perf: 给基础体构造加计时，定位到 DrawPipe/DrawCone/DrawLoft 共占 33%
继续追"剩余效率问题"。上一轮已确认 Parasolid 的 93.4% 发生在我们的视口绘制里， 本轮机位下一层：给 GiDrawUtility 的基础体构造函数加计时。 新增 inc/PerfCounterLite.h：header-only 的轻量计数器。 AecRxBase.arx 不链接 CoreLibrary（静态库只被 CoreCommand 链接），拿不到 SectionPerf 的符号，所以单独做一份；只在 Debug + AEC_PERF_COUNTER=1 时生效，Release 零开销。 统计写 %TEMP%\aec_prim_perf.txt（独立文件，避免提前写出 aec_section_perf.txt 让自动化脚本误判"开图完成"）。 接入 GiDrawUtility：DrawPipe 三个重载、DrawLoft 两个重载、DrawCone、DrawCylinder。 实测（big1.dwg，93 s 开图 / 81 s CPU）： DrawPipe(region) 7201 次 13.2 s 平均 1837us 最长 130ms DrawLoft(reg) 883 次 5.27 s 平均 5969us DrawLoft(curve) 883 次 5.90 s 平均 6679us DrawCone 811 次 4.64 s 平均 5717us DrawPipe(pts)/Cylinder 2411 次 4.07 s 平均 1688us 去重后约 27 s，占 81 s CPU 的 33%。 结论（详见 perf-primitive-cost.md）： - 7201 次管道构建 vs 约 2260 个管类实体 = 3.2 次/实体，护栏/爬梯/设备细节也在造管道 - 单次成本差一个量级：圆柱 1.7ms 最便宜（已是优化过的快路径），圆锥 5.7ms、放样 6.0~6.7ms 最贵 - 长尾明显：DrawPipe 最长 130ms（平均的 70 倍），存在个别大件 - 都在"首次构建"路径上，没有重复重建，所以只能少建 / 建得更便宜 / 跨实体复用同参数几何 验证：R230 Debug 编译通过，tools/run_acad_perf.ps1 自动开图取数； 截面缓存与回写仍为正常/0（未受本次插桩影响）。
This commit is contained in:
company-desktop committed
1 parent
dd9b49e5a4
commit
47d7b96cad
3 files changed
+219
No files matched your search
@@ -1,5 +1,6 @@
|
||||
#include "StdAfx.h"
|
||||
#include "GiDrawutility.h"
|
||||
#include "PerfCounterLite.h"
|
||||
#include "EarThwormDef.h"
|
||||
|
||||
namespace
|
||||
@@ -262,6 +263,7 @@ AECDLLIMPEXP AcDbEntity* GiDrawUtility::DrawLoft(const AcDbRegion *pTop
|
||||
, const AcDbRegion *pBottom
|
||||
, const AcDbCurve *pLine)
|
||||
{
|
||||
perf_lite::Scope perfScope("GDU::DrawLoft(reg)");
|
||||
AcDbRegion *_pTop = (AcDbRegion*)pTop->clone();
|
||||
AcDbRegion *_pBottom = (AcDbRegion*)pBottom->clone();
|
||||
|
||||
@@ -297,6 +299,7 @@ AECDLLIMPEXP AcDbEntity* GiDrawUtility::DrawLoft(const AcDbCurve *pTop
|
||||
, const AcDbCurve *pBottom
|
||||
, const AcDbCurve *pLine)
|
||||
{
|
||||
perf_lite::Scope perfScope("GDU::DrawLoft(curve)");
|
||||
if (!pTop || !pBottom || !pLine) return NULL;
|
||||
|
||||
AcDbCurve *_pTop = (AcDbCurve*)pTop->clone();
|
||||
@@ -385,6 +388,7 @@ AECDLLIMPEXP AcDbEntity* GiDrawUtility::DrawCone(const AcGePoint3d &ptPos
|
||||
, const double &H
|
||||
, const double &xOffset)
|
||||
{
|
||||
perf_lite::Scope perfScope("GDU::DrawCone");
|
||||
AcGePoint3d ptBottom(AcGePoint3d::kOrigin);
|
||||
AcGePoint3d ptTop(ptBottom + xOffset * AcGeVector3d::kXAxis);
|
||||
ptTop.z = H;
|
||||
@@ -684,6 +688,7 @@ AECDLLIMPEXP AcDbEntity* GiDrawUtility::DrawCylinder(const AcGePoint3d &ptFr
|
||||
, const AcGePoint3d &ptTo
|
||||
, const double &dRadius)
|
||||
{
|
||||
perf_lite::Scope perfScope("GDU::DrawCylinder");
|
||||
return DrawPipe(ptFr, ptTo, dRadius, 0.);
|
||||
}
|
||||
|
||||
@@ -782,6 +787,7 @@ AcDbEntity* GiDrawUtility::DrawPipe(const AcGePoint3d& ptFrom
|
||||
, const double &dThick
|
||||
, const int &nSegs /*= 20*/)
|
||||
{
|
||||
perf_lite::Scope perfScope("GDU::DrawPipe(pts)");
|
||||
// 零长度路径无法扫掠
|
||||
if (EQUAL_ZERO(ptFrom.distanceTo(ptTo))) return NULL;
|
||||
|
||||
@@ -844,6 +850,7 @@ AcDbEntity* GiDrawUtility::DrawPipe(const AcDbCurve* pCurve
|
||||
|
||||
AcDbEntity* GiDrawUtility::DrawPipe(const AcDbCurve *pCurve, double dR)
|
||||
{
|
||||
perf_lite::Scope perfScope("GDU::DrawPipe(curve,dR)");
|
||||
if (!pCurve) return NULL;
|
||||
if (DOWNEQUAL_ZERO(dR)) return NULL;
|
||||
|
||||
@@ -940,6 +947,7 @@ AcDbEntity *GiDrawUtility::DrawPipe(
|
||||
, const AcDbRegion *pRegion
|
||||
, const AcGePoint3d &ptCenInRgn)
|
||||
{
|
||||
perf_lite::Scope perfScope("GDU::DrawPipe(region)");
|
||||
if (!pRegion || !pCurve) return NULL;
|
||||
|
||||
AcGePoint3d ptStart(AcGePoint3d::kOrigin);
|
||||
|
||||
@@ -0,0 +1,147 @@
|
||||
// PerfCounterLite.h —— 给 AecRxBase 这类"不链接 CoreLibrary"的模块用的轻量计数器
|
||||
//
|
||||
// 为什么不复用 SectionPerf：CoreLibrary 是静态库且只被 CoreCommand 链接，
|
||||
// AecRxBase.arx 里拿不到它的符号；所以这里做成 header-only。
|
||||
//
|
||||
// 行为：
|
||||
// - 仅在 Debug（NDEBUG 未定义）且 AEC_PERF_COUNTER=1 时开启，Release 零开销
|
||||
// - 每个 tag 统计：调用次数、累计耗时、平均/最长
|
||||
// - 每 200 次调用重写一次 %TEMP%\aec_prim_perf.txt（快照）
|
||||
// —— 故意不写 aec_section_perf.txt：那个文件是自动化脚本判断"开图完成"的信号，
|
||||
// 提前写会让脚本误判并提前关闭 CAD
|
||||
|
||||
#ifndef PERF_COUNTER_LITE_H
|
||||
#define PERF_COUNTER_LITE_H
|
||||
|
||||
#include <windows.h>
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include <string.h>
|
||||
#include <map>
|
||||
#include <string>
|
||||
|
||||
namespace perf_lite
|
||||
{
|
||||
struct Entry
|
||||
{
|
||||
Entry() : n(0), us(0), maxus(0) {}
|
||||
long long n;
|
||||
long long us;
|
||||
long long maxus;
|
||||
};
|
||||
|
||||
inline bool Enabled()
|
||||
{
|
||||
#ifndef NDEBUG
|
||||
static int s_nOn = -1;
|
||||
if (s_nOn < 0)
|
||||
{
|
||||
const char* psz = ::getenv("AEC_PERF_COUNTER");
|
||||
s_nOn = (psz != NULL && psz[0] != '\0' && psz[0] != '0') ? 1 : 0;
|
||||
}
|
||||
return s_nOn == 1;
|
||||
#else
|
||||
return false;
|
||||
#endif
|
||||
}
|
||||
|
||||
inline std::map<std::string, Entry>& Stats()
|
||||
{
|
||||
static std::map<std::string, Entry> s_map;
|
||||
return s_map;
|
||||
}
|
||||
|
||||
inline long long QpcNow()
|
||||
{
|
||||
LARGE_INTEGER li;
|
||||
::QueryPerformanceCounter(&li);
|
||||
return (long long)li.QuadPart;
|
||||
}
|
||||
|
||||
inline long long QpcFreq()
|
||||
{
|
||||
static long long s_nFreq = 0;
|
||||
if (s_nFreq == 0)
|
||||
{
|
||||
LARGE_INTEGER li;
|
||||
if (!::QueryPerformanceFrequency(&li) || li.QuadPart == 0)
|
||||
s_nFreq = 1;
|
||||
else
|
||||
s_nFreq = (long long)li.QuadPart;
|
||||
}
|
||||
return s_nFreq;
|
||||
}
|
||||
|
||||
inline void Dump()
|
||||
{
|
||||
char szTmp[MAX_PATH + 1];
|
||||
::memset(szTmp, 0, sizeof(szTmp));
|
||||
if (::GetTempPathA(MAX_PATH, szTmp) == 0)
|
||||
return;
|
||||
std::string sPath(szTmp);
|
||||
sPath += "aec_prim_perf.txt";
|
||||
|
||||
FILE* fp = NULL;
|
||||
if (fopen_s(&fp, sPath.c_str(), "w") != 0 || fp == NULL)
|
||||
return;
|
||||
|
||||
fprintf(fp, "---- primitive builders (per tag) ----\n");
|
||||
fprintf(fp, " count total_ms avg_us max_us tag\n");
|
||||
const std::map<std::string, Entry>& m = Stats();
|
||||
for (std::map<std::string, Entry>::const_iterator it = m.begin(); it != m.end(); ++it)
|
||||
{
|
||||
const Entry& e = it->second;
|
||||
const double dAvg = e.n ? ((double)e.us / (double)e.n) : 0.0;
|
||||
fprintf(fp, " %-9lld %10.1f %10.1f %10lld %s\n",
|
||||
e.n, (double)e.us / 1000.0, dAvg, e.maxus, it->first.c_str());
|
||||
}
|
||||
fclose(fp);
|
||||
}
|
||||
|
||||
inline void Add(const char* pszTag, long long nStart)
|
||||
{
|
||||
const long long nUs = (QpcNow() - nStart) * 1000000LL / QpcFreq();
|
||||
Entry& e = Stats()[std::string(pszTag != NULL ? pszTag : "?")];
|
||||
++e.n;
|
||||
e.us += nUs;
|
||||
if (nUs > e.maxus)
|
||||
e.maxus = nUs;
|
||||
|
||||
// 每 200 次落一次快照，开图结束后文件里就是全量统计
|
||||
static long long s_nTotal = 0;
|
||||
if (++s_nTotal % 200 == 0)
|
||||
Dump();
|
||||
}
|
||||
|
||||
// RAII：函数入口构造即可
|
||||
class Scope
|
||||
{
|
||||
public:
|
||||
explicit Scope(const char* pszTag) : m_pszTag(pszTag), m_nStart(0)
|
||||
{
|
||||
if (Enabled())
|
||||
m_nStart = QpcNow();
|
||||
}
|
||||
~Scope()
|
||||
{
|
||||
if (m_nStart != 0)
|
||||
Add(m_pszTag, m_nStart);
|
||||
}
|
||||
|
||||
private:
|
||||
Scope(const Scope&);
|
||||
Scope& operator=(const Scope&);
|
||||
|
||||
const char* m_pszTag;
|
||||
long long m_nStart;
|
||||
};
|
||||
|
||||
// 在卸载时再补一次，确保最后的零头也落盘
|
||||
inline void Flush()
|
||||
{
|
||||
if (Enabled())
|
||||
Dump();
|
||||
}
|
||||
}
|
||||
|
||||
#endif // PERF_COUNTER_LITE_H
|
||||
Reference in new issue
Block a user