[x265] [PATCH 2 of 4] limitTU : use spatial and temporal CUs' TU depth to limit recursion
bhavna at multicorewareinc.com
bhavna at multicorewareinc.com
Fri Nov 18 13:06:27 CET 2016
# HG changeset patch
# User Bhavna Hariharan <bhavna at multicorewareinc.com>
# Date 1479450968 -19800
# Fri Nov 18 12:06:08 2016 +0530
# Node ID 40a0a322b26fc0516a72d4de9a941e18b5bb97b9
# Parent c5295126f248411481a8361acfd2bc8b0636cedc
limitTU : use spatial and temporal CUs' TU depth to limit recursion
diff -r c5295126f248 -r 40a0a322b26f doc/reST/cli.rst
--- a/doc/reST/cli.rst Fri Nov 18 11:49:05 2016 +0530
+++ b/doc/reST/cli.rst Fri Nov 18 12:06:08 2016 +0530
@@ -869,13 +869,15 @@
partitions, in which case a TU split is implied and thus the
residual quad-tree begins one layer below the CU quad-tree.
-.. option:: --limit-tu <0|1|2>
+.. option:: --limit-tu <0..3>
Enables early exit from TU depth recursion, for inter coded blocks.
Level 1 - decides to recurse to next higher depth based on cost
comparison of full size TU and split TU.
Level 2 - based on first split subTU's depth, limits recursion of
other split subTUs.
+ Level 3 - based on the average depth of the co-located and the neighbor
+ CUs' TU depth, limits recursion of the current CU.
Default: 0
diff -r c5295126f248 -r 40a0a322b26f source/common/cudata.cpp
--- a/source/common/cudata.cpp Fri Nov 18 11:49:05 2016 +0530
+++ b/source/common/cudata.cpp Fri Nov 18 12:06:08 2016 +0530
@@ -296,6 +296,9 @@
/* initialize the remaining CU data in one memset */
memset(m_cuDepth, 0, (frame.m_param->internalCsp == X265_CSP_I400 ? BytesPerPartition - 11 : BytesPerPartition - 7) * m_numPartitions);
+ for (int8_t i = 0; i < NUM_TU_DEPTH; i++)
+ m_refTuDepth[i] = -1;
+
uint32_t widthInCU = m_slice->m_sps->numCuInWidth;
m_cuLeft = (m_cuAddr % widthInCU) ? m_encData->getPicCTU(m_cuAddr - 1) : NULL;
m_cuAbove = (m_cuAddr >= widthInCU) && !m_bFirstRowInSlice ? m_encData->getPicCTU(m_cuAddr - widthInCU) : NULL;
diff -r c5295126f248 -r 40a0a322b26f source/common/cudata.h
--- a/source/common/cudata.h Fri Nov 18 11:49:05 2016 +0530
+++ b/source/common/cudata.h Fri Nov 18 12:06:08 2016 +0530
@@ -28,6 +28,8 @@
#include "slice.h"
#include "mv.h"
+#define NUM_TU_DEPTH 21
+
namespace X265_NS {
// private namespace
@@ -204,6 +206,7 @@
enum { BytesPerPartition = 21 }; // combined sizeof() of all per-part data
coeff_t* m_trCoeff[3]; // transformed coefficient buffer per plane
+ int8_t m_refTuDepth[NUM_TU_DEPTH]; // TU depth of CU at depths 0, 1 and 2
MV* m_mv[2]; // array of motion vectors per list
MV* m_mvd[2]; // array of coded motion vector deltas per list
diff -r c5295126f248 -r 40a0a322b26f source/common/param.cpp
--- a/source/common/param.cpp Fri Nov 18 11:49:05 2016 +0530
+++ b/source/common/param.cpp Fri Nov 18 12:06:08 2016 +0530
@@ -1126,7 +1126,7 @@
"QuadtreeTUMaxDepthInter must be less than or equal to the difference between log2(maxCUSize) and QuadtreeTULog2MinSize plus 1");
CHECK((param->maxTUSize != 32 && param->maxTUSize != 16 && param->maxTUSize != 8 && param->maxTUSize != 4),
"max TU size must be 4, 8, 16, or 32");
- CHECK(param->limitTU > 2, "Invalid limit-tu option, limit-TU must be 0, 1 or 2");
+ CHECK(param->limitTU > 3, "Invalid limit-tu option, limit-TU must be between 0 and 3");
CHECK(param->maxNumMergeCand < 1, "MaxNumMergeCand must be 1 or greater.");
CHECK(param->maxNumMergeCand > 5, "MaxNumMergeCand must be 5 or smaller.");
diff -r c5295126f248 -r 40a0a322b26f source/encoder/analysis.cpp
--- a/source/encoder/analysis.cpp Fri Nov 18 11:49:05 2016 +0530
+++ b/source/encoder/analysis.cpp Fri Nov 18 12:06:08 2016 +0530
@@ -203,6 +203,57 @@
return *m_modeDepth[0].bestMode;
}
+int32_t Analysis::loadTUDepth(CUGeom cuGeom, CUData parentCTU)
+{
+ float predDepth = 0;
+ CUData* neighbourCU;
+ uint8_t count = 0;
+ int32_t maxTUDepth = -1;
+ neighbourCU = m_slice->m_refFrameList[0][0]->m_encData->m_picCTU;
+ predDepth += neighbourCU->m_refTuDepth[cuGeom.geomRecurId];
+ count++;
+ if (m_slice->isInterB())
+ {
+ neighbourCU = m_slice->m_refFrameList[1][0]->m_encData->m_picCTU;
+ predDepth += neighbourCU->m_refTuDepth[cuGeom.geomRecurId];
+ count++;
+ }
+ if (parentCTU.m_cuAbove)
+ {
+ predDepth += parentCTU.m_cuAbove->m_refTuDepth[cuGeom.geomRecurId];
+ count++;
+ if (parentCTU.m_cuAboveLeft)
+ {
+ predDepth += parentCTU.m_cuAboveLeft->m_refTuDepth[cuGeom.geomRecurId];
+ count++;
+ }
+ if (parentCTU.m_cuAboveRight)
+ {
+ predDepth += parentCTU.m_cuAboveRight->m_refTuDepth[cuGeom.geomRecurId];
+ count++;
+ }
+ }
+ if (parentCTU.m_cuLeft)
+ {
+ predDepth += parentCTU.m_cuLeft->m_refTuDepth[cuGeom.geomRecurId];
+ count++;
+ }
+ predDepth /= count;
+
+ if (predDepth == 0)
+ maxTUDepth = 0;
+ else if (predDepth < 1)
+ maxTUDepth = 1;
+ else if (predDepth >= 1 && predDepth <= 1.5)
+ maxTUDepth = 2;
+ else if (predDepth > 1.5 && predDepth <= 2.5)
+ maxTUDepth = 3;
+ else
+ maxTUDepth = -1;
+
+ return maxTUDepth;
+}
+
void Analysis::tryLossless(const CUGeom& cuGeom)
{
ModeDepth& md = m_modeDepth[cuGeom.depth];
@@ -326,6 +377,15 @@
checkBestMode(md.pred[PRED_INTRA_NxN], depth);
}
+ if ((m_limitTU & X265_TU_LIMIT_NEIGH) && cuGeom.log2CUSize >= 4)
+ {
+ CUData* ctu = md.bestMode->cu.m_encData->getPicCTU(parentCTU.m_cuAddr);
+ int8_t maxTUDepth = -1;
+ for (uint32_t i = 0; i < cuGeom.numPartitions; i++)
+ maxTUDepth = X265_MAX(maxTUDepth, md.bestMode->cu.m_tuDepth[i]);
+ ctu->m_refTuDepth[cuGeom.geomRecurId] = maxTUDepth;
+ }
+
if (m_bTryLossless)
tryLossless(cuGeom);
@@ -894,6 +954,9 @@
bool skipRectAmp = false;
bool chooseMerge = false;
+ if ((m_limitTU & X265_TU_LIMIT_NEIGH) && cuGeom.log2CUSize >= 4)
+ m_maxTUDepth = loadTUDepth(cuGeom, parentCTU);
+
SplitData splitData[4];
splitData[0].initSplitCUData();
splitData[1].initSplitCUData();
@@ -1400,6 +1463,18 @@
if (m_param->rdLevel)
md.bestMode->reconYuv.copyToPicYuv(reconPic, cuAddr, cuGeom.absPartIdx);
+ if ((m_limitTU & X265_TU_LIMIT_NEIGH) && cuGeom.log2CUSize >= 4)
+ {
+ if (mightNotSplit)
+ {
+ CUData* ctu = md.bestMode->cu.m_encData->getPicCTU(parentCTU.m_cuAddr);
+ int8_t maxTUDepth = -1;
+ for (uint32_t i = 0; i < cuGeom.numPartitions; i++)
+ maxTUDepth = X265_MAX(maxTUDepth, md.bestMode->cu.m_tuDepth[i]);
+ ctu->m_refTuDepth[cuGeom.geomRecurId] = maxTUDepth;
+ }
+ }
+
return splitCUData;
}
@@ -1424,6 +1499,9 @@
md.pred[PRED_2Nx2N].rdCost = 0;
}
+ if ((m_limitTU & X265_TU_LIMIT_NEIGH) && cuGeom.log2CUSize >= 4)
+ m_maxTUDepth = loadTUDepth(cuGeom, parentCTU);
+
SplitData splitData[4];
splitData[0].initSplitCUData();
splitData[1].initSplitCUData();
@@ -1751,6 +1829,18 @@
addSplitFlagCost(*md.bestMode, cuGeom.depth);
}
+ if ((m_limitTU & X265_TU_LIMIT_NEIGH) && cuGeom.log2CUSize >= 4)
+ {
+ if (mightNotSplit)
+ {
+ CUData* ctu = md.bestMode->cu.m_encData->getPicCTU(parentCTU.m_cuAddr);
+ int8_t maxTUDepth = -1;
+ for (uint32_t i = 0; i < cuGeom.numPartitions; i++)
+ maxTUDepth = X265_MAX(maxTUDepth, md.bestMode->cu.m_tuDepth[i]);
+ ctu->m_refTuDepth[cuGeom.geomRecurId] = maxTUDepth;
+ }
+ }
+
/* compare split RD cost against best cost */
if (mightSplit && !skipRecursion)
checkBestMode(md.pred[PRED_SPLIT], depth);
diff -r c5295126f248 -r 40a0a322b26f source/encoder/analysis.h
--- a/source/encoder/analysis.h Fri Nov 18 11:49:05 2016 +0530
+++ b/source/encoder/analysis.h Fri Nov 18 12:06:08 2016 +0530
@@ -116,6 +116,7 @@
void destroy();
Mode& compressCTU(CUData& ctu, Frame& frame, const CUGeom& cuGeom, const Entropy& initialContext);
+ int32_t loadTUDepth(CUGeom cuGeom, CUData parentCTU);
protected:
/* Analysis data for save/load mode, writes/reads data based on absPartIdx */
diff -r c5295126f248 -r 40a0a322b26f source/encoder/search.cpp
--- a/source/encoder/search.cpp Fri Nov 18 11:49:05 2016 +0530
+++ b/source/encoder/search.cpp Fri Nov 18 12:06:08 2016 +0530
@@ -67,7 +67,7 @@
m_param = NULL;
m_slice = NULL;
m_frame = NULL;
- m_maxTUDepth = 0;
+ m_maxTUDepth = -1;
}
bool Search::initSearch(const x265_param& param, ScalingList& scalingList)
@@ -97,10 +97,12 @@
m_limitTU = 0;
if (m_param->limitTU)
{
- if (m_param->limitTU == 1)
- m_limitTU = X265_TU_LIMIT_BFS;
- else if (m_param->limitTU == 2)
+ if (m_param->limitTU == 1)
+ m_limitTU = X265_TU_LIMIT_BFS;
+ else if (m_param->limitTU == 2)
m_limitTU = X265_TU_LIMIT_DFS;
+ else if (m_param->limitTU == 3)
+ m_limitTU = X265_TU_LIMIT_NEIGH;
}
/* these are indexed by qtLayer (log2size - 2) so nominally 0=4x4, 1=8x8, 2=16x16, 3=32x32
@@ -2628,11 +2630,17 @@
uint32_t tuDepthRange[2];
cu.getInterTUQtDepthRange(tuDepthRange, 0);
+ if (m_limitTU & X265_TU_LIMIT_NEIGH)
+ {
+ int32_t maxLog2CUSize = g_log2Size[m_param->maxCUSize];
+ m_maxTUDepth = x265_clip3(maxLog2CUSize - (int32_t)tuDepthRange[1], maxLog2CUSize - (int32_t)tuDepthRange[0], m_maxTUDepth);
+ }
+
m_entropyCoder.load(m_rqt[depth].cur);
- if (m_param->limitTU & X265_TU_LIMIT_DFS)
- m_maxTUDepth = 0;
- else if (m_param->limitTU & X265_TU_LIMIT_BFS)
+ if (m_limitTU & X265_TU_LIMIT_DFS)
+ m_maxTUDepth = -1;
+ else if (m_limitTU & X265_TU_LIMIT_BFS)
memset(&m_cacheTU, 0, sizeof(TUInfoCache));
Cost costs;
@@ -2895,7 +2903,7 @@
uint32_t ycbf = 0, ucbf = 0, vcbf = 0;
for (uint32_t qIdx = 0, qPartIdx = absPartIdx; qIdx < 4; ++qIdx, qPartIdx += qNumParts)
{
- if ((m_param->limitTU & X265_TU_LIMIT_DFS) && tuDepth == 0 && qIdx == 1)
+ if ((m_limitTU & X265_TU_LIMIT_DFS) && tuDepth == 0 && qIdx == 1)
{
// Fetch maximum TU depth of first sub partition to limit recursion of others
for (uint32_t i = 0; i < cuGeom.numPartitions / 4; i++)
@@ -2946,12 +2954,7 @@
bool bSaveTUData = false, bLoadTUData = false;
uint32_t idx = 0;
- if ((m_param->limitTU & X265_TU_LIMIT_DFS) && m_maxTUDepth)
- {
- uint32_t log2MaxTrSize = cuGeom.log2CUSize - m_maxTUDepth;
- bCheckSplit = log2TrSize > log2MaxTrSize;
- }
- else if ((m_param->limitTU & X265_TU_LIMIT_BFS) && splitMore >= 0)
+ if ((m_limitTU & X265_TU_LIMIT_BFS) && splitMore >= 0)
{
if (bCheckSplit && bCheckFull && tuDepth)
{
@@ -2970,6 +2973,14 @@
}
}
}
+ else if (m_limitTU & X265_TU_LIMIT_DFS || m_limitTU & X265_TU_LIMIT_NEIGH)
+ {
+ if (bCheckSplit && m_maxTUDepth >= 0)
+ {
+ uint32_t log2MaxTrSize = cuGeom.log2CUSize - m_maxTUDepth;
+ bCheckSplit = log2TrSize > log2MaxTrSize;
+ }
+ }
bool bSplitPresentFlag = bCheckSplit && bCheckFull;
@@ -3497,7 +3508,7 @@
{
if (splitCost.rdcost < fullCost.rdcost)
{
- if (m_param->limitTU & X265_TU_LIMIT_BFS)
+ if (m_limitTU & X265_TU_LIMIT_BFS)
{
uint32_t nextlog2TrSize = cuGeom.log2CUSize - (tuDepth + 1);
bool nextSplit = nextlog2TrSize > depthRange[0];
diff -r c5295126f248 -r 40a0a322b26f source/encoder/search.h
--- a/source/encoder/search.h Fri Nov 18 11:49:05 2016 +0530
+++ b/source/encoder/search.h Fri Nov 18 12:06:08 2016 +0530
@@ -277,7 +277,7 @@
uint32_t m_numLayers;
uint32_t m_refLagPixels;
- uint32_t m_maxTUDepth;
+ int32_t m_maxTUDepth;
uint16_t m_limitTU;
int16_t m_sliceMaxY;
diff -r c5295126f248 -r 40a0a322b26f source/x265.h
--- a/source/x265.h Fri Nov 18 11:49:05 2016 +0530
+++ b/source/x265.h Fri Nov 18 12:06:08 2016 +0530
@@ -357,6 +357,7 @@
#define X265_TU_LIMIT_BFS 1
#define X265_TU_LIMIT_DFS 2
+#define X265_TU_LIMIT_NEIGH 4
#define X265_BFRAME_MAX 16
#define X265_MAX_FRAME_THREADS 16
More information about the x265-devel
mailing list