// Copyright 2011 Google Inc. All Rights Reserved. // // Use of this source code is governed by a BSD-style license // that can be found in the COPYING file in the root of the source // tree. An additional intellectual property rights grant can be found // in the file PATENTS. All contributing project authors may // be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // Quantization // // Author: Skal (pascal.massimino@gmail.com)
#include <assert.h> #include <math.h> #include <stdlib.h> // for abs()
#define DO_TRELLIS_I4 1 #define DO_TRELLIS_I16 1 // not a huge gain, but ok at low bitrate. #define DO_TRELLIS_UV 0 // disable trellis for UV. Risky. Not worth. #define USE_TDISTO 1
#define MID_ALPHA 64 // neutral value for susceptibility #define MIN_ALPHA 30 // lowest usable value for susceptibility #define MAX_ALPHA 100 // higher meaningful value for susceptibility
#define SNS_TO_DQ 0.9 // Scaling constant between the sns value and the QP // power-law modulation. Must be strictly less than 1.
// number of non-zero coeffs below which we consider the block very flat // (and apply a penalty to complex predictions) #define FLATNESS_LIMIT_I16 0 // I16 mode (special case) #define FLATNESS_LIMIT_I4 3 // I4 mode #define FLATNESS_LIMIT_UV 2 // UV mode #define FLATNESS_PENALTY 140 // roughly ~1bit per block
#define MULT_8B(a, b) (((a) * (b) + 128) >> 8)
#define RD_DISTO_MULT 256 // distortion multiplier (equivalent of lambda)
// Sharpening by (slightly) raising the hi-frequency coeffs. // Hack-ish but helpful for mid-bitrate range. Use with care. #define SHARPEN_BITS 11 // number of descaling bits for sharpening bias staticconst uint8_t kFreqSharpening[16] = {
0, 30, 60, 90,
30, 60, 90, 90,
60, 90, 90, 90,
90, 90, 90, 90
};
//------------------------------------------------------------------------------ // Initialize quantization parameters in VP8Matrix
// Returns the average quantizer staticint ExpandMatrix(VP8Matrix* const m, int type) { int i, sum; for (i = 0; i < 2; ++i) { constint is_ac_coeff = (i > 0); constint bias = kBiasMatrices[type][is_ac_coeff];
m->iq_[i] = (1 << QFIX) / m->q_[i];
m->bias_[i] = BIAS(bias); // zthresh_ is the exact value such that QUANTDIV(coeff, iQ, B) is: // * zero if coeff <= zthresh // * non-zero if coeff > zthresh
m->zthresh_[i] = ((1 << QFIX) - 1 - m->bias_[i]) / m->iq_[i];
} for (i = 2; i < 16; ++i) {
m->q_[i] = m->q_[1];
m->iq_[i] = m->iq_[1];
m->bias_[i] = m->bias_[1];
m->zthresh_[i] = m->zthresh_[1];
} for (sum = 0, i = 0; i < 16; ++i) { if (type == 0) { // we only use sharpening for AC luma coeffs
m->sharpen_[i] = (kFreqSharpening[i] * m->q_[i]) >> SHARPEN_BITS;
} else {
m->sharpen_[i] = 0;
}
sum += m->q_[i];
} return (sum + 8) >> 4;
}
// none of these constants should be < 1
CheckLambdaValue(&m->lambda_i4_);
CheckLambdaValue(&m->lambda_i16_);
CheckLambdaValue(&m->lambda_uv_);
CheckLambdaValue(&m->lambda_mode_);
CheckLambdaValue(&m->lambda_trellis_i4_);
CheckLambdaValue(&m->lambda_trellis_i16_);
CheckLambdaValue(&m->lambda_trellis_uv_);
CheckLambdaValue(&m->tlambda_);
// Very small filter-strength values have close to no visual effect. So we can // save a little decoding-CPU by turning filtering off for these. #define FSTRENGTH_CUTOFF 2
staticvoid SetupFilterStrength(VP8Encoder* const enc) { int i; // level0 is in [0..500]. Using '-f 50' as filter_strength is mid-filtering. constint level0 = 5 * enc->config_->filter_strength; for (i = 0; i < NUM_MB_SEGMENTS; ++i) {
VP8SegmentInfo* const m = &enc->dqm_[i]; // We focus on the quantization of AC coeffs. constint qstep = kAcTable[clip(m->quant_, 0, 127)] >> 2; constint base_strength =
VP8FilterStrengthFromDelta(enc->filter_hdr_.sharpness_, qstep); // Segments with lower complexity ('beta') will be less filtered. constint f = base_strength * level0 / (256 + m->beta_);
m->fstrength_ = (f < FSTRENGTH_CUTOFF) ? 0 : (f > 63) ? 63 : f;
} // We record the initial strength (mainly for the case of 1-segment only).
enc->filter_hdr_.level_ = enc->dqm_[0].fstrength_;
enc->filter_hdr_.simple_ = (enc->config_->filter_type == 0);
enc->filter_hdr_.sharpness_ = enc->config_->filter_sharpness;
}
// Note: if you change the values below, remember that the max range // allowed by the syntax for DQ_UV is [-16,16]. #define MAX_DQ_UV (6) #define MIN_DQ_UV (-4)
// We want to emulate jpeg-like behaviour where the expected "good" quality // is around q=75. Internally, our "good" middle is around c=50. So we // map accordingly using linear piece-wise function staticdouble QualityToCompression(double c) { constdouble linear_c = (c < 0.75) ? c * (2. / 3.) : 2. * c - 1.; // The file size roughly scales as pow(quantizer, 3.). Actually, the // exponent is somewhere between 2.8 and 3.2, but we're mostly interested // in the mid-quant range. So we scale the compressibility inversely to // this power-law: quant ~= compression ^ 1/3. This law holds well for // low quant. Finer modeling for high-quant would make use of kAcTable[] // more explicitly. constdouble v = pow(linear_c, 1 / 3.); return v;
}
staticdouble QualityToJPEGCompression(double c, double alpha) { // We map the complexity 'alpha' and quality setting 'c' to a compression // exponent empirically matched to the compression curve of libjpeg6b. // On average, the WebP output size will be roughly similar to that of a // JPEG file compressed with same quality factor. constdouble amin = 0.30; constdouble amax = 0.85; constdouble exp_min = 0.4; constdouble exp_max = 0.9; constdouble slope = (exp_min - exp_max) / (amax - amin); // Linearly interpolate 'expn' from exp_min to exp_max // in the [amin, amax] range. constdouble expn = (alpha > amax) ? exp_min
: (alpha < amin) ? exp_max
: exp_max + slope * (alpha - amin); constdouble v = pow(c, expn); return v;
}
staticvoid SimplifySegments(VP8Encoder* const enc) { int map[NUM_MB_SEGMENTS] = { 0, 1, 2, 3 }; // 'num_segments_' is previously validated and <= NUM_MB_SEGMENTS, but an // explicit check is needed to avoid a spurious warning about 'i' exceeding // array bounds of 'dqm_' with some compilers (noticed with gcc-4.9). constint num_segments = (enc->segment_hdr_.num_segments_ < NUM_MB_SEGMENTS)
? enc->segment_hdr_.num_segments_
: NUM_MB_SEGMENTS; int num_final_segments = 1; int s1, s2; for (s1 = 1; s1 < num_segments; ++s1) { // find similar segments const VP8SegmentInfo* const S1 = &enc->dqm_[s1]; int found = 0; // check if we already have similar segment for (s2 = 0; s2 < num_final_segments; ++s2) { const VP8SegmentInfo* const S2 = &enc->dqm_[s2]; if (SegmentsAreEquivalent(S1, S2)) {
found = 1; break;
}
}
map[s1] = s2; if (!found) { if (num_final_segments != s1) {
enc->dqm_[num_final_segments] = enc->dqm_[s1];
}
++num_final_segments;
}
} if (num_final_segments < num_segments) { // Remap int i = enc->mb_w_ * enc->mb_h_; while (i-- > 0) enc->mb_info_[i].segment_ = map[enc->mb_info_[i].segment_];
enc->segment_hdr_.num_segments_ = num_final_segments; // Replicate the trailing segment infos (it's mostly cosmetics) for (i = num_final_segments; i < num_segments; ++i) {
enc->dqm_[i] = enc->dqm_[num_final_segments - 1];
}
}
}
void VP8SetSegmentParams(VP8Encoder* const enc, float quality) { int i; int dq_uv_ac, dq_uv_dc; constint num_segments = enc->segment_hdr_.num_segments_; constdouble amp = SNS_TO_DQ * enc->config_->sns_strength / 100. / 128.; constdouble Q = quality / 100.; constdouble c_base = enc->config_->emulate_jpeg_size ?
QualityToJPEGCompression(Q, enc->alpha_ / 255.) :
QualityToCompression(Q); for (i = 0; i < num_segments; ++i) { // We modulate the base coefficient to accommodate for the quantization // susceptibility and allow denser segments to be quantized more. constdouble expn = 1. - amp * enc->dqm_[i].alpha_; constdouble c = pow(c_base, expn); constint q = (int)(127. * (1. - c));
assert(expn > 0.);
enc->dqm_[i].quant_ = clip(q, 0, 127);
}
// purely indicative in the bitstream (except for the 1-segment case)
enc->base_quant_ = enc->dqm_[0].quant_;
// fill-in values for the unused segments (required by the syntax) for (i = num_segments; i < NUM_MB_SEGMENTS; ++i) {
enc->dqm_[i].quant_ = enc->base_quant_;
}
// uv_alpha_ is normally spread around ~60. The useful range is // typically ~30 (quite bad) to ~100 (ok to decimate UV more). // We map it to the safe maximal range of MAX/MIN_DQ_UV for dq_uv.
dq_uv_ac = (enc->uv_alpha_ - MID_ALPHA) * (MAX_DQ_UV - MIN_DQ_UV)
/ (MAX_ALPHA - MIN_ALPHA); // we rescale by the user-defined strength of adaptation
dq_uv_ac = dq_uv_ac * enc->config_->sns_strength / 100; // and make it safe.
dq_uv_ac = clip(dq_uv_ac, MIN_DQ_UV, MAX_DQ_UV); // We also boost the dc-uv-quant a little, based on sns-strength, since // U/V channels are quite more reactive to high quants (flat DC-blocks // tend to appear, and are unpleasant).
dq_uv_dc = -4 * enc->config_->sns_strength / 100;
dq_uv_dc = clip(dq_uv_dc, -15, 15); // 4bit-signed max allowed
// Form all the ten Intra4x4 predictions in the yuv_p_ cache // for the 4x4 block it->i4_ staticvoid MakeIntra4Preds(const VP8EncIterator* const it) {
VP8EncPredLuma4(it->yuv_p_, it->i4_top_);
}
// If a coefficient was quantized to a value Q (using a neutral bias), // we test all alternate possibilities between [Q-MIN_DELTA, Q+MAX_DELTA] // We don't test negative values though. #define MIN_DELTA 0 // how much lower level to try #define MAX_DELTA 1 // how much higher #define NUM_NODES (MIN_DELTA + 1 + MAX_DELTA) #define NODE(n, l) (nodes[(n)][(l) + MIN_DELTA]) #define SCORE_STATE(n, l) (score_states[n][(l) + MIN_DELTA])
// compute the position of the last interesting coefficient
last = first - 1; for (n = 15; n >= first; --n) { constint j = kZigzag[n]; constint err = in[j] * in[j]; if (err > thresh) {
last = n; break;
}
} // we don't need to go inspect up to n = 16 coeffs. We can just go up // to last + 1 (inclusive) without losing much. if (last < 15) ++last;
// compute 'skip' score. This is the max score one can do.
cost = VP8BitCost(0, last_proba);
best_score = RDScoreTrellis(lambda, cost, 0);
{ // Compute delta_error = how much coding this level will // subtract to max_error as distortion. // Here, distortion = sum of (|coeff_i| - level_i * Q_i)^2 constint new_error = coeff0 - level * Q; constint delta_error =
kWeightTrellis[j] * (new_error * new_error - coeff0 * coeff0);
base_score = RDScoreTrellis(lambda, 0, delta_error);
}
// Inspect all possible non-dead predecessors. Retain only the best one. // The base_score is added to all scores so it is only added for the final // value after the loop.
cost = VP8LevelCost(ss_prev[-MIN_DELTA].costs, level);
best_cur_score =
ss_prev[-MIN_DELTA].score + RDScoreTrellis(lambda, cost, 0);
best_prev = -MIN_DELTA; for (p = -MIN_DELTA + 1; p <= MAX_DELTA; ++p) { // Dead nodes (with ss_prev[p].score >= MAX_COST) are automatically // eliminated since their score can't be better than the current best.
cost = VP8LevelCost(ss_prev[p].costs, level); // Examine node assuming it's a non-terminal one.
score = ss_prev[p].score + RDScoreTrellis(lambda, cost, 0); if (score < best_cur_score) {
best_cur_score = score;
best_prev = p;
}
}
best_cur_score += base_score; // Store best finding in current node.
cur->sign = sign;
cur->level = level;
cur->prev = best_prev;
ss_cur[m].score = best_cur_score;
// Now, record best terminal node (and thus best entry in the graph). if (level != 0 && best_cur_score < best_score) { const score_t last_pos_cost =
(n < 15) ? VP8BitCost(0, probas[band][ctx][0]) : 0; const score_t last_pos_score = RDScoreTrellis(lambda, last_pos_cost, 0);
score = best_cur_score + last_pos_score; if (score < best_score) {
best_score = score;
best_path[0] = n; // best eob position
best_path[1] = m; // best node index
best_path[2] = best_prev; // best predecessor
}
}
}
}
// Fresh start // Beware! We must preserve in[0]/out[0] value for TYPE_I16_AC case. if (coeff_type == TYPE_I16_AC) {
memset(in + 1, 0, 15 * sizeof(*in));
memset(out + 1, 0, 15 * sizeof(*out));
} else {
memset(in, 0, 16 * sizeof(*in));
memset(out, 0, 16 * sizeof(*out));
} if (best_path[0] == -1) { return 0; // skip!
}
{ // Unwind the best path. // Note: best-prev on terminal node is not necessarily equal to the // best_prev for non-terminal. So we patch best_path[2] in. int nz = 0; int best_node = best_path[1];
n = best_path[0];
NODE(n, best_node).prev = best_path[2]; // force best-prev for terminal
//------------------------------------------------------------------------------ // Performs: difference, transform, quantize, back-transform, add // all at once. Output is the reconstructed block in *yuv_out, and the // quantized levels in *levels.
// Diffusion weights. We under-correct a bit (15/16th of the error is actually // diffused) to avoid 'rainbow' chessboard pattern of blocks at q~=0. #define C1 7 // fraction of error sent to the 4x4 block below #define C2 8 // fraction of error sent to the 4x4 block on the right #define DSHIFT 4 #define DSCALE 1 // storage descaling, needed to make the error fit int8_t
// Quantize as usual, but also compute and return the quantization error. // Error is already divided by DSHIFT. staticint QuantizeSingle(int16_t* WEBP_RESTRICT const v, const VP8Matrix* WEBP_RESTRICT const mtx) { int V = *v; constint sign = (V < 0); if (sign) V = -V; if (V > (int)mtx->zthresh_[0]) { constint qV = QUANTDIV(V, mtx->iq_[0], mtx->bias_[0]) * mtx->q_[0]; constint err = (V - qV);
*v = sign ? -qV : qV; return (sign ? -err : err) >> DSCALE;
}
*v = 0; return (sign ? -V : V) >> DSCALE;
}
staticvoid CorrectDCValues(const VP8EncIterator* WEBP_RESTRICT const it, const VP8Matrix* WEBP_RESTRICT const mtx,
int16_t tmp[][16],
VP8ModeScore* WEBP_RESTRICT const rd) { // | top[0] | top[1] // --------+--------+--------- // left[0] | tmp[0] tmp[1] <-> err0 err1 // left[1] | tmp[2] tmp[3] err2 err3 // // Final errors {err1,err2,err3} are preserved and later restored // as top[]/left[] on the next block. int ch; for (ch = 0; ch <= 1; ++ch) { const int8_t* const top = it->top_derr_[it->x_][ch]; const int8_t* const left = it->left_derr_[ch];
int16_t (* const c)[16] = &tmp[ch * 4]; int err0, err1, err2, err3;
c[0][0] += (C1 * top[0] + C2 * left[0]) >> (DSHIFT - DSCALE);
err0 = QuantizeSingle(&c[0][0], mtx);
c[1][0] += (C1 * top[1] + C2 * err0) >> (DSHIFT - DSCALE);
err1 = QuantizeSingle(&c[1][0], mtx);
c[2][0] += (C1 * err0 + C2 * left[1]) >> (DSHIFT - DSCALE);
err2 = QuantizeSingle(&c[2][0], mtx);
c[3][0] += (C1 * err1 + C2 * err2) >> (DSHIFT - DSCALE);
err3 = QuantizeSingle(&c[3][0], mtx); // error 'err' is bounded by mtx->q_[0] which is 132 at max. Hence // err >> DSCALE will fit in an int8_t type if DSCALE>=1.
assert(abs(err1) <= 127 && abs(err2) <= 127 && abs(err3) <= 127);
rd->derr[ch][0] = (int8_t)err1;
rd->derr[ch][1] = (int8_t)err2;
rd->derr[ch][2] = (int8_t)err3;
}
}
for (n = 0; n < 8; n += 2) {
VP8ITransform(ref + VP8ScanUV[n], tmp[n], yuv_out + VP8ScanUV[n], 1);
} return (nz << 16);
}
//------------------------------------------------------------------------------ // RD-opt decision. Reconstruct each modes, evalue distortion and bit-cost. // Pick the mode is lower RD-cost = Rate + lambda * Distortion.
staticvoid StoreMaxDelta(VP8SegmentInfo* const dqm, const int16_t DCs[16]) { // We look at the first three AC coefficients to determine what is the average // delta between each sub-4x4 block. constint v0 = abs(DCs[1]); constint v1 = abs(DCs[2]); constint v2 = abs(DCs[4]); int max_v = (v1 > v0) ? v1 : v0;
max_v = (v2 > max_v) ? v2 : max_v; if (max_v > dqm->max_edge_) dqm->max_edge_ = max_v;
}
staticvoid SwapModeScore(VP8ModeScore** a, VP8ModeScore** b) {
VP8ModeScore* const tmp = *a;
*a = *b;
*b = tmp;
}
staticvoid SwapPtr(uint8_t** a, uint8_t** b) {
uint8_t* const tmp = *a;
*a = *b;
*b = tmp;
}
// Measure RD-score
rd_cur->D = VP8SSE16x16(src, tmp_dst);
rd_cur->SD =
tlambda ? MULT_8B(tlambda, VP8TDisto16x16(src, tmp_dst, kWeightY)) : 0;
rd_cur->H = VP8FixedCostsI16[mode];
rd_cur->R = VP8GetCostLuma16(it, rd_cur); if (is_flat) { // refine the first impression (which was in pixel space)
is_flat = IsFlat(rd_cur->y_ac_levels[0], kNumBlocks, FLATNESS_LIMIT_I16); if (is_flat) { // Block is very flat. We put emphasis on the distortion being very low!
rd_cur->D *= 2;
rd_cur->SD *= 2;
}
}
// Since we always examine Intra16 first, we can overwrite *rd directly.
SetRDScore(lambda, rd_cur); if (mode == 0 || rd_cur->score < rd_best->score) {
SwapModeScore(&rd_cur, &rd_best);
SwapOut(it);
}
} if (rd_best != rd) {
memcpy(rd, rd_best, sizeof(*rd));
}
SetRDScore(dqm->lambda_mode_, rd); // finalize score for mode decision.
VP8SetIntra16Mode(it, rd->mode_i16);
// we have a blocky macroblock (only DCs are non-zero) with fairly high // distortion, record max delta so we can later adjust the minimal filtering // strength needed to smooth these blocks out. if ((rd->nz & 0x100ffff) == 0x1000000 && rd->D > dqm->min_disto_) {
StoreMaxDelta(dqm, rd->y_dc_levels);
}
}
if (score < best_score) {
best_mode = mode;
best_score = score;
}
} if (it->x_ == 0 || it->y_ == 0) { // avoid starting a checkerboard resonance from the border. See bug #432. if (IsFlatSource16(src)) {
best_mode = (it->x_ == 0) ? 0 : 2;
try_both_modes = 0; // stick to i16
}
}
VP8SetIntra16Mode(it, best_mode); // we'll reconstruct later, if i16 mode actually gets selected
}
// Next, evaluate Intra4 if (try_both_modes || !is_i16) { // We don't evaluate the rate here, but just account for it through a // constant penalty (i4 mode usually needs more bits compared to i16).
is_i16 = 0;
VP8IteratorStartI4(it); do { int best_i4_mode = -1;
score_t best_i4_score = MAX_COST; const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC + VP8Scan[it->i4_]; const uint16_t* const mode_costs = GetCostModeI4(it, rd->modes_i4);
//------------------------------------------------------------------------------ // Entry point
int VP8Decimate(VP8EncIterator* WEBP_RESTRICT const it,
VP8ModeScore* WEBP_RESTRICT const rd,
VP8RDLevel rd_opt) { int is_skipped; constint method = it->enc_->method_;
InitScore(rd);
// We can perform predictions for Luma16x16 and Chroma8x8 already. // Luma4x4 predictions needs to be done as-we-go.
VP8MakeLuma16Preds(it);
VP8MakeChroma8Preds(it);
if (rd_opt > RD_OPT_NONE) {
it->do_trellis_ = (rd_opt >= RD_OPT_TRELLIS_ALL);
PickBestIntra16(it, rd); if (method >= 2) {
PickBestIntra4(it, rd);
}
PickBestUV(it, rd); if (rd_opt == RD_OPT_TRELLIS) { // finish off with trellis-optim now
it->do_trellis_ = 1;
SimpleQuantize(it, rd);
}
} else { // At this point we have heuristically decided intra16 / intra4. // For method >= 2, pick the best intra4/intra16 based on SSE (~tad slower). // For method <= 1, we don't re-examine the decision but just go ahead with // quantization/reconstruction.
RefineUsingDistortion(it, (method >= 2), (method >= 1), rd);
}
is_skipped = (rd->nz == 0);
VP8SetSkip(it, is_skipped); return is_skipped;
}
Messung V0.5
¤ Dauer der Verarbeitung: 0.18 Sekunden
(vorverarbeitet)
¤
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.