[x265] [PATCH] sao: only load/save entropy state of sao whenapplying sao
Ximing Cheng
chengximing1989 at foxmail.com
Mon Nov 20 18:20:43 CET 2017
A generic function copyFrom() will copy MAX_OFF_CTX_MOD bytes values, which MAX_OFF_CTX_MOD == 157, but SAO only need two bytes,
load function in release mode, we can get its disassembly code of memcpy:
void Entropy::copyFrom(const Entropy& src)
{
X265_CHECK(src.m_valid, "invalid copy source context\n");
copyState(src);
000000013F4334D0 mov eax,dword ptr [rdx+0B0h]
000000013F4334D6 mov dword ptr [rcx+0B0h],eax
000000013F4334DC mov eax,dword ptr [rdx+0B4h]
000000013F4334E2 mov dword ptr [rcx+0B4h],eax
000000013F4334E8 mov eax,dword ptr [rdx+0C0h]
000000013F4334EE mov dword ptr [rcx+0C0h],eax
000000013F4334F4 mov eax,dword ptr [rdx+0B8h]
000000013F4334FA mov dword ptr [rcx+0B8h],eax
000000013F433500 mov eax,dword ptr [rdx+0BCh]
000000013F433506 mov dword ptr [rcx+0BCh],eax
000000013F43350C mov rax,qword ptr [rdx+0C8h]
000000013F433513 mov qword ptr [rcx+0C8h],rax
memcpy(m_contextState, src.m_contextState, MAX_OFF_CTX_MOD * sizeof(uint8_t));
000000013F43351A movups xmm0,xmmword ptr [rdx+10h]
000000013F43351E movups xmmword ptr [rcx+10h],xmm0
000000013F433522 movups xmm1,xmmword ptr [rdx+20h]
000000013F433526 movups xmmword ptr [rcx+20h],xmm1
000000013F43352A movups xmm0,xmmword ptr [rdx+30h]
000000013F43352E movups xmmword ptr [rcx+30h],xmm0
000000013F433532 movups xmm1,xmmword ptr [rdx+40h]
000000013F433536 movups xmmword ptr [rcx+40h],xmm1
000000013F43353A movups xmm0,xmmword ptr [rdx+50h]
000000013F43353E movups xmmword ptr [rcx+50h],xmm0
000000013F433542 movups xmm1,xmmword ptr [rdx+60h]
000000013F433546 movups xmmword ptr [rcx+60h],xmm1
000000013F43354A movups xmm0,xmmword ptr [rdx+70h]
000000013F43354E movups xmmword ptr [rcx+70h],xmm0
000000013F433552 movups xmm1,xmmword ptr [rdx+80h]
000000013F433559 movups xmmword ptr [rcx+80h],xmm1
000000013F433560 movups xmm0,xmmword ptr [rdx+90h]
000000013F433567 movups xmmword ptr [rcx+90h],xmm0
000000013F43356E movsd xmm1,mmword ptr [rdx+0A0h]
000000013F433576 movsd mmword ptr [rcx+0A0h],xmm1
000000013F43357E mov eax,dword ptr [rdx+0A8h]
000000013F433584 mov dword ptr [rcx+0A8h],eax
000000013F43358A movzx eax,byte ptr [rdx+0ACh]
000000013F433591 mov byte ptr [rcx+0ACh],al
markValid();
}
We can see that, VC++ compiler will get a lot of movups to copy data. We can also get loadSAO disassembly code:
void Entropy::copyFromSAO(const Entropy& src)
{
X265_CHECK(src.m_valid, "invalid copy source context\n");
copyState(src);
000000013F4335A0 mov eax,dword ptr [rdx+0B0h]
000000013F4335A6 mov dword ptr [rcx+0B0h],eax
000000013F4335AC mov eax,dword ptr [rdx+0B4h]
000000013F4335B2 mov dword ptr [rcx+0B4h],eax
000000013F4335B8 mov eax,dword ptr [rdx+0C0h]
000000013F4335BE mov dword ptr [rcx+0C0h],eax
000000013F4335C4 mov eax,dword ptr [rdx+0B8h]
000000013F4335CA mov dword ptr [rcx+0B8h],eax
000000013F4335D0 mov eax,dword ptr [rdx+0BCh]
000000013F4335D6 mov dword ptr [rcx+0BCh],eax
000000013F4335DC mov rax,qword ptr [rdx+0C8h]
000000013F4335E3 mov qword ptr [rcx+0C8h],rax
// only copy OFF_SAO_MERGE_FLAG_CTX and OFF_SAO_TYPE_IDX_CTX, two bytes
uint16_t* srcEntropy = (uint16_t*)(src.m_contextState + OFF_SAO_MERGE_FLAG_CTX);
uint16_t* dstEntropy = (uint16_t*)(m_contextState + OFF_SAO_MERGE_FLAG_CTX);
*dstEntropy = *srcEntropy;
000000013F4335EA movzx eax,word ptr [rdx+0A8h]
000000013F4335F1 mov word ptr [rcx+0A8h],ax
markValid();
}
We can see that only one movzx and one mov instruction as only copy two bytes.
------------------ Original ------------------
From: "Ashok Kumar Mishra";<ashok at multicorewareinc.com>;
Date: Mon, Nov 20, 2017 05:27 PM
To: "Development for x265"<x265-devel at videolan.org>;
Subject: Re: [x265] [PATCH] sao: only load/save entropy state of sao whenapplying sao
On Sat, Nov 18, 2017 at 6:25 PM, Ximing Cheng <chengximing1989 at foxmail.com> wrote:
# HG changeset patch
# User Ximing Cheng <ximingcheng at tencent.com>
# Date 1511009612 -28800
# Sat Nov 18 20:53:32 2017 +0800
# Node ID a5b805430a4b0cb797604ab7f0176538d93e8d9d
# Parent 06979c0423504a324ea05ca3de59769c6d0fba0d
sao: only load/save entropy state of sao when applying sao
diff -r 06979c042350 -r a5b805430a4b source/encoder/entropy.cpp
--- a/source/encoder/entropy.cpp Thu Nov 16 20:23:14 2017 +0530
+++ b/source/encoder/entropy.cpp Sat Nov 18 20:53:32 2017 +0800
@@ -1506,6 +1506,19 @@
markValid();
}
We use a generic function copyFrom() for loading and storing the context during analysis. copyFrom() use memcopy(), so it is faster.
It is not much beneficial for writing a separate function for SAO, though only two bytes are changing.
+void Entropy::copyFromSAO(const Entropy& src)
+{
+ X265_CHECK(src.m_valid, "invalid copy source context\n");
+
+ copyState(src);
+
+ // only copy OFF_SAO_MERGE_FLAG_CTX and OFF_SAO_TYPE_IDX_CTX, two bytes
+ uint16_t* srcEntropy = (uint16_t*)(src.m_contextState + OFF_SAO_MERGE_FLAG_CTX);
+ uint16_t* dstEntropy = (uint16_t*)(m_contextState + OFF_SAO_MERGE_FLAG_CTX);
+ *dstEntropy = *srcEntropy;
+ markValid();
+}
+
void Entropy::codePartSize(const CUData& cu, uint32_t absPartIdx, uint32_t depth)
{
PartSize partSize = (PartSize)cu.m_partSize[absPartIdx];
diff -r 06979c042350 -r a5b805430a4b source/encoder/entropy.h
--- a/source/encoder/entropy.h Thu Nov 16 20:23:14 2017 +0530
+++ b/source/encoder/entropy.h Sat Nov 18 20:53:32 2017 +0800
@@ -136,7 +136,9 @@
// SBAC RD
void load(const Entropy& src) { copyFrom(src); }
+ void loadSAO(const Entropy& src) { copyFromSAO(src); }
void store(Entropy& dest) const { dest.copyFrom(*this); }
+ void storeSAO(Entropy& dest) const { dest.copyFromSAO(*this); }
void loadContexts(const Entropy& src) { copyContextsFrom(src); }
void loadIntraDirModeLuma(const Entropy& src);
void copyState(const Entropy& other);
@@ -254,6 +256,7 @@
bool& bCodeDQP, const uint32_t depthRange[2]);
void copyFrom(const Entropy& src);
+ void copyFromSAO(const Entropy& src);
void copyContextsFrom(const Entropy& src);
};
}
diff -r 06979c042350 -r a5b805430a4b source/encoder/sao.cpp
--- a/source/encoder/sao.cpp Thu Nov 16 20:23:14 2017 +0530
+++ b/source/encoder/sao.cpp Sat Nov 18 20:53:32 2017 +0800
@@ -242,9 +242,9 @@
break;
}
- m_entropyCoder.load(initState);
- m_rdContexts.next.load(initState);
- m_rdContexts.cur.load(initState);
+ m_entropyCoder.loadSAO(initState);
+ m_rdContexts.next.loadSAO(initState);
+ m_rdContexts.cur.loadSAO(initState);
SAOParam* saoParam = frame->m_encData->m_saoParam;
if (!saoParam)
@@ -1262,13 +1262,13 @@
for (int i = 0; i < planes; i++)
saoParam->ctuParam[i][addr].reset();
// SAO distortion calculation
- m_entropyCoder.load(m_rdContexts.cur);
+ m_entropyCoder.loadSAO(m_rdContexts.cur);
m_entropyCoder.resetBits();
if (allowMerge[0])
m_entropyCoder.codeSaoMerge(0);
if (allowMerge[1])
m_entropyCoder.codeSaoMerge(0);
- m_entropyCoder.store(m_rdContexts.temp);
+ m_entropyCoder.storeSAO(m_rdContexts.temp);
memset(m_offset, 0, sizeof(m_offset));
int64_t bestCost = 0;
int64_t rateDist = 0;
@@ -1333,7 +1333,7 @@
mergeDist += (estDist << 8) / lambda[!!plane];
}
- m_entropyCoder.load(m_rdContexts.cur);
+ m_entropyCoder.loadSAO(m_rdContexts.cur);
m_entropyCoder.resetBits();
if (allowMerge[0])
m_entropyCoder.codeSaoMerge(1 - mergeIdx);
@@ -1346,7 +1346,7 @@
{
SaoMergeMode mergeMode = mergeIdx ? SAO_MERGE_UP : SAO_MERGE_LEFT;
bestCost = mergeCost;
- m_entropyCoder.store(m_rdContexts.temp);
+ m_entropyCoder.storeSAO(m_rdContexts.temp);
for (int plane = 0; plane < planes; plane++)
{
if (saoParam->bSaoFlag[plane > 0])
@@ -1368,8 +1368,8 @@
m_numNoSao[0]++;
if (chroma && saoParam->ctuParam[1][addr].typeIdx < 0)
m_numNoSao[1]++;
- m_entropyCoder.load(m_rdContexts.temp);
- m_entropyCoder.store(m_rdContexts.cur);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
+ m_entropyCoder.storeSAO(m_rdContexts.cur);
}
}
@@ -1488,7 +1488,7 @@
int64_t costClasses[MAX_NUM_SAO_CLASS];
// RDO SAO_NA
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
m_entropyCoder.resetBits();
m_entropyCoder.codeSaoType(0);
int64_t costPartBest = calcSaoRdoCost(0, m_entropyCoder.getNumberOfWrittenBits(), lambda[0]);
@@ -1518,7 +1518,7 @@
estDist += distClasses[classIdx];
}
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
m_entropyCoder.resetBits();
m_entropyCoder.codeSaoOffsetEO(m_offset[0][typeIdx] + 1, typeIdx, 0);
@@ -1576,7 +1576,7 @@
for (int classIdx = bestClassBO; classIdx < bestClassBO + SAO_NUM_OFFSET; classIdx++)
estDist += distClasses[classIdx];
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
m_entropyCoder.resetBits();
m_entropyCoder.codeSaoOffsetBO(m_offset[0][SAO_BO] + bestClassBO, bestClassBO, 0);
@@ -1595,9 +1595,9 @@
}
rateDist = (bestDist << 8) / lambda[0];
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
m_entropyCoder.codeSaoOffset(*lclCtuParam, 0);
- m_entropyCoder.store(m_rdContexts.temp);
+ m_entropyCoder.storeSAO(m_rdContexts.temp);
if (m_param->internalCsp == X265_CSP_I400)
{
@@ -1616,7 +1616,7 @@
int32_t distClasses[MAX_NUM_SAO_CLASS];
int32_t bestClassBO[2] = { 0, 0 };
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
m_entropyCoder.resetBits();
m_entropyCoder.codeSaoType(0);
@@ -1651,7 +1651,7 @@
}
}
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
m_entropyCoder.resetBits();
for (int compIdx = 0; compIdx < 2; compIdx++)
@@ -1715,7 +1715,7 @@
estDist[compIdx - 1] += distClasses[classIdx];
}
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
m_entropyCoder.resetBits();
for (int compIdx = 0; compIdx < 2; compIdx++)
@@ -1740,13 +1740,13 @@
}
rateDist += (bestDist << 8) / lambda[1];
- m_entropyCoder.load(m_rdContexts.temp);
+ m_entropyCoder.loadSAO(m_rdContexts.temp);
if (saoParam->bSaoFlag[1])
{
m_entropyCoder.codeSaoOffset(*lclCtuParam[0], 1);
m_entropyCoder.codeSaoOffset(*lclCtuParam[1], 2);
- m_entropyCoder.store(m_rdContexts.temp);
+ m_entropyCoder.storeSAO(m_rdContexts.temp);
uint32_t rate = m_entropyCoder.getNumberOfWrittenBits();
bestCost = rateDist + rate;
_______________________________________________
x265-devel mailing list
x265-devel at videolan.org
https://mailman.videolan.org/listinfo/x265-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20171121/b34e4d0d/attachment-0001.html>
More information about the x265-devel
mailing list