perf: 增加截面 Region 缓存,开图恢复耗时降 56%

一次开图要做 5641 次截面恢复(22.4 s),其中大量内容重复,
而每次恢复都是"写临时 SAT 文件 + 从盘 acisIn + 删文件",单次 2.7~5.5 ms,
还要过一次杀软。本提交加一级母体缓存,命中直接 clone 返回。

新增 CoreLibrary/SectionRegionCache.{h,cpp},三个通道按 key 的正确性强度分开:
- kParam:纯参数 hash,用于 AecDbEwFitting::GetSection 的两个重载
  (key = dD/dt/HLR 标志 或 W/H/dt/HLR 标志,同参数必同几何)
- kAttr:类名 + spec XData(排除编号,并跳过 Id/点/向量/矩阵/二进制类型),
  只给 AecDbEwPipe 用。AecDbEwPipeBase 是很多类的基类,其余派生类实例间的
  区别就是截面,不满足"同数据必同截面",所以一律走内容 hash
- kText:SAT 文本内容 hash,用于 RegionAcisInDwgInFields 的其它类
  与 CoreTools::StringToRegion 的两个重载

约定:母体由缓存持有、永不外泄,Acquire/Put 都返回 clone,调用方拥有返回值;
母体不入库;每条通道 4096 条上限(AEC_SECTION_CACHE_MAX 可调),超出按 FIFO 淘汰;
AEC_SECTION_CACHE=0 可整体关闭做 A/B;卸载时 ClearAll() 释放,避免 DLL 析构顺序问题。

关键发现:SAT 文本第 2 行带存盘时间戳
  16 Autodesk AutoCAD 19 ASM 219.0.0.5600 NT 24 Thu Oct 22 10:19:20 2026
同一几何存盘时间不同,整份文本就不同(采样里两份 2596 字节的文本只差
10:19:20 与 10:19:16),这正是内容 hash 命中率只有 54% 的原因,与截面无关。
因此新增 HashSatText():算 hash 时跳过第 2 行,仅当该行含 "Autodesk" 时才跳过。
计数器也改用同一口径,避免两套统计互相矛盾。

实测(big1.dwg / AutoCAD 2020 / R230 Debug,AEC_SECTION_CACHE 关 vs 开):
  开图墙钟        108~113 s      -> 98 s
  截面恢复总耗时  22.4 s         -> 9.87 s(-56%)
    StringToRegion  14.05 s      -> 7.73 s
    RegionAcisIn     8.38 s      -> 2.14 s
  临时文件        6773 次/9.34MB -> 2875 次/3.61MB
  命中率          text 65.5% / attr 84.1% / param 88.8%
  clone 开销      命中 58us、未命中 73us(合计 0.37 s,可忽略)

顺带:
- tools/run_acad_perf.ps1 增加 -CacheOff 开关,用于同一构建上的 A/B 对比
- SectionPerfCounter 增加 AppendReport(),让缓存统计也写进同一份报告文件
- 诊断口子 AEC_SAT_SAMPLE=n 可导出未命中的 SAT 原文,用于定位这类差异

验证:R230 Debug 编译通过并部署;用脚本自动开图各跑一次(缓存关/开),
数据已写入 pipe-section-region-performance-plan.md。
尚未做:R190 编译(本机缺 v100 工具集)、Release 构建的收益复测、
开图期间 1140 次 RegionAcisout 回写(与本缓存无关,需单独立项)。
This commit is contained in:
company-desktop committed 2026-10-28 20:41:05 +08:00
1 parent fd5bbb60ab
commit f8d10d5a3b
11 files changed
+594 -2

No files matched your search

@@ -0,0 +1,344 @@
#include "stdafx.h"
#include "SectionRegionCache.h"
#include "AecDbObjXDataMap.h"
#include "SectionPerfCounter.h"
#include <windows.h>
#include <stdlib.h>
#include <string>
#include <map>
#include <list>
namespace
{
// 每个通道的条目上限。实测大图纸唯一 SAT 文本约 2600 条,2048 会把热条目挤掉,
// 所以默认给 4096;可用 AEC_SECTION_CACHE_MAX 覆盖(内存吃紧时调小)。
const size_t kDefaultMaxEntries = 4096;
size_t GetMaxEntries()
{
static size_t s_nMax = 0;
if (s_nMax == 0)
{
const char* psz = ::getenv("AEC_SECTION_CACHE_MAX");
s_nMax = (psz != NULL) ? (size_t)atoi(psz) : 0;
if (s_nMax < 16)
s_nMax = kDefaultMaxEntries;
}
return s_nMax;
}
static const char* kChannelName[] = { "text", "attr", "param" };
struct ChannelData
{
ChannelData() : nHits(0), nMisses(0), nEvicted(0) {}
std::map<unsigned long long, AcDbRegion*> map;
std::list<unsigned long long> fifo;
unsigned long long nHits;
unsigned long long nMisses;
unsigned long long nEvicted;
void ReleaseAll()
{
for (std::map<unsigned long long, AcDbRegion*>::iterator it = map.begin(); it != map.end(); ++it)
{
AcDbRegion* p = it->second;
if (p != NULL)
delete p;
}
map.clear();
fifo.clear();
nEvicted = 0;
}
};
struct State
{
State() : bInited(false), bEnabled(true) {}
bool bInited;
bool bEnabled;
ChannelData channels[SectionRegionCache::kChannelCount];
CRITICAL_SECTION cs;
};
// 函数内 static:CoreLibrary 是静态库且只被 CoreCommand 链接,
// 所以 CoreCommand.arx 内只有这一份缓存。
// 注意:不在析构里 delete 母体(DLL 卸载时 modeler 可能已先关,
// 析构顺序不安全);由 ClearAll() 在 kUnloadAppMsg 里显式释放。
State& GetState()
{
static State s_state;
if (!s_state.bInited)
{
::InitializeCriticalSection(&s_state.cs);
const char* psz = ::getenv("AEC_SECTION_CACHE");
s_state.bEnabled = !(psz != NULL && psz[0] == '0');
s_state.bInited = true;
}
return s_state;
}
struct AutoLock
{
explicit AutoLock(CRITICAL_SECTION* pcs) : m_pcs(pcs) { ::EnterCriticalSection(m_pcs); }
~AutoLock() { ::LeaveCriticalSection(m_pcs); }
CRITICAL_SECTION* m_pcs;
};
unsigned long long HashAdd(unsigned long long h, const void* pData, size_t nLen)
{
if (pData == NULL || nLen == 0)
return h;
const unsigned char* p = (const unsigned char*)pData;
for (size_t i = 0; i < nLen; ++i)
{
h ^= (unsigned long long)p[i];
h *= 1099511628211ULL;
}
return h;
}
}
namespace {
// 诊断用:AEC_SAT_SAMPLE=n 时把前 n 个"未命中的 SAT 文本"导出,便于比对差异来源
void DumpSample(const void* pData, size_t nLen)
{
static int s_nRemain = -1;
if (s_nRemain < 0)
{
const char* psz = ::getenv("AEC_SAT_SAMPLE");
s_nRemain = (psz != NULL) ? atoi(psz) : 0;
}
if (s_nRemain <= 0)
return;
--s_nRemain;
char szTmp[MAX_PATH + 1];
::memset(szTmp, 0, sizeof(szTmp));
if (::GetTempPathA(MAX_PATH, szTmp) == 0)
return;
std::string sPath(szTmp);
sPath += "aec_sat_samples.txt";
FILE* fp = NULL;
if (fopen_s(&fp, sPath.c_str(), "ab") == 0 && fp != NULL)
{
fprintf(fp, "===== SAMPLE len=%llu =====\n", (unsigned long long)nLen);
fwrite(pData, 1, nLen, fp);
fprintf(fp, "\n===== END =====\n");
fclose(fp);
}
}
}
namespace SectionRegionCache
{
bool Enabled()
{
return GetState().bEnabled;
}
unsigned long long HashBytes(const void* pData, size_t nLen)
{
return HashAdd(1469598103934665603ULL, pData, nLen);
}
unsigned long long HashSatText(const void* pData, size_t nLen)
{
if (pData == NULL || nLen == 0)
return 1469598103934665603ULL;
const char* p = (const char*)pData;
const char* pEnd = p + nLen;
// 找第 2 行(SAT 的 16 号记录:Autodesk 版本 + 存盘时间戳)
const char* pNL1 = (const char*)memchr(p, '\n', nLen);
if (pNL1 != NULL && pNL1 + 1 < pEnd)
{
const char* pLine2 = pNL1 + 1;
const char* pNL2 = (const char*)memchr(pLine2, '\n', (size_t)(pEnd - pLine2));
if (pNL2 != NULL)
{
// 只在第 2 行确实是 Autodesk 头时才跳过,避免误伤别的格式
bool bHeader = false;
for (const char* q = pLine2; q + 9 <= pNL2; ++q)
{
if (_strnicmp(q, "Autodesk", 8) == 0)
{
bHeader = true;
break;
}
}
if (bHeader)
{
unsigned long long h = HashBytes(p, (size_t)(pNL1 - p));
return HashAdd(h, pNL2 + 1, (size_t)(pEnd - (pNL2 + 1)));
}
}
}
return HashBytes(pData, nLen);
}
unsigned long long BuildSpecAttrHash(const AecDbObjXDataMap& map, const TCHAR* pszClass)
{
unsigned long long h = 1469598103934665603ULL;
// 类名参与:不同实体类即使数据相同也绝不能共用截面
CStringA sCls(pszClass != NULL ? pszClass : _T("?"));
h = HashAdd(h, (LPCSTR)sCls, strlen((LPCSTR)sCls));
// std::map 按 key 排序,遍历顺序稳定
for (AecDbObjXDataMap::It it = const_cast<AecDbObjXDataMap&>(map).Begin();
it != const_cast<AecDbObjXDataMap&>(map).End(); ++it)
{
const CString& sKey = it->first;
const AecXData& xd = it->second;
// 编号每根管子唯一,必须排除
if (sKey.CompareNoCase(_T("No.")) == 0)
continue;
h = HashAdd(h, sKey.GetString(), (size_t)sKey.GetLength() * sizeof(TCHAR));
h = HashAdd(h, &xd.nVt, sizeof(xd.nVt));
switch (xd.nVt)
{
case AecXData::kTypeShort:
h = HashAdd(h, &xd.iVal, sizeof(xd.iVal));
break;
case AecXData::kTypeInt:
case AecXData::kTypeInt64:
h = HashAdd(h, &xd.nVal, sizeof(xd.nVal));
break;
case AecXData::kTypeDouble:
h = HashAdd(h, &xd.dVal[0], sizeof(double));
break;
case AecXData::kTypeString:
if (xd.pszVal != NULL)
h = HashAdd(h, xd.pszVal, _tcslen(xd.pszVal) * sizeof(TCHAR));
break;
default:
// Id / 点 / 向量 / 矩阵 / 二进制:与截面无关或随实体变化,跳过
break;
}
}
return h;
}
AcDbRegion* Acquire(Channel ch, unsigned long long nKey, const void* pSampleData, size_t nSampleLen)
{
State& s = GetState();
if (!s.bEnabled || ch < 0 || ch >= kChannelCount)
return NULL;
AutoLock lock(&s.cs);
ChannelData& cd = s.channels[ch];
std::map<unsigned long long, AcDbRegion*>::iterator it = cd.map.find(nKey);
if (it == cd.map.end() || it->second == NULL)
{
++cd.nMisses;
if (ch == SectionRegionCache::kText && pSampleData != NULL)
DumpSample(pSampleData, nSampleLen);
return NULL;
}
++cd.nHits;
SectionPerf::Scope hitScope("cache.hit.clone");
return (AcDbRegion*)it->second->clone();
}
AcDbRegion* Put(Channel ch, unsigned long long nKey, AcDbRegion* pMaster)
{
State& s = GetState();
if (pMaster == NULL)
return NULL;
if (!s.bEnabled || ch < 0 || ch >= kChannelCount)
return pMaster; // 关闭时原样返回,调用方照旧拥有
AcDbRegion* pForCaller = NULL;
{
SectionPerf::Scope putScope("cache.miss.clone");
pForCaller = (AcDbRegion*)pMaster->clone();
}
AutoLock lock(&s.cs);
ChannelData& cd = s.channels[ch];
std::map<unsigned long long, AcDbRegion*>::iterator it = cd.map.find(nKey);
if (it != cd.map.end())
{
// 已有同 key(并发或重复入库):丢弃新的母体
delete pMaster;
return pForCaller;
}
// 超出上限先淘汰最旧的
while (cd.map.size() >= GetMaxEntries() && !cd.fifo.empty())
{
const unsigned long long nOld = cd.fifo.front();
cd.fifo.pop_front();
std::map<unsigned long long, AcDbRegion*>::iterator itOld = cd.map.find(nOld);
if (itOld != cd.map.end())
{
AcDbRegion* pOld = itOld->second;
cd.map.erase(itOld);
delete pOld;
++cd.nEvicted;
}
}
cd.map[nKey] = pMaster;
cd.fifo.push_back(nKey);
return pForCaller;
}
void Dump()
{
State& s = GetState();
std::string out;
char szLine[512];
out += "\n---- section region cache ----\n";
out += " channel entries hits misses hit_rate evicted\n";
for (int i = 0; i < (int)kChannelCount; ++i)
{
const ChannelData& cd = s.channels[i];
const unsigned long long nTotal = cd.nHits + cd.nMisses;
const double rate = nTotal ? (100.0 * (double)cd.nHits / (double)nTotal) : 0.0;
sprintf_s(szLine, sizeof(szLine), " %-9s %-9llu %-7llu %-7llu %7.1f%% %llu\n",
kChannelName[i], (unsigned long long)cd.map.size(),
cd.nHits, cd.nMisses, rate, cd.nEvicted);
out += szLine;
}
const char* psz = out.c_str();
while (*psz != '\0')
{
const char* pszEnd = strchr(psz, '\n');
const size_t nSeg = (pszEnd != NULL) ? (size_t)(pszEnd - psz) : strlen(psz);
if (nSeg > 0 && nSeg < sizeof(szLine))
{
memcpy(szLine, psz, nSeg);
szLine[nSeg] = '\0';
acutPrintf(_T("\n%hs"), szLine);
}
if (pszEnd == NULL)
break;
psz = pszEnd + 1;
}
// 同时写进与 SectionPerf 相同的报告文件,便于脚本回收
SectionPerf::AppendReport(out.c_str());
}
void ClearAll()
{
State& s = GetState();
AutoLock lock(&s.cs);
for (int i = 0; i < (int)kChannelCount; ++i)
s.channels[i].ReleaseAll();
}
}