我正在修改一个开源项目(粉末玩具)并始终使用/O2(最大化速度)选项进行编译，启用 SSE2 代码生成并且刚刚检查了这个:
void membwand(void * destv, void * srcv, size_t destsize, size_t srcsize)
{
size_t i;
unsigned char * dest = (unsigned char*)destv;
unsigned char * src = (unsigned char*)srcv;
for(i = 0; i < destsize; i++){
dest[i] = dest[i] & src[i%srcsize];
}
}
这里我替换了下面几行
membwand(gravy, gravmask, size_dst, size_src);
membwand(gravx, gravmask, size_dst, size_src);
// gravy, gravx and gravmask are 1MB each
与
membwand2(gravy,gravx, gravmask, size_dst, size_src);
// sizes are same, why not put all in a single function?
实现为:
void membwand2(void * destv1, void * destv2,void * srcv, size_t destsize, size_t srcsize)
{
if(destsize!=srcsize)
{
size_t i;
unsigned char * dest1 = (unsigned char*)destv1;
unsigned char * dest2 = (unsigned char*)destv2;
unsigned char * src = (unsigned char*)srcv;
for(i = 0; i < destsize; i++)
{
dest1[i] = dest1[i] & src[i%srcsize];
dest2[i] = dest2[i] & src[i%srcsize];
}
}
else
{
size_t i;
unsigned char * dest1 = (unsigned char*)destv1;unsigned char * dest2 = (unsigned char*)destv2;
unsigned char * src = (unsigned char*)srcv;
unsigned char tmpChar0=0;unsigned char tmpChar1=0;unsigned char tmpChar2=0;
unsigned char tmpChar3=0;unsigned char tmpChar4=0;unsigned char tmpChar5=0;
unsigned char tmpChar6=0;unsigned char tmpChar7=0;unsigned char tmpChar8=0;
unsigned char tmpChar9=0;unsigned char tmpChar10=0;unsigned char tmpChar11=0;
unsigned char tmpChar12=0;unsigned char tmpChar13=0;unsigned char tmpChar14=0;
unsigned char tmpChar15=0;unsigned char tmpChar16=0;unsigned char tmpChar17=0;
unsigned char tmpChar18=0;unsigned char tmpChar19=0;unsigned char tmpChar20=0;
unsigned char tmpChar21=0;unsigned char tmpChar22=0;unsigned char tmpChar23=0;
unsigned char tmpChar24=0;unsigned char tmpChar25=0;unsigned char tmpChar26=0;
unsigned char tmpChar27=0;unsigned char tmpChar28=0;unsigned char tmpChar29=0;
unsigned char tmpChar30=0;unsigned char tmpChar31=0;
for(i = 0; i < destsize; i+=32)
{
tmpChar0=src[i];tmpChar1=src[i+1];tmpChar2=src[i+2];tmpChar3=src[i+3];
tmpChar4=src[i+4];tmpChar5=src[i+5];tmpChar6=src[i+6];tmpChar7=src[i+7];
tmpChar8=src[i+8];tmpChar9=src[i+9];tmpChar10=src[i+10];tmpChar11=src[i+11];
tmpChar12=src[i+12];tmpChar13=src[i+13];tmpChar14=src[i+14];tmpChar15=src[i+15];
tmpChar16=src[i+16];tmpChar17=src[i+17];tmpChar18=src[i+18];tmpChar19=src[i+19];
tmpChar20=src[i+20];tmpChar21=src[i+21];tmpChar22=src[i+22];tmpChar23=src[i+23];
tmpChar24=src[i+24];tmpChar25=src[i+25];tmpChar26=src[i+26];tmpChar27=src[i+27];
tmpChar28=src[i+28];tmpChar29=src[i+29];tmpChar30=src[i+30];tmpChar31=src[i+31];
dest1[i] = dest1[i] & tmpChar0;
dest1[i+1] = dest1[i+1] & tmpChar1;
dest1[i+2] = dest1[i+2] & tmpChar2;
dest1[i+3] = dest1[i+3] & tmpChar3;
dest1[i+4] = dest1[i+4] & tmpChar4;
dest1[i+5] = dest1[i+5] & tmpChar5;
dest1[i+6] = dest1[i+6] & tmpChar6;
dest1[i+7] = dest1[i+7] & tmpChar7;
dest1[i+8] = dest1[i+8] & tmpChar8;
dest1[i+9] = dest1[i+9] & tmpChar9;
dest1[i+10] = dest1[i+10] & tmpChar10;
dest1[i+11] = dest1[i+11] & tmpChar11;
dest1[i+12] = dest1[i+12] & tmpChar12;
dest1[i+13] = dest1[i+13] & tmpChar13;
dest1[i+14] = dest1[i+14] & tmpChar14;
dest1[i+15] = dest1[i+15] & tmpChar15;
dest1[i+16] = dest1[i+16] & tmpChar16;
dest1[i+17] = dest1[i+17] & tmpChar17;
dest1[i+18] = dest1[i+18] & tmpChar18;
dest1[i+19] = dest1[i+19] & tmpChar19;
dest1[i+20] = dest1[i+20] & tmpChar20;
dest1[i+21] = dest1[i+21] & tmpChar21;
dest1[i+22] = dest1[i+22] & tmpChar22;
dest1[i+23] = dest1[i+23] & tmpChar23;
dest1[i+24] = dest1[i+24] & tmpChar24;
dest1[i+25] = dest1[i+25] & tmpChar25;
dest1[i+26] = dest1[i+26] & tmpChar26;
dest1[i+27] = dest1[i+27] & tmpChar27;
dest1[i+28] = dest1[i+28] & tmpChar28;
dest1[i+29] = dest1[i+29] & tmpChar29;
dest1[i+30] = dest1[i+30] & tmpChar30;
dest1[i+31] = dest1[i+31] & tmpChar31;
dest2[i] = dest2[i] & tmpChar0;
dest2[i+1] = dest2[i+1] & tmpChar1;
dest2[i+2] = dest2[i+2] & tmpChar2;
dest2[i+3] = dest2[i+3] & tmpChar3;
dest2[i+4] = dest2[i+4] & tmpChar4;
dest2[i+5] = dest2[i+5] & tmpChar5;
dest2[i+6] = dest2[i+6] & tmpChar6;
dest2[i+7] = dest2[i+7] & tmpChar7;
dest2[i+8] = dest2[i+8] & tmpChar8;
dest2[i+9] = dest2[i+9] & tmpChar9;
dest2[i+10] = dest2[i+10] & tmpChar10;
dest2[i+11] = dest2[i+11] & tmpChar11;
dest2[i+12] = dest2[i+12] & tmpChar12;
dest2[i+13] = dest2[i+13] & tmpChar13;
dest2[i+14] = dest2[i+14] & tmpChar14;
dest2[i+15] = dest2[i+15] & tmpChar15;
dest2[i+16] = dest2[i+16] & tmpChar16;
dest2[i+17] = dest2[i+17] & tmpChar17;
dest2[i+18] = dest2[i+18] & tmpChar18;
dest2[i+19] = dest2[i+19] & tmpChar19;
dest2[i+20] = dest2[i+20] & tmpChar20;
dest2[i+21] = dest2[i+21] & tmpChar21;
dest2[i+22] = dest2[i+22] & tmpChar22;
dest2[i+23] = dest2[i+23] & tmpChar23;
dest2[i+24] = dest2[i+24] & tmpChar24;
dest2[i+25] = dest2[i+25] & tmpChar25;
dest2[i+26] = dest2[i+26] & tmpChar26;
dest2[i+27] = dest2[i+27] & tmpChar27;
dest2[i+28] = dest2[i+28] & tmpChar28;
dest2[i+29] = dest2[i+29] & tmpChar29;
dest2[i+30] = dest2[i+30] & tmpChar30;
dest2[i+31] = dest2[i+31] & tmpChar31;
}
}
}
两行版本的 Omp 挂钟计时为 17 毫秒，单行版本为 1.5 毫秒。 (经过数千次迭代)
问题:是什么导致了 9 倍的加速？ src[] 或缓存利用的数据重用？也许代码之前不适合矢量化，并且展开启用了它，这可能是索引缺少的模数操作吗？我应该在循环顶部添加#pragma omp 吗？
编辑:#pragma omp parallel for num_threads(4) 以某种方式使情况变得更糟。也许只有 1MB 并不能隐藏多线程开销？
Edit2: 省略模运算符使其为 2.5 毫秒，但性能从 2.5 毫秒增加到 1.5 毫秒(~%60 增加)一定来自展开/缓存/...
注意:启用“favor fast code”、启用整个程序优化、完全优化(/Ox)和启用内在函数并没有改变速度(也许启用 SSE2 和/O2 就足够了)
中央处理器:FX8150 IDE:MSVC
最佳答案
关于C++ 数组访问优化，我们在Stack Overflow上找到一个类似的问题： https://stackoverflow.com/questions/21753744/