private unsafe void AdjustFilterStrength()
{
if (this.filterStrength > 0)
{
int maxLevel = 0;
for (int s = 0; s < WebpConstants.NumMbSegments; s++)
{
Vp8SegmentInfo dqm = this.SegmentInfos[s];
// this '>> 3' accounts for some inverse WHT scaling
int delta = (dqm.MaxEdge * dqm.Y2.Q[1]) >> 3;
int level = this.FilterStrengthFromDelta(this.FilterHeader.Sharpness, delta);
if (level > dqm.FStrength)
{
dqm.FStrength = level;
}
if (maxLevel < dqm.FStrength)
{
maxLevel = dqm.FStrength;
}
}
this.FilterHeader.FilterLevel = maxLevel;
}
}
private void SetupFilterStrength()
{
int filterSharpness = 0; // TODO: filterSharpness is hardcoded
int filterType = 1; // TODO: filterType is hardcoded
// level0 is in [0..500]. Using '-f 50' as filter_strength is mid-filtering.
int level0 = 5 * this.filterStrength;
for (int i = 0; i < WebpConstants.NumMbSegments; i++)
{
Vp8SegmentInfo m = this.SegmentInfos[i];
// We focus on the quantization of AC coeffs.
int qstep = WebpLookupTables.AcTable[Numerics.Clamp(m.Quant, 0, 127)] >> 2;
int baseStrength = this.FilterStrengthFromDelta(this.FilterHeader.Sharpness, qstep);
// Segments with lower complexity ('beta') will be less filtered.
int f = baseStrength * level0 / (256 + m.Beta);
m.FStrength = f < WebpConstants.FilterStrengthCutoff ? 0 : f > 63 ? 63 : f;
}
// We record the initial strength (mainly for the case of 1-segment only).
this.FilterHeader.FilterLevel = this.SegmentInfos[0].FStrength;
this.FilterHeader.Simple = filterType == 0;
this.FilterHeader.Sharpness = filterSharpness;
}
public static int ReconstructUv(Vp8EncIterator it, Vp8SegmentInfo dqm, Vp8ModeScore rd, Span <byte> yuvOut, int mode)
{
Span <byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8UvModeOffsets[mode]);
Span <byte> src = it.YuvIn.AsSpan(Vp8EncIterator.UOffEnc);
int nz = 0;
int n;
Span <short> tmp = it.Scratch2.AsSpan(0, 8 * 16);
Span <int> scratch = it.Scratch3.AsSpan(0, 16);
for (n = 0; n < 8; n += 2)
{
Vp8Encoding.FTransform2(
src.Slice(WebpLookupTables.Vp8ScanUv[n]),
reference.Slice(WebpLookupTables.Vp8ScanUv[n]),
tmp.Slice(n * 16, 16),
tmp.Slice((n + 1) * 16, 16),
scratch);
}
CorrectDcValues(it, ref dqm.Uv, tmp, rd);
for (n = 0; n < 8; n += 2)
{
nz |= Quantize2Blocks(tmp.Slice(n * 16, 32), rd.UvLevels.AsSpan(n * 16, 32), ref dqm.Uv) << n;
}
for (n = 0; n < 8; n += 2)
{
Vp8Encoding.ITransformTwo(reference.Slice(WebpLookupTables.Vp8ScanUv[n]), tmp.Slice(n * 16, 32), yuvOut.Slice(WebpLookupTables.Vp8ScanUv[n]), scratch);
}
return(nz << 16);
}
public static int ReconstructIntra4(Vp8EncIterator it, Vp8SegmentInfo dqm, Span <short> levels, Span <byte> src, Span <byte> yuvOut, int mode)
{
Span <byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8I4ModeOffsets[mode]);
Span <short> tmp = it.Scratch2.AsSpan(0, 16);
Span <int> scratch = it.Scratch3.AsSpan(0, 16);
Vp8Encoding.FTransform(src, reference, tmp, scratch);
int nz = QuantizeBlock(tmp, levels, ref dqm.Y1);
Vp8Encoding.ITransformOne(reference, tmp, yuvOut, scratch);
return(nz);
}
public static int ReconstructIntra16(Vp8EncIterator it, Vp8SegmentInfo dqm, Vp8ModeScore rd, Span <byte> yuvOut, int mode)
{
Span <byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8I16ModeOffsets[mode]);
Span <byte> src = it.YuvIn.AsSpan(Vp8EncIterator.YOffEnc);
int nz = 0;
int n;
Span <short> shortScratchSpan = it.Scratch2.AsSpan();
Span <int> scratch = it.Scratch3.AsSpan(0, 16);
shortScratchSpan.Clear();
scratch.Clear();
Span <short> dcTmp = shortScratchSpan.Slice(0, 16);
Span <short> tmp = shortScratchSpan.Slice(16, 16 * 16);
for (n = 0; n < 16; n += 2)
{
Vp8Encoding.FTransform2(
src.Slice(WebpLookupTables.Vp8Scan[n]),
reference.Slice(WebpLookupTables.Vp8Scan[n]),
tmp.Slice(n * 16, 16),
tmp.Slice((n + 1) * 16, 16),
scratch);
}
Vp8Encoding.FTransformWht(tmp, dcTmp, scratch);
nz |= QuantizeBlock(dcTmp, rd.YDcLevels, ref dqm.Y2) << 24;
for (n = 0; n < 16; n += 2)
{
// Zero-out the first coeff, so that: a) nz is correct below, and
// b) finding 'last' non-zero coeffs in SetResidualCoeffs() is simplified.
tmp[n * 16] = tmp[(n + 1) * 16] = 0;
nz |= Quantize2Blocks(tmp.Slice(n * 16, 32), rd.YAcLevels.AsSpan(n * 16, 32), ref dqm.Y1) << n;
}
// Transform back.
LossyUtils.TransformWht(dcTmp, tmp, scratch);
for (n = 0; n < 16; n += 2)
{
Vp8Encoding.ITransformTwo(reference.Slice(WebpLookupTables.Vp8Scan[n]), tmp.Slice(n * 16, 32), yuvOut.Slice(WebpLookupTables.Vp8Scan[n]), scratch);
}
return(nz);
}
private unsafe void SetupMatrices(Vp8SegmentInfo[] dqm)
{
int tlambdaScale = this.method >= WebpEncodingMethod.Default ? this.spatialNoiseShaping : 0;
for (int i = 0; i < dqm.Length; i++)
{
Vp8SegmentInfo m = dqm[i];
int q = m.Quant;
m.Y1.Q[0] = WebpLookupTables.DcTable[Numerics.Clamp(q + this.DqY1Dc, 0, 127)];
m.Y1.Q[1] = WebpLookupTables.AcTable[Numerics.Clamp(q, 0, 127)];
m.Y2.Q[0] = (ushort)(WebpLookupTables.DcTable[Numerics.Clamp(q + this.DqY2Dc, 0, 127)] * 2);
m.Y2.Q[1] = WebpLookupTables.AcTable2[Numerics.Clamp(q + this.DqY2Ac, 0, 127)];
m.Uv.Q[0] = WebpLookupTables.DcTable[Numerics.Clamp(q + this.DqUvDc, 0, 117)];
m.Uv.Q[1] = WebpLookupTables.AcTable[Numerics.Clamp(q + this.DqUvAc, 0, 127)];
int qi4 = m.Y1.Expand(0);
int qi16 = m.Y2.Expand(1);
int quv = m.Uv.Expand(2);
m.LambdaI16 = 3 * qi16 * qi16;
m.LambdaI4 = (3 * qi4 * qi4) >> 7;
m.LambdaUv = (3 * quv * quv) >> 6;
m.LambdaMode = (1 * qi4 * qi4) >> 7;
m.TLambda = (tlambdaScale * qi4) >> 5;
// none of these constants should be < 1.
m.LambdaI16 = m.LambdaI16 < 1 ? 1 : m.LambdaI16;
m.LambdaI4 = m.LambdaI4 < 1 ? 1 : m.LambdaI4;
m.LambdaUv = m.LambdaUv < 1 ? 1 : m.LambdaUv;
m.LambdaMode = m.LambdaMode < 1 ? 1 : m.LambdaMode;
m.TLambda = m.TLambda < 1 ? 1 : m.TLambda;
m.MinDisto = 20 * m.Y1.Q[0];
m.MaxEdge = 0;
m.I4Penalty = 1000 * qi4 * qi4;
}
}
private const int DSCALE = 1; // storage descaling, needed to make the error fit byte
public static void PickBestIntra16(Vp8EncIterator it, ref Vp8ModeScore rd, Vp8SegmentInfo[] segmentInfos, Vp8EncProba proba)
{
const int numBlocks = 16;
Vp8SegmentInfo dqm = segmentInfos[it.CurrentMacroBlockInfo.Segment];
int lambda = dqm.LambdaI16;
int tlambda = dqm.TLambda;
Span <byte> src = it.YuvIn.AsSpan(Vp8EncIterator.YOffEnc);
Span <int> scratch = it.Scratch3;
var rdTmp = new Vp8ModeScore();
var res = new Vp8Residual();
Vp8ModeScore rdCur = rdTmp;
Vp8ModeScore rdBest = rd;
int mode;
bool isFlat = IsFlatSource16(src);
rd.ModeI16 = -1;
for (mode = 0; mode < WebpConstants.NumPredModes; ++mode)
{
// Scratch buffer.
Span <byte> tmpDst = it.YuvOut2.AsSpan(Vp8EncIterator.YOffEnc);
rdCur.ModeI16 = mode;
// Reconstruct.
rdCur.Nz = (uint)ReconstructIntra16(it, dqm, rdCur, tmpDst, mode);
// Measure RD-score.
rdCur.D = LossyUtils.Vp8_Sse16X16(src, tmpDst);
rdCur.SD = tlambda != 0 ? Mult8B(tlambda, LossyUtils.Vp8Disto16X16(src, tmpDst, WeightY, scratch)) : 0;
rdCur.H = WebpConstants.Vp8FixedCostsI16[mode];
rdCur.R = it.GetCostLuma16(rdCur, proba, res);
if (isFlat)
{
// Refine the first impression (which was in pixel space).
isFlat = IsFlat(rdCur.YAcLevels, numBlocks, WebpConstants.FlatnessLimitI16);
if (isFlat)
{
// Block is very flat. We put emphasis on the distortion being very low!
rdCur.D *= 2;
rdCur.SD *= 2;
}
}
// Since we always examine Intra16 first, we can overwrite *rd directly.
rdCur.SetRdScore(lambda);
if (mode == 0 || rdCur.Score < rdBest.Score)
{
Vp8ModeScore tmp = rdCur;
rdCur = rdBest;
rdBest = tmp;
it.SwapOut();
}
}
if (rdBest != rd)
{
rd = rdBest;
}
// Finalize score for mode decision.
rd.SetRdScore(dqm.LambdaMode);
it.SetIntra16Mode(rd.ModeI16);
// 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.MinDisto)
{
dqm.StoreMaxDelta(rd.YDcLevels);
}
}
// Refine intra16/intra4 sub-modes based on distortion only (not rate).
public static void RefineUsingDistortion(Vp8EncIterator it, Vp8SegmentInfo[] segmentInfos, Vp8ModeScore rd, bool tryBothModes, bool refineUvMode, int mbHeaderLimit)
{
long bestScore = Vp8ModeScore.MaxCost;
int nz = 0;
int mode;
bool isI16 = tryBothModes || it.CurrentMacroBlockInfo.MacroBlockType == Vp8MacroBlockType.I16X16;
Vp8SegmentInfo dqm = segmentInfos[it.CurrentMacroBlockInfo.Segment];
// Some empiric constants, of approximate order of magnitude.
const int lambdaDi16 = 106;
const int lambdaDi4 = 11;
const int lambdaDuv = 120;
long scoreI4 = dqm.I4Penalty;
long i4BitSum = 0;
long bitLimit = tryBothModes
? mbHeaderLimit
: Vp8ModeScore.MaxCost; // no early-out allowed.
if (isI16)
{
int bestMode = -1;
Span <byte> src = it.YuvIn.AsSpan(Vp8EncIterator.YOffEnc);
for (mode = 0; mode < WebpConstants.NumPredModes; ++mode)
{
Span <byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8I16ModeOffsets[mode]);
long score = (LossyUtils.Vp8_Sse16X16(src, reference) * WebpConstants.RdDistoMult) + (WebpConstants.Vp8FixedCostsI16[mode] * lambdaDi16);
if (mode > 0 && WebpConstants.Vp8FixedCostsI16[mode] > bitLimit)
{
continue;
}
if (score < bestScore)
{
bestMode = mode;
bestScore = score;
}
}
if (it.X == 0 || it.Y == 0)
{
// Avoid starting a checkerboard resonance from the border. See bug #432 of libwebp.
if (IsFlatSource16(src))
{
bestMode = it.X == 0 ? 0 : 2;
tryBothModes = false; // Stick to i16.
}
}
it.SetIntra16Mode(bestMode);
// We'll reconstruct later, if i16 mode actually gets selected.
}
// Next, evaluate Intra4.
if (tryBothModes || !isI16)
{
// 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).
isI16 = false;
it.StartI4();
do
{
int bestI4Mode = -1;
long bestI4Score = Vp8ModeScore.MaxCost;
Span <byte> src = it.YuvIn.AsSpan(Vp8EncIterator.YOffEnc + WebpLookupTables.Vp8Scan[it.I4]);
short[] modeCosts = it.GetCostModeI4(rd.ModesI4);
it.MakeIntra4Preds();
for (mode = 0; mode < WebpConstants.NumBModes; ++mode)
{
Span <byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8I4ModeOffsets[mode]);
long score = (LossyUtils.Vp8_Sse4X4(src, reference) * WebpConstants.RdDistoMult) + (modeCosts[mode] * lambdaDi4);
if (score < bestI4Score)
{
bestI4Mode = mode;
bestI4Score = score;
}
}
i4BitSum += modeCosts[bestI4Mode];
rd.ModesI4[it.I4] = (byte)bestI4Mode;
scoreI4 += bestI4Score;
if (scoreI4 >= bestScore || i4BitSum > bitLimit)
{
// Intra4 won't be better than Intra16. Bail out and pick Intra16.
isI16 = true;
break;
}
else
{
// Reconstruct partial block inside YuvOut2 buffer
Span <byte> tmpDst = it.YuvOut2.AsSpan(Vp8EncIterator.YOffEnc + WebpLookupTables.Vp8Scan[it.I4]);
nz |= ReconstructIntra4(it, dqm, rd.YAcLevels.AsSpan(it.I4 * 16, 16), src, tmpDst, bestI4Mode) << it.I4;
}
}while (it.RotateI4(it.YuvOut2.AsSpan(Vp8EncIterator.YOffEnc)));
}
// Final reconstruction, depending on which mode is selected.
if (!isI16)
{
it.SetIntra4Mode(rd.ModesI4);
it.SwapOut();
bestScore = scoreI4;
}
else
{
int intra16Mode = it.Preds[it.PredIdx];
nz = ReconstructIntra16(it, dqm, rd, it.YuvOut.AsSpan(Vp8EncIterator.YOffEnc), intra16Mode);
}
// ... and UV!
if (refineUvMode)
{
int bestMode = -1;
long bestUvScore = Vp8ModeScore.MaxCost;
Span <byte> src = it.YuvIn.AsSpan(Vp8EncIterator.UOffEnc);
for (mode = 0; mode < WebpConstants.NumPredModes; ++mode)
{
Span <byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8UvModeOffsets[mode]);
long score = (LossyUtils.Vp8_Sse16X8(src, reference) * WebpConstants.RdDistoMult) + (WebpConstants.Vp8FixedCostsUv[mode] * lambdaDuv);
if (score < bestUvScore)
{
bestMode = mode;
bestUvScore = score;
}
}
it.SetIntraUvMode(bestMode);
}
nz |= ReconstructUv(it, dqm, rd, it.YuvOut.AsSpan(Vp8EncIterator.UOffEnc), it.CurrentMacroBlockInfo.UvMode);
rd.Nz = (uint)nz;
rd.Score = bestScore;
}
public static void PickBestUv(Vp8EncIterator it, ref Vp8ModeScore rd, Vp8SegmentInfo[] segmentInfos, Vp8EncProba proba)
{
const int numBlocks = 8;
Vp8SegmentInfo dqm = segmentInfos[it.CurrentMacroBlockInfo.Segment];
int lambda = dqm.LambdaUv;
Span <byte> src = it.YuvIn.AsSpan(Vp8EncIterator.UOffEnc);
Span <byte> tmpDst = it.YuvOut2.AsSpan(Vp8EncIterator.UOffEnc);
Span <byte> dst0 = it.YuvOut.AsSpan(Vp8EncIterator.UOffEnc);
Span <byte> dst = dst0;
var rdBest = new Vp8ModeScore();
var rdUv = new Vp8ModeScore();
var res = new Vp8Residual();
int mode;
rd.ModeUv = -1;
rdBest.InitScore();
for (mode = 0; mode < WebpConstants.NumPredModes; ++mode)
{
rdUv.Clear();
// Reconstruct
rdUv.Nz = (uint)ReconstructUv(it, dqm, rdUv, tmpDst, mode);
// Compute RD-score
rdUv.D = LossyUtils.Vp8_Sse16X8(src, tmpDst);
rdUv.SD = 0; // not calling TDisto here: it tends to flatten areas.
rdUv.H = WebpConstants.Vp8FixedCostsUv[mode];
rdUv.R = it.GetCostUv(rdUv, proba, res);
if (mode > 0 && IsFlat(rdUv.UvLevels, numBlocks, WebpConstants.FlatnessLimitIUv))
{
rdUv.R += WebpConstants.FlatnessPenality * numBlocks;
}
rdUv.SetRdScore(lambda);
if (mode == 0 || rdUv.Score < rdBest.Score)
{
rdBest.CopyScore(rdUv);
rd.ModeUv = mode;
rdUv.UvLevels.CopyTo(rd.UvLevels.AsSpan());
for (int i = 0; i < 2; i++)
{
rd.Derr[i, 0] = rdUv.Derr[i, 0];
rd.Derr[i, 1] = rdUv.Derr[i, 1];
rd.Derr[i, 2] = rdUv.Derr[i, 2];
}
Span <byte> tmp = dst;
dst = tmpDst;
tmpDst = tmp;
}
}
it.SetIntraUvMode(rd.ModeUv);
rd.AddScore(rdBest);
if (dst != dst0)
{
// copy 16x8 block if needed.
LossyUtils.Vp8Copy16X8(dst, dst0);
}
// Store diffusion errors for next block.
it.StoreDiffusionErrors(rd);
}
public static bool PickBestIntra4(Vp8EncIterator it, ref Vp8ModeScore rd, Vp8SegmentInfo[] segmentInfos, Vp8EncProba proba, int maxI4HeaderBits)
{
Vp8SegmentInfo dqm = segmentInfos[it.CurrentMacroBlockInfo.Segment];
int lambda = dqm.LambdaI4;
int tlambda = dqm.TLambda;
Span <byte> src0 = it.YuvIn.AsSpan(Vp8EncIterator.YOffEnc);
Span <byte> bestBlocks = it.YuvOut2.AsSpan(Vp8EncIterator.YOffEnc);
Span <int> scratch = it.Scratch3;
int totalHeaderBits = 0;
var rdBest = new Vp8ModeScore();
if (maxI4HeaderBits == 0)
{
return(false);
}
rdBest.InitScore();
rdBest.H = 211; // '211' is the value of VP8BitCost(0, 145)
rdBest.SetRdScore(dqm.LambdaMode);
it.StartI4();
var rdi4 = new Vp8ModeScore();
var rdTmp = new Vp8ModeScore();
var res = new Vp8Residual();
Span <short> tmpLevels = new short[16];
do
{
int numBlocks = 1;
rdi4.Clear();
int mode;
int bestMode = -1;
Span <byte> src = src0.Slice(WebpLookupTables.Vp8Scan[it.I4]);
short[] modeCosts = it.GetCostModeI4(rd.ModesI4);
Span <byte> bestBlock = bestBlocks.Slice(WebpLookupTables.Vp8Scan[it.I4]);
Span <byte> tmpDst = it.Scratch.AsSpan();
tmpDst.Clear();
rdi4.InitScore();
it.MakeIntra4Preds();
for (mode = 0; mode < WebpConstants.NumBModes; ++mode)
{
rdTmp.Clear();
tmpLevels.Clear();
// Reconstruct.
rdTmp.Nz = (uint)ReconstructIntra4(it, dqm, tmpLevels, src, tmpDst, mode);
// Compute RD-score.
rdTmp.D = LossyUtils.Vp8_Sse4X4(src, tmpDst);
rdTmp.SD = tlambda != 0 ? Mult8B(tlambda, LossyUtils.Vp8Disto4X4(src, tmpDst, WeightY, scratch)) : 0;
rdTmp.H = modeCosts[mode];
// Add flatness penalty, to avoid flat area to be mispredicted by a complex mode.
if (mode > 0 && IsFlat(tmpLevels, numBlocks, WebpConstants.FlatnessLimitI4))
{
rdTmp.R = WebpConstants.FlatnessPenality * numBlocks;
}
else
{
rdTmp.R = 0;
}
// Early-out check.
rdTmp.SetRdScore(lambda);
if (bestMode >= 0 && rdTmp.Score >= rdi4.Score)
{
continue;
}
// Finish computing score.
rdTmp.R += it.GetCostLuma4(tmpLevels, proba, res);
rdTmp.SetRdScore(lambda);
if (bestMode < 0 || rdTmp.Score < rdi4.Score)
{
rdi4.CopyScore(rdTmp);
bestMode = mode;
Span <byte> tmp = tmpDst;
tmpDst = bestBlock;
bestBlock = tmp;
tmpLevels.CopyTo(rdBest.YAcLevels.AsSpan(it.I4 * 16, 16));
}
}
rdi4.SetRdScore(dqm.LambdaMode);
rdBest.AddScore(rdi4);
if (rdBest.Score >= rd.Score)
{
return(false);
}
totalHeaderBits += (int)rdi4.H; // <- equal to modeCosts[bestMode];
if (totalHeaderBits > maxI4HeaderBits)
{
return(false);
}
// Copy selected samples to the right place.
LossyUtils.Vp8Copy4X4(bestBlock, bestBlocks.Slice(WebpLookupTables.Vp8Scan[it.I4]));
rd.ModesI4[it.I4] = (byte)bestMode;
it.TopNz[it.I4 & 3] = it.LeftNz[it.I4 >> 2] = rdi4.Nz != 0 ? 1 : 0;
}while (it.RotateI4(bestBlocks));
// Finalize state.
rd.CopyScore(rdBest);
it.SetIntra4Mode(rd.ModesI4);
it.SwapOut();
rdBest.YAcLevels.AsSpan().CopyTo(rd.YAcLevels);
// Select intra4x4 over intra16x16.
return(true);
}
