perflab實驗報告材料

1、HUNAN UNIVERSITY課程實驗報告課程名稱：計算機組成與結構實驗項目名稱： perflab-handout專業(yè)班級：姓名：學號：指導教師：科華宀兀成時間：2016 年 5 月 27日信息科學與工程學院實驗題目：perflab程序性能調優(yōu) 實驗目的：理解編譯器，學習程序優(yōu)化，從優(yōu)化程序代碼和程序執(zhí)行速度兩方面著 手。實驗要求：本次實驗，要求針對每個函數(shù)、每個人均至少寫出3種優(yōu)化版本、并根據(jù)driver報告的結果進行性能分析實驗環(huán)境：ubuntu-15.10、x32 系統(tǒng)、VMware workstation實驗容及操作步驟：將下載下來的 kernels.c中的rotate、smooth

2、函數(shù)進行優(yōu)化。rotate函數(shù)的作 用是將圖像逆時針旋轉90°, smooth函數(shù)的作用是對于圖像中的每一個像素點， 取 它和周圍的像素點的平均值，讓圖片變得模糊。下面對代碼進行逐一優(yōu)化。源代碼的CPE測試:Rotate: Versto仃 DtnYou r CPEs Baseline CPEsSpeedup=naive rotate: 643, fl14.7S + 012&4.140.19.7Naive2S67,54446.2baseline512 11*7 65 + 9 5.6i 們 plernentation: 1024IS-3夕4 于5.2MeanRotate: Ver

3、sion DtmYour CPE5Baseline CPE&Speedup=rotate:匚urrent working version:2 561235121024Mean7* 5 CJ "16 52 5* - 2 2 4 2 9 4Smooth: Verston DimYour CPEs Baseline CPEsSpeedupsmooth: Current working version:12&57, 3702012* 33257.2 695. & 1226457.3 69S.012.22S658*4 717.& 12*351261.2722.0

4、11.8Mean12.1Smooth: Version DtmYour CPEs Baseline CPEsSpeedup=natve3257269 5.912*2smooth:6457*2698.012.2Naive128 57* 1 702.612,3b-asetine25658.4717.012.3implenentatton:512Mean62.8 722.011.512.1Summary of Your Best Scores:Rotate: 6* 1 (ntve_rotate: Natve baseline inplenenttton)Smooth: 121 (snooth二 Cu

5、rrent working version)1. Naive_rotate1) 源代碼：char naive rotate descr= "naive rotate: Naive baseline implementation"void naive_rotate(int dim, pixel *src, pixel *dst)int i, j;for (i = 0; i < dim; i+)for (j = 0; j < dim; j+)dstRIDX(dim-1-j, i, dim) = srcRIDX(i, j, dim);2) 分析：這段代碼的作用就是用一

6、個雙層循環(huán)將所有的像素進行行列調位、導致整幅圖畫進 行了 90度旋轉。然而分析一下代碼就能發(fā)現(xiàn)一個十分簡單的優(yōu)化方法：因為在最 層循環(huán)中，j的值每次都會改變，所以每執(zhí)行一次賦值就要計算一次dim-1-j，算多了自然就慢了。我們可以利用簡單的數(shù)學技巧改寫公式，將賦值語句改成 dstRIDX(i, j, dim) = srcRIDX(j, dim-i-1, dim);這樣就不用每次都計算了。3) 優(yōu)化代碼1如下：char naive_rotate_descr2= "naive_rotate2:only change the place of i andj"void naive_

7、rotate2(int dim, pixel *src, pixel *dst)int i, j;for (i = 0; i < dim; i+)for (j = 0; j < dim; j+)dstRIDX(i, j, dim) = srcRIDX(j, dim-i-1, dim);/i change less優(yōu)化結果如下：Rotate: Ver&ton=natve_rotate:NaivebaselineimpleFientatton:64'1232565121024Mea nYour CPE55*39*212,719*SBaseline CPEs14.740.

8、146.465.994.5Speedup畀26,95,15*24*85*4Rotatei Version-naiverotated:onlychange the placeof I and jiDin64_1282565121024Your CPEsE.52.83.76.515.7Baseline CPEs14,746.146*465.994 t 5Speedup14.412.410A氣09.1這是一種最為簡單的優(yōu)化方案，由圖可知，速度提升不大，性能優(yōu)化結果也不是很好。再分析 源代碼，從cache友好性來分析，這個代碼的效率機會很低，所以按照cache的大小，應在存儲的時候進行 32個像素依次

9、存儲（列存儲）。做到cache友好這樣就可以可以大幅度提高效率。4) 優(yōu)化代碼2如下：char rotate_descr2 = "rotate2: versi on2break into 4*4 blocks"void rotate2(int dim, pixel *src, pixel*dst)int i, j,ii,jj;for(ii=0; ii < dim; ii+=4)for(jj=0; jj < dim; jj+=4)for(i=ii; i < ii+4; i+)for(j=jj; j < jj+4; j+)dstRIDX(dim-1-j,

10、i,dim) = srcRIDX(i,j,dim);優(yōu)化結果如下：Rot3Version= otate; Current working version:Dim641282565121024Mea nYour CPEs4-67.011.517.8Baseline CPEs14*740*146*465*994*5Speedup3-29.96.75.75.S5.8Rotate: Version=rotate2: version?break into4*4 blocks:Dim6412825 65121024MeanYour CPEs3,23.86.910.4Baseline CPEs14.740.

11、146.4&59鄉(xiāng)4.5Speedup4.510.58.859.18*2用分塊的方式，進行優(yōu)化。將整個程序分成4*4的小塊，提高空間局部性5) 優(yōu)化代碼3如下：char rotate_descr3 = "rotate3: version3 break into 32*32 blocks" void rotate3(int dim, pixel *src, pixel *dst)int i, j,ii,jj;for(ii=0; ii < dim; ii+=32)for(jj=0; jj < dim; jj+=32)for(i=ii; i < ii+3

12、2; i+) for(j=jj; j < jj+32; j+)dstRIDX(dim-1-j,i,dim) = srcRIDX(i,j,dim);優(yōu)化結果如下：Rotate: Version=rotate3: version3break Into32*32blocks:Din6412825峪5121024Your CPEs3.12 + 94.19.415.3Baseline CPEs14.740,146e465.994.5Speedup4.713.911.47.06.28.6分成32*32塊，提高空間局部性6) 優(yōu)化代碼4如下：char rotate_descr4 = "rot

13、ate4:Current working version,usingpointerrather tha n computi ng address"void rotate4(int dim, pixel *src, pixel *dst)int i;int j;int tmp1=dim*dim;int tmp2=dim *31;int tmp3=tmp1-dim;int tmp4=tmp1+32;int tmp5=dim+31;dst+=tmp3;for(i=0; i< dim; i+=32) for(j=0;j<dim;j+)*dst=*src; dst+;src+=di

14、m;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=di

15、m;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=di

16、m;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim;*dst=*src; dst+;src+=dim; *dst=*src; dst+;src+=dim;*dst=*src; src+; src-=tmp2;dst-=tmp5;src+=tmp2; dst+=tmp4;優(yōu)化結果如下：Rotate: Version=rotated: Current working version,usingpointer rather than conputin

17、g address:Din6412&2505121血4MeanYour CPEs2.32.42.44.47.9Baseline CPEs14.746.146.455.994,5Speedup&317.019,014.912.e12.9用循環(huán)展開，分成32路并行來寫2. Naive_smooth1) 源代碼char naive_smooth_descr = "naive_smooth: Naive baseline implementation" void naive_smooth(int dim, pixel *src, pixel *dst)int i,

18、j;for (i = 0; i < dim; i+)for (j = 0; j < dim; j+)dstRIDX(i, j, dim) = avg(dim, i, j, src);cpeS生能如下：Srooth: Version=smoothi: Current worktng version:Dim326412Q256E12Mea nYour CPEs57*957.85S.064.364.0Baseline CPEs65-S702.0717.0722.0Speedup1Z-1化111.111.311 -1Srooth: Version=naive_smooth:Naive ba

19、selineimplementation:Dim3264128256512MeanYour CPEs6S. iS7.857-861.964,8BseLtne CPEs695*0698.0762.8717722,0Speedup10.712.112.111.611.111,5Surnnary of YourBest Scores:Rotate: 12.4(rotate2:Current:working)version f usingpointer rting address)Smooth: 117(snooth:Currentworkingversion)2) 分析這段代碼很多次地調用 avg函

20、數(shù)，而 avg函數(shù)也頻繁調用 initialize_pixel_sum、accumulate_sum、assign_sum_to_pixel 這幾個函數(shù)，且含有2層for循環(huán)。雖然會以損害程序的模塊性為代價，但消除函數(shù)調用的時 間開銷，得到的代碼運行速度會快得多。所以，需要改寫代碼，不調用 avg函數(shù)。3) 優(yōu)化代碼1如下：char smooth_descr1 = "smoothl: with less func call and grossly simplified calculation for central parts"void smooth1(int dim, p

21、ixel *src, pixel *dst)int i, j, ii, jj;pixel_sum sum;pixel curre nt_pixel, cp;for (j = 0; j < dim; j+)dstRIDX(0, j, dim) = avg(dim, 0, j, src);dstRIDX(dim-1, j, dim) = avg(dim, dim-1, j, src);for (i = 0; i < dim; i+)dstRIDX(i, 0, dim) = avg(dim, i, 0, src);dstRIDX(i, dim-1, dim) = avg(dim, i,

22、dim-1, src);for (i = 1; i < dim-1; i+)for (j = 1; j < dim-1; j+)sum.red = sum.gree n = sum.blue = 0;for(ii = max(i-1,0); ii <= min (i+1, dim-1); ii+) for(jj = max(j-1,0); jj <= min(j+1, dim-1); jj+) cp=srcRIDX(ii,jj,dim);sum.red += cp.red;sum.gree n += cp.gree n; sum.blue += cp.blue;curr

23、e nt_pixel.red = sum.red/9;curre nt_pixel.gree n = sum.gree n/9;curre nt_pixel.blue = sum.blue/9; dstRIDX(i, j, dim) = curre nt_pixel;優(yōu)化結果如下：Smooth: Version=smooth: Current working version:Dim3264128256512MeanYour CPEs75*067 >466*667*971*6Baseline CPEs695.D698.0702.0717.0722.0Speedup9.316.410.S10

24、.610.110.1Smooth: Verston二 snoothl: withlessfunc call呂nd grossly simplified calculatton for central parts:Dim326412825&512leanYour CPEs49.047.346.849.350.0Baseline CPEs695698.0702.0717.0722.9Speedup14.214.815x014x514x414.64) 優(yōu)化代碼2如下：char smooth_descr2 = "smooth2: test vers ion"void smo

25、oth2(i nt dim, pixel *src, pixel *dst)int i,j;/no using avg()/cornersdstRIDX(0,0,dim).red=(srcRIDX(0,0,dim).red+srcRIDX(1,0,dim).red+sr cRIDX(0,1,dim).red+srcRIDX(1,1,dim).red)>>2;dstRIDX(0,0,dim).blue=(srcRIDX(0,0,dim).blue+srcRIDX(1,0,dim).blue +srcRIDX(0,1,dim).blue+srcRIDX(1,1,dim).blue)&g

26、t;>2;dstRIDX(0,0,dim).gree n=(srcRIDX(0,0,dim).gree n+srcRIDX(1,0,dim).gr ee n+srcRIDX(0,1,dim).gree n+srcRIDX(1,1,dim).gree n)>>2;dstRIDX(0,dim-1,dim).red=(srcRIDX(0,dim-1,dim).red+srcRIDX(1,dim-1, dim).red+srcRIDX(0,dim-2,dim).red+srcRIDX(1,dim-2,dim).red)>>2;dstRIDX(0,dim-1,dim).bl

27、ue=(srcRIDX(0,dim-1,dim).blue+srcRIDX(1,dim-1, dim).blue+srcRIDX(0,dim-2,dim).blue+srcRIDX(1,dim-2,dim).blue)>>2JdstRIDX(0,dim-1,dim).gree n=(srcRIDX(0,dim-1,dim).gree n+srcRIDX(1,di m-1,dim).gree n+srcRIDX(0,dim-2,dim).gree n+srcRIDX(1,dim-2,dim).gree n)>>2;dstRIDX(dim-1,0,dim).red=(src

28、RIDX(dim-1,0,dim).red+srcRIDX(dim-2,0, dim).red+srcRIDX(dim-1,1,dim).red+srcRIDX(dim-2,1,dim).red)>>2;dstRIDX(dim-1,0,dim).blue=(srcRIDX(dim-1,0,dim).blue+srcRIDX(dim-2, 0,dim).blue+srcRIDX(dim-1,1,dim).blue+srcRIDX(dim-2,1,dim).blue)>>2JdstRIDX(dim-1,0,dim).gree n=(srcRIDX(dim-1,0,dim).

29、gree n+srcRIDX(dim- 2,0,dim).gree n+srcRIDX(dim-1,1,dim).gree n+srcRIDX(dim-2,1,dim).gree n)>>2;dstRIDX(dim-1,dim-1,dim).red=(srcRIDX(dim-1,dim-1,dim).red+srcRIDX( dim-1,dim-2,dim).red+srcRIDX(dim-2,dim-1,dim).red+srcRIDX(dim-2,dim-2, dim).red)>>2;dstRIDX(dim-1,dim-1,dim).blue=(srcRIDX(d

30、im-1,dim-1,dim).blue+srcRID X(dim-1,dim-2,dim).blue+srcRIDX(dim-2,dim-1,dim).blue+srcRIDX(dim-2, dim-2,dim).blue)>>2;dstRIDX(dim-1,dim-1,dim).gree n=(srcRIDX(dim-1,dim-1,dim).gree n+srcR IDX(dim-1,dim-2,dim).gree n+srcRIDX(dim-2,dim-1,dim).gree n+srcRIDX(di m-2,dim-2,dim).gree n)>>2;/boa

31、rderfor(i=1;i<dim-1;i+) dstRIDX(i,0,dim).red=(srcRIDX(i,0,dim).red+srcRIDX(i-1,0,dim).red+ srcRIDX(i-1,1,dim).red+srcRIDX(i,1,dim).red+srcRIDX(i+1,0,dim).red +srcRIDX(i+1,1,dim).red)/6;dstRIDX(i,0,dim).blue=(srcRIDX(i,0,dim).blue+srcRIDX(i-1,0,dim).bl ue+srcRIDX(i-1,1,dim).blue+srcRIDX(i,1,dim).b

32、lue+srcRIDX(i+1,0,dim) .blue+srcRIDX(i+1,1,dim).blue)/6;dstRIDX(i,0,dim).gree n=(srcRIDX(i,0,dim).gree n+srcRIDX(i-1,0,dim). gree n+srcRIDX(i-1,1,dim).gree n+srcRIDX(i,1,dim).gree n+srcRIDX(i+1,0 ,dim).gree n+srcRIDX(i+1,1,dim).gree n)/6;for(i=1;i<dim-1;i+) dstRIDX(i,dim-1,dim).red=(srcRIDX(i,dim

33、-1,dim).red+srcRIDX(i-1,dim- 1,dim).red+srcRIDX(i-1,dim-2,dim).red+srcRIDX(i,dim-2,dim).red+src RIDX(i+1,dim-1,dim).red+srcRIDX(i+1,dim-2,dim).red)/6;dstRIDX(i,dim-1,dim).blue=(srcRIDX(i,dim-1,dim).blue+srcRIDX(i-1,di m-1,dim).blue+srcRIDX(i-1,dim-2,dim).blue+srcRIDX(i,dim-2,dim).blue +srcRIDX(i+1,d

34、im-1,dim).blue+srcRIDX(i+1,dim-2,dim).blue)/6;dstRIDX(i,dim-1,dim).gree n=(srcRIDX(i,dim-1,dim).gree n+srcRIDX(i-1, dim-1,dim).gree n+srcRIDX(i-1,dim-2,dim).gree n+srcRIDX(i,dim-2,dim). gree n+srcRIDX(i+1,dim-1,dim).gree n+srcRIDX(i+1,dim-2,dim).gree n)/6;for(j=1;j<dim-1;j+) dstRIDX(0,j,dim).red=

35、(srcRIDX(0,j,dim).red+srcRIDX(0,j-1,dim).red+ srcRIDX(1,j-1,dim).red+srcRIDX(1,j,dim).red+srcRIDX(0,j+1,dim).red +srcRIDX(1,j+1,dim).red)/6;dstRIDX(0,j,dim).blue=(srcRIDX(0,j,dim).blue+srcRIDX(0,j-1,dim).bl ue+srcRIDX(1,j-1,dim).blue+srcRIDX(1,j,dim).blue+srcRIDX(0,j+1,dim) .blue+srcRIDX(1,j+1,dim).

36、blue)/6;dstRIDX(0,j,dim).gree n=(srcRIDX(0,j,dim).gree n+srcRIDX(0,j-1,dim).gree n+srcRIDX(1,j-1,dim).gree n+srcRIDX(1,j,dim).gree n+srcRIDX(0,j+1 ,dim).gree n+srcRIDX(1,j+1,dim).gree n)/6;for(j=1;j<dim-1;j+)dstRIDX(dim-1,i,dim).red=(srcRIDX(dim-1,i,dim).red+srcRIDX(dim-1,i+1,dim).red+srcRIDX(dim

37、-1,j-1,dim).red+srcRIDX(dim-2,j,dim).red+srcRIDX(dim-2,j+1,dim).red+srcRIDX(dim-2,j-1,dim).red)/6;dstRIDX(dim-1,j,dim).blue=(srcRIDX(dim-1,j,dim).blue+srcRIDX(dim-1, j+1,dim).blue+srcRIDX(dim-1,j-1,dim).blue+srcRIDX(dim-2,j,dim).blue +srcRIDX(dim-2,j+1,dim).blue+srcRIDX(dim-2,j-1,dim).blue)/6;dstRID

38、X(dim-1,j,dim).gree n=(srcRIDX(dim-1,j,dim).gree n+srcRIDX(dim- 1,j+1,dim).gree n+srcRIDX(dim-1,j-1,dim).gree n+srcRIDX(dim-2,j,dim). gree n+srcRIDX(dim-2,j+1,dim).gree n+srcRIDX(dim-2,j-1,dim).gree n)/6;/commonfor(i=1;i<dim-1;i+) for(j=1;j<dim-1;j+)dstRIDX(i,j,dim).red=(srcRIDX(i,j,dim).red+s

39、rcRIDX(i+1,j,dim).red+srcRIDX(i-1,j,dim).red+srcRIDX(i,j-1,dim).red+srcRIDX(i+1,j-1,dim) .red+srcRIDX(i-1,j-1,dim).red+srcRIDX(i,j+1,dim).red+srcRIDX(i+1,j+ 1,dim).red+srcRIDX(i-1,j+1,dim).red)/9;dstRIDX(i,j,dim).blue=(srcRIDX(i,j,dim).blue+srcRIDX(i+1,j,dim).bl ue+srcRIDX(i-1,j,dim).blue+srcRIDX(i,j-1,dim).blue+srcRIDX(i+1,j-1, dim).blue+srcRIDX(i-1,j-1,dim).blue+srcRIDX(i,j+1,dim).blue+srcRID