実験：unsigned char – float変換編

画像処理の高速化はshaderでやれというツッコミは置いておいて、(shaderでは色はそもそもfloat値なので関係ない。)unsigned char – float変換はよく使うし、出現頻度が多い。これは精度の都合で 0-255(8bit)を0.0f-1.0fとしてfloat精度で取り扱うことで画質ロスを無くし、かつ高速なfloat演算にあやかるというもの。で、今回は1.0f/255.0f-255.0f/255.0fの計算結果を配列にしておいてunsigned charをインデックスにして取得したら高速になるんじゃない？っていう仮説。あんま意味ないのでこのエントリーは読まないほうがいいｗｗｗ

いわゆる下記のようなケース
この時点で掛算しか使ってないので、意味無さそうだよな・・・。
単純に掛算VS配列の結果呼び出しっていうテストになってる。

//byte から float
unsigned char b=128;
float f=b*0.003921568627451f; // *1.0f/255.0f

適当に配列を用意し、constで計算しておく。
一応if分岐しまくるfloat2Byteも作ったけど面倒なので貼らない。

#pragma once
class KCast
{
public:
 static float b2Float(unsigned char in);
 static unsigned char f2Byte(float in);
 static const float b2fMap[256];
};

const float KCast::b2fMap[256]={
 0.0f/255.0f,1.0f/255.0f,2.0f/255.0f,3.0f/255.0f,4.0f/255.0f,
 5.0f/255.0f,6.0f/255.0f,7.0f/255.0f,8.0f/255.0f,9.0f/255.0f,
 10.0f/255.0f,11.0f/255.0f,12.0f/255.0f,13.0f/255.0f,14.0f/255.0f,
 15.0f/255.0f,16.0f/255.0f,17.0f/255.0f,18.0f/255.0f,19.0f/255.0f,
 20.0f/255.0f,21.0f/255.0f,22.0f/255.0f,23.0f/255.0f,24.0f/255.0f,
 25.0f/255.0f,26.0f/255.0f,27.0f/255.0f,28.0f/255.0f,29.0f/255.0f,
 30.0f/255.0f,31.0f/255.0f,32.0f/255.0f,33.0f/255.0f,34.0f/255.0f,
 35.0f/255.0f,36.0f/255.0f,37.0f/255.0f,38.0f/255.0f,39.0f/255.0f,
 40.0f/255.0f,41.0f/255.0f,42.0f/255.0f,43.0f/255.0f,44.0f/255.0f,
 45.0f/255.0f,46.0f/255.0f,47.0f/255.0f,48.0f/255.0f,49.0f/255.0f,
 50.0f/255.0f,51.0f/255.0f,52.0f/255.0f,53.0f/255.0f,54.0f/255.0f,
 55.0f/255.0f,56.0f/255.0f,57.0f/255.0f,58.0f/255.0f,59.0f/255.0f,
 60.0f/255.0f,61.0f/255.0f,62.0f/255.0f,63.0f/255.0f,64.0f/255.0f,
 65.0f/255.0f,66.0f/255.0f,67.0f/255.0f,68.0f/255.0f,69.0f/255.0f,
 70.0f/255.0f,71.0f/255.0f,72.0f/255.0f,73.0f/255.0f,74.0f/255.0f,
 75.0f/255.0f,76.0f/255.0f,77.0f/255.0f,78.0f/255.0f,79.0f/255.0f,
 80.0f/255.0f,81.0f/255.0f,82.0f/255.0f,83.0f/255.0f,84.0f/255.0f,
 85.0f/255.0f,86.0f/255.0f,87.0f/255.0f,88.0f/255.0f,89.0f/255.0f,
 90.0f/255.0f,91.0f/255.0f,92.0f/255.0f,93.0f/255.0f,94.0f/255.0f,
 95.0f/255.0f,96.0f/255.0f,97.0f/255.0f,98.0f/255.0f,99.0f/255.0f,
 100.0f/255.0f,101.0f/255.0f,102.0f/255.0f,103.0f/255.0f,104.0f/255.0f,
 105.0f/255.0f,106.0f/255.0f,107.0f/255.0f,108.0f/255.0f,109.0f/255.0f,
 110.0f/255.0f,111.0f/255.0f,112.0f/255.0f,113.0f/255.0f,114.0f/255.0f,
 115.0f/255.0f,116.0f/255.0f,117.0f/255.0f,118.0f/255.0f,119.0f/255.0f,
 120.0f/255.0f,121.0f/255.0f,122.0f/255.0f,123.0f/255.0f,124.0f/255.0f,
 125.0f/255.0f,126.0f/255.0f,127.0f/255.0f,128.0f/255.0f,129.0f/255.0f,
 130.0f/255.0f,131.0f/255.0f,132.0f/255.0f,133.0f/255.0f,134.0f/255.0f,
 135.0f/255.0f,136.0f/255.0f,137.0f/255.0f,138.0f/255.0f,139.0f/255.0f,
 140.0f/255.0f,141.0f/255.0f,142.0f/255.0f,143.0f/255.0f,144.0f/255.0f,
 145.0f/255.0f,146.0f/255.0f,147.0f/255.0f,148.0f/255.0f,149.0f/255.0f,
 150.0f/255.0f,151.0f/255.0f,152.0f/255.0f,153.0f/255.0f,154.0f/255.0f,
 155.0f/255.0f,156.0f/255.0f,157.0f/255.0f,158.0f/255.0f,159.0f/255.0f,
 160.0f/255.0f,161.0f/255.0f,162.0f/255.0f,163.0f/255.0f,164.0f/255.0f,
 165.0f/255.0f,166.0f/255.0f,167.0f/255.0f,168.0f/255.0f,169.0f/255.0f,
 170.0f/255.0f,171.0f/255.0f,172.0f/255.0f,173.0f/255.0f,174.0f/255.0f,
 175.0f/255.0f,176.0f/255.0f,177.0f/255.0f,178.0f/255.0f,179.0f/255.0f,
 180.0f/255.0f,181.0f/255.0f,182.0f/255.0f,183.0f/255.0f,184.0f/255.0f,
 185.0f/255.0f,186.0f/255.0f,187.0f/255.0f,188.0f/255.0f,189.0f/255.0f,
 190.0f/255.0f,191.0f/255.0f,192.0f/255.0f,193.0f/255.0f,194.0f/255.0f,
 195.0f/255.0f,196.0f/255.0f,197.0f/255.0f,198.0f/255.0f,199.0f/255.0f,
 200.0f/255.0f,201.0f/255.0f,202.0f/255.0f,203.0f/255.0f,204.0f/255.0f,
 205.0f/255.0f,206.0f/255.0f,207.0f/255.0f,208.0f/255.0f,209.0f/255.0f,
 210.0f/255.0f,211.0f/255.0f,212.0f/255.0f,213.0f/255.0f,214.0f/255.0f,
 215.0f/255.0f,216.0f/255.0f,217.0f/255.0f,218.0f/255.0f,219.0f/255.0f,
 220.0f/255.0f,221.0f/255.0f,222.0f/255.0f,223.0f/255.0f,224.0f/255.0f,
 225.0f/255.0f,226.0f/255.0f,227.0f/255.0f,228.0f/255.0f,229.0f/255.0f,
 230.0f/255.0f,231.0f/255.0f,232.0f/255.0f,233.0f/255.0f,234.0f/255.0f,
 235.0f/255.0f,236.0f/255.0f,237.0f/255.0f,238.0f/255.0f,239.0f/255.0f,
 240.0f/255.0f,241.0f/255.0f,242.0f/255.0f,243.0f/255.0f,244.0f/255.0f,
 245.0f/255.0f,246.0f/255.0f,247.0f/255.0f,248.0f/255.0f,249.0f/255.0f,
 250.0f/255.0f,251.0f/255.0f,252.0f/255.0f,253.0f/255.0f,254.0f/255.0f,
 255.0f/255.0f
};

float KCast::b2Float(unsigned char in){
 return *(b2fMap+in);
}

実験。
適当にたして割るだけのループ。関数キャストと普通の掛算による計算を各10回ずつ。
関数のオーバヘッドもありそうなのでメモリを直接読むテストも加える。

int main(){
  const float div=1.0f/255.0f;
  unsigned long long start;
  float res;
  unsigned char i;

  puts("start");
  for(int x=0;x<10;x++){

    //キャスト関数で処理
    start= ofGetSystemTime( );
    res=0.0f;
    for (int m=0;m<1000000;m++){
      i=0;
      while (i<254){
        res+=KCast::b2Float(i);
        i++;
      }
      res*=div;
    }
    printf("キャスト TIME=%llu:RESULT=%e\n",ofGetSystemTime( )-start,res);

    //直接メモリ読む
    start= ofGetSystemTime( );
    res=0.0f;
    for (int m=0;m<1000000;m++){
      i=0;
      while (i<254){
        res+=*(KCast::b2fMap+i);
        i++;
      }
      res*=div;
    }
    printf("直接メモリ TIME=%llu:RESULT=%e\n",ofGetSystemTime( )-start,res);

    //掛算のみ
    start= ofGetSystemTime( );
    res=0.0f;
    for (int m=0;m<1000000;m++){
      i=0;
      while (i<254){
        res+=i*div;
        i++;
      }
      res*=div;
    }
    printf("かけざん TIME=%llu:RESULT=%e\n",ofGetSystemTime( )-start,res);
  }
  return 1;
}

結果

start
キャスト TIME=2459:RESULT=4.960788e-001
直接メモリ TIME=729:RESULT=4.960788e-001
かけざん TIME=724:RESULT=4.960788e-001
キャスト TIME=2435:RESULT=4.960788e-001
直接メモリ TIME=722:RESULT=4.960788e-001
かけざん TIME=716:RESULT=4.960788e-001
キャスト TIME=2396:RESULT=4.960788e-001
直接メモリ TIME=721:RESULT=4.960788e-001
かけざん TIME=747:RESULT=4.960788e-001
キャスト TIME=2419:RESULT=4.960788e-001
直接メモリ TIME=718:RESULT=4.960788e-001
かけざん TIME=716:RESULT=4.960788e-001
キャスト TIME=2398:RESULT=4.960788e-001
直接メモリ TIME=717:RESULT=4.960788e-001
かけざん TIME=718:RESULT=4.960788e-001
キャスト TIME=2395:RESULT=4.960788e-001
直接メモリ TIME=716:RESULT=4.960788e-001
かけざん TIME=718:RESULT=4.960788e-001
キャスト TIME=2403:RESULT=4.960788e-001
直接メモリ TIME=726:RESULT=4.960788e-001
かけざん TIME=721:RESULT=4.960788e-001
キャスト TIME=2411:RESULT=4.960788e-001
直接メモリ TIME=717:RESULT=4.960788e-001
かけざん TIME=715:RESULT=4.960788e-001
キャスト TIME=2397:RESULT=4.960788e-001
直接メモリ TIME=717:RESULT=4.960788e-001
かけざん TIME=720:RESULT=4.960788e-001
キャスト TIME=2395:RESULT=4.960788e-001
直接メモリ TIME=715:RESULT=4.960788e-001
かけざん TIME=717:RESULT=4.960788e-001

トリプルスコアで完敗ｗｗｗ

結論 :

掛算は早いから計算したほうがよっぽど良い。まず関数のオーバーヘッドは大きい印象。inlineとかでこのへん少しは改善できるのかな。メモリ読み込みも検討しているけど、そもそもポインタのリードは遅い。このテストでは１－２５５まで順に読ませているけど連続するメモリ読み出しはかなり高速になるので、ランダムにアクセスされた場合はほぼ使えないと考えられる。

まぁ、こういう実験するとポインタ呼び出しの速度とか、身にしみて覚えるからいいよっていう話でした。
m9(^Д^)ﾌﾟｷﾞｬｰ