<div dir="ltr">Thanks, I like initAqQPs since it gets rid of the messy section, but I was also hoping we could avoid passing qp around for the compress* functions. <br><br>Since cuGeom now has cuIdx - the analysis functions can just look up the appropriate QP in m_aqQp? In fact, with some interesting calculations on cuGeom->absPartIdx and depth, we dont even need cuIdx?<br></div><div class="gmail_extra"><br><div class="gmail_quote">On Sun, Apr 26, 2015 at 10:51 PM, Steve Borho <span dir="ltr"><<a href="mailto:steve@borho.org" target="_blank">steve@borho.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"># HG changeset patch<br>
# User Steve Borho <<a href="mailto:steve@borho.org">steve@borho.org</a>><br>
# Date 1429909512 18000<br>
# Fri Apr 24 16:05:12 2015 -0500<br>
# Node ID 5644bbd23e71996651f4ed558e0260201a91f70d<br>
# Parent bfd57a0c0875e219d902ff3af6f4a0ddaa16b125<br>
analysis: keep per-CU AQ QPs in cuGeom index order, simplify arguments<br>
<br>
diff -r bfd57a0c0875 -r 5644bbd23e71 source/common/cudata.cpp<br>
--- a/source/common/cudata.cpp Fri Apr 24 15:14:54 2015 -0500<br>
+++ b/source/common/cudata.cpp Fri Apr 24 16:05:12 2015 -0500<br>
@@ -2027,6 +2027,7 @@<br>
uint32_t blockSize = 1 << log2CUSize;<br>
uint32_t sbWidth = 1 << (g_log2Size[maxCUSize] - log2CUSize);<br>
int32_t lastLevelFlag = log2CUSize == g_log2Size[minCUSize];<br>
+<br>
for (uint32_t sbY = 0; sbY < sbWidth; sbY++)<br>
{<br>
for (uint32_t sbX = 0; sbX < sbWidth; sbX++)<br>
@@ -2049,7 +2050,8 @@<br>
cu->childOffset = childIdx - cuIdx;<br>
cu->absPartIdx = g_depthScanIdx[yOffset][xOffset] * 4;<br>
cu->numPartitions = (NUM_4x4_PARTITIONS >> ((g_maxLog2CUSize - cu->log2CUSize) * 2));<br>
- cu->depth = g_log2Size[maxCUSize] - log2CUSize;<br>
+ cu->depth = (uint16_t)(g_log2Size[maxCUSize] - log2CUSize);<br>
+ cu->index = (uint16_t)cuIdx;<br>
<br>
cu->flags = 0;<br>
CU_SET_FLAG(cu->flags, CUGeom::PRESENT, presentFlag);<br>
diff -r bfd57a0c0875 -r 5644bbd23e71 source/common/cudata.h<br>
--- a/source/common/cudata.h Fri Apr 24 15:14:54 2015 -0500<br>
+++ b/source/common/cudata.h Fri Apr 24 16:05:12 2015 -0500<br>
@@ -85,8 +85,8 @@<br>
uint32_t childOffset; // offset of the first child CU from current CU<br>
uint32_t absPartIdx; // Part index of this CU in terms of 4x4 blocks.<br>
uint32_t numPartitions; // Number of 4x4 blocks in the CU<br>
- uint32_t depth; // depth of this CU relative from CTU<br>
uint32_t flags; // CU flags.<br>
+ uint16_t depth, index; // depth of this CU relative from CTU, absolute index<br>
};<br>
<br>
struct MVField<br>
diff -r bfd57a0c0875 -r 5644bbd23e71 source/encoder/analysis.cpp<br>
--- a/source/encoder/analysis.cpp Fri Apr 24 15:14:54 2015 -0500<br>
+++ b/source/encoder/analysis.cpp Fri Apr 24 16:05:12 2015 -0500<br>
@@ -75,8 +75,6 @@<br>
m_reuseInterDataCTU = NULL;<br>
m_reuseRef = NULL;<br>
m_reuseBestMergeCand = NULL;<br>
- for (int i = 0; i < NUM_CU_DEPTH; i++)<br>
- m_aqQP[i] = NULL;<br>
}<br>
<br>
bool Analysis::create(ThreadLocalData *tld)<br>
@@ -103,12 +101,9 @@<br>
ok &= md.pred[j].reconYuv.create(cuSize, csp);<br>
md.pred[j].fencYuv = &md.fencYuv;<br>
}<br>
- CHECKED_MALLOC(m_aqQP[depth], int, (size_t)1 << (depth << 1));<br>
}<br>
<br>
return ok;<br>
-fail:<br>
- return false;<br>
}<br>
<br>
void Analysis::destroy()<br>
@@ -123,7 +118,17 @@<br>
m_modeDepth[i].pred[j].predYuv.destroy();<br>
m_modeDepth[i].pred[j].reconYuv.destroy();<br>
}<br>
- X265_FREE(m_aqQP[i]);<br>
+ }<br>
+}<br>
+<br>
+void Analysis::initAqQPs(uint32_t depth, const CUData& ctu, const CUGeom* rootGeom)<br>
+{<br>
+ for (int d0 = 0; d0 < 4; d0++)<br>
+ {<br>
+ m_aqQP[rootGeom->index + d0] = calculateQpforCuSize(ctu, rootGeom[d0]);<br>
+<br>
+ if (m_slice->m_pps->maxCuDQPDepth > depth)<br>
+ initAqQPs(depth + 1, ctu, &rootGeom[d0] + rootGeom[d0].childOffset);<br>
}<br>
}<br>
<br>
@@ -141,32 +146,16 @@<br>
<br>
if (m_slice->m_pps->bUseDQP)<br>
{<br>
- /* TODO: In future, we could extend this to 8x8 QGs as well, since that's the minimum size<br>
- * allowed by the HEVC standard. The AQ offset calculation will need to be at 8x8 granularity.<br>
- * And this messy section will need to be reworked */<br>
- m_aqQP[0][0] = calculateQpforCuSize(ctu, cuGeom);<br>
+ m_aqQP[0] = calculateQpforCuSize(ctu, cuGeom);<br>
+ setLambdaFromQP(*m_slice, m_aqQP[0]);<br>
+ m_aqQP[0] = x265_clip3(QP_MIN, QP_MAX_SPEC, m_aqQP[0]);<br>
+ ctu.setQPSubParts((int8_t)m_aqQP[0], 0, 0);<br>
<br>
- const CUGeom* rootGeom = &cuGeom + 1;<br>
- if (m_slice->m_pps->maxCuDQPDepth >= 1)<br>
- {<br>
- for (int d0 = 0; d0 < 4; d0++)<br>
- {<br>
- m_aqQP[1][d0] = calculateQpforCuSize(ctu, rootGeom[d0]);<br>
- if (m_slice->m_pps->maxCuDQPDepth == 2)<br>
- {<br>
- const CUGeom* curGeom = &rootGeom[d0] + rootGeom[d0].childOffset;<br>
- for (int d1 = 0; d1 < 4; d1++)<br>
- m_aqQP[2][d0 * 4 + d1] = calculateQpforCuSize(ctu, curGeom[d1]);<br>
- }<br>
- }<br>
- }<br>
-<br>
- setLambdaFromQP(*m_slice, m_aqQP[0][0]);<br>
- m_aqQP[0][0] = x265_clip3(QP_MIN, QP_MAX_SPEC, m_aqQP[0][0]);<br>
- ctu.setQPSubParts((int8_t)m_aqQP[0][0], 0, 0);<br>
+ if (m_slice->m_pps->maxCuDQPDepth)<br>
+ initAqQPs(1, ctu, &cuGeom + 1);<br>
}<br>
else<br>
- m_aqQP[0][0] = m_slice->m_sliceQp;<br>
+ m_aqQP[0] = m_slice->m_sliceQp;<br>
<br>
m_quant.setQPforQuant(ctu);<br>
m_rqt[0].cur.load(initialContext);<br>
@@ -191,7 +180,7 @@<br>
uint32_t zOrder = 0;<br>
if (m_slice->m_sliceType == I_SLICE)<br>
{<br>
- compressIntraCU(ctu, cuGeom, zOrder, m_aqQP[0][0], 0);<br>
+ compressIntraCU(ctu, cuGeom, zOrder, m_aqQP[0]);<br>
if (m_param->analysisMode == X265_ANALYSIS_SAVE && m_frame->m_analysisData.intraData)<br>
{<br>
CUData* bestCU = &m_modeDepth[0].bestMode->cu;<br>
@@ -209,18 +198,18 @@<br>
* they are available for intra predictions */<br>
m_modeDepth[0].fencYuv.copyToPicYuv(*m_frame->m_reconPic, ctu.m_cuAddr, 0);<br>
<br>
- compressInterCU_rd0_4(ctu, cuGeom, m_aqQP[0][0], 0);<br>
+ compressInterCU_rd0_4(ctu, cuGeom, m_aqQP[0]);<br>
<br>
/* generate residual for entire CTU at once and copy to reconPic */<br>
encodeResidue(ctu, cuGeom);<br>
}<br>
else if (m_param->bDistributeModeAnalysis && m_param->rdLevel >= 2)<br>
- compressInterCU_dist(ctu, cuGeom, m_aqQP[0][0], 0);<br>
+ compressInterCU_dist(ctu, cuGeom, m_aqQP[0]);<br>
else if (m_param->rdLevel <= 4)<br>
- compressInterCU_rd0_4(ctu, cuGeom, m_aqQP[0][0], 0);<br>
+ compressInterCU_rd0_4(ctu, cuGeom, m_aqQP[0]);<br>
else<br>
{<br>
- compressInterCU_rd5_6(ctu, cuGeom, zOrder, m_aqQP[0][0], 0);<br>
+ compressInterCU_rd5_6(ctu, cuGeom, zOrder, m_aqQP[0]);<br>
if (m_param->analysisMode == X265_ANALYSIS_SAVE && m_frame->m_analysisData.interData)<br>
{<br>
CUData* bestCU = &m_modeDepth[0].bestMode->cu;<br>
@@ -259,7 +248,7 @@<br>
}<br>
}<br>
<br>
-void Analysis::compressIntraCU(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t& zOrder, int32_t qp, uint32_t partIdx)<br>
+void Analysis::compressIntraCU(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t& zOrder, int32_t qp)<br>
{<br>
uint32_t depth = cuGeom.depth;<br>
ModeDepth& md = m_modeDepth[depth];<br>
@@ -270,7 +259,6 @@<br>
<br>
if (m_slice->m_pps->bUseDQP && depth && depth <= m_slice->m_pps->maxCuDQPDepth)<br>
{<br>
- qp = m_aqQP[depth][partIdx];<br>
setLambdaFromQP(*m_slice, qp);<br>
qp = x265_clip3(QP_MIN, QP_MAX_SPEC, qp);<br>
}<br>
@@ -342,7 +330,10 @@<br>
{<br>
m_modeDepth[0].fencYuv.copyPartToYuv(nd.fencYuv, childGeom.absPartIdx);<br>
m_rqt[nextDepth].cur.load(*nextContext);<br>
- compressIntraCU(parentCTU, childGeom, zOrder, qp, partIdx * 4 + subPartIdx);<br>
+<br>
+ int32_t nextQP = m_slice->m_pps->bUseDQP && nextDepth <= m_slice->m_pps->maxCuDQPDepth ?<br>
+ m_aqQP[childGeom.index] : qp;<br>
+ compressIntraCU(parentCTU, childGeom, zOrder, nextQP);<br>
<br>
// Save best CU and pred data for this sub CU<br>
splitCU->copyPartFrom(nd.bestMode->cu, childGeom, subPartIdx);<br>
@@ -530,7 +521,7 @@<br>
while (task >= 0);<br>
}<br>
<br>
-void Analysis::compressInterCU_dist(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qp, uint32_t partIdx)<br>
+void Analysis::compressInterCU_dist(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qp)<br>
{<br>
uint32_t depth = cuGeom.depth;<br>
uint32_t cuAddr = parentCTU.m_cuAddr;<br>
@@ -545,7 +536,6 @@<br>
<br>
if (m_slice->m_pps->bUseDQP && depth && depth <= m_slice->m_pps->maxCuDQPDepth)<br>
{<br>
- qp = m_aqQP[depth][partIdx];<br>
setLambdaFromQP(*m_slice, qp);<br>
qp = x265_clip3(QP_MIN, QP_MAX_SPEC, qp);<br>
}<br>
@@ -749,7 +739,9 @@<br>
{<br>
m_modeDepth[0].fencYuv.copyPartToYuv(nd.fencYuv, childGeom.absPartIdx);<br>
m_rqt[nextDepth].cur.load(*nextContext);<br>
- compressInterCU_dist(parentCTU, childGeom, qp, partIdx * 4 + subPartIdx);<br>
+ int32_t nextQP = m_slice->m_pps->bUseDQP && nextDepth <= m_slice->m_pps->maxCuDQPDepth ?<br>
+ m_aqQP[childGeom.index] : qp;<br>
+ compressInterCU_dist(parentCTU, childGeom, nextQP);<br>
<br>
// Save best CU and pred data for this sub CU<br>
splitCU->copyPartFrom(nd.bestMode->cu, childGeom, subPartIdx);<br>
@@ -788,7 +780,7 @@<br>
md.bestMode->reconYuv.copyToPicYuv(*m_frame->m_reconPic, cuAddr, cuGeom.absPartIdx);<br>
}<br>
<br>
-void Analysis::compressInterCU_rd0_4(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qp, uint32_t partIdx)<br>
+void Analysis::compressInterCU_rd0_4(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qp)<br>
{<br>
uint32_t depth = cuGeom.depth;<br>
uint32_t cuAddr = parentCTU.m_cuAddr;<br>
@@ -801,7 +793,6 @@<br>
<br>
if (m_slice->m_pps->bUseDQP && depth && depth <= m_slice->m_pps->maxCuDQPDepth)<br>
{<br>
- qp = m_aqQP[depth][partIdx];<br>
setLambdaFromQP(*m_slice, qp);<br>
qp = x265_clip3(QP_MIN, QP_MAX_SPEC, qp);<br>
}<br>
@@ -1028,7 +1019,9 @@<br>
{<br>
m_modeDepth[0].fencYuv.copyPartToYuv(nd.fencYuv, childGeom.absPartIdx);<br>
m_rqt[nextDepth].cur.load(*nextContext);<br>
- compressInterCU_rd0_4(parentCTU, childGeom, qp, partIdx * 4 + subPartIdx);<br>
+ int32_t nextQP = m_slice->m_pps->bUseDQP && nextDepth <= m_slice->m_pps->maxCuDQPDepth ?<br>
+ m_aqQP[childGeom.index] : qp;<br>
+ compressInterCU_rd0_4(parentCTU, childGeom, nextQP);<br>
<br>
// Save best CU and pred data for this sub CU<br>
splitCU->copyPartFrom(nd.bestMode->cu, childGeom, subPartIdx);<br>
@@ -1079,7 +1072,7 @@<br>
md.bestMode->reconYuv.copyToPicYuv(*m_frame->m_reconPic, cuAddr, cuGeom.absPartIdx);<br>
}<br>
<br>
-void Analysis::compressInterCU_rd5_6(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t &zOrder, int32_t qp, uint32_t partIdx)<br>
+void Analysis::compressInterCU_rd5_6(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t &zOrder, int32_t qp)<br>
{<br>
uint32_t depth = cuGeom.depth;<br>
ModeDepth& md = m_modeDepth[depth];<br>
@@ -1090,7 +1083,6 @@<br>
<br>
if (m_slice->m_pps->bUseDQP && depth && depth <= m_slice->m_pps->maxCuDQPDepth)<br>
{<br>
- qp = m_aqQP[depth][partIdx];<br>
setLambdaFromQP(*m_slice, qp);<br>
qp = x265_clip3(QP_MIN, QP_MAX_SPEC, qp);<br>
}<br>
@@ -1234,7 +1226,9 @@<br>
{<br>
m_modeDepth[0].fencYuv.copyPartToYuv(nd.fencYuv, childGeom.absPartIdx);<br>
m_rqt[nextDepth].cur.load(*nextContext);<br>
- compressInterCU_rd5_6(parentCTU, childGeom, zOrder, qp, partIdx * 4 + subPartIdx);<br>
+ int32_t nextQP = m_slice->m_pps->bUseDQP && nextDepth <= m_slice->m_pps->maxCuDQPDepth ?<br>
+ m_aqQP[childGeom.index] : qp;<br>
+ compressInterCU_rd5_6(parentCTU, childGeom, zOrder, nextQP);<br>
<br>
// Save best CU and pred data for this sub CU<br>
splitCU->copyPartFrom(nd.bestMode->cu, childGeom, subPartIdx);<br>
diff -r bfd57a0c0875 -r 5644bbd23e71 source/encoder/analysis.h<br>
--- a/source/encoder/analysis.h Fri Apr 24 15:14:54 2015 -0500<br>
+++ b/source/encoder/analysis.h Fri Apr 24 16:05:12 2015 -0500<br>
@@ -90,7 +90,7 @@<br>
void processPmode(PMODE& pmode, Analysis& slave);<br>
<br>
ModeDepth m_modeDepth[NUM_CU_DEPTH];<br>
- int* m_aqQP[NUM_CU_DEPTH];<br>
+ int m_aqQP[CUGeom::MAX_GEOMS];<br>
bool m_bTryLossless;<br>
bool m_bChromaSa8d;<br>
<br>
@@ -109,13 +109,15 @@<br>
int32_t* m_reuseRef;<br>
uint32_t* m_reuseBestMergeCand;<br>
<br>
+ void initAqQPs(uint32_t depth, const CUData& ctu, const CUGeom* rootGeom);<br>
+<br>
/* full analysis for an I-slice CU */<br>
- void compressIntraCU(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t &zOrder, int32_t qpDepth, uint32_t partIdx);<br>
+ void compressIntraCU(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t &zOrder, int32_t qp);<br>
<br>
/* full analysis for a P or B slice CU */<br>
- void compressInterCU_dist(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qpDepth, uint32_t partIdx);<br>
- void compressInterCU_rd0_4(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qpDepth, uint32_t partIdx);<br>
- void compressInterCU_rd5_6(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t &zOrder, int32_t qpDepth, uint32_t partIdx);<br>
+ void compressInterCU_dist(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qp);<br>
+ void compressInterCU_rd0_4(const CUData& parentCTU, const CUGeom& cuGeom, int32_t qp);<br>
+ void compressInterCU_rd5_6(const CUData& parentCTU, const CUGeom& cuGeom, uint32_t &zOrder, int32_t qp);<br>
<br>
/* measure merge and skip */<br>
void checkMerge2Nx2N_rd0_4(Mode& skip, Mode& merge, const CUGeom& cuGeom);<br>
_______________________________________________<br>
x265-devel mailing list<br>
<a href="mailto:x265-devel@videolan.org">x265-devel@videolan.org</a><br>
<a href="https://mailman.videolan.org/listinfo/x265-devel" target="_blank">https://mailman.videolan.org/listinfo/x265-devel</a><br>
</blockquote></div><br></div>