private bool Decimate(Vp8EncIterator it, ref Vp8ModeScore rd, Vp8RdLevel rdOpt)
{
rd.InitScore();
// We can perform predictions for Luma16x16 and Chroma8x8 already.
// Luma4x4 predictions needs to be done as-we-go.
it.MakeLuma16Preds();
it.MakeChroma8Preds();
if (rdOpt > Vp8RdLevel.RdOptNone)
{
QuantEnc.PickBestIntra16(it, ref rd, this.SegmentInfos, this.Proba);
if (this.method >= WebpEncodingMethod.Level2)
{
QuantEnc.PickBestIntra4(it, ref rd, this.SegmentInfos, this.Proba, this.maxI4HeaderBits);
}
QuantEnc.PickBestUv(it, ref rd, this.SegmentInfos, this.Proba);
}
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.
QuantEnc.RefineUsingDistortion(it, this.SegmentInfos, rd, this.method >= WebpEncodingMethod.Level2, this.method >= WebpEncodingMethod.Level1, this.MbHeaderLimit);
}
bool isSkipped = rd.Nz == 0;
it.SetSkip(isSkipped);
return(isSkipped);
}
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);
}