From c64511d9d46aa0f126ac3060b0c4afe095763b49 Mon Sep 17 00:00:00 2001
From: Vadim Seregin <vseregin@qti.qualcomm.com>
Date: Wed, 3 Nov 2021 03:16:22 +0000
Subject: [PATCH] Remove leftover shift in SSE and add initialization for all
tIDs for MTT depth
---
source/App/EncoderApp/EncAppCfg.cpp | 7 ++++++-
source/Lib/CommonLib/TypeDef.h | 4 ++--
source/Lib/CommonLib/x86/RdCostX86.h | 2 +-
3 files changed, 9 insertions(+), 4 deletions(-)
diff --git a/source/App/EncoderApp/EncAppCfg.cpp b/source/App/EncoderApp/EncAppCfg.cpp
index bbc0ef67c..b0e1889da 100644
--- a/source/App/EncoderApp/EncAppCfg.cpp
+++ b/source/App/EncoderApp/EncAppCfg.cpp
@@ -2694,11 +2694,16 @@ bool EncAppCfg::parseCfg( int argc, char* argv[] )
#if JVET_X0144_MAX_MTT_DEPTH_TID
CHECK( m_sMaxMTTHierarchyDepthByTid.size() > MAX_TLAYER, "MaxMTTHierarchyDepthByTid is greater than MAX_TLAYER" );
- for( int i = 0; i < m_sMaxMTTHierarchyDepthByTid.size(); i++ )
+ for( int i = 0; i < (int)m_sMaxMTTHierarchyDepthByTid.size(); i++ )
{
CHECK( i >= MAX_TLAYER, "Index exceeds MAX_TLAYER" );
m_maxMTTHierarchyDepthByTid[i] = std::stoul( m_sMaxMTTHierarchyDepthByTid.substr( i, 1 ) );
}
+
+ for( int i = (int)m_sMaxMTTHierarchyDepthByTid.size(); i < MAX_TLAYER; i++ )
+ {
+ m_maxMTTHierarchyDepthByTid[i] = m_uiMaxMTTHierarchyDepth;
+ }
#endif
// check validity of input parameters
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index 38d9110f4..a0834f899 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -116,7 +116,7 @@
#define JVET_X0124_TMP_SIGNAL 1 // JVET-X0124: cleanup on signalling of intra template matching
#endif
#define JVET_V0130_INTRA_TMP 1 // JVET-V0130: template matching prediction
-#define JVET_W0069_TMP_BOUNDARY 1 // JVET-W0069: boundary handling for TMP
+#define JVET_W0069_TMP_BOUNDARY 1 // JVET-W0069: boundary handling for TMP
#define JVET_W0123_TIMD_FUSION 1 // JVET-W0123: Template based intra mode derivation and fusion
#if JVET_W0123_TIMD_FUSION
#define JVET_X0148_TIMD_PDPC 1 // JVET-X0148: PDPC handling for TIMD
@@ -189,7 +189,7 @@
#define ENABLE_SIMD_SIGN_PREDICTION 1
#endif
#if JVET_V0130_INTRA_TMP
-#define ENABLE_SIMD_TMP 1
+#define ENABLE_SIMD_TMP 1
#endif
#if JVET_V0094_BILATERAL_FILTER
#define ENABLE_SIMD_BILATERAL_FILTER 1
diff --git a/source/Lib/CommonLib/x86/RdCostX86.h b/source/Lib/CommonLib/x86/RdCostX86.h
index 0bbbe46e5..6225c46a6 100644
--- a/source/Lib/CommonLib/x86/RdCostX86.h
+++ b/source/Lib/CommonLib/x86/RdCostX86.h
@@ -421,7 +421,7 @@ Distortion RdCost::xGetSSE_NxN_SIMD(const DistParam &rcDtParam)
Sum = _mm256_add_epi64( _mm256_unpacklo_epi32( Sum, vzero ), _mm256_unpackhi_epi32( Sum, vzero ) );
Sum = _mm256_add_epi64( Sum, _mm256_permute4x64_epi64( Sum, 14 ) );
Sum = _mm256_add_epi64( Sum, _mm256_permute4x64_epi64( Sum, 1 ) );
- uiRet += _mm_cvtsi128_si64( _mm256_castsi256_si128( Sum ) ) >> uiShift;
+ uiRet += _mm_cvtsi128_si64( _mm256_castsi256_si128( Sum ) );
}
uiRet >>= uiShift;
#endif
--
GitLab