源代码，是你的
【深入探讨用位掩码代替分支（4）：VC2010速度测试 】
http://blog.csdn.net/zyl910/article/details/7403497
没有任何改动源代码
Intel parallel studio 2011 xe + 最新的 update ...
编译器选项： （默认的选项，然后，在 属性里面的优化选项，
+ 使用了qxhost---使用本机最高支持的sse的版本，
其它的qvec等，一大堆选项，还没试。。。。
/Zi /nologo /W3 /O2 /Oi /Ot /Qipo /D "WIN32" /D "NDEBUG" /D "_CONSOLE" /D "_UNICODE" /D "UNICODE" /EHsc /GS /Gy /fp:precise /QxHost /Zc:wchar_t /Zc:forScope /Fp"x64\Release\noifVC2010.pch" /FAs /Fa"x64\Release\" /Fo"x64\Release\" /Fd"x64\Release\vc100.pdb"
== noif:VC2010(64) on 64bit ==<Press any key to continue>
f0_if[1]: 906
f0_if[2]: 872
f0_if[3]: 879
f1_min[1]: 274
f1_min[2]: 274
f1_min[3]: 271
f2_neg[1]: 437
f2_neg[2]: 441
f2_neg[3]: 442
f3_sar[1]: 410
f3_sar[2]: 367
f3_sar[3]: 366
<Press any key to exit>
下面的是，VC 2010 的默认设置。
/Zi /nologo /W3 /WX- /O2 /Oi /Ot /GL /D "WIN32" /D "NDEBUG" /D "_CONSOLE" /D "_UNICODE" /D "UNICODE" /Gm- /EHsc /GS /Gy /fp:precise /Zc:wchar_t /Zc:forScope /Fp"x64\Release\noifVC2010.pch" /FAs /Fa"x64\Release\" /Fo"x64\Release\" /Fd"x64\Release\vc100.pdb" /Gd /errorReport:queue
== noif:VC2010(64) on 64bit ==<Press any key to continue>
f0_if[1]: 1443
f0_if[2]: 1440
f0_if[3]: 1428
f1_min[1]: 1553
f1_min[2]: 1575
f1_min[3]: 1554
f2_neg[1]: 383
f2_neg[2]: 380
f2_neg[3]: 385
f3_sar[1]: 246
f3_sar[2]: 245
f3_sar[3]: 246
<Press any key to exit>
/Zi /nologo /W3 /WX- /Ox /Oi /Ot /GL /D "WIN32"
/D "NDEBUG" /D "_CONSOLE" /D "_UNICODE" /D "UNICODE" /Gm- /EHsc
/GS /Gy /arch:SSE2 /fp:precise /Zc:wchar_t /Zc:forScope
/Fp"x64\Release\noifVC2010.pch" /FAs /Fa"x64\Release\"
/Fo"x64\Release\" /Fd"x64\Release\vc100.pdb" /Gd /errorReport:queue
== noif:VC2010(64) on 64bit ==<Press any key to continue>
f0_if[1]: 1460
f0_if[2]: 1453
f0_if[3]: 1449
f1_min[1]: 1572
f1_min[2]: 1573
f1_min[3]: 1566
f2_neg[1]: 388
f2_neg[2]: 386
f2_neg[3]: 384
f3_sar[1]: 248
f3_sar[2]: 248
f3_sar[3]: 248
<Press any key to exit>
而intel的
f0_if[1]: 906
f0_if[2]: 872
f0_if[3]: 879
f1_min[1]: 274
f1_min[2]: 274
f1_min[3]: 271
f2_neg[1]: 437
f2_neg[2]: 441
f2_neg[3]: 442
f3_sar[1]: 410
f3_sar[2]: 367
f3_sar[3]: 366
////////////////////////////////////////////////////////////
// noifVC2010.cpp : VC2010饱和处理速度测试
// Author: zyl910
// Blog: http://www.cnblogs.com/zyl910
// URL: http://www.cnblogs.com/zyl910/archive/2012/03/28/noifopex4.html
// Version: V1.1
// Updata: 2012-05-03
//
// [2012-05-03] V1.1
// 增加SIMD指令集测试：f4_mmx、f5_sse。
//
// [2012-03-28] V1.0
// 发布1.0版。
//
////////////////////////////////////////////////////////////
#include "stdafx.h"
#include <Windows.h>
#include <stdio.h>
#include <time.h>
#include <conio.h>
// MMX, SSE, SSE2
#include <emmintrin.h>
// 用位掩码做饱和处理.用求负生成掩码
#define LIMITSU_FAST(n, bits) ( (n) & -((n) >= 0) | -((n) >= (1<<(bits))) )
#define LIMITSU_SAFE(n, bits) ( (LIMITSU_FAST(n, bits)) & ((1<<(bits)) - 1) )
#define LIMITSU_BYTE(n) ((BYTE)(LIMITSU_FAST(n, 8)))
// 用位掩码做饱和处理.用带符号右移生成掩码
#define LIMITSW_FAST(n, bits) ( ( (n) | ((signed short)((1<<(bits)) - 1 - (n)) >> 15) ) & ~((signed short)(n) >> 15) )
#define LIMITSW_SAFE(n, bits) ( (LIMITSW_FAST(n, bits)) & ((1<<(bits)) - 1) )
#define LIMITSW_BYTE(n) ((BYTE)(LIMITSW_FAST(n, 8)))
// 数据规模
#define DATASIZE 16384 // 128KB / (sizeof(signed short) * 4)
// 缓冲区。SSE需要按128位对齐
__declspec(align(16)) signed short bufS[DATASIZE*4]; // 源缓冲区。64位的颜色（4通道，每通道16位）
__declspec(align(16)) BYTE bufD[DATASIZE*4]; // 目标缓冲区。32位的颜色（4通道，每通道8位）
// 测试时的函数类型
typedef void (*TESTPROC)(BYTE* pbufD, const signed short* pbufS, int cnt);
// 用if分支做饱和处理
void f0_if(BYTE* pbufD, const signed short* pbufS, int cnt)
{
const signed short* pS = pbufS;
BYTE* pD = pbufD;
int i;
for(i=0; i<cnt; ++i)
{
// 分别对4个通道做饱和处理
pD[0] = (pS[0]<0) ? 0 : ( (pS[0]>255) ? 255 : (BYTE)pS[0] );
pD[1] = (pS[1]<0) ? 0 : ( (pS[1]>255) ? 255 : (BYTE)pS[1] );
pD[2] = (pS[2]<0) ? 0 : ( (pS[2]>255) ? 255 : (BYTE)pS[2] );
pD[3] = (pS[3]<0) ? 0 : ( (pS[3]>255) ? 255 : (BYTE)pS[3] );
// next
pS += 4;
pD += 4;
}
}
// 用min、max饱和处理
void f1_min(BYTE* pbufD, const signed short* pbufS, int cnt)
{
const signed short* pS = pbufS;
BYTE* pD = pbufD;
int i;
for(i=0; i<cnt; ++i)
{
// 分别对4个通道做饱和处理
pD[0] = min(max(0, pS[0]), 255);
pD[1] = min(max(0, pS[1]), 255);
pD[2] = min(max(0, pS[2]), 255);
pD[3] = min(max(0, pS[3]), 255);
// next
pS += 4;
pD += 4;
}
}
// 用位掩码做饱和处理.用求负生成掩码
void f2_neg(BYTE* pbufD, const signed short* pbufS, int cnt)
{
const signed short* pS = pbufS;
BYTE* pD = pbufD;
int i;
for(i=0; i<cnt; ++i)
{
// 分别对4个通道做饱和处理
pD[0] = LIMITSU_BYTE(pS[0]);
pD[1] = LIMITSU_BYTE(pS[1]);
pD[2] = LIMITSU_BYTE(pS[2]);
pD[3] = LIMITSU_BYTE(pS[3]);
// next
pS += 4;
pD += 4;
}
}
// 用位掩码做饱和处理.用带符号右移生成掩码
void f3_sar(BYTE* pbufD, const signed short* pbufS, int cnt)
{
const signed short* pS = pbufS;
BYTE* pD = pbufD;
int i;
for(i=0; i<cnt; ++i)
{
// 分别对4个通道做饱和处理
pD[0] = LIMITSW_BYTE(pS[0]);
pD[1] = LIMITSW_BYTE(pS[1]);
pD[2] = LIMITSW_BYTE(pS[2]);
pD[3] = LIMITSW_BYTE(pS[3]);
// next
pS += 4;
pD += 4;
}
}
#if defined(_WIN64)
// 64位下不支持MMX指令集
#else
// 饱和处理MMX版
void f4_mmx(BYTE* pbufD, const signed short* pbufS, int cnt)
{
//const signed short* pS = pbufS;
//BYTE* pD = pbufD;
const __m64* pS = (const __m64*)pbufS;
__m64* pD = (__m64*)pbufD;
int i;
for(i=0; i<cnt; i+=2)
{
// 同时对两个像素做饱和处理。即 将两个64位像素（4通道，每分量为带符号16位） 转为 两个32位像素（每分量为无符号8位）。
pD[0] = _mm_packs_pu16(pS[0], pS[1]); // 饱和方式数据打包（带符号16位->无符号8位）。等价于 for(i=0;i<4;++i){ pD[0].uB[i]=SU(pS[0].iW[i]); pD[0].uB[4+i]=SU(pS[1].iW[i]); }
// next
pS += 2;
pD += 1;
}
// MMX状态置空
_mm_empty();
}
#endif
// 饱和处理SSE版
void f5_sse(BYTE* pbufD, const signed short* pbufS, int cnt)
{
//const signed short* pS = pbufS;
//BYTE* pD = pbufD;
const __m128i* pS = (const __m128i*)pbufS;
__m128i* pD = (__m128i*)pbufD;
int i;
for(i=0; i<cnt; i+=4)
{
// 同时对四个像素做饱和处理。即 将四个64位像素（4通道，每分量为带符号16位） 转为 四个32位像素（每分量为无符号8位）。
pD[0] = _mm_packus_epi16(pS[0], pS[1]); // 饱和方式数据打包（带符号16位->无符号8位）。等价于 for(i=0;i<8;++i){ pD[0].uB[i]=SU(pS[0].iW[i]); r.uB[8+i]=SU(pS[1].iW[i]); }
// next
pS += 2;
pD += 1;
}
}
// 进行测试
void runTest(_TCHAR* szname, TESTPROC proc)
{
int i,j;
DWORD tm0, tm1; // 存储时间
for(i=1; i<=3; ++i) // 多次测试
{
//tm0 = GetTickCount();
tm0 = timeGetTime();
// main
for(j=1; j<=4000; ++j) // 重复运算几次延长时间，避免计时精度问题
{
proc(bufD, bufS, DATASIZE);
}
// show
//tm1 = GetTickCount() - tm0;
tm1 = timeGetTime() - tm0;
_tprintf(_T("%s[%d]:\t%u\n"), szname, i, tm1);
}
}
// 进行测试（测试SIMD指令集专用）
void runTest2(_TCHAR* szname, TESTPROC proc)
{
const int nLoop = 16; // 使用MMX/SSE指令时速度太快了，只好再多循环几次
int i,j,k;
DWORD tm0, tm1; // 存储时间
for(i=1; i<=3; ++i) // 多次测试
{
//tm0 = GetTickCount();
tm0 = timeGetTime();
// main
for(k=1; k<=nLoop; ++k)
{
for(j=1; j<=4000; ++j) // 重复运算几次延长时间，避免计时精度问题
{
proc(bufD, bufS, DATASIZE);
}
}
// show
//tm1 = GetTickCount() - tm0;
tm1 = timeGetTime() - tm0;
_tprintf(_T("%s[%d]:\t%.1f\n"), szname, i, (double)tm1/nLoop);
}
}
// 获取程序位数（被编译为多少位的代码）
int GetProgramBits()
{
return sizeof(int*) * 8;
}
// 安全的取得真实系统信息
VOID SafeGetNativeSystemInfo(__out LPSYSTEM_INFO lpSystemInfo)
{
if (NULL==lpSystemInfo) return;
typedef VOID (WINAPI *LPFN_GetNativeSystemInfo)(LPSYSTEM_INFO lpSystemInfo);
LPFN_GetNativeSystemInfo fnGetNativeSystemInfo = (LPFN_GetNativeSystemInfo)GetProcAddress( GetModuleHandle(_T("kernel32")), "GetNativeSystemInfo");;
if (NULL != fnGetNativeSystemInfo)
{
fnGetNativeSystemInfo(lpSystemInfo);
}
else
{
GetSystemInfo(lpSystemInfo);
}
}
// 获取操作系统位数
int GetSystemBits()
{
SYSTEM_INFO si;
SafeGetNativeSystemInfo(&si);
if (si.wProcessorArchitecture == PROCESSOR_ARCHITECTURE_AMD64 ||
si.wProcessorArchitecture == PROCESSOR_ARCHITECTURE_IA64 )
{
return 64;
}
return 32;
}
int _tmain(int argc, _TCHAR* argv[])
{
int i; // 循环变量
//printf("Hello World!\n");
const int nBitCode = GetProgramBits();
const int nBitSys = GetSystemBits();
_tprintf(_T("== noif:VC2010(%d) on %dbit =="), nBitCode, nBitSys);
// 初始化
srand( (unsigned)time( NULL ) );
for(i=0; i<DATASIZE*4; ++i)
{
bufS[i] = (signed short)((rand()&0x1FF) - 128); // 使数值在 [-128, 383] 区间
}
// 准备开始。可以将将进程优先级设为实时
if (argc<=1)
{
_tprintf(_T("<Press any key to continue>"));
_getch();
}
_tprintf(_T("\n"));
// 进行测试
runTest(_T("f0_if"), f0_if);
runTest(_T("f1_min"), f1_min);
runTest(_T("f2_neg"), f2_neg);
runTest(_T("f3_sar"), f3_sar);
#if defined(_WIN64)
// 64位下不支持MMX指令集
#else
runTest2(_T("f4_mmx"), f4_mmx);
#endif
runTest2(_T("f5_sse"), f5_sse);
// 结束前提示
if (argc<=1)
{
_tprintf(_T("<Press any key to exit>"));
_getch();
}
_tprintf(_T("\n"));
return 0;
}
== noif:VC2010(32) on 64bit ==<Press any key to continue>
f0_if[1]: 1716
f0_if[2]: 1701
f0_if[3]: 1685
f1_min[1]: 2074
f1_min[2]: 2091
f1_min[3]: 2090
f2_neg[1]: 515
f2_neg[2]: 499
f2_neg[3]: 515
f3_sar[1]: 437
f3_sar[2]: 437
f3_sar[3]: 436
f4_mmx[1]: 36.1
f4_mmx[2]: 36.1
f4_mmx[3]: 36.1
f5_sse[1]: 25.4
f5_sse[2]: 25.3
f5_sse[3]: 25.4
== noif:VC2010(64) on 64bit ==<Press any key to continue>
f0_if[1]: 1622
f0_if[2]: 1623
f0_if[3]: 1606
f1_min[1]: 1997
f1_min[2]: 1981
f1_min[3]: 1997
f2_neg[1]: 421
f2_neg[2]: 437
f2_neg[3]: 421
f3_sar[1]: 328
f3_sar[2]: 312
f3_sar[3]: 328
f5_sse[1]: 25.3
f5_sse[2]: 25.4
f5_sse[3]: 25.3
_TEXT SEGMENT 'CODE'
; COMDAT ?f1_min@@YAXPEAEPEBFH@Z
TXTST2:
; -- Begin ?f1_min@@YAXPEAEPEBFH@Z
; mark_begin;
ALIGN 16
PUBLIC ?f1_min@@YAXPEAEPEBFH@Z
?f1_min@@YAXPEAEPEBFH@Z PROC
; parameter 1(pbufD): rcx
; parameter 2(pbufS): rdx
; parameter 3(cnt): r8d
.B3.1:: ; Preds .B3.0
;;; {
$LN501:
movsxd r10, r8d ;60.1
$LN502:
;;; const signed short* pS = pbufS;
;;; BYTE* pD = pbufD;
;;; int i;
;;; for(i=0; i<cnt; ++i)
xor r8d, r8d ;64.2
$LN503:
test r10, r10 ;64.2
$LN504:
jle .B3.5 ; Prob 10% ;64.2
$LN505:
; LOE rdx rcx rbx rbp rsi rdi r8 r10 r12 r13 r14 r15 xmm6 xmm7 xmm8 xmm9 xmm10 xmm11 xmm12 xmm13 xmm14 xmm15
.B3.2:: ; Preds .B3.1
$LN506:
;;; {
;;; // 分别对4个通道做饱和处理
;;; pD[0] = min(max(0, pS[0]), 255);
xor r9d, r9d ;67.3
$LN507:
mov eax, 255 ;67.3
ALIGN 16
$LN508:
; LOE rdx rcx rbx rbp rsi rdi r8 r10 r12 r13 r14 r15 eax r9d xmm6 xmm7 xmm8 xmm9 xmm10 xmm11 xmm12 xmm13 xmm14 xmm15
.B3.3:: ; Preds .B3.3 .B3.2
$LN509:
movsx r11d, WORD PTR [rdx+r8*8] ;67.3
$LN510:
test r11d, r11d ;67.3
$LN511:
cmovl r11d, r9d ;67.3
$LN512:
cmp r11d, 255 ;67.3
$LN513:
cmovge r11d, eax ;67.3
$LN514:
mov BYTE PTR [rcx+r8*4], r11b ;67.3
$LN515:
movsx r11d, WORD PTR [2+rdx+r8*8] ;67.3
$LN516:
test r11d, r11d ;67.3
$LN517:
cmovl r11d, r9d ;67.3
$LN518:
cmp r11d, 255 ;67.3
$LN519:
cmovge r11d, eax ;67.3
$LN520:
mov BYTE PTR [1+rcx+r8*4], r11b ;67.3
$LN521:
movsx r11d, WORD PTR [4+rdx+r8*8] ;67.3
$LN522:
test r11d, r11d ;67.3
$LN523:
cmovl r11d, r9d ;67.3
$LN524:
cmp r11d, 255 ;67.3
$LN525:
cmovge r11d, eax ;67.3
$LN526:
mov BYTE PTR [2+rcx+r8*4], r11b ;67.3
$LN527:
movsx r11d, WORD PTR [6+rdx+r8*8] ;67.3
$LN528:
test r11d, r11d ;67.3
$LN529:
cmovl r11d, r9d ;67.3
$LN530:
cmp r11d, 255 ;67.3
$LN531:
cmovge r11d, eax ;67.3
$LN532:
mov BYTE PTR [3+rcx+r8*4], r11b ;67.3
$LN533:
inc r8 ;64.2
$LN534:
cmp r8, r10 ;64.2
$LN535:
jb .B3.3 ; Prob 78% ;64.2
$LN536:
; LOE rdx rcx rbx rbp rsi rdi r8 r10 r12 r13 r14 r15 eax r9d xmm6 xmm7 xmm8 xmm9 xmm10 xmm11 xmm12 xmm13 xmm14 xmm15
.B3.5:: ; Preds .B3.3 .B3.1
;;; pD[1] = min(max(0, pS[1]), 255);
;;; pD[2] = min(max(0, pS[2]), 255);
;;; pD[3] = min(max(0, pS[3]), 255);
;;; // next
;;; pS += 4;
;;; pD += 4;
;;; }
;;; }
$LN537:
ret ;75.1
ALIGN 16
$LN538:
; LOE
.B3.6::
$LN539:
; mark_end;
?f1_min@@YAXPEAEPEBFH@Z ENDP
$LN?f1_min@@YAXPEAEPEBFH@Z$540:
$LN?f1_min@@YAXPEAEPEBFH@Z$541:
;?f1_min@@YAXPEAEPEBFH@Z ENDS
PUBLIC ?f1_min@@YAXPEAEPEBFH@Z ; f1_min
; Function compile flags: /Ogtpy
; COMDAT ?f1_min@@YAXPEAEPEBFH@Z
_TEXT SEGMENT
pbufD$ = 8
pbufS$ = 16
cnt$ = 24
?f1_min@@YAXPEAEPEBFH@Z PROC ; f1_min, COMDAT
; 62 : BYTE* pD = pbufD;
; 63 : int i;
; 64 : for(i=0; i<cnt; ++i)
test r8d, r8d
jle $LN1@f1_min
; 61 : const signed short* pS = pbufS;
add rdx, 4
mov r10d, r8d
lea r9, QWORD PTR [rcx+2]
xor r8d, r8d
npad 9
$LL3@f1_min:
; 65 : {
; 66 : // mmMak分别对4个通道做饱和处理
; 67 : pD[0] = min(max(0, pS[0]), 255);
movzx ecx, WORD PTR [rdx-4]
cmp r8w, cx
jg SHORT $LN37@f1_min
cmp cx, 255 ; 000000ffH
jl SHORT $LN32@f1_min
mov eax, 255 ; 000000ffH
jmp SHORT $LN11@f1_min
$LN32@f1_min:
cmp r8w, cx
jle SHORT $LN8@f1_min
$LN37@f1_min:
mov eax, r8d
jmp SHORT $LN11@f1_min
$LN8@f1_min:
movsx eax, cx
$LN11@f1_min:
mov BYTE PTR [r9-2], al
; 68 : pD[1] = min(max(0, pS[1]), 255);
movzx ecx, WORD PTR [rdx-2]
cmp r8w, cx
jg SHORT $LN39@f1_min
cmp cx, 255 ; 000000ffH
jl SHORT $LN33@f1_min
mov eax, 255 ; 000000ffH
jmp SHORT $LN17@f1_min
$LN33@f1_min:
cmp r8w, cx
jle SHORT $LN14@f1_min
$LN39@f1_min:
mov eax, r8d
jmp SHORT $LN17@f1_min
$LN14@f1_min:
movsx eax, cx
$LN17@f1_min:
mov BYTE PTR [r9-1], al
; 69 : pD[2] = min(max(0, pS[2]), 255);
movzx ecx, WORD PTR [rdx]
cmp r8w, cx
jg SHORT $LN41@f1_min
cmp cx, 255 ; 000000ffH
jl SHORT $LN34@f1_min
mov eax, 255 ; 000000ffH
jmp SHORT $LN23@f1_min
$LN34@f1_min:
cmp r8w, cx
jle SHORT $LN20@f1_min
$LN41@f1_min:
mov eax, r8d
jmp SHORT $LN23@f1_min
$LN20@f1_min:
movsx eax, cx
$LN23@f1_min:
mov BYTE PTR [r9], al
; 70 : pD[3] = min(max(0, pS[3]), 255);
movzx ecx, WORD PTR [rdx+2]
cmp r8w, cx
jg SHORT $LN43@f1_min
cmp cx, 255 ; 000000ffH
jl SHORT $LN35@f1_min
mov eax, 255 ; 000000ffH
jmp SHORT $LN29@f1_min
$LN35@f1_min:
cmp r8w, cx
jle SHORT $LN26@f1_min
$LN43@f1_min:
mov eax, r8d
jmp SHORT $LN29@f1_min
$LN26@f1_min:
movsx eax, cx
$LN29@f1_min:
mov BYTE PTR [r9+1], al
; 71 : // next
; 72 : pS += 4;
; 73 : pD += 4;
add r9, 4
add rdx, 8
dec r10
jne $LL3@f1_min
$LN1@f1_min:
; 74 : }
; 75 : }
fatret 0
?f1_min@@YAXPEAEPEBFH@Z ENDP ; f1_min
_TEXT ENDS
if ((t1|t2|t3)==0)
{
t4=1;
}
//汇编 t1 - edi, t2-ebx, t3-ebp,t4-eax
or edi,ebx
or edi,ebp
jne L1
mov eax,1
L1:
这个代码，分支预测率 92%...
如果使用 cmov
mov ecx,1
or edi,ebx
or edi,ebp
cmove eax,ecx
性能从 6.2s 提升到 4.9s
----------------------
visual studio 2010
== noif:VC2010(32) on 64bit ==<Press any key to continue>
f0_if[1]: 1526
f0_if[2]: 1531
f0_if[3]: 1503
f1_min[1]: 2190
f1_min[2]: 2221
f1_min[3]: 2205
f2_neg[1]: 421
f2_neg[2]: 423
f2_neg[3]: 414
f3_sar[1]: 321
f3_sar[2]: 327
f3_sar[3]: 329
<Press any key to exit>
-----------------------
intel parallel studio 2011 xe
== noif:VC2010(32) on 64bit ==<Press any key to continue>
f0_if[1]: 900
f0_if[2]: 876
f0_if[3]: 874
f1_min[1]: 271
f1_min[2]: 270
f1_min[3]: 267
f2_neg[1]: 462
f2_neg[2]: 453
f2_neg[3]: 458
f3_sar[1]: 379
f3_sar[2]: 373
f3_sar[3]: 380
<Press any key to exit>
VC 2010
PUBLIC ?f1_min@@YAXPAEPBFH@Z ; f1_min
; Function compile flags: /Ogtp
; COMDAT ?f1_min@@YAXPAEPBFH@Z
_TEXT SEGMENT
_pbufD$ = 8 ; size = 4
_pbufS$ = 12 ; size = 4
_cnt$ = 16 ; size = 4
?f1_min@@YAXPAEPBFH@Z PROC ; f1_min, COMDAT
; 60 : {
push ebp
mov ebp, esp
; 62 : BYTE* pD = pbufD;
; 63 : int i;
; 64 : for(i=0; i<cnt; ++i)
mov eax, DWORD PTR _cnt$[ebp]
test eax, eax
jle $LN1@f1_min
; 61 : const signed short* pS = pbufS;
mov ecx, DWORD PTR _pbufD$[ebp]
push esi
mov esi, DWORD PTR _pbufS$[ebp]
push edi
add ecx, 2
add esi, 4
mov edi, eax
npad 2
$LL3@f1_min:
; 65 : {
; 66 : // 分别对4个通道做饱和处理
; 67 : pD[0] = min(max(0, pS[0]), 255);
movzx eax, WORD PTR [esi-4]
xor edx, edx
cmp dx, ax
jg SHORT $LN32@f1_min
cmp ax, 255 ; 000000ffH
jge SHORT $LN10@f1_min
$LN32@f1_min:
xor edx, edx
cmp dx, ax
jle SHORT $LN8@f1_min
xor eax, eax
jmp SHORT $LN11@f1_min
$LN8@f1_min:
cwde
jmp SHORT $LN11@f1_min
$LN10@f1_min:
mov eax, 255 ; 000000ffH
$LN11@f1_min:
mov BYTE PTR [ecx-2], al
; 68 : pD[1] = min(max(0, pS[1]), 255);
movzx eax, WORD PTR [esi-2]
xor edx, edx
cmp dx, ax
jg SHORT $LN33@f1_min
cmp ax, 255 ; 000000ffH
jge SHORT $LN16@f1_min
$LN33@f1_min:
xor edx, edx
cmp dx, ax
jle SHORT $LN14@f1_min
xor eax, eax
jmp SHORT $LN17@f1_min
$LN14@f1_min:
cwde
jmp SHORT $LN17@f1_min
$LN16@f1_min:
mov eax, 255 ; 000000ffH
$LN17@f1_min:
mov BYTE PTR [ecx-1], al
; 69 : pD[2] = min(max(0, pS[2]), 255);
movzx eax, WORD PTR [esi]
xor edx, edx
cmp dx, ax
jg SHORT $LN34@f1_min
cmp ax, 255 ; 000000ffH
jge SHORT $LN22@f1_min
$LN34@f1_min:
xor edx, edx
cmp dx, ax
jle SHORT $LN20@f1_min
xor eax, eax
jmp SHORT $LN23@f1_min
$LN20@f1_min:
cwde
jmp SHORT $LN23@f1_min
$LN22@f1_min:
mov eax, 255 ; 000000ffH
$LN23@f1_min:
mov BYTE PTR [ecx], al
; 70 : pD[3] = min(max(0, pS[3]), 255);
movzx eax, WORD PTR [esi+2]
xor edx, edx
cmp dx, ax
jg SHORT $LN35@f1_min
cmp ax, 255 ; 000000ffH
jge SHORT $LN28@f1_min
$LN35@f1_min:
xor edx, edx
cmp dx, ax
jle SHORT $LN26@f1_min
xor eax, eax
jmp SHORT $LN29@f1_min
$LN26@f1_min:
cwde
jmp SHORT $LN29@f1_min
$LN28@f1_min:
mov eax, 255 ; 000000ffH
$LN29@f1_min:
mov BYTE PTR [ecx+1], al
; 71 : // next
; 72 : pS += 4;
add esi, 8
; 73 : pD += 4;
add ecx, 4
dec edi
jne $LL3@f1_min
pop edi
pop esi
$LN1@f1_min:
; 74 : }
; 75 : }
pop ebp
ret 0
?f1_min@@YAXPAEPBFH@Z ENDP ; f1_min
_TEXT ENDS
_TEXT SEGMENT PARA PUBLIC FLAT 'CODE'
; COMDAT ?f1_min@@YAXPAEPBFH@Z
TXTST2:
; -- Begin ?f1_min@@YAXPAEPBFH@Z
; mark_begin;
ALIGN 16
PUBLIC ?f1_min@@YAXPAEPBFH@Z
?f1_min@@YAXPAEPBFH@Z PROC NEAR
; parameter 1(pbufD): 20 + esp
; parameter 2(pbufS): 24 + esp
; parameter 3(cnt): 28 + esp
.B3.1: ; Preds .B3.0
;;; {
$LN501:
sub esp, 16 ;60.1
$LN502:
mov edx, DWORD PTR [28+esp] ;59.6
$LN503:
;;; const signed short* pS = pbufS;
;;; BYTE* pD = pbufD;
;;; int i;
;;; for(i=0; i<cnt; ++i)
test edx, edx ;64.2
$LN504:
jle .B3.5 ; Prob 10% ;64.2
$LN505:
; LOE edx ebx ebp esi edi
.B3.2: ; Preds .B3.1
xor ecx, ecx ;
mov DWORD PTR [12+esp], esi ;
xor eax, eax ;
mov DWORD PTR [8+esp], edi ;
mov DWORD PTR [4+esp], ebx ;
mov DWORD PTR [esp], ebp ;
mov ebp, 255 ;
mov esi, DWORD PTR [24+esp] ;
mov edi, DWORD PTR [20+esp] ;
ALIGN 16
$LN506:
; LOE eax edx ecx ebp esi edi
.B3.3: ; Preds .B3.3 .B3.2
$LN507:
;;; {
;;; // 分别对4个通道做饱和处理
;;; pD[0] = min(max(0, pS[0]), 255);
movsx ebx, WORD PTR [esi+ecx*8] ;67.3
$LN508:
test ebx, ebx ;67.3
$LN509:
cmovl ebx, eax ;67.3
$LN510:
cmp ebx, 255 ;67.3
$LN511:
cmovge ebx, ebp ;67.3
$LN512:
mov BYTE PTR [edi+ecx*4], bl ;67.3
$LN513:
movsx ebx, WORD PTR [2+esi+ecx*8] ;67.3
$LN514:
test ebx, ebx ;67.3
$LN515:
cmovl ebx, eax ;67.3
$LN516:
cmp ebx, 255 ;67.3
$LN517:
cmovge ebx, ebp ;67.3
$LN518:
mov BYTE PTR [1+edi+ecx*4], bl ;67.3
$LN519:
movsx ebx, WORD PTR [4+esi+ecx*8] ;67.3
$LN520:
test ebx, ebx ;67.3
$LN521:
cmovl ebx, eax ;67.3
$LN522:
cmp ebx, 255 ;67.3
$LN523:
cmovge ebx, ebp ;67.3
$LN524:
mov BYTE PTR [2+edi+ecx*4], bl ;67.3
$LN525:
movsx ebx, WORD PTR [6+esi+ecx*8] ;67.3
$LN526:
test ebx, ebx ;67.3
$LN527:
cmovl ebx, eax ;67.3
$LN528:
cmp ebx, 255 ;67.3
$LN529:
cmovge ebx, ebp ;67.3
$LN530:
mov BYTE PTR [3+edi+ecx*4], bl ;67.3
$LN531:
inc ecx ;64.2
$LN532:
cmp ecx, edx ;64.2
$LN533:
jb .B3.3 ; Prob 78% ;64.2
$LN534:
; LOE eax edx ecx ebp esi edi
.B3.4: ; Preds .B3.3
mov esi, DWORD PTR [12+esp] ;
mov edi, DWORD PTR [8+esp] ;
mov ebx, DWORD PTR [4+esp] ;
mov ebp, DWORD PTR [esp] ;
$LN535:
; LOE ebx ebp esi edi
.B3.5: ; Preds .B3.4 .B3.1
;;; pD[1] = min(max(0, pS[1]), 255);
;;; pD[2] = min(max(0, pS[2]), 255);
;;; pD[3] = min(max(0, pS[3]), 255);
;;; // next
;;; pS += 4;
;;; pD += 4;
;;; }
;;; }
$LN536:
add esp, 16 ;75.1
$LN537:
ret ;75.1
ALIGN 16
$LN538:
; LOE
$LN539:
; mark_end;
?f1_min@@YAXPAEPBFH@Z ENDP
$LN?f1_min@@YAXPAEPBFH@Z$540:
$LN?f1_min@@YAXPAEPBFH@Z$541:
;?f1_min@@YAXPAEPBFH@Z ENDS
_TEXT ENDS
_DATA SEGMENT DWORD PUBLIC FLAT 'DATA'
_DATA ENDS
; -- End ?f1_min@@YAXPAEPBFH@Z
###########intel
== noif:VC2010(64) on 64bit ==<Press any key to continue>
f0_if[1]: 921
f0_if[2]: 903
f0_if[3]: 920
f1_min[1]: 277
f1_min[2]: 280
f1_min[3]: 290
f2_neg[1]: 457
f2_neg[2]: 447
f2_neg[3]: 446
f3_sar[1]: 367
f3_sar[2]: 388
f3_sar[3]: 378
f5_sse[1]: 25.5
f5_sse[2]: 25.2
f5_sse[3]: 25.4
<Press any key to exit>
##########VC 2010
== noif:VC2010(64) on 64bit ==<Press any key to continue>
f0_if[1]: 1465
f0_if[2]: 1439
f0_if[3]: 1453
f1_min[1]: 1583
f1_min[2]: 1554
f1_min[3]: 1580
f2_neg[1]: 379
f2_neg[2]: 378
f2_neg[3]: 380
f3_sar[1]: 244
f3_sar[2]: 252
f3_sar[3]: 250
f5_sse[1]: 26.0
f5_sse[2]: 25.8
f5_sse[3]: 25.3
<Press any key to exit>
//matlab 代码
tic; //计时开始
datSz = 16384;
// 数据规模 ，和LZ 的#define DATASIZE 16384
tDat = round(-128 + (383--128).*rand(datSz,4));
//生成数据， 随机数据，范围 -128 ~ 383，参照LZ的
tDat(tDat<0)=0; //小于0的归0
tDat(tDat>255)=255; //大于255的归255
tDat= fix(tDat);
//小数的变为整数。 因为matlab是 double类型的数据。 如果需转为int16等，也可转
toc //计时结束
Elapsed time is 0.007347 seconds. //耗时 0.007 秒
最后检查看看数据， 抽取若干个，
tDat(1:3000:end,:)
ans =
20 255 93 0
17 0 255 0
183 0 236 158
171 255 0 255
0 171 102 0
140 60 161 255
Intel® Integrated Performance Primitives Reference Manual, Volume 2: Image and Video Processing
4 Image Data Exchange and 4 Initialization Function
Convert
Converts image pixel values from one data type to another.
Syntax
Case 1: Conversion to increase bit depth and change signed to unsigned type
……
Case 2: Conversion to reduce bit depth and change unsigned to signed type:
integer to integer type
IppStatus ippiConvert_8u1u_C1R(const Ipp8u* pSrc, int srcStep, Ipp8u* pDst, int dstStep,
int dstBitOffset, IppiSize roiSize, Ipp8u threshold);
IppStatus ippiConvert_<mod>(const Ipp<scrDatatype>* pSrc, int srcStep, Ipp<dstDatatype>*
pDst, int dstStep, IppiSize roiSize);
Supported values for mod:
16u8u_C1R 16s8u_C1R 32s8u_C1R 32s8s_C1R
16u8u_C3R 16s8u_C3R 32s8u_C3R 32s8s_C3R
16u8u_C4R 16s8u_C4R 32s8u_C4R 32s8s_C4R
16u8u_AC4R 16s8u_AC4R 32s8u_AC4R 32s8s_AC4R
……
Description
The function ippiConvert is declared in the ippi.h file. It operates with ROI (see Regions of Interest in Intel IPP).
This function converts pixel values in the source image ROI pSrc to a different data type and writes them to the destination image ROI pDst.
The result of integer operations is always saturated to the destination data type range. It means that if the value
of the source pixel is out of the data range of the destination image, the value of the corresponding destination pixel is set to the value of the lower or upper bound (minimum or maximum) of the destination data range:
x = pSrc[i,j]
if (x > MAX_VAL) x = MAX_VAL
if (x < MIN_VAL) x = MIN_VAL
pDst[i,j] = (CASTING)x
