实现以下目标的最佳算法是什么:
0010 0000 => 0000 0100
转换从MSB-> LSB到LSB-> MSB.所有位必须反转; 也就是说,这不是字节顺序交换.
注意:下面的所有算法都是用C语言编写的,但是应该可以移植到您选择的语言中(只是当它们不那么快时不要看我:)
低内存(32位int
,32位机器)(从这里):
unsigned int reverse(register unsigned int x) { x = (((x & 0xaaaaaaaa) >> 1) | ((x & 0x55555555) << 1)); x = (((x & 0xcccccccc) >> 2) | ((x & 0x33333333) << 2)); x = (((x & 0xf0f0f0f0) >> 4) | ((x & 0x0f0f0f0f) << 4)); x = (((x & 0xff00ff00) >> 8) | ((x & 0x00ff00ff) << 8)); return((x >> 16) | (x << 16)); }
来自着名的Bit Twiddling Hacks页面:
最快(查找表):
static const unsigned char BitReverseTable256[] = { 0x00, 0x80, 0x40, 0xC0, 0x20, 0xA0, 0x60, 0xE0, 0x10, 0x90, 0x50, 0xD0, 0x30, 0xB0, 0x70, 0xF0, 0x08, 0x88, 0x48, 0xC8, 0x28, 0xA8, 0x68, 0xE8, 0x18, 0x98, 0x58, 0xD8, 0x38, 0xB8, 0x78, 0xF8, 0x04, 0x84, 0x44, 0xC4, 0x24, 0xA4, 0x64, 0xE4, 0x14, 0x94, 0x54, 0xD4, 0x34, 0xB4, 0x74, 0xF4, 0x0C, 0x8C, 0x4C, 0xCC, 0x2C, 0xAC, 0x6C, 0xEC, 0x1C, 0x9C, 0x5C, 0xDC, 0x3C, 0xBC, 0x7C, 0xFC, 0x02, 0x82, 0x42, 0xC2, 0x22, 0xA2, 0x62, 0xE2, 0x12, 0x92, 0x52, 0xD2, 0x32, 0xB2, 0x72, 0xF2, 0x0A, 0x8A, 0x4A, 0xCA, 0x2A, 0xAA, 0x6A, 0xEA, 0x1A, 0x9A, 0x5A, 0xDA, 0x3A, 0xBA, 0x7A, 0xFA, 0x06, 0x86, 0x46, 0xC6, 0x26, 0xA6, 0x66, 0xE6, 0x16, 0x96, 0x56, 0xD6, 0x36, 0xB6, 0x76, 0xF6, 0x0E, 0x8E, 0x4E, 0xCE, 0x2E, 0xAE, 0x6E, 0xEE, 0x1E, 0x9E, 0x5E, 0xDE, 0x3E, 0xBE, 0x7E, 0xFE, 0x01, 0x81, 0x41, 0xC1, 0x21, 0xA1, 0x61, 0xE1, 0x11, 0x91, 0x51, 0xD1, 0x31, 0xB1, 0x71, 0xF1, 0x09, 0x89, 0x49, 0xC9, 0x29, 0xA9, 0x69, 0xE9, 0x19, 0x99, 0x59, 0xD9, 0x39, 0xB9, 0x79, 0xF9, 0x05, 0x85, 0x45, 0xC5, 0x25, 0xA5, 0x65, 0xE5, 0x15, 0x95, 0x55, 0xD5, 0x35, 0xB5, 0x75, 0xF5, 0x0D, 0x8D, 0x4D, 0xCD, 0x2D, 0xAD, 0x6D, 0xED, 0x1D, 0x9D, 0x5D, 0xDD, 0x3D, 0xBD, 0x7D, 0xFD, 0x03, 0x83, 0x43, 0xC3, 0x23, 0xA3, 0x63, 0xE3, 0x13, 0x93, 0x53, 0xD3, 0x33, 0xB3, 0x73, 0xF3, 0x0B, 0x8B, 0x4B, 0xCB, 0x2B, 0xAB, 0x6B, 0xEB, 0x1B, 0x9B, 0x5B, 0xDB, 0x3B, 0xBB, 0x7B, 0xFB, 0x07, 0x87, 0x47, 0xC7, 0x27, 0xA7, 0x67, 0xE7, 0x17, 0x97, 0x57, 0xD7, 0x37, 0xB7, 0x77, 0xF7, 0x0F, 0x8F, 0x4F, 0xCF, 0x2F, 0xAF, 0x6F, 0xEF, 0x1F, 0x9F, 0x5F, 0xDF, 0x3F, 0xBF, 0x7F, 0xFF }; unsigned int v; // reverse 32-bit value, 8 bits at time unsigned int c; // c will get v reversed // Option 1: c = (BitReverseTable256[v & 0xff] << 24) | (BitReverseTable256[(v >> 8) & 0xff] << 16) | (BitReverseTable256[(v >> 16) & 0xff] << 8) | (BitReverseTable256[(v >> 24) & 0xff]); // Option 2: unsigned char * p = (unsigned char *) &v; unsigned char * q = (unsigned char *) &c; q[3] = BitReverseTable256[p[0]]; q[2] = BitReverseTable256[p[1]]; q[1] = BitReverseTable256[p[2]]; q[0] = BitReverseTable256[p[3]];
您可以将此想法扩展到64位int
,或者为了速度而牺牲内存(假设您的L1数据缓存足够大),并使用64K条目查找表一次反向16位.
简单
unsigned int v; // input bits to be reversed unsigned int r = v & 1; // r will be reversed bits of v; first get LSB of v int s = sizeof(v) * CHAR_BIT - 1; // extra shift needed at end for (v >>= 1; v; v >>= 1) { r <<= 1; r |= v & 1; s--; } r <<= s; // shift when v's highest bits are zero
更快(32位处理器)
unsigned char b = x; b = ((b * 0x0802LU & 0x22110LU) | (b * 0x8020LU & 0x88440LU)) * 0x10101LU >> 16;
更快(64位处理器)
unsigned char b; // reverse this (8-bit) byte b = (b * 0x0202020202ULL & 0x010884422010ULL) % 1023;
如果要在32位上执行此操作int
,只需反转每个字节中的位,并反转字节的顺序.那是:
unsigned int toReverse; unsigned int reversed; unsigned char inByte0 = (toReverse & 0xFF); unsigned char inByte1 = (toReverse & 0xFF00) >> 8; unsigned char inByte2 = (toReverse & 0xFF0000) >> 16; unsigned char inByte3 = (toReverse & 0xFF000000) >> 24; reversed = (reverseBits(inByte0) << 24) | (reverseBits(inByte1) << 16) | (reverseBits(inByte2) << 8) | (reverseBits(inByte3);
我对两个最有前途的解决方案,查找表和按位AND(第一个)进行了基准测试.该测试机是一台配备4GB DDR2-800和Core 2 Duo T7500 @ 2.4GHz,4MB L2 Cache的笔记本电脑; 因人而异.我在64位Linux上使用了gcc 4.3.2.OpenMP(和GCC绑定)用于高分辨率计时器.
reverse.c
#include#include #include unsigned int reverse(register unsigned int x) { x = (((x & 0xaaaaaaaa) >> 1) | ((x & 0x55555555) << 1)); x = (((x & 0xcccccccc) >> 2) | ((x & 0x33333333) << 2)); x = (((x & 0xf0f0f0f0) >> 4) | ((x & 0x0f0f0f0f) << 4)); x = (((x & 0xff00ff00) >> 8) | ((x & 0x00ff00ff) << 8)); return((x >> 16) | (x << 16)); } int main() { unsigned int *ints = malloc(100000000*sizeof(unsigned int)); unsigned int *ints2 = malloc(100000000*sizeof(unsigned int)); for(unsigned int i = 0; i < 100000000; i++) ints[i] = rand(); unsigned int *inptr = ints; unsigned int *outptr = ints2; unsigned int *endptr = ints + 100000000; // Starting the time measurement double start = omp_get_wtime(); // Computations to be measured while(inptr != endptr) { (*outptr) = reverse(*inptr); inptr++; outptr++; } // Measuring the elapsed time double end = omp_get_wtime(); // Time calculation (in seconds) printf("Time: %f seconds\n", end-start); free(ints); free(ints2); return 0; }
reverse_lookup.c
#include#include #include static const unsigned char BitReverseTable256[] = { 0x00, 0x80, 0x40, 0xC0, 0x20, 0xA0, 0x60, 0xE0, 0x10, 0x90, 0x50, 0xD0, 0x30, 0xB0, 0x70, 0xF0, 0x08, 0x88, 0x48, 0xC8, 0x28, 0xA8, 0x68, 0xE8, 0x18, 0x98, 0x58, 0xD8, 0x38, 0xB8, 0x78, 0xF8, 0x04, 0x84, 0x44, 0xC4, 0x24, 0xA4, 0x64, 0xE4, 0x14, 0x94, 0x54, 0xD4, 0x34, 0xB4, 0x74, 0xF4, 0x0C, 0x8C, 0x4C, 0xCC, 0x2C, 0xAC, 0x6C, 0xEC, 0x1C, 0x9C, 0x5C, 0xDC, 0x3C, 0xBC, 0x7C, 0xFC, 0x02, 0x82, 0x42, 0xC2, 0x22, 0xA2, 0x62, 0xE2, 0x12, 0x92, 0x52, 0xD2, 0x32, 0xB2, 0x72, 0xF2, 0x0A, 0x8A, 0x4A, 0xCA, 0x2A, 0xAA, 0x6A, 0xEA, 0x1A, 0x9A, 0x5A, 0xDA, 0x3A, 0xBA, 0x7A, 0xFA, 0x06, 0x86, 0x46, 0xC6, 0x26, 0xA6, 0x66, 0xE6, 0x16, 0x96, 0x56, 0xD6, 0x36, 0xB6, 0x76, 0xF6, 0x0E, 0x8E, 0x4E, 0xCE, 0x2E, 0xAE, 0x6E, 0xEE, 0x1E, 0x9E, 0x5E, 0xDE, 0x3E, 0xBE, 0x7E, 0xFE, 0x01, 0x81, 0x41, 0xC1, 0x21, 0xA1, 0x61, 0xE1, 0x11, 0x91, 0x51, 0xD1, 0x31, 0xB1, 0x71, 0xF1, 0x09, 0x89, 0x49, 0xC9, 0x29, 0xA9, 0x69, 0xE9, 0x19, 0x99, 0x59, 0xD9, 0x39, 0xB9, 0x79, 0xF9, 0x05, 0x85, 0x45, 0xC5, 0x25, 0xA5, 0x65, 0xE5, 0x15, 0x95, 0x55, 0xD5, 0x35, 0xB5, 0x75, 0xF5, 0x0D, 0x8D, 0x4D, 0xCD, 0x2D, 0xAD, 0x6D, 0xED, 0x1D, 0x9D, 0x5D, 0xDD, 0x3D, 0xBD, 0x7D, 0xFD, 0x03, 0x83, 0x43, 0xC3, 0x23, 0xA3, 0x63, 0xE3, 0x13, 0x93, 0x53, 0xD3, 0x33, 0xB3, 0x73, 0xF3, 0x0B, 0x8B, 0x4B, 0xCB, 0x2B, 0xAB, 0x6B, 0xEB, 0x1B, 0x9B, 0x5B, 0xDB, 0x3B, 0xBB, 0x7B, 0xFB, 0x07, 0x87, 0x47, 0xC7, 0x27, 0xA7, 0x67, 0xE7, 0x17, 0x97, 0x57, 0xD7, 0x37, 0xB7, 0x77, 0xF7, 0x0F, 0x8F, 0x4F, 0xCF, 0x2F, 0xAF, 0x6F, 0xEF, 0x1F, 0x9F, 0x5F, 0xDF, 0x3F, 0xBF, 0x7F, 0xFF }; int main() { unsigned int *ints = malloc(100000000*sizeof(unsigned int)); unsigned int *ints2 = malloc(100000000*sizeof(unsigned int)); for(unsigned int i = 0; i < 100000000; i++) ints[i] = rand(); unsigned int *inptr = ints; unsigned int *outptr = ints2; unsigned int *endptr = ints + 100000000; // Starting the time measurement double start = omp_get_wtime(); // Computations to be measured while(inptr != endptr) { unsigned int in = *inptr; // Option 1: //*outptr = (BitReverseTable256[in & 0xff] << 24) | // (BitReverseTable256[(in >> 8) & 0xff] << 16) | // (BitReverseTable256[(in >> 16) & 0xff] << 8) | // (BitReverseTable256[(in >> 24) & 0xff]); // Option 2: unsigned char * p = (unsigned char *) &(*inptr); unsigned char * q = (unsigned char *) &(*outptr); q[3] = BitReverseTable256[p[0]]; q[2] = BitReverseTable256[p[1]]; q[1] = BitReverseTable256[p[2]]; q[0] = BitReverseTable256[p[3]]; inptr++; outptr++; } // Measuring the elapsed time double end = omp_get_wtime(); // Time calculation (in seconds) printf("Time: %f seconds\n", end-start); free(ints); free(ints2); return 0; }
我在几种不同的优化中尝试了两种方法,在每个级别进行了3次试验,每次试验随机地逆转了1亿次unsigned ints
.对于查找表选项,我尝试了在按位黑客页面上给出的两个方案(选项1和2).结果如下所示.
按位AND
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -o reverse reverse.c mrj10@mjlap:~/code$ ./reverse Time: 2.000593 seconds mrj10@mjlap:~/code$ ./reverse Time: 1.938893 seconds mrj10@mjlap:~/code$ ./reverse Time: 1.936365 seconds mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O2 -o reverse reverse.c mrj10@mjlap:~/code$ ./reverse Time: 0.942709 seconds mrj10@mjlap:~/code$ ./reverse Time: 0.991104 seconds mrj10@mjlap:~/code$ ./reverse Time: 0.947203 seconds mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O3 -o reverse reverse.c mrj10@mjlap:~/code$ ./reverse Time: 0.922639 seconds mrj10@mjlap:~/code$ ./reverse Time: 0.892372 seconds mrj10@mjlap:~/code$ ./reverse Time: 0.891688 seconds
查找表(选项1)
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -o reverse_lookup reverse_lookup.c mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.201127 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.196129 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.235972 seconds mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O2 -o reverse_lookup reverse_lookup.c mrj10@mjlap:~/code$ ./reverse_lookup Time: 0.633042 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 0.655880 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 0.633390 seconds mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O3 -o reverse_lookup reverse_lookup.c mrj10@mjlap:~/code$ ./reverse_lookup Time: 0.652322 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 0.631739 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 0.652431 seconds
查找表(选项2)
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -o reverse_lookup reverse_lookup.c mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.671537 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.688173 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.664662 seconds mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O2 -o reverse_lookup reverse_lookup.c mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.049851 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.048403 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.085086 seconds mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O3 -o reverse_lookup reverse_lookup.c mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.082223 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.053431 seconds mrj10@mjlap:~/code$ ./reverse_lookup Time: 1.081224 seconds
如果您关注性能,请使用查找表,选项1(字节寻址速度慢得令人惊讶).如果你需要从系统中挤出每个最后一个字节的内存(如果你关心位反转的性能你可能会这样做),那么bit-AND方法的优化版本也不会太破旧.
是的,我知道基准代码是一个完整的黑客.关于如何改进它的建议非常受欢迎.我所知道的事情:
我无法访问ICC.这可能会更快(如果你可以测试一下,请回复评论).
对于具有大型L1D的一些现代微体系结构,64K查找表可能表现良好.
-mtune = native对-O2/-O3不起作用(ld
引发一些疯狂的符号重定义错误),所以我不相信生成的代码会针对我的微体系结构进行调整.
使用SSE可能有一种方法可以更快地完成此操作.我不知道如何,但是通过快速复制,压缩按位AND和调配指令,必须有一些东西.
我知道只有足够的x86组件是危险的; 这是在-O3上为选项1生成的代码GCC,所以比我自己知道更多的人可以查看它:
32位
.L3: movl (%r12,%rsi), %ecx movzbl %cl, %eax movzbl BitReverseTable256(%rax), %edx movl %ecx, %eax shrl $24, %eax mov %eax, %eax movzbl BitReverseTable256(%rax), %eax sall $24, %edx orl %eax, %edx movzbl %ch, %eax shrl $16, %ecx movzbl BitReverseTable256(%rax), %eax movzbl %cl, %ecx sall $16, %eax orl %eax, %edx movzbl BitReverseTable256(%rcx), %eax sall $8, %eax orl %eax, %edx movl %edx, (%r13,%rsi) addq $4, %rsi cmpq $400000000, %rsi jne .L3
编辑:我也尝试uint64_t
在我的机器上使用类型来查看是否有任何性能提升.性能比32位快10%左右,无论你是一次使用64位类型来反转两个32位int
类型的位,还是实际上将位数反转为64位 -位值.汇编代码如下所示(对于前一种情况,一次反转两个32位int
类型的位):
.L3: movq (%r12,%rsi), %rdx movq %rdx, %rax shrq $24, %rax andl $255, %eax movzbl BitReverseTable256(%rax), %ecx movzbq %dl,%rax movzbl BitReverseTable256(%rax), %eax salq $24, %rax orq %rax, %rcx movq %rdx, %rax shrq $56, %rax movzbl BitReverseTable256(%rax), %eax salq $32, %rax orq %rax, %rcx movzbl %dh, %eax shrq $16, %rdx movzbl BitReverseTable256(%rax), %eax salq $16, %rax orq %rax, %rcx movzbq %dl,%rax shrq $16, %rdx movzbl BitReverseTable256(%rax), %eax salq $8, %rax orq %rax, %rcx movzbq %dl,%rax shrq $8, %rdx movzbl BitReverseTable256(%rax), %eax salq $56, %rax orq %rax, %rcx movzbq %dl,%rax shrq $8, %rdx movzbl BitReverseTable256(%rax), %eax andl $255, %edx salq $48, %rax orq %rax, %rcx movzbl BitReverseTable256(%rdx), %eax salq $40, %rax orq %rax, %rcx movq %rcx, (%r13,%rsi) addq $8, %rsi cmpq $400000000, %rsi jne .L3
这个线程引起了我的注意,因为它处理一个简单的问题,即使对于现代CPU也需要大量的工作(CPU周期).有一天,我也站在那里,遇到同样的¤#%"#"问题.我不得不翻转数百万字节.但是我知道我所有的目标系统都是现代的基于英特尔的,所以让我们开始优化到极致!
所以我用Matt J的查找代码作为基础.我正在进行基准测试的系统是i7 haswell 4700eq.
Matt J的查找位翻转400 000 000字节:大约0.272秒.
然后我继续尝试看看英特尔的ISPC编译器是否可以在反向转换算法中使用算法.
因为我尝试了很多东西来帮助编译器找到东西,所以我最终得到了大约0.15秒的性能来bitflip 400 000 000字节.这是一个很大的减少,但对于我的应用程序仍然太慢..
所以人们让我展示了世界上最快的基于英特尔的bitflipper.时钟:
bitflip 400000000字节的时间:0.050082秒!!!!!
// Bitflip using AVX2 - The fastest Intel based bitflip in the world!! // Made by Anders Cedronius 2014 (anders.cedronius (you know what) gmail.com) #include#include #include #include using namespace std; #define DISPLAY_HEIGHT 4 #define DISPLAY_WIDTH 32 #define NUM_DATA_BYTES 400000000 // Constants (first we got the mask, then the high order nibble look up table and last we got the low order nibble lookup table) __attribute__ ((aligned(32))) static unsigned char k1[32*3]={ 0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f, 0x00,0x08,0x04,0x0c,0x02,0x0a,0x06,0x0e,0x01,0x09,0x05,0x0d,0x03,0x0b,0x07,0x0f,0x00,0x08,0x04,0x0c,0x02,0x0a,0x06,0x0e,0x01,0x09,0x05,0x0d,0x03,0x0b,0x07,0x0f, 0x00,0x80,0x40,0xc0,0x20,0xa0,0x60,0xe0,0x10,0x90,0x50,0xd0,0x30,0xb0,0x70,0xf0,0x00,0x80,0x40,0xc0,0x20,0xa0,0x60,0xe0,0x10,0x90,0x50,0xd0,0x30,0xb0,0x70,0xf0 }; // The data to be bitflipped (+32 to avoid the quantization out of memory problem) __attribute__ ((aligned(32))) static unsigned char data[NUM_DATA_BYTES+32]={}; extern "C" { void bitflipbyte(unsigned char[],unsigned int,unsigned char[]); } int main() { for(unsigned int i = 0; i < NUM_DATA_BYTES; i++) { data[i] = rand(); } printf ("\r\nData in(start):\r\n"); for (unsigned int j = 0; j < 4; j++) { for (unsigned int i = 0; i < DISPLAY_WIDTH; i++) { printf ("0x%02x,",data[i+(j*DISPLAY_WIDTH)]); } printf ("\r\n"); } printf ("\r\nNumber of 32-byte chunks to convert: %d\r\n",(unsigned int)ceil(NUM_DATA_BYTES/32.0)); double start_time = omp_get_wtime(); bitflipbyte(data,(unsigned int)ceil(NUM_DATA_BYTES/32.0),k1); double end_time = omp_get_wtime(); printf ("\r\nData out:\r\n"); for (unsigned int j = 0; j < 4; j++) { for (unsigned int i = 0; i < DISPLAY_WIDTH; i++) { printf ("0x%02x,",data[i+(j*DISPLAY_WIDTH)]); } printf ("\r\n"); } printf("\r\n\r\nTime to bitflip %d bytes: %f seconds\r\n\r\n",NUM_DATA_BYTES, end_time-start_time); // return with no errors return 0; }
printf用于调试..
这是主力:
bits 64 global bitflipbyte bitflipbyte: vmovdqa ymm2, [rdx] add rdx, 20h vmovdqa ymm3, [rdx] add rdx, 20h vmovdqa ymm4, [rdx] bitflipp_loop: vmovdqa ymm0, [rdi] vpand ymm1, ymm2, ymm0 vpandn ymm0, ymm2, ymm0 vpsrld ymm0, ymm0, 4h vpshufb ymm1, ymm4, ymm1 vpshufb ymm0, ymm3, ymm0 vpor ymm0, ymm0, ymm1 vmovdqa [rdi], ymm0 add rdi, 20h dec rsi jnz bitflipp_loop ret
代码占用32个字节,然后屏蔽掉半字节.高半字节右移4.然后我使用vpshufb和ymm4/ymm3作为查找表.我可以使用单个查找表但是我必须在将半字节再次进行ORing之前向左移动.
有更快的方法来翻转位.但我必须使用单线程和CPU,因此这是我能实现的最快速度.你能制作更快的版本吗?
请不要对使用英特尔C/C++编译器内在等效命令发表评论......
对于喜欢递归的人来说,这是另一种解决方案.
这个想法很简单.将输入分为两半并交换两半,继续直到达到单个位.
Illustrated in the example below. Ex : If Input is 00101010 ==> Expected output is 01010100 1. Divide the input into 2 halves 0010 --- 1010 2. Swap the 2 Halves 1010 0010 3. Repeat the same for each half. 10 -- 10 --- 00 -- 10 10 10 10 00 1-0 -- 1-0 --- 1-0 -- 0-0 0 1 0 1 0 1 0 0 Done! Output is 01010100
这是一个解决它的递归函数.(注意我使用了无符号整数,因此它可以用于最大sizeof(unsigned int)*8位的输入.
递归函数需要2个参数 - 需要反转其位的值和值中的位数.
int reverse_bits_recursive(unsigned int num, unsigned int numBits) { unsigned int reversedNum;; unsigned int mask = 0; mask = (0x1 << (numBits/2)) - 1; if (numBits == 1) return num; reversedNum = reverse_bits_recursive(num >> numBits/2, numBits/2) | reverse_bits_recursive((num & mask), numBits/2) << numBits/2; return reversedNum; } int main() { unsigned int reversedNum; unsigned int num; num = 0x55; reversedNum = reverse_bits_recursive(num, 8); printf ("Bit Reversal Input = 0x%x Output = 0x%x\n", num, reversedNum); num = 0xabcd; reversedNum = reverse_bits_recursive(num, 16); printf ("Bit Reversal Input = 0x%x Output = 0x%x\n", num, reversedNum); num = 0x123456; reversedNum = reverse_bits_recursive(num, 24); printf ("Bit Reversal Input = 0x%x Output = 0x%x\n", num, reversedNum); num = 0x11223344; reversedNum = reverse_bits_recursive(num,32); printf ("Bit Reversal Input = 0x%x Output = 0x%x\n", num, reversedNum); }
这是输出:
Bit Reversal Input = 0x55 Output = 0xaa Bit Reversal Input = 0xabcd Output = 0xb3d5 Bit Reversal Input = 0x123456 Output = 0x651690 Bit Reversal Input = 0x11223344 Output = 0x22cc4488
那么这肯定不会像Matt J那样是一个答案,但希望它仍然有用.
size_t reverse(size_t n, unsigned int bytes) { __asm__("BSWAP %0" : "=r"(n) : "0"(n)); n >>= ((sizeof(size_t) - bytes) * 8); n = ((n & 0xaaaaaaaaaaaaaaaa) >> 1) | ((n & 0x5555555555555555) << 1); n = ((n & 0xcccccccccccccccc) >> 2) | ((n & 0x3333333333333333) << 2); n = ((n & 0xf0f0f0f0f0f0f0f0) >> 4) | ((n & 0x0f0f0f0f0f0f0f0f) << 4); return n; }
这与Matt最好的算法完全相同,只是有一个叫做BSWAP的小指令可以交换64位数字的字节(而不是位).因此b7,b6,b5,b4,b3,b2,b1,b0变为b0,b1,b2,b3,b4,b5,b6,b7.由于我们使用的是32位数字,因此我们需要将字节交换数字向下移位32位.这让我们完成了交换每个字节的8位的任务,这就完成了!我们完成了.
时间:在我的机器上,Matt的算法在每次试验中运行约0.52秒.我的试验每次试验大约0.42秒.我认为快20%也不错.
如果您担心指令的可用性,BSWAP 维基百科会将指令BSWAP列为1989年推出的80846.应该注意维基百科还指出该指令仅适用于32位寄存器,这显然不是在我的机器上,它非常适用于64位寄存器.
此方法对于任何整数数据类型都同样有效,因此可以通过传递所需的字节数来简单地推广该方法:
size_t reverse(size_t n, unsigned int bytes) { __asm__("BSWAP %0" : "=r"(n) : "0"(n)); n >>= ((sizeof(size_t) - bytes) * 8); n = ((n & 0xaaaaaaaaaaaaaaaa) >> 1) | ((n & 0x5555555555555555) << 1); n = ((n & 0xcccccccccccccccc) >> 2) | ((n & 0x3333333333333333) << 2); n = ((n & 0xf0f0f0f0f0f0f0f0) >> 4) | ((n & 0x0f0f0f0f0f0f0f0f) << 4); return n; }
然后可以这样调用:
n = reverse(n, sizeof(char));//only reverse 8 bits n = reverse(n, sizeof(short));//reverse 16 bits n = reverse(n, sizeof(int));//reverse 32 bits n = reverse(n, sizeof(size_t));//reverse 64 bits
编译器应该能够优化额外的参数(假设编译器内联函数),对于这种sizeof(size_t)
情况,右移将完全删除.请注意,GCC至少不能删除BSWAP并且如果通过则右移sizeof(char)
.
Anders Cedronius的回答为拥有支持AVX2的x86 CPU 的用户提供了一个很好的解决方案.对于没有AVX支持或非x86平台的x86平台,以下任一实现都应该可以正常工作.
第一个代码是经典二进制分区方法的一种变体,经编码以最大限度地利用在各种ARM处理器上有用的shift-plus-logic惯用法.此外,它使用动态掩码生成,这对于RISC处理器是有益的,否则需要多个指令来加载每个32位掩码值.x86平台的编译器应该使用常量传播来在编译时而不是运行时计算所有掩码.
/* Classic binary partitioning algorithm */
inline uint32_t brev_classic (uint32_t a)
{
uint32_t m;
a = (a >> 16) | (a << 16); // swap halfwords
m = 0x00ff00ff; a = ((a >> 8) & m) | ((a << 8) & ~m); // swap bytes
m = m^(m << 4); a = ((a >> 4) & m) | ((a << 4) & ~m); // swap nibbles
m = m^(m << 2); a = ((a >> 2) & m) | ((a << 2) & ~m);
m = m^(m << 1); a = ((a >> 1) & m) | ((a << 1) & ~m);
return a;
}
在"计算机编程艺术"的第4A卷中,D.Knuth展示了一种巧妙的反转位的方法,这些方法在某种程度上令人惊讶地需要比传统二进制分区算法更少的操作.用于32位操作数的一种算法,我在TAOCP中找不到,在Hacker's Delight网站上的文档中显示.
/* Knuth's algorithm from http://www.hackersdelight.org/revisions.pdf. Retrieved 8/19/2015 */
inline uint32_t brev_knuth (uint32_t a)
{
uint32_t t;
a = (a << 15) | (a >> 17);
t = (a ^ (a >> 10)) & 0x003f801f;
a = (t + (t << 10)) ^ a;
t = (a ^ (a >> 4)) & 0x0e038421;
a = (t + (t << 4)) ^ a;
t = (a ^ (a >> 2)) & 0x22488842;
a = (t + (t << 2)) ^ a;
return a;
}
使用英特尔编译器C/C++编译器13.1.3.198,上述两个函数都可以很好地自动向量化目标XMM
寄存器.它们也可以手动矢量化而无需花费太多精力.
在我的IvyBridge Xeon E3 1270v2上,使用自动矢量化代码,uint32_t
在0.070秒内使用1亿个字位反转,使用brev_classic()
0.068秒brev_knuth()
.我注意确保我的基准测试不受系统内存带宽的限制.
假设你有一个位数组,那么:1.从MSB开始,逐位将位推入堆栈.2.从该堆栈弹出位到另一个阵列(如果你想节省空间,则是相同的阵列),将第一个弹出位置于MSB中,然后从那里继续执行不太重要的位.
Stack stack = new Stack(); Bit[] bits = new Bit[] { 0, 0, 1, 0, 0, 0, 0, 0 }; for (int i = 0; i < bits.Length; i++) { stack.push(bits[i]); } for (int i = 0; i < bits.Length; i++) { bits[i] = stack.pop(); }
本机ARM指令"rbit"可以用1个cpu周期和1个额外的cpu寄存器来实现,不可能被击败.
这不是一个人的工作! ......但对于机器来说是完美的
这是2015年,也就是第一次提出这个问题的6年.编辑器从此成为我们的主人,而我们作为人类的工作只是为了帮助他们.那么将我们的意图用于机器的最佳方式是什么?
位反转是如此常见,以至于你不得不想知道为什么x86不断增长的ISA不包含一次性指令.
原因是:如果你给编译器提供真正简洁的意图,那么位反转应该只需要20个CPU周期.让我告诉你如何制作reverse()并使用它:
#include#include uint64_t reverse(const uint64_t n, const uint64_t k) { uint64_t r, i; for (r = 0, i = 0; i < k; ++i) r |= ((n >> i) & 1) << (k - i - 1); return r; } int main() { const uint64_t size = 64; uint64_t sum = 0; uint64_t a; for (a = 0; a < (uint64_t)1 << 30; ++a) sum += reverse(a, size); printf("%" PRIu64 "\n", sum); return 0; }
使用Clang版本> = 3.6,-O3,-march = native(使用Haswell测试)编译此示例程序,使用新的AVX2指令提供图形质量的代码,运行时间为11秒,处理~10亿reverse()s.每个反向()大约10 ns,假设2 GHz,0.5 ns的CPU周期使我们处于甜蜜的20个CPU周期.
对于单个大型阵列,您可以在访问RAM一次所需的时间内安装10个reverse()!
您可以在两次访问L2缓存LUT所需的时间内调整1 reverse().
警告:这个示例代码应该在几年内作为一个不错的基准,但是一旦编译器足够聪明以优化main()以仅仅打印最终结果而不是真正计算任何东西,它最终将开始显示它的年龄.但是现在它可以展示reverse().
当然,这里有一个明显的尖端黑客来源:http: //graphics.stanford.edu/~seander/bithacks.html#BitReverseObvious
我知道它不是C而是asm:
var1 dw 0f0f0 clc push ax push cx mov cx 16 loop1: shl var1 shr ax loop loop1 pop ax pop cx
这适用于进位,因此您也可以保存标志