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);
}
}
private static void RunVp8Sse16X16Test()
{
// arrange
byte[] a =
{
154, 154, 151, 151, 149, 148, 151, 157, 163, 163, 154, 132, 102, 98, 104, 108, 107, 104, 104, 103,
101, 106, 123, 119, 170, 171, 172, 171, 168, 175, 171, 173, 151, 151, 149, 150, 147, 147, 146, 159,
164, 165, 154, 129, 92, 90, 101, 105, 104, 103, 104, 101, 100, 105, 123, 117, 172, 172, 172, 168,
170, 177, 170, 175, 151, 149, 150, 150, 147, 147, 156, 161, 161, 161, 151, 126, 93, 90, 102, 107,
104, 103, 104, 101, 104, 104, 122, 117, 172, 172, 170, 168, 170, 177, 172, 175, 150, 149, 152, 151,
148, 151, 160, 159, 157, 157, 148, 133, 96, 90, 103, 107, 104, 104, 101, 100, 102, 102, 121, 117,
170, 170, 169, 171, 171, 179, 173, 175, 149, 151, 152, 151, 148, 154, 162, 157, 154, 154, 151, 132,
92, 89, 101, 108, 104, 102, 101, 101, 103, 103, 123, 118, 171, 168, 177, 173, 171, 178, 172, 176,
152, 152, 152, 151, 154, 162, 161, 155, 149, 157, 156, 129, 92, 87, 101, 107, 102, 100, 107, 100,
101, 102, 123, 118, 170, 175, 182, 172, 171, 179, 173, 175, 152, 151, 154, 155, 160, 162, 161, 153,
150, 156, 153, 129, 92, 91, 102, 106, 100, 109, 115, 99, 101, 102, 124, 120, 171, 179, 178, 172,
171, 181, 171, 173, 154, 154, 154, 162, 160, 158, 156, 152, 153, 157, 151, 128, 86, 86, 102, 105,
102, 122, 114, 99, 101, 102, 125, 120, 178, 173, 177, 172, 171, 180, 172, 173, 154, 152, 158, 163,
150, 148, 148, 156, 151, 158, 152, 129, 87, 87, 101, 105, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 154, 151, 165, 156, 141, 137, 146, 158, 152, 159, 152, 133,
90, 88, 99, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
154, 160, 164, 150, 126, 127, 149, 159, 155, 161, 153, 131, 84, 86, 97, 103, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 157, 167, 157, 137, 102, 128, 155, 161,
157, 159, 154, 134, 84, 82, 97, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 163, 163, 150, 113, 78, 132, 156, 162, 159, 160, 154, 132, 83, 78, 91, 97, 204,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 163, 157, 137, 80, 78,
131, 154, 163, 157, 159, 149, 131, 82, 77, 94, 100, 204, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 159, 151, 108, 72, 88, 132, 156, 162, 159, 157, 151, 130, 79, 78,
95, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 151, 130,
82, 82, 89, 134, 154, 161, 161, 157, 152, 129, 81, 77, 95, 102, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 204
};
byte[] b =
{
150, 150, 150, 150, 146, 149, 152, 154, 164, 166, 154, 132, 99, 92, 106, 112, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 150, 150, 150, 150, 146, 149, 152, 154,
161, 164, 151, 130, 93, 86, 100, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 150, 150, 150, 150, 146, 149, 152, 154, 158, 161, 148, 127, 93, 86, 100, 106,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 150, 150, 150, 150,
146, 149, 152, 154, 156, 159, 146, 125, 99, 92, 106, 112, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 148, 148, 148, 148, 149, 158, 162, 159, 155, 155, 153, 129,
94, 87, 101, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
151, 151, 151, 151, 152, 159, 161, 156, 155, 155, 153, 129, 94, 87, 101, 106, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 154, 154, 154, 154, 156, 161, 159, 152,
155, 155, 153, 129, 94, 87, 101, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 156, 156, 156, 156, 159, 162, 158, 149, 155, 155, 153, 129, 94, 87, 101, 106,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 152, 153, 157, 162,
150, 149, 149, 151, 155, 160, 150, 131, 91, 90, 104, 104, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 152, 156, 158, 157, 140, 137, 145, 159, 155, 160, 150, 131,
89, 88, 102, 101, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
153, 161, 160, 149, 118, 128, 147, 162, 155, 160, 150, 131, 86, 85, 99, 98, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 154, 165, 161, 144, 96, 128, 154, 159, 155,
160, 150, 131, 83, 82, 97, 96, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 161, 160, 149, 105, 78, 127, 156, 170, 156, 156, 154, 130, 81, 77, 95, 102, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 160, 160, 133, 85, 81, 129, 155,
167, 156, 156, 154, 130, 81, 77, 95, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 156, 147, 109, 76, 85, 130, 153, 163, 156, 156, 154, 130, 81, 77, 95, 102,
204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 152, 128, 87, 83,
88, 132, 152, 159, 156, 156, 154, 130, 81, 77, 95, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204,
204, 204, 204, 204, 204, 204, 204
};
int expected = 2063;
// act
int actual = LossyUtils.Vp8_Sse16X16(a, b);
// assert
Assert.Equal(expected, actual);
}
// 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;
}