bx >> hshift[c];
      size_t sby = by >> vshift[c];
      if (JXL_UNLIKELY((sbx << hshift[c] != bx) || (sby << vshift[c] != by))) {
        continue;
      }

      for (size_t pass = 0; JXL_UNLIKELY(pass < num_passes); pass++) {
        auto decode_ac_varblock =
            decoders[pass].UsesLZ77()
                ? (ac_type == ACType::k16 ? DecodeACVarBlock<ACType::k16, 1>
                                          : DecodeACVarBlock<ACType::k32, 1>)
                : (ac_type == ACType::k16 ? DecodeACVarBlock<ACType::k16, 0>
                                          : DecodeACVarBlock<ACType::k32, 0>);
        JXL_RETURN_IF_ERROR(decode_ac_varblock(
            ctx_offset[pass], log2_covered_blocks, row_nzeros[pass][c],
            row_nzeros_top[pass][c], nzeros_stride, c, sbx, sby, bx, acs,
            &coeff_orders[pass * coeff_order_size], readers[pass],
            &decoders[pass], context_map[pass], quant_dc_row, qf_row,
            *block_ctx_map, block[c], shift_for_pass[pass]));
      }
    }
    return true;
  }

  Status Init(const FrameHeader& frame_header,
              BitReader* JXL_RESTRICT* JXL_RESTRICT readers, size_t num_passes,
              size_t group_idx, size_t histo_selector_bits, const Rect& rect,
              GroupDecCache* JXL_RESTRICT group_dec_cache,
              PassesDecoderState* dec_state, size_t first_pass) {
    for (size_t i = 0; i < 3; i++) {
      hshift[i] = frame_header.chroma_subsampling.HShift(i);
      vshift[i] = frame_header.chroma_subsampling.VShift(i);
    }
    this->coeff_order_size = dec_state->shared->coeff_order_size;
    this->coeff_orders =
        dec_state->shared->coeff_orders.data() + first_pass * coeff_order_size;
    this->context_map = dec_state->context_map.data() + first_pass;
    this->readers = readers;
    this->num_passes = num_passes;
    this->shift_for_pass = frame_header.passes.shift + first_pass;
    this->group_dec_cache = group_dec_cache;
    this->rect = rect;
    block_ctx_map = &dec_state->shared->block_ctx_map;
    qf = &dec_state->shared->raw_quant_field;
    quant_dc = &dec_state->shared->quant_dc;

    for (size_t pass = 0; pass < num_passes; pass++) {
      // Select which histogram set to use among those of the current pass.
      size_t cur_histogram = 0;
      if (histo_selector_bits != 0) {
        cur_histogram = readers[pass]->ReadBits(histo_selector_bits);
      }
      if (cur_histogram >= dec_state->shared->num_histograms) {
        return JXL_FAILURE("Invalid histogram selector");
      }
      ctx_offset[pass] = cur_histogram * block_ctx_map->NumACContexts();

      JXL_ASSIGN_OR_RETURN(
          decoders[pass],
          ANSSymbolReader::Create(&dec_state->code[pass + first_pass],
                                  readers[pass]));
    }
    nzeros_stride = group_dec_cache->num_nzeroes[0].PixelsPerRow();
    for (size_t i = 0; i < num_passes; i++) {
      JXL_ENSURE(
          nzeros_stride ==
          static_cast<size_t>(group_dec_cache->num_nzeroes[i].PixelsPerRow()));
    }
    return true;
  }

  const uint32_t* shift_for_pass = nullptr;  // not owned
  const coeff_order_t* JXL_RESTRICT coeff_orders;
  size_t coeff_order_size;
  const std::vector<uint8_t>* JXL_RESTRICT context_map;
  ANSSymbolReader decoders[kMaxNumPasses];
  BitReader* JXL_RESTRICT* JXL_RESTRICT readers;
  size_t num_passes;
  size_t ctx_offset[kMaxNumPasses];
  size_t nzeros_stride;
  int32_t* JXL_RESTRICT row_nzeros[kMaxNumPasses][3];
  const int32_t* JXL_RESTRICT row_nzeros_top[kMaxNumPasses][3];
  GroupDecCache* JXL_RESTRICT group_dec_cache;
  const BlockCtxMap* block_ctx_map;
  const ImageI* qf;
  const ImageB* quant_dc;
  const int32_t* qf_row;
  const uint8_t* quant_dc_row;
  Rect rect;
  size_t hshift[3], vshift[3];
};

struct GetBlockFromEncoder : public GetBlock {
  void StartRow(size_t by) override {}

  Status LoadBlock(size_t bx, size_t by, const AcStrategy& acs, size_t size,
                   size_t log2_covered_blocks, ACPtr block[3],
                   ACType ac_type) override {
    JXL_ENSURE(ac_type == ACType::k32);
    for (size_t c = 0; c < 3; c++) {
      // for each pass
      for (size_t i = 0; i < quantized_ac->size(); i++) {
        for (size_t k = 0; k < size; k++) {
          // TODO(veluca): SIMD.
          block[c].ptr32[k] +=
              rows[i][c][offset + k] * (1 << shift_for_pass[i]);
        }
      }
    }
    offset += size;
    return true;
  }

  static StatusOr<GetBlockFromEncoder> Create(
      const std::vector<std::unique_ptr<ACImage>>& ac, size_t group_idx,
      const uint32_t* shift_for_pass) {
    GetBlockFromEncoder result(ac, group_idx, shift_for_pass);
    // TODO(veluca): not supported with chroma subsampling.
    for (size_t i = 0; i < ac.size(); i++) {
      JXL_ENSURE(ac[i]->Type() == ACType::k32);
      for (size_t c = 0; c < 3; c++) {
        result.rows[i][c] = ac[i]->PlaneRow(c, group_idx, 0).ptr32;
      }
    }
    return result;
  }

  const std::vector<std::unique_ptr<ACImage>>* JXL_RESTRICT quantized_ac;
  size_t offset = 0;
  const int32_t* JXL_RESTRICT rows[kMaxNumPasses][3];
  const uint32_t* shift_for_pass = nullptr;  // not owned

 private:
  GetBlockFromEncoder(const std::vector<std::unique_ptr<ACImage>>& ac,
                      size_t group_idx, const uint32_t* shift_for_pass)
      : quantized_ac(&ac), shift_for_pass(shift_for_pass) {}
};

HWY_EXPORT(DecodeGroupImpl);

}  // namespace

Status DecodeGroup(const FrameHeader& frame_header,
                   BitReader* JXL_RESTRICT* JXL_RESTRICT readers,
                   size_t num_passes, size_t group_idx,
                   PassesDecoderState* JXL_RESTRICT dec_state,
                   GroupDecCache* JXL_RESTRICT group_dec_cache, size_t thread,
                   RenderPipelineInput& render_pipeline_input,
                   jpeg::JPEGData* JXL_RESTRICT jpeg_data, size_t first_pass,
                   bool force_draw, bool dc_only, bool* should_run_pipeline) {
  JxlMemoryManager* memory_manager = dec_state->memory_manager();
  DrawMode draw =
      (num_passes + first_pass == frame_header.passes.num_passes) || force_draw
          ? kDraw
          : kDontDraw;

  if (should_run_pipeline) {
    *should_run_pipeline = draw != kDontDraw;
  }

  if (draw == kDraw && num_passes == 0 && first_pass == 0) {
    JXL_RETURN_IF_ERROR(group_dec_cache->InitDCBufferOnce(memory_manager));
    const YCbCrChromaSubsampling& cs = frame_header.chroma_subsampling;
    for (size_t c : {0, 1, 2}) {
      size_t hs = cs.HShift(c);
      size_t vs = cs.VShift(c);
      // We reuse filter_input_storage here as it is not currently in use.
      const Rect src_rect_precs =
          dec_state->shared->frame_dim.BlockGroupRect(group_idx);
      const Rect src_rect =
          Rect(src_rect_precs.x0() >> hs, src_rect_precs.y0() >> vs,
               src_rect_precs.xsize() >> hs, src_rect_precs.ysize() >> vs);
      const Rect copy_rect(kRenderPipelineXOffset, 2, src_rect.xsize(),
                           src_rect.ysize());
      JXL_RETURN_IF_ERROR(
          CopyImageToWithPadding(src_rect, dec_state->shared->dc->Plane(c), 2,
                                 copy_rect, &group_dec_cache->dc_buffer));
      // Mirrorpad. Interleaving left and right padding ensures that padding
      // works out correctly even for images with DC size of 1.
      for (size_t y = 0; y < src_rect.ysize() + 4; y++) {
        size_t xend = kRenderPipelineXOffset +
                      (dec_state->shared->dc->Plane(c).xsize() >> hs) -
                      src_rect.x0();
        for (size_t ix = 0; ix < 2; ix++) {
          if (src_rect.x0() == 0) {
            group_dec_cache->dc_buffer.Row(y)[kRenderPipelineXOffset - ix - 1] =
                group_dec_cache->dc_buffer.Row(y)[kRenderPipelineXOffset + ix];
          }
          if (src_rect.x0() + src_rect.xsize() + 2 >=
              (dec_state->shared->dc->xsize() >> hs)) {
            group_dec_cache->dc_buffer.Row(y)[xend + ix] =
                group_dec_cache->dc_buffer.Row(y)[xend - ix - 1];
          }
        }
      }
      const auto& buffer = render_pipeline_input.GetBuffer(c);
      Rect dst_rect = buffer.second;
      ImageF* upsampling_dst = buffer.first;
      JXL_ENSURE(dst_rect.IsInside(*upsampling_dst));

      RenderPipelineStage::RowInfo input_rows(1, std::vector<float*>(5));
      RenderPipelineStage::RowInfo output_rows(1, std::vector<float*>(8));
      for (size_t y = src_rect.y0(); y < src_rect.y0() + src_rect.ysize();
           y++) {
        for (ssize_t iy = 0; iy < 5; iy++) {
          input_rows[0][iy] = group_dec_cache->dc_buffer.Row(
              Mirror(static_cast<ssize_t>(y) + iy - 2,
                     dec_state->shared->dc->Plane(c).ysize() >> vs) +
              2 - src_rect.y0());
        }
        for (size_t iy = 0; iy < 8; iy++) {
          output_rows[0][iy] =
              dst_rect.Row(upsampling_dst, ((y - src_rect.y0()) << 3) + iy) -
              kRenderPipelineXOffset;
        }
        // Arguments set to 0/nullptr are not used.
        JXL_RETURN_IF_ERROR(dec_state->upsampler8x->ProcessRow(
            input_rows, output_rows,
            /*xextra=*/0, src_rect.xsize(), 0, 0, thread));
      }
    }
    return true;
  }

  size_t histo_selector_bits = 0;
  if (dc_only) {
    JXL_ENSURE(num_passes == 0);
  } else {
    JXL_ENSURE(dec_state->shared->num_histograms > 0);
    histo_selector_bits = CeilLog2Nonzero(dec_state->shared->num_histograms);
  }

  auto get_block = jxl::make_unique<GetBlockFromBitstream>();
  JXL_RETURN_IF_ERROR(get_block->Init(
      frame_header, readers, num_passes, group_idx, histo_selector_bits,
      dec_state->shared->frame_dim.BlockGroupRect(group_idx), group_dec_cache,
      dec_state, first_pass));

  JXL_RETURN_IF_ERROR(HWY_DYNAMIC_DISPATCH(DecodeGroupImpl)(
      frame_header, get_block.get(), group_dec_cache, dec_state, thread,
      group_idx, render_pipeline_input, jpeg_data, draw));

  for (size_t pass = 0; pass < num_passes; pass++) {
    if (!get_block->decoders[pass].CheckANSFinalState()) {
      return JXL_FAILURE("ANS checksum failure.");
    }
  }
  return true;
}

Status DecodeGroupForRoundtrip(const FrameHeader& frame_header,
                               const std::vector<std::unique_ptr<ACImage>>& ac,
                               size_t group_idx,
                               PassesDecoderState* JXL_RESTRICT dec_state,
                               GroupDecCache* JXL_RESTRICT group_dec_cache,
                               size_t thread,
                               RenderPipelineInput& render_pipeline_input,
                               jpeg::JPEGData* JXL_RESTRICT jpeg_data,
                               AuxOut* aux_out) {
  JxlMemoryManager* memory_manager = dec_state->memory_manager();
  JXL_ASSIGN_OR_RETURN(
      GetBlockFromEncoder get_block,
      GetBlockFromEncoder::Create(ac, group_idx, frame_header.passes.shift));
  JXL_RETURN_IF_ERROR(group_dec_cache->InitOnce(
      memory_manager,
      /*num_passes=*/0,
      /*used_acs=*/(1u << AcStrategy::kNumValidStrategies) - 1));

  return HWY_DYNAMIC_DISPATCH(DecodeGroupImpl)(
      frame_header, &get_block, group_dec_cache, dec_state, thread, group_idx,
      render_pipeline_input, jpeg_data, kDraw);
}

}  // namespace jxl
#endif  // HWY_ONCE