[x265] [PATCH] psy-rd: fix bug in chroma psyEnergy for intra 4x4

[deepthi at multicorewareinc.com]

# HG changeset patch
# User Deepthi Nandakumar <deepthi at multicorewareinc.com>
# Date 1411356220 -19800
#      Mon Sep 22 08:53:40 2014 +0530
# Node ID d1c2b82de4db401e6e85316e5db4dd738a5f36a5
# Parent  817abe294c8b3e60f88093841d208fbb96bf8403
psy-rd: fix bug in chroma psyEnergy for intra 4x4

Also add TODO, for all psyCost calculations

diff -r 817abe294c8b -r d1c2b82de4db source/encoder/analysis.cpp
--- a/source/encoder/analysis.cpp	Sun Sep 21 23:18:49 2014 +0530
+++ b/source/encoder/analysis.cpp	Mon Sep 22 08:53:40 2014 +0530
@@ -724,6 +724,7 @@
     outTempCU->m_totalBits = m_entropyCoder->getNumberOfWrittenBits();
     outTempCU->m_coeffBits = outTempCU->m_totalBits - outTempCU->m_mvBits;
 
+    /* TODO: add chroma psyEnergy also to psyCost*/
     if (m_rdCost.m_psyRd)
     {
         int part = outTempCU->getLog2CUSize(0) - 2;
diff -r 817abe294c8b -r d1c2b82de4db source/encoder/search.cpp
--- a/source/encoder/search.cpp	Sun Sep 21 23:18:49 2014 +0530
+++ b/source/encoder/search.cpp	Mon Sep 22 08:53:40 2014 +0530
@@ -841,7 +841,8 @@
 }
 
 /* returns distortion */
-uint32_t Search::xRecurIntraChromaCodingQT(TComDataCU* cu, uint32_t trDepth, uint32_t absPartIdx, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv)
+uint32_t Search::xRecurIntraChromaCodingQT(TComDataCU* cu, uint32_t trDepth, uint32_t absPartIdx, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv,
+                                           uint32_t& psyEnergy)
 {
     uint32_t fullDepth = cu->getDepth(0) + trDepth;
     uint32_t trMode    = cu->getTransformIdx(absPartIdx);
@@ -1016,7 +1017,7 @@
             if (splitIntoSubTUs)
                 offsetSubTUCBFs(cu, (TextType)chromaId, trDepth, absPartIdx);
         }
-        cu->m_psyEnergy = singlePsyEnergy;
+        psyEnergy = singlePsyEnergy;
     }
     else
     {
@@ -1027,13 +1028,14 @@
         uint32_t absPartIdxSub = absPartIdx;
         for (uint32_t part = 0; part < 4; part++, absPartIdxSub += qPartsDiv)
         {
-            outDist += xRecurIntraChromaCodingQT(cu, trDepth + 1, absPartIdxSub, fencYuv, predYuv, resiYuv);
-            splitPsyEnergy += cu->m_psyEnergy;
+            uint32_t psyEnergyTemp = 0;
+            outDist += xRecurIntraChromaCodingQT(cu, trDepth + 1, absPartIdxSub, fencYuv, predYuv, resiYuv, psyEnergyTemp);
+            splitPsyEnergy += psyEnergyTemp;
             splitCbfU |= cu->getCbf(absPartIdxSub, TEXT_CHROMA_U, trDepth + 1);
             splitCbfV |= cu->getCbf(absPartIdxSub, TEXT_CHROMA_V, trDepth + 1);
         }
 
-        cu->m_psyEnergy = splitPsyEnergy;
+        psyEnergy = splitPsyEnergy;
         for (uint32_t offs = 0; offs < 4 * qPartsDiv; offs++)
         {
             cu->getCbf(TEXT_CHROMA_U)[absPartIdx + offs] |= (splitCbfU << trDepth);
@@ -1529,7 +1531,8 @@
             // chroma coding
             cu->setChromIntraDirSubParts(modeList[mode], absPartIdxC, depth + initTrDepth);
 
-            uint32_t dist = xRecurIntraChromaCodingQT(cu, initTrDepth, absPartIdxC, fencYuv, predYuv, resiYuv);
+            uint32_t psyEnergy = 0;
+            uint32_t dist = xRecurIntraChromaCodingQT(cu, initTrDepth, absPartIdxC, fencYuv, predYuv, resiYuv, psyEnergy);
 
             if (cu->m_slice->m_pps->bTransformSkipEnabled)
                 m_entropyCoder->load(m_rdEntropyCoders[depth][CI_CURR_BEST]);
@@ -1537,7 +1540,7 @@
             uint32_t bits = xGetIntraBitsQTChroma(cu, initTrDepth, absPartIdxC, tuIterator.absPartIdxStep);
             uint64_t cost = 0; 
             if (m_rdCost.m_psyRd)
-                cost = m_rdCost.calcPsyRdCost(dist, bits, cu->m_psyEnergy);
+                cost = m_rdCost.calcPsyRdCost(dist, bits, psyEnergy);
             else
                 cost = m_rdCost.calcRdCost(dist, bits);
 
diff -r 817abe294c8b -r d1c2b82de4db source/encoder/search.h
--- a/source/encoder/search.h	Sun Sep 21 23:18:49 2014 +0530
+++ b/source/encoder/search.h	Mon Sep 22 08:53:40 2014 +0530
@@ -118,7 +118,8 @@
     uint32_t xRecurIntraCodingQT(TComDataCU* cu, uint32_t trDepth, uint32_t absPartIdx, TComYuv* fencYuv, TComYuv* predYuv,
                                  ShortYuv* resiYuv, bool bAllowRQTSplit, uint64_t& dRDCost, uint32_t& puBits, uint32_t depthRange[2]);
 
-    uint32_t xRecurIntraChromaCodingQT(TComDataCU* cu, uint32_t trDepth, uint32_t absPartIdx, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv);
+    uint32_t xRecurIntraChromaCodingQT(TComDataCU* cu, uint32_t trDepth, uint32_t absPartIdx, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv,
+                                       uint32_t& psyEnergy);
 
     uint32_t xIntraCodingChromaBlk(TComDataCU* cu, uint32_t absPartIdx, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv,
                                    int16_t* reconQt, uint32_t reconQtStride, coeff_t* coeff, uint32_t& cbf, uint32_t chromaId, uint32_t log2TrSizeC);