IFP and RC harmonization, fixes for ARM, early GCC 11 (#350)

.github/workflows/Build.yml

@@ -73,6 +73,6 @@ jobs:
       run: |
         mkdir build
         cd build
-        cmake ..  -DCMAKE_BUILD_TYPE=Release -A "${{ matrix.config.msvc_arch }}"
+        cmake ..  -DCMAKE_BUILD_TYPE=Release -DVVENC_OVERRIDE_COMPILER_CHECK=ON -A "${{ matrix.config.msvc_arch }}"
         cmake --build . --config Release
       shell: cmd

.gitlab-ci-internal.yml

@@ -214,13 +214,15 @@ test_vc193x_Win32:
   extends: .build_test_msvc_template
   variables:
      MSVC_ARCH: Win32
+     CONFIG_OPTIONS: "-DVVENC_OVERRIDE_COMPILER_CHECK=1"
   tags:
     - vc193x
 
 test_vc193x:
   extends: .build_test_msvc_template
   variables:
      MSVC_ARCH: x64
+     CONFIG_OPTIONS: "-DVVENC_OVERRIDE_COMPILER_CHECK=1"
   tags:
     - vc193x
 

CMakeLists.txt

@@ -31,10 +31,31 @@ if( ( "${CMAKE_SYSTEM_PROCESSOR}" MATCHES "aarch64\|arm"
 endif()
 
 # we enable x86 intrinsics for all target architectures, because they are implemented through simd-everywhere on non-x86
-set( VVENC_ENABLE_X86_SIMD TRUE CACHE BOOL "enable x86 intrinsics" )
+set( VVENC_ENABLE_X86_SIMD TRUE                      CACHE BOOL "enable x86 intrinsics" )
 set( VVENC_ENABLE_ARM_SIMD ${VVENC_ARM_SIMD_DEFAULT} CACHE BOOL "enable ARM intrinsics" )
 
 include( vvencCompilerSupport )
+check_problematic_compiler( VVENC_PROBLEMATIC_COMPILER "MSVC" 19.38 "" )
+
+if( VVENC_PROBLEMATIC_COMPILER )
+  set( VVENC_OVERRIDE_COMPILER_CHECK      OFF        CACHE BOOL "Build with known problematic compiler version" )
+
+  if( VVENC_OVERRIDE_COMPILER_CHECK )
+    set( VVENC_PROBLEMATIC_COMPILER_MSG_TYPE     WARNING )
+    set( VVENC_PROBLEMATIC_COMPILER_MSG_OVERRIDE
+         "The performance will not be optimal due to workarounds." )
+  else()
+    set( VVENC_PROBLEMATIC_COMPILER_MSG_TYPE FATAL_ERROR )
+    set( VVENC_PROBLEMATIC_COMPILER_MSG_OVERRIDE
+         "Set -DVVENC_OVERRIDE_COMPILER_CHECK=ON to build with this compiler anyways, which enables workarounds impacting performance.")
+  endif()
+
+  message( ${VVENC_PROBLEMATIC_COMPILER_MSG_TYPE}
+          "Binaries compiled with ${CMAKE_CXX_COMPILER} version ${CMAKE_CXX_COMPILER_VERSION} are known not to behave as intended. "
+          "The problematic version range is ${VVENC_PROBLEMATIC_COMPILER_VERSION_RANGE}. Please consider using a different compiler.\n"
+          ${VVENC_PROBLEMATIC_COMPILER_MSG_OVERRIDE} )
+
+endif()
 
 # enable sse4.1 build for all source files for gcc and clang
 if( VVENC_ENABLE_X86_SIMD )
@@ -81,14 +102,14 @@ endif()
 # enable install target
 set( VVENC_ENABLE_INSTALL                   ON  CACHE BOOL   "Enable or disable install target" )
 
-# enable postfix                                             
+# enable postfix
 set( VVENC_ENABLE_BUILD_TYPE_POSTFIX        OFF CACHE BOOL   "Enable or disable build type postfix for apps and libs" )
 
 set( VVENC_ENABLE_LINK_TIME_OPT             ON  CACHE BOOL   "Enable link time optimization for release and profile builds" )
 
 set( VVENC_ENABLE_THIRDPARTY_JSON           ON  CACHE BOOL   "Enable use of thirdparty json library" )
 
-set( VVENC_INSTALL_FULLFEATURE_APP		    OFF CACHE BOOL   "Install the full-feature app: vvencFFapp" )
+set( VVENC_INSTALL_FULLFEATURE_APP          OFF CACHE BOOL   "Install the full-feature app: vvencFFapp" )
 
 if( CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR
     CMAKE_CXX_COMPILER_ID STREQUAL "Clang" )
@@ -300,7 +321,7 @@ if( VVENC_ENABLE_INSTALL )
 
     set( CMAKE_INSTALL_RPATH ${RPATH_BASE} ${RPATH_BASE}/${RPATH_REL_DIR} )
     message( STATUS "CMAKE_INSTALL_RPATH=${CMAKE_INSTALL_RPATH}" )
-  endif()  
+  endif()
 endif()
 
 

Makefile

@@ -74,6 +74,10 @@ ifneq ($(install-ffapp),)
 CONFIG_OPTIONS += -DVVENC_INSTALL_FULLFEATURE_APP=$(install-ffapp)
 endif
 
+ifneq ($(override-compiler-check),)
+CONFIG_OPTIONS += -DVVENC_OVERRIDE_COMPILER_CHECK=$(override-compiler-check)
+endif
+
 ifeq ($(j),)
 # Query cmake for the number of cores
 NUM_JOBS := $(shell cmake -P cmake/modules/vvencNumCores.cmake)

cmake/modules/vvencCompilerSupport.cmake

@@ -83,3 +83,21 @@ function( _emscripten_enable_wasm_simd128 )
     set( CMAKE_REQUIRED_FLAGS -msimd128 PARENT_SCOPE )
   endif()
 endfunction()
+
+function( check_problematic_compiler output_var compiler_id first_bad_version first_fixed_version )
+  if( CMAKE_CXX_COMPILER_ID STREQUAL "${compiler_id}"
+      AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL "${first_bad_version}"
+      AND (
+        NOT "${first_fixed_version}"
+        OR CMAKE_CXX_COMPILER_VERSION VERSION_LESS "${first_fixed_version}" ) )
+
+    set( ${output_var} TRUE PARENT_SCOPE )
+
+    if( "${first_fixed_version}" )
+      set( ${output_var}_VERSION_RANGE "(${first_bad_version}...${first_fixed_version}]" PARENT_SCOPE )
+    else()
+      set( ${output_var}_VERSION_RANGE "(${first_bad_version}...)"                       PARENT_SCOPE )
+    endif()
+
+  endif()
+endfunction()

include/vvenc/vvencCfg.h

@@ -430,7 +430,7 @@ typedef struct vvenc_config
   int                 m_framesToBeEncoded;                                               // number of encoded frames (default: 0, all)
   int                 m_inputBitDepth[ 2 ];                                              // bit-depth of input pictures (2d array for luma,chroma)
 
-  int                 m_numThreads;                                                      // number of worker threads ( if <0: <720p 4threads, else 8threads (limited to available cores))
+  int                 m_numThreads;                                                      // number of worker threads ( if <0: <720p 4threads, <5K 2880p 8threads, else 12threads (limited to available cores))
 
   int                 m_QP;                                                              // QP value of key-picture (0-63, default: 32)
   int                 m_RCTargetBitrate;                                                 // target bitrate in bps (default: 0 (RC disabled))
@@ -761,7 +761,7 @@ typedef struct vvenc_config
   bool                m_picReordering;
   bool                m_reservedFlag;
   bool                m_poc0idr;
-  int8_t              m_fppLinesSynchro;
+  int8_t              m_ifpLines;
   bool                m_blockImportanceMapping;
   bool                m_saoScc;
   bool                m_addGOP32refPics;
@@ -774,8 +774,14 @@ typedef struct vvenc_config
                                                                                          // if negative, the absolute value is interpreted as a 4-bit fixed point multiplier of the target bitrate).
                                                                                          // -24, i.e. -1.1000 binary, means the maxrate would be set to be the 1.5x of the target bitrate.
                                                                                          // for convenience use VVENC_SET_MAXRATE_FACTOR, e.g. VVENC_SET_MAXRATE_FACTOR(1.5), to set the multiplier
-  int                 m_forceScc;
-  double              m_reservedDouble[9];
+  int8_t              m_forceScc;
+  bool                m_ifp;
+
+  int8_t              m_reservedInt8[2];
+
+  int                 m_minIntraDist;
+  int                 m_reservedInt;
+  double              m_reservedDouble[8];
 
   // internal state variables
   bool                m_configDone;                                                      // state variable, Private context used for internal data ( do not change )

source/Lib/CommonLib/CommonDef.h

@@ -501,8 +501,6 @@ static constexpr uint8_t MAX_TMP_BUFS = 6;
 
 static constexpr int QPA_MAX_NOISE_LEVELS = 8;
 
-static constexpr int FPPLS_ALF_DERIVE_LINES   = 1; ///< number of CTU lines for ALF filter derivation
-static constexpr int FPPLS_CCALF_DERIVE_LINES = 1; ///< number of CTU lines for CCALF filter derivation
 
 
 // ====================================================================================================================

source/Lib/CommonLib/DepQuant.cpp

@@ -1436,7 +1436,7 @@ void DepQuant::quant( TransformUnit& tu, const ComponentID compID, const CCoeffB
     const uint32_t    log2TrHeight    = Log2(height);
     const bool isLfnstApplied         = tu.cu->lfnstIdx > 0 && (CU::isSepTree(*tu.cu) ? true : isLuma(compID));
     const bool enableScalingLists     = getUseScalingList(width, height, (tu.mtsIdx[compID] == MTS_SKIP), isLfnstApplied);
-    static_cast<DQIntern::DepQuant*>(p)->quant( tu, pSrc, compID, cQP, Quant::m_dLambda, ctx, uiAbsSum, enableScalingLists, Quant::getQuantCoeff(scalingListType, qpRem, log2TrWidth, log2TrHeight) );
+    p->quant( tu, pSrc, compID, cQP, Quant::m_dLambda, ctx, uiAbsSum, enableScalingLists, Quant::getQuantCoeff(scalingListType, qpRem, log2TrWidth, log2TrHeight) );
   }
   else
   {
@@ -1460,7 +1460,7 @@ void DepQuant::dequant( const TransformUnit& tu, CoeffBuf& dstCoeff, const Compo
     const uint32_t    log2TrHeight   = Log2(height);
     const bool isLfnstApplied        = tu.cu->lfnstIdx > 0 && (CU::isSepTree(*tu.cu) ? true : isLuma(compID));
     const bool enableScalingLists    = getUseScalingList(width, height, (tu.mtsIdx[compID] == MTS_SKIP), isLfnstApplied);
-    static_cast<DQIntern::DepQuant*>(p)->dequant( tu, dstCoeff, compID, cQP, enableScalingLists, Quant::getDequantCoeff(scalingListType, qpRem, log2TrWidth, log2TrHeight) );
+    p->dequant( tu, dstCoeff, compID, cQP, enableScalingLists, Quant::getDequantCoeff(scalingListType, qpRem, log2TrWidth, log2TrHeight) );
   }
   else
   {
@@ -1472,7 +1472,7 @@ void DepQuant::init( int rdoq, bool useRDOQTS, int thrVal )
 {
   QuantRDOQ2::init( rdoq, useRDOQTS, thrVal );
 
-  static_cast<DQIntern::DepQuant*>(p)->init( thrVal );
+  p->init( thrVal );
 }
 
 } // namespace vvenc

source/Lib/CommonLib/DepQuant.h

@@ -241,8 +241,8 @@ class DepQuantImpl
 public:
   virtual ~DepQuantImpl() {}
   virtual void quant   ( TransformUnit& tu, const CCoeffBuf& srcCoeff, const ComponentID compID, const QpParam& cQP, const double lambda, const Ctx& ctx, TCoeff& absSum, bool enableScalingLists, int* quantCoeff ) = 0;
-  virtual void dequant ( const TransformUnit& tu,  CoeffBuf& recCoeff, const ComponentID compID, const QpParam& cQP,                                                      bool enableScalingLists, int* quantCoeff );
-  virtual void init    ( int dqTrVal );
+  void         dequant ( const TransformUnit& tu,  CoeffBuf& recCoeff, const ComponentID compID, const QpParam& cQP,                                                      bool enableScalingLists, int* quantCoeff );
+  void         init    ( int dqTrVal );
 
 protected:
   DQIntern::Quantizer  m_quant;

source/Lib/CommonLib/InterPrediction.cpp

@@ -246,7 +246,7 @@ void InterPrediction::destroy()
   m_IBCBuffer.destroy();
 }
 
-void InterPrediction::init( RdCost* pcRdCost, ChromaFormat chFormat, const int ctuSize, const int fppLinesSynchro )
+void InterPrediction::init( RdCost* pcRdCost, ChromaFormat chFormat, const int ctuSize, const int ifpLines )
 {
   // if it has been initialised before, but the chroma format has changed, release the memory and start again.
   if( m_yuvPred[L0].getOrigin( COMP_Y ) != nullptr && m_currChromaFormat != chFormat )
@@ -279,7 +279,7 @@ void InterPrediction::init( RdCost* pcRdCost, ChromaFormat chFormat, const int c
     m_IBCBufferWidth = g_IBCBufferSize / ctuSize;
     m_IBCBuffer.create(UnitArea(chFormat, Area(0, 0, m_IBCBufferWidth, ctuSize)));
   }
-  InterPredInterpolation::m_fppLinesSynchro = fppLinesSynchro;
+  InterPredInterpolation::m_ifpLines = ifpLines;
 }
 
 // ====================================================================================================================
@@ -615,7 +615,7 @@ InterPredInterpolation::InterPredInterpolation()
   , m_skipPROF(false)
   , m_encOnly(false)
   , m_isBi(false)
-  , m_fppLinesSynchro(0)
+  , m_ifpLines(0)
 {
 
 }
@@ -727,7 +727,7 @@ void InterPredInterpolation::xPredInterBlk ( const ComponentID compID, const Cod
 
   bool  wrapRef = false;
   Mv    mv(_mv);
-  CHECKD( m_fppLinesSynchro && !srcPadBuf && !CU::isMvInRangeFPP( cu[compID].y, cu[compID].height, mv.ver, m_fppLinesSynchro, *cu.cs->pcv, getComponentScaleY(compID, chFmt) ), "xPredInterBlk: CTU line-wise FPP MV restriction failed!\n" );
+  CHECKD( m_ifpLines && !srcPadBuf && cu.cs->picture != refPic && !CU::isMvInRangeFPP( cu[compID].y, cu[compID].height, mv.ver, m_ifpLines, *cu.cs->pcv, getComponentScaleY(compID, chFmt) ), "xPredInterBlk: CTU line-wise FPP MV restriction failed!\n" );
   if( !isIBC && cu.cs->pcv->wrapArround )
   {
     wrapRef = wrapClipMv( mv, cu.blocks[0].pos(), cu.blocks[0].size(), *cu.cs);
@@ -1796,7 +1796,7 @@ void InterPredInterpolation::xPredAffineBlk(const ComponentID compID, const Codi
         iMvScaleTmpVer = curMv.ver;
       }
 
-      CHECKD( m_fppLinesSynchro && !CU::isMvInRangeFPP( puY + h, blockHeight, iMvScaleTmpVer, m_fppLinesSynchro, *pps.pcv, iScaleY ), "xPredAffineBlk: FPP MV restriction failed!\n" );
+      CHECKD( m_ifpLines && !CU::isMvInRangeFPP( puY + h, blockHeight, iMvScaleTmpVer, m_ifpLines, *pps.pcv, iScaleY ), "xPredAffineBlk: FPP MV restriction failed!\n" );
       // get the MV in high precision
       int xFrac, yFrac, xInt, yInt;
 
@@ -1896,10 +1896,10 @@ void InterPredInterpolation::xPredAffineBlk(const ComponentID compID, const Codi
 
 }
 
-bool InterPredInterpolation::xIsAffineMvInRangeFPP( const CodingUnit &cu, const Mv* _mv, const int fppLinesSynchro, const int mvPrecShift )
+bool InterPredInterpolation::xIsAffineMvInRangeFPP( const CodingUnit &cu, const Mv* _mv, const int ifpLines, const int mvPrecShift )
 {
   const PreCalcValues& pcv = *cu.cs->pcv;
-  if( cu.ly() >= ( ( pcv.heightInCtus - 1 - fppLinesSynchro ) << pcv.maxCUSizeLog2 ) )
+  if( cu.ly() >= ( ( pcv.heightInCtus - 1 - ifpLines ) << pcv.maxCUSizeLog2 ) )
     return true;
 
   const ChromaFormat chFmt = cu.chromaFormat;
@@ -1941,7 +1941,7 @@ bool InterPredInterpolation::xIsAffineMvInRangeFPP( const CodingUnit &cu, const
   }
   const bool subblkMVSpreadOverLimit = InterPrediction::isSubblockVectorSpreadOverLimit(iDMvHorX, iDMvHorY, iDMvVerX, iDMvVerY, cu.interDir);
 
-  const int yRefMax     = ( ( ( cu.ly() >> pcv.maxCUSizeLog2 ) + fppLinesSynchro + 1 ) << pcv.maxCUSizeLog2 ) - 1;
+  const int yRefMax     = ( ( ( cu.ly() >> pcv.maxCUSizeLog2 ) + ifpLines + 1 ) << pcv.maxCUSizeLog2 ) - 1;
   const int dctifMarginVerBot = 4;
 
   auto roundMvVal = [&](int mvVal, int shift)

source/Lib/CommonLib/InterPrediction.h

@@ -82,7 +82,7 @@ class InterPredInterpolation
   InterpolationFilter  m_if;
   Pel*                 m_filteredBlock        [LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS_SIGNAL][LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS_SIGNAL][MAX_NUM_COMP];
   Pel*                 m_filteredBlockTmp     [LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS_SIGNAL][MAX_NUM_COMP];
-  int                  m_fppLinesSynchro;
+  int                  m_ifpLines;
 
   int  xRightShiftMSB         ( int numer, int denom );
   void xApplyBDOF             ( PelBuf& yuvDst, const ClpRng& clpRng );
@@ -122,7 +122,7 @@ class InterPredInterpolation
                             PelUnitBuf &predDst, PelUnitBuf &predSrc0, PelUnitBuf &predSrc1);
 
   static bool isSubblockVectorSpreadOverLimit(int a, int b, int c, int d, int predType);
-  bool xIsAffineMvInRangeFPP (const CodingUnit& cu, const Mv* _mv, const int fppLinesSynchro, const int mvPrecShift = MV_FRACTIONAL_BITS_INTERNAL);
+  bool xIsAffineMvInRangeFPP (const CodingUnit& cu, const Mv* _mv, const int ifpLines, const int mvPrecShift = MV_FRACTIONAL_BITS_INTERNAL);
 };
 
 class DMVR : public InterPredInterpolation
@@ -171,7 +171,7 @@ class InterPrediction : public DMVR
   InterPrediction();
   virtual ~InterPrediction();
 
-  void    init                  ( RdCost* pcRdCost, ChromaFormat chromaFormatIDC, const int ctuSize, const int fppLinesSynchro = 0 );
+  void    init                  ( RdCost* pcRdCost, ChromaFormat chromaFormatIDC, const int ctuSize, const int ifpLines = 0 );
   void    destroy               ();
 
   // inter

source/Lib/CommonLib/Picture.cpp

@@ -161,6 +161,7 @@ Picture::Picture()
     , isFinished        ( false )
     , isLongTerm        ( false )
     , isFlush           ( false )
+    , isInProcessList   ( false )
     , precedingDRAP     ( false )
     , gopEntry          ( nullptr )
     , refCounter        ( 0 )
@@ -226,6 +227,7 @@ void Picture::reset()
   isFinished          = false;
   isLongTerm          = false;
   isFlush             = false;
+  isInProcessList     = false;
   isMeanQPLimited     = false;
   precedingDRAP       = false;
 
@@ -236,6 +238,9 @@ void Picture::reset()
   gopAdaptedQP        = 0;
   actualHeadBits      = 0;
   actualTotalBits     = 0;
+  encRCPic            = nullptr;
+  picApsGlobal        = nullptr;
+  refApsGlobal        = nullptr;
 
   std::fill_n( m_sharedBufs, (int)NUM_PIC_TYPES, nullptr );
   std::fill_n( m_bufsOrigPrev, NUM_QPA_PREV_FRAMES, nullptr );

source/Lib/CommonLib/Picture.h

@@ -114,6 +114,7 @@ struct PicApsGlobal{
   int      poc;
   unsigned tid;
   bool     initalized = false;
+  int      refCnt = 0;
   ParameterSetMap<APS> apsMap;
   PicApsGlobal( int _p ) : poc(_p), tid(MAX_UINT), apsMap( MAX_NUM_APS * MAX_NUM_APS_TYPE ) {}
   PicApsGlobal( int _p, unsigned _t ) : poc(_p), tid(_t), apsMap( MAX_NUM_APS * MAX_NUM_APS_TYPE ) {}
@@ -216,6 +217,7 @@ struct Picture : public UnitArea
   bool                          isFinished;
   bool                          isLongTerm;
   bool                          isFlush;
+  bool                          isInProcessList;
   bool                          precedingDRAP; // preceding a DRAP picture in decoding order
 
   const GOPEntry*               gopEntry;

source/Lib/CommonLib/Slice.cpp

@@ -443,7 +443,7 @@ void Slice::updateRefPicCounter( int step )
   }
 }
 
-bool Slice::checkRefPicsReconstructed() const
+bool Slice::checkAllRefPicsReconstructed() const
 {
   for ( int refList = 0; refList < NUM_REF_PIC_LIST_01; refList++ )
   {
@@ -460,6 +460,23 @@ bool Slice::checkRefPicsReconstructed() const
   return true;
 }
 
+bool Slice::checkAllRefPicsAccessible() const
+{
+  for ( int refList = 0; refList < NUM_REF_PIC_LIST_01; refList++ )
+  {
+    int numOfActiveRef = numRefIdx[ refList ];
+    for ( int i = 0; i < numOfActiveRef; i++ )
+    {
+      if ( ! refPicList[ refList ][ i ]->isInProcessList )
+      {
+        return false;
+      }
+    }
+  }
+
+  return true;
+}
+
 void Slice::checkColRefIdx(uint32_t curSliceSegmentIdx, const Picture* pic) const
 {
   Slice* curSlice   = pic->slices[ curSliceSegmentIdx ];

source/Lib/CommonLib/Slice.h

@@ -1265,7 +1265,8 @@ class Slice
   void                        resetSlicePart();
   void                        constructRefPicList(const PicList& rcListPic, bool extBorder, const bool usingLongTerm = true);
   void                        updateRefPicCounter( int step );
-  bool                        checkRefPicsReconstructed() const;
+  bool                        checkAllRefPicsAccessible() const;
+  bool                        checkAllRefPicsReconstructed() const;
   void                        setRefPOCList();
   void                        setSMVDParam();
   void                        checkColRefIdx(uint32_t curSliceSegmentIdx, const Picture* pic) const;

source/Lib/CommonLib/TimeProfiler.h

@@ -91,6 +91,7 @@ namespace vvenc {
   E_( P_INTRA_CHROMA            ) \
   E_( P_INTRA                   ) \
   E_( P_QUANT                   ) \
+  E_( P_DEQUANT                 ) \
   E_( P_TRAFO                   ) \
   E_( P_RESHAPER                ) \
   E_( P_DEBLOCK_FILTER          ) \

source/Lib/CommonLib/TrQuant.cpp

@@ -292,7 +292,7 @@ void TrQuant::xDeQuant(const TransformUnit& tu,
                        const ComponentID   &compID,
                        const QpParam       &cQP)
 {
-  PROFILER_SCOPE_AND_STAGE( 1, _TPROF, P_QUANT );
+  PROFILER_SCOPE_AND_STAGE( 1, _TPROF, P_DEQUANT );
   m_quant->dequant( tu, dstCoeff, compID, cQP );
 }