Source code
Revision control
Copy as Markdown
Other Tools
// 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_group.h"
#include <jxl/memory_manager.h>
#include "lib/jxl/base/status.h"
#include "lib/jxl/memory_manager_internal.h"
#undef HWY_TARGET_INCLUDE
#define HWY_TARGET_INCLUDE "lib/jxl/enc_group.cc"
#include <hwy/foreach_target.h>
#include <hwy/highway.h>
#include "lib/jxl/ac_strategy.h"
#include "lib/jxl/base/bits.h"
#include "lib/jxl/base/compiler_specific.h"
#include "lib/jxl/base/rect.h"
#include "lib/jxl/common.h" // kMaxNumPasses
#include "lib/jxl/dct_util.h"
#include "lib/jxl/dec_transforms-inl.h"
#include "lib/jxl/enc_aux_out.h"
#include "lib/jxl/enc_cache.h"
#include "lib/jxl/enc_params.h"
#include "lib/jxl/enc_transforms-inl.h"
#include "lib/jxl/image.h"
#include "lib/jxl/quantizer-inl.h"
#include "lib/jxl/quantizer.h"
#include "lib/jxl/simd_util.h"
HWY_BEFORE_NAMESPACE();
namespace jxl {
namespace HWY_NAMESPACE {
// These templates are not found via ADL.
using hwy::HWY_NAMESPACE::Abs;
using hwy::HWY_NAMESPACE::Ge;
using hwy::HWY_NAMESPACE::IfThenElse;
using hwy::HWY_NAMESPACE::IfThenElseZero;
using hwy::HWY_NAMESPACE::MaskFromVec;
using hwy::HWY_NAMESPACE::Round;
// NOTE: caller takes care of extracting quant from rect of RawQuantField.
void QuantizeBlockAC(const Quantizer& quantizer, const bool error_diffusion,
size_t c, float qm_multiplier, AcStrategyType quant_kind,
size_t xsize, size_t ysize, float* thresholds,
const float* JXL_RESTRICT block_in, const int32_t* quant,
int32_t* JXL_RESTRICT block_out) {
const float* JXL_RESTRICT qm = quantizer.InvDequantMatrix(quant_kind, c);
float qac = quantizer.Scale() * (*quant);
// Not SIMD-ified for now.
if (c != 1 && xsize * ysize >= 4) {
for (int i = 0; i < 4; ++i) {
thresholds[i] -= 0.00744f * xsize * ysize;
if (thresholds[i] < 0.5) {
thresholds[i] = 0.5;
}
}
}
HWY_CAPPED(float, kBlockDim) df;
HWY_CAPPED(int32_t, kBlockDim) di;
HWY_CAPPED(uint32_t, kBlockDim) du;
const auto quantv = Set(df, qac * qm_multiplier);
for (size_t y = 0; y < ysize * kBlockDim; y++) {
size_t yfix = static_cast<size_t>(y >= ysize * kBlockDim / 2) * 2;
const size_t off = y * kBlockDim * xsize;
for (size_t x = 0; x < xsize * kBlockDim; x += Lanes(df)) {
auto threshold = Zero(df);
if (xsize == 1) {
HWY_ALIGN uint32_t kMask[kBlockDim] = {0, 0, 0, 0, ~0u, ~0u, ~0u, ~0u};
const auto mask = MaskFromVec(BitCast(df, Load(du, kMask + x)));
threshold = IfThenElse(mask, Set(df, thresholds[yfix + 1]),
Set(df, thresholds[yfix]));
} else {
// Same for all lanes in the vector.
threshold = Set(
df,
thresholds[yfix + static_cast<size_t>(x >= xsize * kBlockDim / 2)]);
}
const auto q = Mul(Load(df, qm + off + x), quantv);
const auto in = Load(df, block_in + off + x);
const auto val = Mul(q, in);
const auto nzero_mask = Ge(Abs(val), threshold);
const auto v = ConvertTo(di, IfThenElseZero(nzero_mask, Round(val)));
Store(v, di, block_out + off + x);
}
}
}
void AdjustQuantBlockAC(const Quantizer& quantizer, size_t c,
float qm_multiplier, AcStrategyType quant_kind,
size_t xsize, size_t ysize, float* thresholds,
const float* JXL_RESTRICT block_in, int32_t* quant) {
// No quantization adjusting for these small blocks.
// Quantization adjusting attempts to fix some known issues
// with larger blocks and on the 8x8 dct's emerging 8x8 blockiness
// when there are not many non-zeros.
constexpr size_t kPartialBlockKinds =
(1 << static_cast<size_t>(AcStrategyType::IDENTITY)) |
(1 << static_cast<size_t>(AcStrategyType::DCT2X2)) |
(1 << static_cast<size_t>(AcStrategyType::DCT4X4)) |
(1 << static_cast<size_t>(AcStrategyType::DCT4X8)) |
(1 << static_cast<size_t>(AcStrategyType::DCT8X4)) |
(1 << static_cast<size_t>(AcStrategyType::AFV0)) |
(1 << static_cast<size_t>(AcStrategyType::AFV1)) |
(1 << static_cast<size_t>(AcStrategyType::AFV2)) |
(1 << static_cast<size_t>(AcStrategyType::AFV3));
if ((1 << static_cast<size_t>(quant_kind)) & kPartialBlockKinds) {
return;
}
const float* JXL_RESTRICT qm = quantizer.InvDequantMatrix(quant_kind, c);
float qac = quantizer.Scale() * (*quant);
if (xsize > 1 || ysize > 1) {
for (int i = 0; i < 4; ++i) {
thresholds[i] -= Clamp1(0.003f * xsize * ysize, 0.f, 0.08f);
if (thresholds[i] < 0.54) {
thresholds[i] = 0.54;
}
}
}
float sum_of_highest_freq_row_and_column = 0;
float sum_of_error = 0;
float sum_of_vals = 0;
float hfNonZeros[4] = {};
float hfMaxError[4] = {};
for (size_t y = 0; y < ysize * kBlockDim; y++) {
for (size_t x = 0; x < xsize * kBlockDim; x++) {
const size_t pos = y * kBlockDim * xsize + x;
if (x < xsize && y < ysize) {
continue;
}
const size_t hfix = (static_cast<size_t>(y >= ysize * kBlockDim / 2) * 2 +
static_cast<size_t>(x >= xsize * kBlockDim / 2));
const float val = block_in[pos] * (qm[pos] * qac * qm_multiplier);
const float v = (std::abs(val) < thresholds[hfix]) ? 0 : rintf(val);
const float error = std::abs(val - v);
sum_of_error += error;
sum_of_vals += std::abs(v);
if (c == 1 && v == 0) {
if (hfMaxError[hfix] < error) {
hfMaxError[hfix] = error;
}
}
if (v != 0.0f) {
hfNonZeros[hfix] += std::abs(v);
bool in_corner = y >= 7 * ysize && x >= 7 * xsize;
bool on_border =
y == ysize * kBlockDim - 1 || x == xsize * kBlockDim - 1;
bool in_larger_corner = x >= 4 * xsize && y >= 4 * ysize;
if (in_corner || (on_border && in_larger_corner)) {
sum_of_highest_freq_row_and_column += std::abs(val);
}
}
}
}
if (c == 1 && sum_of_vals * 8 < xsize * ysize) {
static const double kLimit[4] = {
0.46,
0.46,
0.46,
0.46,
};
static const double kMul[4] = {
0.9999,
0.9999,
0.9999,
0.9999,
};
const int32_t orig_quant = *quant;
int32_t new_quant = *quant;
for (int i = 1; i < 4; ++i) {
if (hfNonZeros[i] == 0.0 && hfMaxError[i] > kLimit[i]) {
new_quant = orig_quant + 1;
break;
}
}
*quant = new_quant;
if (hfNonZeros[3] == 0.0 && hfMaxError[3] > kLimit[3]) {
thresholds[3] = kMul[3] * hfMaxError[3] * new_quant / orig_quant;
} else if ((hfNonZeros[1] == 0.0 && hfMaxError[1] > kLimit[1]) ||
(hfNonZeros[2] == 0.0 && hfMaxError[2] > kLimit[2])) {
thresholds[1] = kMul[1] * std::max(hfMaxError[1], hfMaxError[2]) *
new_quant / orig_quant;
thresholds[2] = thresholds[1];
} else if (hfNonZeros[0] == 0.0 && hfMaxError[0] > kLimit[0]) {
thresholds[0] = kMul[0] * hfMaxError[0] * new_quant / orig_quant;
}
}
// Heuristic for improving accuracy of high-frequency patterns
// occurring in an environment with no medium-frequency masking
// patterns.
{
float all =
hfNonZeros[0] + hfNonZeros[1] + hfNonZeros[2] + hfNonZeros[3] + 1;
float mul[3] = {70, 30, 60};
if (mul[c] * sum_of_highest_freq_row_and_column >= all) {
*quant += mul[c] * sum_of_highest_freq_row_and_column / all;
if (*quant >= Quantizer::kQuantMax) {
*quant = Quantizer::kQuantMax - 1;
}
}
}
if (quant_kind == AcStrategyType::DCT) {
// If this 8x8 block is too flat, increase the adaptive quantization level
// a bit to reduce visible block boundaries and requantize the block.
if (hfNonZeros[0] + hfNonZeros[1] + hfNonZeros[2] + hfNonZeros[3] < 11) {
*quant += 1;
if (*quant >= Quantizer::kQuantMax) {
*quant = Quantizer::kQuantMax - 1;
}
}
}
{
static const double kMul1[4][3] = {
{
0.22080615753848404,
0.45797479824262011,
0.29859235095977965,
},
{
0.70109486510286834,
0.16185281305512639,
0.14387691730035473,
},
{
0.114985964456218638,
0.44656840441027695,
0.10587658215149048,
},
{
0.46849665264409396,
0.41239077937781954,
0.088667407767185444,
},
};
static const double kMul2[4][3] = {
{
0.27450281941822197,
1.1255766549984996,
0.98950459134128388,
},
{
0.4652168675598285,
0.40945807983455818,
0.36581899811751367,
},
{
0.28034972424715715,
0.9182653201929738,
1.5581531543057416,
},
{
0.26873118114033728,
0.68863712390392484,
1.2082185408666786,
},
};
static const double kQuantNormalizer = 2.2942708343284721;
sum_of_error *= kQuantNormalizer;
sum_of_vals *= kQuantNormalizer;
if (quant_kind >= AcStrategyType::DCT16X16) {
int ix = 3;
if (quant_kind == AcStrategyType::DCT32X16 ||
quant_kind == AcStrategyType::DCT16X32) {
ix = 1;
} else if (quant_kind == AcStrategyType::DCT16X16) {
ix = 0;
} else if (quant_kind == AcStrategyType::DCT32X32) {
ix = 2;
}
int step =
sum_of_error / (kMul1[ix][c] * xsize * ysize * kBlockDim * kBlockDim +
kMul2[ix][c] * sum_of_vals);
if (step >= 2) {
step = 2;
}
if (step < 0) {
step = 0;
}
if (sum_of_error > kMul1[ix][c] * xsize * ysize * kBlockDim * kBlockDim +
kMul2[ix][c] * sum_of_vals) {
*quant += step;
if (*quant >= Quantizer::kQuantMax) {
*quant = Quantizer::kQuantMax - 1;
}
}
}
}
{
// Reduce quant in highly active areas.
int32_t div = (xsize * ysize);
int32_t activity = (static_cast<int32_t>(hfNonZeros[0]) + div / 2) / div;
int32_t orig_qp_limit = std::max(4, *quant / 2);
for (int i = 1; i < 4; ++i) {
activity = std::min(
activity, (static_cast<int32_t>(hfNonZeros[i]) + div / 2) / div);
}
if (activity >= 15) {
activity = 15;
}
int32_t qp = *quant - activity;
if (c == 1) {
for (int i = 1; i < 4; ++i) {
thresholds[i] += 0.01 * activity;
}
}
if (qp < orig_qp_limit) {
qp = orig_qp_limit;
}
*quant = qp;
}
}
// NOTE: caller takes care of extracting quant from rect of RawQuantField.
void QuantizeRoundtripYBlockAC(PassesEncoderState* enc_state, const size_t size,
const Quantizer& quantizer,
const bool error_diffusion,
AcStrategyType quant_kind, size_t xsize,
size_t ysize, const float* JXL_RESTRICT biases,
int32_t* quant, float* JXL_RESTRICT inout,
int32_t* JXL_RESTRICT quantized) {
float thres_y[4] = {0.58f, 0.64f, 0.64f, 0.64f};
if (enc_state->cparams.speed_tier <= SpeedTier::kHare) {
int32_t max_quant = 0;
int quant_orig = *quant;
float val[3] = {enc_state->x_qm_multiplier, 1.0f,
enc_state->b_qm_multiplier};
for (int c : {1, 0, 2}) {
float thres[4] = {0.58f, 0.64f, 0.64f, 0.64f};
*quant = quant_orig;
AdjustQuantBlockAC(quantizer, c, val[c], quant_kind, xsize, ysize,
&thres[0], inout + c * size, quant);
// Dead zone adjustment
if (c == 1) {
for (int k = 0; k < 4; ++k) {
thres_y[k] = thres[k];
}
}
max_quant = std::max(*quant, max_quant);
}
*quant = max_quant;
} else {
thres_y[0] = 0.56;
thres_y[1] = 0.62;
thres_y[2] = 0.62;
thres_y[3] = 0.62;
}
QuantizeBlockAC(quantizer, error_diffusion, 1, 1.0f, quant_kind, xsize, ysize,
&thres_y[0], inout + size, quant, quantized + size);
const float* JXL_RESTRICT dequant_matrix =
quantizer.DequantMatrix(quant_kind, 1);
HWY_CAPPED(float, kDCTBlockSize) df;
HWY_CAPPED(int32_t, kDCTBlockSize) di;
const auto inv_qac = Set(df, quantizer.inv_quant_ac(*quant));
for (size_t k = 0; k < kDCTBlockSize * xsize * ysize; k += Lanes(df)) {
const auto quant = Load(di, quantized + size + k);
const auto adj_quant = AdjustQuantBias(di, 1, quant, biases);
const auto dequantm = Load(df, dequant_matrix + k);
Store(Mul(Mul(adj_quant, dequantm), inv_qac), df, inout + size + k);
}
}
Status ComputeCoefficients(size_t group_idx, PassesEncoderState* enc_state,
const Image3F& opsin, const Rect& rect,
Image3F* dc) {
JxlMemoryManager* memory_manager = opsin.memory_manager();
const Rect block_group_rect =
enc_state->shared.frame_dim.BlockGroupRect(group_idx);
const Rect cmap_rect(
block_group_rect.x0() / kColorTileDimInBlocks,
block_group_rect.y0() / kColorTileDimInBlocks,
DivCeil(block_group_rect.xsize(), kColorTileDimInBlocks),
DivCeil(block_group_rect.ysize(), kColorTileDimInBlocks));
const Rect group_rect =
enc_state->shared.frame_dim.GroupRect(group_idx).Translate(rect.x0(),
rect.y0());
const size_t xsize_blocks = block_group_rect.xsize();
const size_t ysize_blocks = block_group_rect.ysize();
const size_t dc_stride = static_cast<size_t>(dc->PixelsPerRow());
const size_t opsin_stride = static_cast<size_t>(opsin.PixelsPerRow());
ImageI& full_quant_field = enc_state->shared.raw_quant_field;
const CompressParams& cparams = enc_state->cparams;
const size_t dct_scratch_size =
3 * (MaxVectorSize() / sizeof(float)) * AcStrategy::kMaxBlockDim;
// TODO(veluca): consider strategies to reduce this memory.
size_t mem_bytes = 3 * AcStrategy::kMaxCoeffArea * sizeof(int32_t);
JXL_ASSIGN_OR_RETURN(auto mem,
AlignedMemory::Create(memory_manager, mem_bytes));
size_t fmem_bytes =
(5 * AcStrategy::kMaxCoeffArea + dct_scratch_size) * sizeof(float);
JXL_ASSIGN_OR_RETURN(auto fmem,
AlignedMemory::Create(memory_manager, fmem_bytes));
float* JXL_RESTRICT scratch_space =
fmem.address<float>() + 3 * AcStrategy::kMaxCoeffArea;
{
// Only use error diffusion in Squirrel mode or slower.
const bool error_diffusion = cparams.speed_tier <= SpeedTier::kSquirrel;
constexpr HWY_CAPPED(float, kDCTBlockSize) d;
int32_t* JXL_RESTRICT coeffs[3][kMaxNumPasses] = {};
size_t num_passes = enc_state->progressive_splitter.GetNumPasses();
JXL_ENSURE(num_passes > 0);
for (size_t i = 0; i < num_passes; i++) {
// TODO(veluca): 16-bit quantized coeffs are not implemented yet.
JXL_ENSURE(enc_state->coeffs[i]->Type() == ACType::k32);
for (size_t c = 0; c < 3; c++) {
coeffs[c][i] = enc_state->coeffs[i]->PlaneRow(c, group_idx, 0).ptr32;
}
}
HWY_ALIGN float* coeffs_in = fmem.address<float>();
HWY_ALIGN int32_t* quantized = mem.address<int32_t>();
for (size_t by = 0; by < ysize_blocks; ++by) {
int32_t* JXL_RESTRICT row_quant_ac =
block_group_rect.Row(&full_quant_field, by);
size_t ty = by / kColorTileDimInBlocks;
const int8_t* JXL_RESTRICT row_cmap[3] = {
cmap_rect.ConstRow(enc_state->shared.cmap.ytox_map, ty),
nullptr,
cmap_rect.ConstRow(enc_state->shared.cmap.ytob_map, ty),
};
const float* JXL_RESTRICT opsin_rows[3] = {
group_rect.ConstPlaneRow(opsin, 0, by * kBlockDim),
group_rect.ConstPlaneRow(opsin, 1, by * kBlockDim),
group_rect.ConstPlaneRow(opsin, 2, by * kBlockDim),
};
float* JXL_RESTRICT dc_rows[3] = {
block_group_rect.PlaneRow(dc, 0, by),
block_group_rect.PlaneRow(dc, 1, by),
block_group_rect.PlaneRow(dc, 2, by),
};
AcStrategyRow ac_strategy_row =
enc_state->shared.ac_strategy.ConstRow(block_group_rect, by);
for (size_t tx = 0; tx < DivCeil(xsize_blocks, kColorTileDimInBlocks);
tx++) {
const auto x_factor =
Set(d, enc_state->shared.cmap.base().YtoXRatio(row_cmap[0][tx]));
const auto b_factor =
Set(d, enc_state->shared.cmap.base().YtoBRatio(row_cmap[2][tx]));
for (size_t bx = tx * kColorTileDimInBlocks;
bx < xsize_blocks && bx < (tx + 1) * kColorTileDimInBlocks; ++bx) {
const AcStrategy acs = ac_strategy_row[bx];
if (!acs.IsFirstBlock()) continue;
size_t xblocks = acs.covered_blocks_x();
size_t yblocks = acs.covered_blocks_y();
CoefficientLayout(&yblocks, &xblocks);
size_t size = kDCTBlockSize * xblocks * yblocks;
// DCT Y channel, roundtrip-quantize it and set DC.
int32_t quant_ac = row_quant_ac[bx];
for (size_t c : {0, 1, 2}) {
TransformFromPixels(acs.Strategy(), opsin_rows[c] + bx * kBlockDim,
opsin_stride, coeffs_in + c * size,
scratch_space);
}
DCFromLowestFrequencies(acs.Strategy(), coeffs_in + size,
dc_rows[1] + bx, dc_stride);
QuantizeRoundtripYBlockAC(
enc_state, size, enc_state->shared.quantizer, error_diffusion,
acs.Strategy(), xblocks, yblocks, kDefaultQuantBias, &quant_ac,
coeffs_in, quantized);
// Unapply color correlation
for (size_t k = 0; k < size; k += Lanes(d)) {
const auto in_x = Load(d, coeffs_in + k);
const auto in_y = Load(d, coeffs_in + size + k);
const auto in_b = Load(d, coeffs_in + 2 * size + k);
const auto out_x = NegMulAdd(x_factor, in_y, in_x);
const auto out_b = NegMulAdd(b_factor, in_y, in_b);
Store(out_x, d, coeffs_in + k);
Store(out_b, d, coeffs_in + 2 * size + k);
}
// Quantize X and B channels and set DC.
for (size_t c : {0, 2}) {
float thres[4] = {0.58f, 0.62f, 0.62f, 0.62f};
QuantizeBlockAC(enc_state->shared.quantizer, error_diffusion, c,
c == 0 ? enc_state->x_qm_multiplier
: enc_state->b_qm_multiplier,
acs.Strategy(), xblocks, yblocks, &thres[0],
coeffs_in + c * size, &quant_ac,
quantized + c * size);
DCFromLowestFrequencies(acs.Strategy(), coeffs_in + c * size,
dc_rows[c] + bx, dc_stride);
}
row_quant_ac[bx] = quant_ac;
for (size_t c = 0; c < 3; c++) {
enc_state->progressive_splitter.SplitACCoefficients(
quantized + c * size, acs, bx, by, coeffs[c]);
for (size_t p = 0; p < num_passes; p++) {
coeffs[c][p] += size;
}
}
}
}
}
}
return true;
}
// NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE
} // namespace jxl
HWY_AFTER_NAMESPACE();
#if HWY_ONCE
namespace jxl {
HWY_EXPORT(ComputeCoefficients);
Status ComputeCoefficients(size_t group_idx, PassesEncoderState* enc_state,
const Image3F& opsin, const Rect& rect,
Image3F* dc) {
return HWY_DYNAMIC_DISPATCH(ComputeCoefficients)(group_idx, enc_state, opsin,
rect, dc);
}
Status EncodeGroupTokenizedCoefficients(size_t group_idx, size_t pass_idx,
size_t histogram_idx,
const PassesEncoderState& enc_state,
BitWriter* writer, AuxOut* aux_out) {
// Select which histogram to use among those of the current pass.
const size_t num_histograms = enc_state.shared.num_histograms;
// num_histograms is 0 only for lossless.
JXL_ENSURE(num_histograms == 0 || histogram_idx < num_histograms);
size_t histo_selector_bits = CeilLog2Nonzero(num_histograms);
if (histo_selector_bits != 0) {
JXL_RETURN_IF_ERROR(
writer->WithMaxBits(histo_selector_bits, LayerType::Ac, aux_out, [&] {
writer->Write(histo_selector_bits, histogram_idx);
return true;
}));
}
size_t context_offset =
histogram_idx * enc_state.shared.block_ctx_map.NumACContexts();
JXL_RETURN_IF_ERROR(WriteTokens(
enc_state.passes[pass_idx].ac_tokens[group_idx],
enc_state.passes[pass_idx].codes, enc_state.passes[pass_idx].context_map,
context_offset, writer, LayerType::AcTokens, aux_out));
return true;
}
} // namespace jxl
#endif // HWY_ONCE