Overview
Request 4640 (accepted)
- Update to version 3.2.1
* Fix output change in analysis load
* Fix encoder crash with zones and add test for zones
* Fix: Performance drop in aq-mode 4 This patch moves the
memory handling part of the edge information required for
aq-mode 4 to the Frame class-level in that way it can be
reused by the threads.
* Fix: Performance drop in aq-mode 4 This patch moves the
memory handling part of the edge information required for
aq-mode 4 to the Frame class-level in that way it can be
reused by the threads.
* Fix output change observed during analysis-load for
inter-refine levels 2 and 3.
* Adaptive Frame duplication This patch does the following. 1.
Replaces 2-3 near-identical frames with one frame and sets
pic_struct based on frame doubling / tripling. 2. Add option
"--frame-dup" and "--dup-threshold' to enable frame
duplication and to set threshold for frame similarity
(optional).
* Fix: AQ mode 4 commit (21db162) introduces slowdown even is
not used AQ mode 4.
x265.changes
Changed
x
1
2
-------------------------------------------------------------------3
+Sun Dec 1 10:44:50 UTC 2019 - Luigi Baldoni <aloisio@gmx.com>4
+5
+- Update to version 3.2.16
+ * Fix output change in analysis load7
+ * Fix encoder crash with zones and add test for zones8
+ * Fix: Performance drop in aq-mode 4 This patch moves the9
+ memory handling part of the edge information required for10
+ aq-mode 4 to the Frame class-level in that way it can be11
+ reused by the threads.12
+ * Fix: Performance drop in aq-mode 4 This patch moves the13
+ memory handling part of the edge information required for14
+ aq-mode 4 to the Frame class-level in that way it can be15
+ reused by the threads.16
+ * Fix output change observed during analysis-load for17
+ inter-refine levels 2 and 3.18
+ * Adaptive Frame duplication This patch does the following. 1.19
+ Replaces 2-3 near-identical frames with one frame and sets20
+ pic_struct based on frame doubling / tripling. 2. Add option21
+ "--frame-dup" and "--dup-threshold' to enable frame22
+ duplication and to set threshold for frame similarity23
+ (optional).24
+ * Fix: AQ mode 4 commit (21db162) introduces slowdown even is25
+ not used AQ mode 4.26
+27
+-------------------------------------------------------------------28
Tue Oct 01 12:21:19 UTC 2019 - enzokiel@kabelmail.de29
30
- Update to version 3.231
x265.spec
Changed
10
1
2
%define libname lib%{name}3
%define libsoname %{libname}-%{soname}4
Name: x2655
-Version: 3.26
+Version: 3.2.17
Release: 08
Summary: A free h265/HEVC encoder - encoder binary9
License: GPL-2.0-or-later10
x265_3.2.tar.gz/.hg_archival.txt -> x265_3.2.1.tar.gz/.hg_archival.txt
Changed
9
1
2
repo: 09fe40627f03a0f9c3e6ac78b22ac93da23f9fdf3
-node: 353572437201d551381002aebf20d244bd49ef174
+node: b5c86a64bbbede216b25092def72272ecde5523a5
branch: Release_3.26
-tag: 3.27
+latesttag: 3.2.18
+latesttagdistance: 19
x265_3.2.tar.gz/.hgtags -> x265_3.2.1.tar.gz/.hgtags
Changed
7
1
2
113518629fa54ffb491dd479e15c1f00dd39d376 3.1_RC13
b4e38ce16d7c4b37a6482dc7ae61fd31071b6ff1 3.1_RC24
20c9994e8bfbeb9443851b2b3a050cd98c8b147b 3.2_RC15
+353572437201d551381002aebf20d244bd49ef17 3.26
+7fa570ead8d361bf6055cd2a881a8e15f12110ae 3.2.17
x265_3.2.tar.gz/source/common/frame.cpp -> x265_3.2.1.tar.gz/source/common/frame.cpp
Changed
44
1
2
m_addOnPrevChange = NULL;3
m_classifyFrame = false;4
m_fieldNum = 0;5
+ m_edgePic = NULL;6
+ m_gaussianPic = NULL;7
+ m_thetaPic = NULL;8
}9
10
bool Frame::create(x265_param *param, float* quantOffsets)11
12
CHECKED_MALLOC_ZERO(m_classifyCount, uint32_t, size);13
}14
15
+ if (param->rc.aqMode == X265_AQ_EDGE || (param->rc.zonefileCount && param->rc.aqMode != 0))16
+ {17
+ uint32_t numCuInWidth = (param->sourceWidth + param->maxCUSize - 1) / param->maxCUSize;18
+ uint32_t numCuInHeight = (param->sourceHeight + param->maxCUSize - 1) / param->maxCUSize;19
+ uint32_t m_lumaMarginX = param->maxCUSize + 32; // search margin and 8-tap filter half-length, padded for 32-byte alignment20
+ uint32_t m_lumaMarginY = param->maxCUSize + 16; // margin for 8-tap filter and infinite padding21
+ intptr_t m_stride = (numCuInWidth * param->maxCUSize) + (m_lumaMarginX << 1);22
+ int maxHeight = numCuInHeight * param->maxCUSize;23
+24
+ m_edgePic = X265_MALLOC(pixel, m_stride * (maxHeight + (m_lumaMarginY * 2)));25
+ m_gaussianPic = X265_MALLOC(pixel, m_stride * (maxHeight + (m_lumaMarginY * 2)));26
+ m_thetaPic = X265_MALLOC(pixel, m_stride * (maxHeight + (m_lumaMarginY * 2)));27
+ }28
+29
if (m_fencPic->create(param, !!m_param->bCopyPicToFrame) && m_lowres.create(param, m_fencPic, param->rc.qgSize))30
{31
X265_CHECK((m_reconColCount == NULL), "m_reconColCount was initialized");32
33
X265_FREE_ZERO(m_classifyVariance);34
X265_FREE_ZERO(m_classifyCount);35
}36
+37
+ if (m_param->rc.aqMode == X265_AQ_EDGE || (m_param->rc.zonefileCount && m_param->rc.aqMode != 0))38
+ {39
+ X265_FREE(m_edgePic);40
+ X265_FREE(m_gaussianPic);41
+ X265_FREE(m_thetaPic);42
+ }43
}44
x265_3.2.tar.gz/source/common/frame.h -> x265_3.2.1.tar.gz/source/common/frame.h
Changed
13
1
2
bool m_classifyFrame;3
int m_fieldNum;4
5
+ /* aq-mode 4 : Gaussian, edge and theta frames for edge information */6
+ pixel* m_edgePic;7
+ pixel* m_gaussianPic;8
+ pixel* m_thetaPic;9
+10
Frame();11
12
bool create(x265_param *param, float* quantOffsets);13
x265_3.2.tar.gz/source/encoder/analysis.cpp -> x265_3.2.1.tar.gz/source/encoder/analysis.cpp
Changed
19
1
2
}3
if (!mode.cu.m_mergeFlag[pu.puAbsPartIdx])4
{5
- if (m_param->mvRefine || m_param->interRefine == 1)6
+ if (m_param->interRefine == 1)7
m_me.setSourcePU(*mode.fencYuv, pu.ctuAddr, pu.cuAbsPartIdx, pu.puAbsPartIdx, pu.width, pu.height, m_param->searchMethod, m_param->subpelRefine, false);8
//AMVP9
MV mvc[(MD_ABOVE_LEFT + 1) * 2 + 2];10
11
12
int numMvc = mode.cu.getPMV(mode.interNeighbours, list, ref, mode.amvpCand[list][ref], mvc);13
mvp = mode.amvpCand[list][ref][mode.cu.m_mvpIdx[list][pu.puAbsPartIdx]];14
- if (m_param->mvRefine || m_param->interRefine == 1)15
+ if (m_param->interRefine == 1)16
{17
MV outmv, mvpSelect[3];18
mvpSelect[0] = interDataCTU->mv[list][cuIdx + part].word;19
x265_3.2.tar.gz/source/encoder/api.cpp -> x265_3.2.1.tar.gz/source/encoder/api.cpp
Changed
9
1
2
int zoneCount = p->rc.zonefileCount ? p->rc.zonefileCount : p->rc.zoneCount;3
param->rc.zones = x265_zone_alloc(zoneCount, !!p->rc.zonefileCount);4
latestParam->rc.zones = x265_zone_alloc(zoneCount, !!p->rc.zonefileCount);5
+ zoneParam->rc.zones = x265_zone_alloc(zoneCount, !!p->rc.zonefileCount);6
}7
8
x265_copy_params(param, p);9
x265_3.2.tar.gz/source/encoder/search.cpp -> x265_3.2.1.tar.gz/source/encoder/search.cpp
Changed
94
1
2
{3
CUData& cu = interMode.cu;4
MV mv, mvmin, mvmax;5
- cu.clipMv(mv);6
int cand = 0, bestcost = INT_MAX;7
- do8
+ while (cand < m_param->mvRefine)9
{10
- if (cand && (mvp[cand] == mvp[cand - 1] || (cand == 2 && mvp[cand] == mvp[cand - 2])))11
+ if ((cand && mvp[cand] == mvp[cand - 1]) || (cand == 2 && (mvp[cand] == mvp[cand - 2] || mvp[cand] == mvp[cand - 1])))12
+ {13
+ cand++;14
continue;15
+ }16
MV bestMV;17
- mv = mvp[cand];18
+ mv = mvp[cand++];19
+ cu.clipMv(mv);20
setSearchRange(cu, mv, m_param->searchRange, mvmin, mvmax);21
int cost = m_me.motionEstimate(&m_slice->m_mref[list][ref], mvmin, mvmax, mv, numMvc, mvc, m_param->searchRange, bestMV, m_param->maxSlices,22
m_param->bSourceReferenceEstimation ? m_slice->m_refFrameList[list][ref]->m_fencPic->getLumaAddr(0) : 0);23
24
bestcost = cost;25
outmv = bestMV;26
}27
- }while (++cand < m_param->mvRefine);28
+ }29
}30
/* find the best inter prediction for each PU of specified mode */31
void Search::predInterSearch(Mode& interMode, const CUGeom& cuGeom, bool bChromaMC, uint32_t refMasks[2])32
33
const MV* amvp = interMode.amvpCand[list][ref];34
int mvpIdx = selectMVP(cu, pu, amvp, list, ref);35
MV mvmin, mvmax, outmv, mvp;36
- mvp = amvp[mvpIdx];37
+ if (useAsMVP)38
+ {39
+ mvp = interDataCTU->mv[list][cuIdx + puIdx].word;40
+ mvpIdx = interDataCTU->mvpIdx[list][cuIdx + puIdx];41
+ }42
+ else43
+ mvp = amvp[mvpIdx];44
if (m_param->searchMethod == X265_SEA)45
{46
int puX = puIdx & 1;47
48
int satdCost;49
if (m_param->analysisMultiPassRefine && m_param->rc.bStatRead && mvpIdx == bestME[list].mvpIdx)50
mvpIn = bestME[list].mv;51
- if (useAsMVP)52
+ if (useAsMVP && m_param->mvRefine > 1)53
{54
MV bestmv, mvpSel[3];55
int mvpIdxSel[3];56
satdCost = m_me.COST_MAX;57
- mvpSel[0] = interDataCTU->mv[list][cuIdx + puIdx].word;58
- mvpIdxSel[0] = interDataCTU->mvpIdx[list][cuIdx + puIdx];59
- if (m_param->mvRefine > 1)60
+ mvpSel[0] = mvp;61
+ mvpIdxSel[0] = mvpIdx;62
+ mvpIdx = selectMVP(cu, pu, amvp, list, ref);63
+ mvpSel[1] = interMode.amvpCand[list][ref][mvpIdx];64
+ mvpIdxSel[1] = mvpIdx;65
+ if (m_param->mvRefine > 2)66
{67
- mvpSel[1] = interMode.amvpCand[list][ref][mvpIdx];68
- mvpIdxSel[1] = mvpIdx;69
- if (m_param->mvRefine > 2)70
- {71
- mvpSel[2] = interMode.amvpCand[list][ref][!mvpIdx];72
- mvpIdxSel[2] = !mvpIdx;73
- }74
+ mvpSel[2] = interMode.amvpCand[list][ref][!mvpIdx];75
+ mvpIdxSel[2] = !mvpIdx;76
}77
for (int cand = 0; cand < m_param->mvRefine; cand++)78
{79
if (cand && (mvpSel[cand] == mvpSel[cand - 1] || (cand == 2 && mvpSel[cand] == mvpSel[cand - 2])))80
continue;81
- setSearchRange(cu, mvp, m_param->searchRange, mvmin, mvmax);82
+ setSearchRange(cu, mvpSel[cand], m_param->searchRange, mvmin, mvmax);83
int bcost = m_me.motionEstimate(&m_slice->m_mref[list][ref], mvmin, mvmax, mvpSel[cand], numMvc, mvc, m_param->searchRange, bestmv, m_param->maxSlices,84
m_param->bSourceReferenceEstimation ? m_slice->m_refFrameList[list][ref]->m_fencPic->getLumaAddr(0) : 0);85
if (satdCost > bcost)86
87
mvpIdx = mvpIdxSel[cand];88
}89
}90
+ mvpIn = mvp;91
}92
else93
{94
x265_3.2.tar.gz/source/encoder/slicetype.cpp -> x265_3.2.1.tar.gz/source/encoder/slicetype.cpp
Changed
108
1
2
3
} // end anonymous namespace4
5
-void edgeFilter(Frame *curFrame, pixel *pic1, pixel *pic2, pixel *pic3, intptr_t stride, int height, int width)6
+void edgeFilter(Frame *curFrame, x265_param* param)7
{8
+ int height = curFrame->m_fencPic->m_picHeight;9
+ int width = curFrame->m_fencPic->m_picWidth;10
+ intptr_t stride = curFrame->m_fencPic->m_stride;11
+ uint32_t numCuInHeight = (height + param->maxCUSize - 1) / param->maxCUSize;12
+ int maxHeight = numCuInHeight * param->maxCUSize;13
+14
+ memset(curFrame->m_edgePic, 0, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));15
+ memset(curFrame->m_gaussianPic, 0, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));16
+ memset(curFrame->m_thetaPic, 0, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));17
+18
pixel *src = (pixel*)curFrame->m_fencPic->m_picOrg[0];19
- pixel *edgePic = pic1 + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;20
- pixel *refPic = pic2 + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;21
- pixel *edgeTheta = pic3 + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;22
+ pixel *edgePic = curFrame->m_edgePic + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;23
+ pixel *refPic = curFrame->m_gaussianPic + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;24
+ pixel *edgeTheta = curFrame->m_thetaPic + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;25
26
for (int i = 0; i < height; i++)27
{28
29
30
//Applying Gaussian filter on the picture31
src = (pixel*)curFrame->m_fencPic->m_picOrg[0];32
- refPic = pic2 + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;33
+ refPic = curFrame->m_gaussianPic + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;34
pixel pixelValue = 0;35
36
for (int rowNum = 0; rowNum < height; rowNum++)37
38
float gradientH = 0, gradientV = 0, radians = 0, theta = 0;39
float gradientMagnitude = 0;40
pixel blackPixel = 0;41
- edgePic = pic1 + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;42
+ edgePic = curFrame->m_edgePic + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;43
//Applying Sobel filter on the gaussian filtered picture44
for (int rowNum = 0; rowNum < height; rowNum++)45
{46
47
angle = sum / (size*size);48
}49
50
-uint32_t LookaheadTLD::edgeDensityCu(Frame* curFrame,pixel *edgeImage, pixel *edgeTheta, uint32_t &avgAngle, uint32_t blockX, uint32_t blockY, uint32_t qgSize)51
+uint32_t LookaheadTLD::edgeDensityCu(Frame* curFrame, uint32_t &avgAngle, uint32_t blockX, uint32_t blockY, uint32_t qgSize)52
{53
+ pixel *edgeImage = curFrame->m_edgePic + curFrame->m_fencPic->m_lumaMarginY * curFrame->m_fencPic->m_stride + curFrame->m_fencPic->m_lumaMarginX;54
+ pixel *edgeTheta = curFrame->m_thetaPic + curFrame->m_fencPic->m_lumaMarginY * curFrame->m_fencPic->m_stride + curFrame->m_fencPic->m_lumaMarginX;55
intptr_t srcStride = curFrame->m_fencPic->m_stride;56
intptr_t blockOffsetLuma = blockX + (blockY * srcStride);57
int plane = 0; // Sobel filter is applied only on Y component58
59
}60
else61
{62
-#define AQ_EDGE_BIAS 0.563
-#define EDGE_INCLINATION 4564
- uint32_t numCuInHeight = (maxRow + param->maxCUSize - 1) / param->maxCUSize;65
- int maxHeight = numCuInHeight * param->maxCUSize;66
- intptr_t stride = curFrame->m_fencPic->m_stride;67
- pixel *edgePic = X265_MALLOC(pixel, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)));68
- pixel *gaussianPic = X265_MALLOC(pixel, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)));69
- pixel *thetaPic = X265_MALLOC(pixel, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)));70
- memset(edgePic, 0, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));71
- memset(gaussianPic, 0, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));72
- memset(thetaPic, 0, stride * (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));73
- if (param->rc.aqMode == X265_AQ_EDGE)74
- edgeFilter(curFrame, edgePic, gaussianPic, thetaPic, stride, maxRow, maxCol);75
-76
int blockXY = 0, inclinedEdge = 0;77
double avg_adj_pow2 = 0, avg_adj = 0, qp_adj = 0;78
double bias_strength = 0.f;79
double strength = 0.f;80
+81
+ if (param->rc.aqMode == X265_AQ_EDGE)82
+ edgeFilter(curFrame, param);83
+84
if (param->rc.aqMode == X265_AQ_AUTO_VARIANCE || param->rc.aqMode == X265_AQ_AUTO_VARIANCE_BIASED || param->rc.aqMode == X265_AQ_EDGE)85
{86
double bit_depth_correction = 1.f / (1 << (2 * (X265_DEPTH - 8)));87
88
energy = acEnergyCu(curFrame, blockX, blockY, param->internalCsp, param->rc.qgSize);89
if (param->rc.aqMode == X265_AQ_EDGE)90
{91
- pixel *edgeImage = edgePic + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;92
- pixel *edgeTheta = thetaPic + curFrame->m_fencPic->m_lumaMarginY * stride + curFrame->m_fencPic->m_lumaMarginX;93
- edgeDensity = edgeDensityCu(curFrame, edgeImage, edgeTheta, avgAngle, blockX, blockY, param->rc.qgSize);94
+ edgeDensity = edgeDensityCu(curFrame, avgAngle, blockX, blockY, param->rc.qgSize);95
if (edgeDensity)96
{97
qp_adj = pow(edgeDensity * bit_depth_correction + 1, 0.1);98
99
else100
strength = param->rc.aqStrength * 1.0397f;101
102
- X265_FREE(edgePic);103
- X265_FREE(gaussianPic);104
- X265_FREE(thetaPic);105
blockXY = 0;106
for (int blockY = 0; blockY < maxRow; blockY += loopIncr)107
{108
x265_3.2.tar.gz/source/encoder/slicetype.h -> x265_3.2.1.tar.gz/source/encoder/slicetype.h
Changed
19
1
2
3
#define LOWRES_COST_MASK ((1 << 14) - 1)4
#define LOWRES_COST_SHIFT 145
+#define AQ_EDGE_BIAS 0.56
+#define EDGE_INCLINATION 457
8
/* Thread local data for lookahead tasks */9
struct LookaheadTLD10
11
protected:12
13
uint32_t acEnergyCu(Frame* curFrame, uint32_t blockX, uint32_t blockY, int csp, uint32_t qgSize);14
- uint32_t edgeDensityCu(Frame*curFrame, pixel *edgeImage, pixel *edgeTheta, uint32_t &avgAngle, uint32_t blockX, uint32_t blockY, uint32_t qgSize);15
+ uint32_t edgeDensityCu(Frame*curFrame, uint32_t &avgAngle, uint32_t blockX, uint32_t blockY, uint32_t qgSize);16
uint32_t lumaSumCu(Frame* curFrame, uint32_t blockX, uint32_t blockY, uint32_t qgSize);17
uint32_t weightCostLuma(Lowres& fenc, Lowres& ref, WeightParam& wp);18
bool allocWeightedRef(Lowres& fenc);19
x265_3.2.tar.gz/source/test/regression-tests.txt -> x265_3.2.1.tar.gz/source/test/regression-tests.txt
Changed
12
1
2
BasketballDrive_1920x1080_50.y4m, --preset medium --no-open-gop --keyint 50 --min-keyint 50 --radl 2 --vbv-maxrate 5000 --vbv-bufsize 50003
big_buck_bunny_360p24.y4m, --bitrate 500 --fades4
720p50_parkrun_ter.y4m,--preset medium --bitrate 400 --hme5
-ducks_take_off_420_1_720p50.y4m,--preset medium --aq-mode 4 --crf 22 --no-cutree6
+ducks_take_off_420_720p50.y4m,--preset medium --aq-mode 4 --crf 22 --no-cutree7
ducks_take_off_420_1_720p50.y4m,--preset medium --selective-sao 4 --sao --crf 208
+Kimono1_1920x1080_24_400.yuv,--preset superfast --qp 28 --zones 0,139,q=329
10
# Main12 intraCost overflow bug test11
720p50_parkrun_ter.y4m,--preset medium12
Refresh
x265
x86_64
aarch64
x86_64
aarch64
armv7l
Refresh
Login required, please
login
or
signup
in order to comment
Request History
olh created request over 5 years ago
- Update to version 3.2.1
* Fix output change in analysis load
* Fix encoder crash with zones and add test for zones
* Fix: Performance drop in aq-mode 4 This patch moves the
memory handling part of the edge information required for
aq-mode 4 to the Frame class-level in that way it can be
reused by the threads.
* Fix: Performance drop in aq-mode 4 This patch moves the
memory handling part of the edge information required for
aq-mode 4 to the Frame class-level in that way it can be
reused by the threads.
* Fix output change observed during analysis-load for
inter-refine levels 2 and 3.
* Adaptive Frame duplication This patch does the following. 1.
Replaces 2-3 near-identical frames with one frame and sets
pic_struct based on frame doubling / tripling. 2. Add option
"--frame-dup" and "--dup-threshold' to enable frame
duplication and to set threshold for frame similarity
(optional).
* Fix: AQ mode 4 commit (21db162) introduces slowdown even is
not used AQ mode 4.
olh accepted request over 5 years ago