图形图像处理－之－高质量的快速的图像缩放补充使用SSE2优化

你的名字 2022-06-12 11:47 240阅读 0赞

转自 [http://blog.csdn.net/housisong/article/details/6318890][http_blog.csdn.net_housisong_article_details_6318890]

**图形图像处理－之－高质量的快速的图像缩放 补充 使用SSE2优化**  
                           [HouSisong@GMail.com ][HouSisong_GMail.com]   2011.04.12

**tag:** 图像缩放,速度优化,线性插值,三次卷积插值,SSE2,scale,bilinear,bicubic,StretchBlt

**摘要:**   
  本文章对线性插值和三次卷积插值(bicubic)的实现做了一些新的优化尝试;  
使用了SSE2的128bit寄存器及相关指令;并预先建立SSE2用到的缩放系数表;   
实现的结果在我的i7电脑上比以前的版本分别快出145%和75%!  
线性插值的速度是StretchBlt的13倍!

**正文:**   
(请先看看我的blog里<高质量的快速的图像缩放>的前3篇文章!)  
支持SSE2指令集的CPU越来越多,CPU的SSE2实现性能也好了很多(以前不比MMX好多少),  
而且软件在64位模式的时候不再支持MMX,所以尝试了SSE2的缩放优化,效果不错!

**速度测试说明:**   
  只测试内存数据到内存数据的缩放  
  测试图片都是800\*600缩放到1024\*768,单线程;fps表示每秒钟的帧数,值越大表示函数越快.

**速度测试对比:** (CPU:i7 920 内存:DDR3 1333 3通道)

(windows)  
StretchBlt  近邻取样           869.09 fps  
StretchBlt  线性插值            44.46 fps //SetStretchBltMode(dc,4);?

PicZoom0:                       95.69 fps  
PicZoom1:                      158.35 fps  
PicZoom2:                      332.78 fps  
PicZoom3:                     1172.79 fps  
PicZoom3\_float:                874.13 fps  
PicZoom3\_Table:               1158.30 fps  
PicZoom3\_SSE:                 1908.40 fps  
PicZoom\_Bilinear0:              28.80 fps  
PicZoom\_Bilinear1:              56.09 fps  
PicZoom\_Bilinear2:              97.09 fps  
PicZoom\_Bilinear\_Common:       119.83 fps  
PicZoom\_Bilinear\_MMX:          180.12 fps  
PicZoom\_Bilinear\_MMX\_Ex:       237.34 fps  
PicZoom\_ftBilinear\_Common:     118.67 fps  
PicZoom\_ftBilinear\_MMX:        213.68 fps  
PicZoom\_ThreeOrder0:             6.11 fps  
PicZoom\_ThreeOrder\_Common:      25.38 fps  
PicZoom\_ThreeOrder\_MMX:         52.32 fps

(SSE2的实现)  
PicZoom\_ftBilinearTable\_SSE2:  588.24 fps  
PicZoom\_ThreeOrderTable\_SSE2:   93.24 fps

**PicZoom\_ftBilinearTable\_SSE2实现代码如下:**

1.      typedef UInt64 TMMXData64;  
2.      //ftBilinearTable\_SSE2(out \[edi+ebx\*4\]; xmm5=v,xmm6=vr,xmm7=0,\[ebp\]=(u,ur),\[edx\]=srx\_x,esi=PSrcLineColor,ecx=PSrcLineColorNext)  
3.      //void \_\_declspec(naked) ftBilinearTable\_SSE2()\{   
4.      \#define  ftBilinearTable\_SSE2()                     /  
5.          asm mov         eax,\[edx+ebx\]                   /  
6.          asm movq        xmm0,qword ptr\[esi+eax\*4\]       /  
7.          asm movq        xmm1,qword ptr\[ecx+eax\*4\]       /  
8.          asm punpcklbw   xmm0,xmm7                       /  
9.          asm punpcklbw   xmm1,xmm7                       /  
10.         asm pmullw      xmm0,mm5                        /  
11.         asm pmullw      xmm1,mm6                        /  
12.         asm paddw       xmm0,xmm1                       /  
13.         asm pmulhw      xmm0,xmmword ptr \[ebp+ebx\*4\] /  
14.         asm movdqa      xmm1,xmm0                       /  
15.         asm punpckhqdq  xmm0,xmm0                       /  
16.         asm paddw       xmm0,xmm1                       /  
17.         asm packuswb    xmm0,xmm7                       /  
18.         asm movd  dword ptr  \[edi+ebx\],xmm0              
19.         //ret //for  \_\_declspec(naked)  
20.         //\}  
21.     //\}  
22.     //void \_\_declspec(naked) ftBilinearTable\_SSE2\_expand2()\{   
23.     \#define  ftBilinearTable\_SSE2\_expand2()             /  
24.         asm mov         eax,\[edx+ebx\]                   /  
25.         asm movq        xmm0,qword ptr\[esi+eax\*4\]       /  
26.         asm movq        xmm1,qword ptr\[ecx+eax\*4\]       /  
27.         asm mov         eax,\[edx+ebx+4\]                 /  
28.         asm movq        xmm2,qword ptr\[esi+eax\*4\]       /  
29.         asm movq        xmm3,qword ptr\[ecx+eax\*4\]       /  
30.         asm punpcklbw   xmm0,xmm7                       /  
31.         asm punpcklbw   xmm1,xmm7                       /  
32.         asm punpcklbw   xmm2,xmm7                       /  
33.         asm punpcklbw   xmm3,xmm7                       /  
34.         asm pmullw      xmm0,mm5                        /  
35.         asm pmullw      xmm1,mm6                        /  
36.         asm pmullw      xmm2,mm5                        /  
37.         asm pmullw      xmm3,mm6                        /  
38.         asm paddw       xmm0,xmm1                       /  
39.         asm paddw       xmm2,xmm3                       /  
40.         asm pmulhw      xmm0,xmmword ptr \[ebp+ebx\*4\] /  
41.         asm pmulhw      xmm2,xmmword ptr \[ebp+ebx\*4+16\] /  
42.         asm movdqa      xmm1,xmm0                       /  
43.         asm punpcklqdq  xmm0,xmm2                       /  
44.         asm punpckhqdq  xmm1,xmm2                       /  
45.         asm paddw       xmm0,xmm1                       /  
46.         asm packuswb    xmm0,xmm7                       /  
47.         asm movq  qword ptr  \[edi+ebx\],xmm0             /  
48.         //ret //for  \_\_declspec(naked)  
49.         //\}  
50.     //\}  
51.      
52. **void** PicZoom\_ftBilinearTable\_SSE2(**const** TPixels32Ref& Dst,**const** TPixels32Ref& Src)  
53. \{  
54.     **if** (  (0==Dst.width)||(0==Dst.height)  
55.         ||(2>Src.width)||(2>Src.height)) **return**;  
56.     **long** xrIntFloat\_16=((Src.width-1)<<16)/Dst.width;  
57.     **long** yrIntFloat\_16=((Src.height-1)<<16)/Dst.height;  
58.     **long** dst\_width=Dst.width;  
59.     UInt8\* \_bufMem=**new** UInt8\[(dst\_width\*2\***sizeof**(TMMXData64)+15)+dst\_width\***sizeof**(Int32)\];  
60.     TMMXData64\* uList=(TMMXData64\*)((((ptrdiff\_t)\_bufMem)+15)>>4<<4); //16byte对齐  
61.     Int32\* xList=(Int32\*)(uList+dst\_width\*2);  
62.     \{ //init u table  
63.         **long** srcx\_16=0;  
64.         **for** (**long** x=0;x<dst\_width\*2;x+=2)\{  
65.             xList\[x>>1\]=(srcx\_16>>16);  
66.             unsigned **long** u=(srcx\_16>>8)&0xFF;  
67.             unsigned **long** ur=(256-u)<<1;  
68.             u=u<<1;  
69.             uList\[x+0\]=(ur|(ur<<16));  
70.             uList\[x+0\]|=uList\[x+0\]<<32;  
71.             uList\[x+1\]=u|(u<<16);  
72.             uList\[x+1\]|=uList\[x+1\]<<32;  
73.             srcx\_16+=xrIntFloat\_16;  
74.         \}  
75.     \}  
76.     Color32\* pDstLine=Dst.pdata;  
77.     **long** srcy\_16=0;  
78.     asm pxor  xmm7,xmm7 //xmm7=0  
79.     **for** (**long** y=0;y<Dst.height;++y)\{  
80.         unsigned **long** v=(srcy\_16>>8) & 0xFF;  
81.         unsigned **long** vr=(256-v)>>1;  
82.         v>>=1;  
83.         Color32\* PSrcLineColor= (Color32\*)((UInt8\*)(Src.pdata)+Src.byte\_width\*(srcy\_16>>16)) ;  
84.         Color32\* PSrcLineColorNext= (Color32\*)((UInt8\*)(PSrcLineColor)+Src.byte\_width) ;  
85.         asm\{  
86.               movd        xmm5,vr  
87.               movd        xmm6,v  
88.               punpcklwd   xmm5,xmm5  
89.               punpcklwd   xmm6,xmm6  
90.               punpckldq   xmm5,xmm5  
91.               punpckldq   xmm6,xmm6  
92.               punpcklqdq  xmm5,xmm5  
93.               punpcklqdq  xmm6,xmm6  
94.              
95.               mov       esi,PSrcLineColor  
96.               mov       ecx,PSrcLineColorNext  
97.               mov       edx,xList //x  
98.               mov       ebx,dst\_width  
99.               mov       edi,pDstLine  
100.               push      ebp  
101.               mov       ebp,uList  
102.               push      ebx  
103.                
104.               and       ebx,(not 1)  
105.               test      ebx,ebx  
106.               jle     end\_loop2  
107.   
108.               lea       ebx,\[ebx\*4\]  
109.               lea       edi,\[edi+ebx\]  
110.               lea       edx,\[edx+ebx\]  
111.               lea       ebp,\[ebp+ebx\*4\]  
112.               neg       ebx  
113.         loop2\_start:  
114.               //call ftBilinearTable\_SSE2\_expand2  
115.               ftBilinearTable\_SSE2\_expand2()  
116.               add       ebx,8  
117.               jnz       loop2\_start  
118.         end\_loop2:  
119.             pop    ebx  
120.             and    ebx,1   
121.             test   ebx,ebx  
122.             jle    end\_write  
123.               lea       ebx,\[ebx\*4\]  
124.               lea       edi,\[edi+ebx\]  
125.               lea       edx,\[edx+ebx\]  
126.               lea       ebp,\[ebp+ebx\*4\]  
127.               neg       ebx  
128.         loop1\_start:  
129.               //call ftBilinearTable\_SSE2  
130.               ftBilinearTable\_SSE2()  
131.               add       ebx,4  
132.               jnz       loop1\_start  
133.         end\_write:  
134.               pop       ebp  
135.         \}  
136.         srcy\_16+=yrIntFloat\_16;  
137.         ((UInt8\*&)pDstLine)+=Dst.byte\_width;  
138.     \}  
139.     delete \[\]\_bufMem;  
140. \}

**PicZoom\_ThreeOrderTable\_SSE2实现代码如下:**

1.      **static** TMMXData64 SinXDivX\_Table64\_MMX\[(2<<8)+1\];  
2.      **class** \_CAutoInti\_SinXDivX\_Table64\_MMX \{  
3.      **private**:  
4.          **void** \_Inti\_SinXDivX\_Table64\_MMX()  
5.          \{  
6.              **for** (**long** i=0;i<=(2<<8);++i)  
7.              \{  
8.                  unsigned **short** t=(unsigned **short**)(0.5+(1<<14)\*SinXDivX(i\*(1.0/(256))));  
9.                  unsigned **long** tl=t|(((unsigned **long**)t)<<16);  
10.                 TMMXData64 tll=tl|(((TMMXData64)tl)<<32);  
11.                 SinXDivX\_Table64\_MMX\[i\]=tll;  
12.             \}  
13.         \};  
14.     **public**:  
15.         \_CAutoInti\_SinXDivX\_Table64\_MMX() \{ \_Inti\_SinXDivX\_Table64\_MMX(); \}  
16.     \};  
17.     **static** \_CAutoInti\_SinXDivX\_Table64\_MMX \_\_tmp\_CAutoInti\_SinXDivX\_Table64\_MMX;  
18.     //void \_\_declspec(naked)  \_private\_ThreeOrderTable\_Fast\_SSE2\_2()\{   
19.     \#define  \_private\_ThreeOrderTable\_Fast\_SSE2\_2() /  
20.         asm movq        xmm0,qword ptr \[eax\]        /  
21.         asm movq        xmm1,qword ptr \[eax+8\]      /  
22.         asm movq        xmm2,qword ptr \[eax+edx\] /  
23.         asm movq        xmm3,qword ptr \[eax+edx+8\] /  
24.         asm punpcklbw   xmm0,xmm7     /  
25.         asm punpcklbw   xmm1,xmm7     /  
26.         asm punpcklbw   xmm2,xmm7     /  
27.         asm punpcklbw   xmm3,xmm7     /  
28.         asm psllw       xmm0,7      /  
29.         asm psllw       xmm1,7      /  
30.         asm psllw       xmm2,7      /  
31.         asm psllw       xmm3,7      /  
32.         asm pmulhw      xmm0,xmmword ptr \[ecx\]  /  
33.         asm pmulhw      xmm1,xmmword ptr \[ecx+16\] /  
34.         asm pmulhw      xmm2,xmmword ptr \[ecx\]  /  
35.         asm pmulhw      xmm3,xmmword ptr \[ecx+16\] /  
36.         asm paddsw      xmm0,xmm1     /  
37.         asm paddsw      xmm2,xmm3     /  
38.         asm pmulhw      xmm0,xmmword ptr \[ebx\]  /  
39.         asm pmulhw      xmm2,xmmword ptr \[ebx+16\] /  
40.         asm paddsw      xmm0,xmm2     /  
41.         //asm ret //for \_\_declspec(naked)  
42.     //\}  
43.     must\_inline UInt32 ThreeOrderTable\_Fast\_SSE2(**const** Color32\* pixel,**long** byte\_width,**const** TMMXData64\* v4,**const** TMMXData64\* u4)\{  
44.         asm mov     eax,pixel  
45.         asm mov     edx,byte\_width  
46.         asm mov     ebx,v4  
47.         asm mov     ecx,u4  
48.         //asm call \_private\_ThreeOrderTable\_Fast\_SSE2\_2  
49.         \_private\_ThreeOrderTable\_Fast\_SSE2\_2();  
50.         asm movdqa  xmm6,xmm0  
51.         asm lea     eax,\[eax+edx\*2\]  //+pic.byte\_width  
52.         asm lea     ebx,\[ebx+32\]  
53.         //asm call \_private\_ThreeOrderTable\_Fast\_SSE2\_2  
54.         \_private\_ThreeOrderTable\_Fast\_SSE2\_2();  
55.         asm paddsw      xmm6,xmm0  
56.         asm movdqa      xmm5,xmm6  
57.         asm psrldq      xmm6,8   //srl 8\*8 bit!  
58.         asm paddsw      xmm5,xmm6  
59.         asm psraw       xmm5,3  
60.         asm packuswb    xmm5,xmm7  
61.         asm movd        eax,xmm5  
62.     \}  
63.     must\_inline **long** getSizeBorder(**long** x,**long** maxx)\{  
64.   **if** (x<=0)  
65.    **return** 0;  
66.   **else** **if** (x>=maxx)  
67.    **return** maxx;  
68.   **else**  
69.    **return** x;  
70.     \}  
71.     must\_inline UInt32 ThreeOrderTable\_Border\_SSE2(**const** TPixels32Ref& pic,**const** **long** x0\_sub1,**const** **long** y0\_sub1,**const** TMMXData64\* v4,**const** TMMXData64\* u4)\{  
72.         Color32 pixel\[16\];  
73.         **long** height\_sub\_1=pic.height-1;  
74.         **long** width\_sub\_1=pic.width-1;  
75.         Color32\* pbuf=pixel;  
76.         **for** (**long** i=0;i<4;++i,pbuf+=4)\{  
77.             **long** y=getSizeBorder(y0\_sub1+i,height\_sub\_1);  
78.             Color32\* pLine=pic.getLinePixels(y);  
79.             pbuf\[0\]=pLine\[getSizeBorder(x0\_sub1+0,width\_sub\_1)\];  
80.             pbuf\[1\]=pLine\[getSizeBorder(x0\_sub1+1,width\_sub\_1)\];  
81.             pbuf\[2\]=pLine\[getSizeBorder(x0\_sub1+2,width\_sub\_1)\];  
82.             pbuf\[3\]=pLine\[getSizeBorder(x0\_sub1+3,width\_sub\_1)\];  
83.         \}  
84.         **return** ThreeOrderTable\_Fast\_SSE2(pixel,4\***sizeof**(Color32),v4,u4);  
85.     \}  
86. **void** PicZoom\_ThreeOrderTable\_SSE2(**const** TPixels32Ref& Dst,**const** TPixels32Ref& Src)  
87. \{  
88.     **if** (  (0==Dst.width)||(0==Dst.height)  
89.         ||(0==Src.width)||(0==Src.height)) **return**;  
90.     **long** dst\_width=Dst.width;  
91.     **long** dst\_height=Dst.height;  
92.     **long** xrIntFloat\_16=((Src.width)<<16)/dst\_width+1;  
93.     **long** yrIntFloat\_16=((Src.height)<<16)/dst\_height+1;  
94.     **const** **long** csDErrorX=-(1<<15)+(xrIntFloat\_16>>1);  
95.     **const** **long** csDErrorY=-(1<<15)+(yrIntFloat\_16>>1);  
96.     //计算出需要特殊处理的边界  
97.     **long** border\_y0=((1<<16)-csDErrorY)/yrIntFloat\_16+1;//y0+y\*yr>=1; y0=csDErrorY => y>=(1-csDErrorY)/yr  
98.     **if** (border\_y0>=dst\_height) border\_y0=dst\_height;  
99.     **long** border\_x0=((1<<16)-csDErrorX)/xrIntFloat\_16+1;  
100.     **if** (border\_x0>=dst\_width ) border\_x0=dst\_width;  
101.     **long** border\_y1=(((Src.height-3)<<16)-csDErrorY)/yrIntFloat\_16+1; //y0+y\*yr<=(height-3) => y<=(height-3-csDErrorY)/yr  
102.     **if** (border\_y1<border\_y0) border\_y1=border\_y0;  
103.     **long** border\_x1=(((Src.width-3)<<16)-csDErrorX)/xrIntFloat\_16+1;;  
104.     **if** (border\_x1<border\_x0) border\_x1=border\_x0;  
105.     UInt8\* \_bufMem=**new** UInt8\[(dst\_width\*4\***sizeof**(TMMXData64)+15)+dst\_width\***sizeof**(Int32)\];  
106.     TMMXData64\* uList=(TMMXData64\*)((((ptrdiff\_t)\_bufMem)+15)>>4<<4); //16byte对齐  
107.     Int32\* xList=(Int32\*)(uList+dst\_width\*4);  
108.     \{ //init u table  
109.         **long** srcx\_16=csDErrorX;  
110.         **for** (**long** x=0;x<dst\_width\*4;x+=4)\{  
111.             xList\[x>>2\]=(srcx\_16>>16)-1;  
112.             **long** u=(srcx\_16>>8)&0xFF;  
113.             uList\[x+0\]=SinXDivX\_Table64\_MMX\[256+u\];  
114.             uList\[x+1\]=SinXDivX\_Table64\_MMX\[u\];  
115.             uList\[x+2\]=SinXDivX\_Table64\_MMX\[256-u\];  
116.             uList\[x+3\]=SinXDivX\_Table64\_MMX\[512-u\];  
117.             srcx\_16+=xrIntFloat\_16;  
118.         \}  
119.     \}  
120.     TMMXData64 \_v4\[8+2\];  
121.     TMMXData64\* v4=(&\_v4\[0\]); v4=(TMMXData64\*)( (((ptrdiff\_t)v4)+15)>>4<<4);  
122.     asm pxor    xmm7,xmm7  
123.     Color32\* pDstLine=Dst.pdata;  
124.     **long** srcy\_16=csDErrorY;  
125.     **for** (**long** y=0;y<dst\_height;++y)\{  
126.         //v table  
127.         **const** **long** srcy\_sub1=(srcy\_16>>16)-1;  
128.         **const** **long** v=(srcy\_16>>8)&0xFF;  
129.         v4\[0\]=SinXDivX\_Table64\_MMX\[256+v\];  
130.         v4\[1\]=v4\[0\];  
131.         v4\[2\]=SinXDivX\_Table64\_MMX\[v\];  
132.         v4\[3\]=v4\[2\];  
133.         v4\[4\]=SinXDivX\_Table64\_MMX\[256-v\];  
134.         v4\[5\]=v4\[4\];  
135.         v4\[6\]=SinXDivX\_Table64\_MMX\[512-v\];  
136.         v4\[7\]=v4\[6\];  
137.         **if** ((y<border\_y0)||(y>=border\_y1))\{  
138.             **for** (**long** x=0;x<dst\_width;++x)  
139.                 pDstLine\[x\].argb=ThreeOrderTable\_Border\_SSE2(Src,xList\[x\],srcy\_sub1,v4,&uList\[x\*4\]); //border  
140.         \}**else**\{  
141.             **for** (**long** x=0;x<border\_x0;++x)  
142.                 pDstLine\[x\].argb=ThreeOrderTable\_Border\_SSE2(Src,xList\[x\],srcy\_sub1,v4,&uList\[x\*4\]);//border  
143.             **const** Color32\* pixelLine=Src.getLinePixels(srcy\_sub1);  
144.             **long** byte\_width=Src.byte\_width;  
145.             **for** (**long** x=border\_x0;x<border\_x1;++x)  
146.                 pDstLine\[x\].argb=ThreeOrderTable\_Fast\_SSE2(&pixelLine\[xList\[x\]\],byte\_width,v4,&uList\[x\*4\]);//fast MMX !  
147.             **for** (**long** x=border\_x1;x<dst\_width;++x)  
148.                 pDstLine\[x\].argb=ThreeOrderTable\_Border\_SSE2(Src,xList\[x\],srcy\_sub1,v4,&uList\[x\*4\]);//border  
149.         \}  
150.         srcy\_16+=yrIntFloat\_16;  
151.         ((UInt8\*&)pDstLine)+=Dst.byte\_width;  
152.     \}  
153.     delete \[\]\_bufMem;  
154. \}

[http_blog.csdn.net_housisong_article_details_6318890]: http://blog.csdn.net/housisong/article/details/6318890
[HouSisong_GMail.com]: mailto:HouSisong@GMail.com