实现以下目的的最有效算法是:
0010 0000 => 0000 0100
转换是从MSB-> LSB到LSB-> MSB。所有位都必须反转;也就是说,这不是字节序交换。
实现以下目的的最有效算法是:
0010 0000 => 0000 0100
转换是从MSB-> LSB到LSB-> MSB。所有位都必须反转;也就是说,这不是字节序交换。
Answers:
注意:以下所有算法均使用C语言,但应可移植到您选择的语言中(当它们不那么快时,请不要看着我:)
低内存(32位int
,32位计算机)(从此处开始):
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));
}
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
,或以内存为代价进行交换(假设您的L1数据高速缓存足够大),并使用64K条目查找表一次反转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,然后反转字节的顺序即可。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进行了基准测试,即查找表和按位与(第一个)。测试机器是一台笔记本电脑,带有4GB DDR2-800和一个Core 2 Duo T7500 @ 2.4GHz,4MB L2缓存;YMMV。我用过gcc在64位Linux上 4.3.2。OpenMP(和GCC绑定)用于高分辨率计时器。
反向
#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
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 <stdlib.h>
#include <stdio.h>
#include <omp.h>
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条件下尝试了这两种方法,在每个级别上进行了3次试验,每个试验都随机抽取了1亿个样本unsigned ints
。对于查找表选项,我尝试了按位hacks页面上给出的两种方案(选项1和2)。结果如下所示。
按位与
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
使用带有选项1的查找表如果您担心性能(字节寻址很慢)。如果您需要从系统中挤出内存的最后一个字节(并且您可能会关心位反转的性能),那么按位“与”方法的优化版本也不会太差劲。
是的,我知道基准代码是一个完整的技巧。我们非常欢迎有关如何改进它的建议。我知道的事情:
ld
炸毁了一些疯狂的符号重新定义错误),因此我不相信为我的微体系结构调整生成的代码。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
在计算机上使用类型,以查看是否有任何性能提升。性能比32位快大约10%,无论您是一次使用64位类型一次反转两个32位int
类型上的位,还是实际上反转了64 位类型的一半,性能几乎都相同。位值。汇编代码如下所示(对于前一种情况,一次反转两种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
该线程引起了我的注意,因为它处理的是一个简单的问题,即使对于现代CPU来说,也需要大量工作(CPU周期)。有一天,我也遇到了同样的¤#%“#”问题。我不得不翻转数百万个字节。但是我知道我所有的目标系统都是基于Intel的现代系统,所以让我们开始进行优化!!!
因此,我使用了Matt J的查找代码作为基础。我基准测试的系统是i7 haswell 4700eq。
Matt J的查找位翻转4000亿个字节:约0.272秒。
然后,我继续尝试查看英特尔的ISPC编译器是否可以反向执行矢量运算。c。
我不会在这里烦恼我的发现,因为我做了很多尝试来帮助编译器查找内容,无论如何,我最终获得了大约0.15秒的性能来翻转400000000字节。这是一个很大的减少,但是对于我的应用程序来说仍然太慢了。
所以人们让我展示世界上最快的基于Intel的bitflipper。时钟在:
翻转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 <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <omp.h>
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。
这是主力军:
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
该代码占用32个字节,然后掩盖了半字节。高半字节右移4。然后我将vpshufb和ymm4 / ymm3用作查找表。我可以使用单个查找表,但是随后我必须先向左移动,然后再对四字节进行“或”运算。
甚至还有更快的方式来翻转位。但是我只能使用单线程和CPU,所以这是我能达到的最快速度。您可以制作一个更快的版本吗?
请不要对使用英特尔C / C ++编译器固有等效命令发表评论...
pshub
,因为毕竟最好的popcount也可以做到!如果不适合您,我会在这里写的。荣誉
popcnt
,tzcnt
以及pext
所有端口1.所以每次pext
或tzcnt
成本你popcnt
吞吐量。如果您的数据在L1D高速缓存中很热,则在AVX2 pshufb上对英特尔CPU进行阵列计数的最快方法是。(Ryzen每时钟popcnt
吞吐量有4个,因此可能是最佳的,但是Bulldozer系列每4个时钟popcnt r64,r64
吞吐量有一个... agner.org/optimize)。
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。将输入除以一半,交换两半,继续直到达到单比特。
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
这是一个递归函数来解决。(请注意,我使用了无符号整数,因此它可以用于最大为sizeof(unsigned int)* 8位的输入。
递归函数采用2个参数-需要反转其位的值以及该值中的位数。
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
numBits
int一样,当您将3乘以2除以函数param时,它将舍入为1?
好吧,这肯定不会像Matt J那样回答,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的最佳算法完全相同,只是有一条称为BSWAP的小指令可以交换64位数字的字节(而不是位)。所以b7,b6,b5,b4,b3,b2,b1,b0变成b0,b1,b2,b3,b4,b5,b6,b7。由于我们使用的是32位数字,因此需要将字节交换的数字下移32位。这就剩下了交换每个字节的8位的任务,瞧!我们完成了。
时间:在我的机器上,Matt的算法每次试用运行时间约为0.52秒。每次审判,矿井跑了大约0.42秒。我认为速度提高20%还不错。
如果您担心指令BSWAP的可用性,那么Wikipedia会将BSSWAP指令添加为1989年推出的80846。在我的机器上的情况下,它仅适用于64位寄存器。
此方法对于任何整数数据类型都将同样有效,因此可以通过传递所需的字节数来简化该方法:
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;
}
然后可以这样称呼:
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掉多余的参数(假设编译器内联了函数),并且对于这种sizeof(size_t)
情况,右移将被完全消除。请注意,如果GCC至少已通过,则无法删除BSWAP并向右移动sizeof(char)
。
Anders Cedronius的答案为拥有x86 CPU和AVX2支持的人们提供了一个很好的解决方案。对于不支持AVX的x86平台或非x86平台,以下两种实现方式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,其编码目的是最大限度地利用对各种ARM处理器有用的shift-plus-logic习惯用法。此外,它使用动态掩码生成,这对于RISC处理器可能是有益的,否则该RISC处理器需要多个指令来加载每个32位掩码值。x86平台的编译器应使用常量传播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“计算机程序设计的艺术”第4A卷中,D。Knuth展示了一种巧妙的方式来反转位,这令人惊讶地比传统的二进制分区算法需要更少的操作。我在TAOCP中找不到这种针对32位操作数的算法,该文档在Hacker's Delight网站上显示。
/* 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Intel编译器C / C ++编译器13.1.3.198,以上两个函数都可以自动向量化,从而很好地定位 XMM
寄存器。也可以手动将其矢量化,而无需付出很多努力。
在我的IvyBridge Xeon E3 1270v2上,使用自动矢量化代码,uint32_t
使用0.070秒brev_classic()
和使用0.068秒,对1亿个单词进行了位反转brev_knuth()
。我注意确保基准不受系统内存带宽的限制。
brev_knuth()
?“黑客的喜悦”中PDF的归属似乎表明这些数字直接来自Knuth本人。我不能声称已经充分理解了Knuth对TAOCP中基本设计原理的描述,以解释常数的派生方式,或者对于任意字长如何推导常数和移位因子。
假设您有一个位数组,该如何做:1.从MSB开始,将位逐个推入堆栈。2.将堆栈中的位弹出到另一个数组(如果要节省空间,则弹出同一数组),将第一个弹出的位放到MSB中,然后从那里继续删除低位。
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();
}
对于人类来说,这不是没有工作! ...但是非常适合一台机器
这是2015年,距第一次提出此问题已有6年。此后,编译器已成为我们的主人,而我们作为人类的工作只是帮助他们。那么,向机器传达意图的最佳方法是什么?
比特反转是如此普遍,以至于您不得不怀疑为什么x86不断增长的ISA没有包含一次性执行指令。
原因是:如果您将真正的简洁意图提供给编译器,则位反转仅需要大约20个CPU周期。让我向您展示如何制作reverse()并使用它:
#include <inttypes.h>
#include <stdio.h>
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(已通过Haswell测试)编译此示例程序,使用新的AVX2指令提供了具有艺术品品质的代码,运行时长为11秒,可处理约10亿个reverse()。每个reverse()约为10 ns,假设2 GHz的CPU周期为0.5 ns,则我们处于20个CPU周期之内。
警告:此示例代码应作为一个不错的基准保持数年,但一旦编译器足够聪明地优化main()以便仅打印最终结果而不真正计算任何内容,它将最终开始显示其年龄。但是现在它可以用来展示reverse()。
Bit-reversal is so common...
我不知道 我几乎每天都在使用处理比特级数据的代码,而我从未想起曾经有这种特定需求。在什么情况下需要?-并不是说它本身不是一个有趣的问题。
当然,这里有很多可疑的骇客来源:http : //graphics.stanford.edu/~seander/bithacks.html#BitReverseObvious
我知道这不是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
这适用于进位位,因此您也可以保存标志
rcl
将CF转换为var1
,而不是shl
不读取标志的。(或 adc dx,dx
)。即使有了该修复程序,使用慢速loop
指令并保留var1
在内存中,这仍然是非常可笑的!实际上,我认为这应该在AX中产生输出,但是它会在结果之上保存/恢复AX的旧值。
嗯,这基本上与第一个“ reverse()”相同,但它是64位,只需要一个立即掩码即可从指令流中加载。GCC创建的代码没有跳转,因此这应该非常快。
#include <stdio.h>
static unsigned long long swap64(unsigned long long val)
{
#define ZZZZ(x,s,m) (((x) >>(s)) & (m)) | (((x) & (m))<<(s));
/* val = (((val) >>16) & 0xFFFF0000FFFF) | (((val) & 0xFFFF0000FFFF)<<16); */
val = ZZZZ(val,32, 0x00000000FFFFFFFFull );
val = ZZZZ(val,16, 0x0000FFFF0000FFFFull );
val = ZZZZ(val,8, 0x00FF00FF00FF00FFull );
val = ZZZZ(val,4, 0x0F0F0F0F0F0F0F0Full );
val = ZZZZ(val,2, 0x3333333333333333ull );
val = ZZZZ(val,1, 0x5555555555555555ull );
return val;
#undef ZZZZ
}
int main(void)
{
unsigned long long val, aaaa[16] =
{ 0xfedcba9876543210,0xedcba9876543210f,0xdcba9876543210fe,0xcba9876543210fed
, 0xba9876543210fedc,0xa9876543210fedcb,0x9876543210fedcba,0x876543210fedcba9
, 0x76543210fedcba98,0x6543210fedcba987,0x543210fedcba9876,0x43210fedcba98765
, 0x3210fedcba987654,0x210fedcba9876543,0x10fedcba98765432,0x0fedcba987654321
};
unsigned iii;
for (iii=0; iii < 16; iii++) {
val = swap64 (aaaa[iii]);
printf("A[]=%016llX Sw=%016llx\n", aaaa[iii], val);
}
return 0;
}
我很好奇明显的原始旋转有多快。在我的机器(i7 @ 2600)上,平均1,500,150,000次迭代为27.28 ns
(在131,071个64位整数的随机集合上)。
优点:所需的内存量很少,并且代码简单。我要说的也不是很大。对于任何输入,所需的时间都是可预测且恒定的(128个算术SHIFT运算+ 64逻辑AND运算+ 64逻辑OR运算)。
我将@Matt J获得的最佳时间进行了比较-他得到了公认的答案。如果我正确地阅读了他的答案,那么他得到的最好结果0.631739
就是1,000,000
迭代的秒数,这导致631 ns
每转的平均值。
我使用的代码段如下所示:
unsigned long long reverse_long(unsigned long long x)
{
return (((x >> 0) & 1) << 63) |
(((x >> 1) & 1) << 62) |
(((x >> 2) & 1) << 61) |
(((x >> 3) & 1) << 60) |
(((x >> 4) & 1) << 59) |
(((x >> 5) & 1) << 58) |
(((x >> 6) & 1) << 57) |
(((x >> 7) & 1) << 56) |
(((x >> 8) & 1) << 55) |
(((x >> 9) & 1) << 54) |
(((x >> 10) & 1) << 53) |
(((x >> 11) & 1) << 52) |
(((x >> 12) & 1) << 51) |
(((x >> 13) & 1) << 50) |
(((x >> 14) & 1) << 49) |
(((x >> 15) & 1) << 48) |
(((x >> 16) & 1) << 47) |
(((x >> 17) & 1) << 46) |
(((x >> 18) & 1) << 45) |
(((x >> 19) & 1) << 44) |
(((x >> 20) & 1) << 43) |
(((x >> 21) & 1) << 42) |
(((x >> 22) & 1) << 41) |
(((x >> 23) & 1) << 40) |
(((x >> 24) & 1) << 39) |
(((x >> 25) & 1) << 38) |
(((x >> 26) & 1) << 37) |
(((x >> 27) & 1) << 36) |
(((x >> 28) & 1) << 35) |
(((x >> 29) & 1) << 34) |
(((x >> 30) & 1) << 33) |
(((x >> 31) & 1) << 32) |
(((x >> 32) & 1) << 31) |
(((x >> 33) & 1) << 30) |
(((x >> 34) & 1) << 29) |
(((x >> 35) & 1) << 28) |
(((x >> 36) & 1) << 27) |
(((x >> 37) & 1) << 26) |
(((x >> 38) & 1) << 25) |
(((x >> 39) & 1) << 24) |
(((x >> 40) & 1) << 23) |
(((x >> 41) & 1) << 22) |
(((x >> 42) & 1) << 21) |
(((x >> 43) & 1) << 20) |
(((x >> 44) & 1) << 19) |
(((x >> 45) & 1) << 18) |
(((x >> 46) & 1) << 17) |
(((x >> 47) & 1) << 16) |
(((x >> 48) & 1) << 15) |
(((x >> 49) & 1) << 14) |
(((x >> 50) & 1) << 13) |
(((x >> 51) & 1) << 12) |
(((x >> 52) & 1) << 11) |
(((x >> 53) & 1) << 10) |
(((x >> 54) & 1) << 9) |
(((x >> 55) & 1) << 8) |
(((x >> 56) & 1) << 7) |
(((x >> 57) & 1) << 6) |
(((x >> 58) & 1) << 5) |
(((x >> 59) & 1) << 4) |
(((x >> 60) & 1) << 3) |
(((x >> 61) & 1) << 2) |
(((x >> 62) & 1) << 1) |
(((x >> 63) & 1) << 0);
}
您可能要使用标准模板库。它可能比上面提到的代码慢。但是,在我看来,它更清晰,更容易理解。
#include<bitset>
#include<iostream>
template<size_t N>
const std::bitset<N> reverse(const std::bitset<N>& ordered)
{
std::bitset<N> reversed;
for(size_t i = 0, j = N - 1; i < N; ++i, --j)
reversed[j] = ordered[i];
return reversed;
};
// test the function
int main()
{
unsigned long num;
const size_t N = sizeof(num)*8;
std::cin >> num;
std::cout << std::showbase << std::hex;
std::cout << "ordered = " << num << std::endl;
std::cout << "reversed = " << reverse<N>(num).to_ulong() << std::endl;
std::cout << "double_reversed = " << reverse<N>(reverse<N>(num)).to_ulong() << std::endl;
}
泛型
C代码。以1字节输入数据num为例。
unsigned char num = 0xaa; // 1010 1010 (aa) -> 0101 0101 (55)
int s = sizeof(num) * 8; // get number of bits
int i, x, y, p;
int var = 0; // make var data type to be equal or larger than num
for (i = 0; i < (s / 2); i++) {
// extract bit on the left, from MSB
p = s - i - 1;
x = num & (1 << p);
x = x >> p;
printf("x: %d\n", x);
// extract bit on the right, from LSB
y = num & (1 << i);
y = y >> i;
printf("y: %d\n", y);
var = var | (x << i); // apply x
var = var | (y << p); // apply y
}
printf("new: 0x%x\n", new);
怎么样:
uint reverseMSBToLSB32ui(uint input)
{
uint output = 0x00000000;
uint toANDVar = 0;
int places = 0;
for (int i = 1; i < 32; i++)
{
places = (32 - i);
toANDVar = (uint)(1 << places);
output |= (uint)(input & (toANDVar)) >> places;
}
return output;
}
小巧易用(不过,仅32位)。
我认为这是逆转位的最简单方法之一。请让我知道此逻辑是否存在任何缺陷。基本上按照这种逻辑,我们检查位的值。如果反转位置的值为1,则设置该位。
void bit_reverse(ui32 *data)
{
ui32 temp = 0;
ui32 i, bit_len;
{
for(i = 0, bit_len = 31; i <= bit_len; i++)
{
temp |= (*data & 1 << i)? (1 << bit_len-i) : 0;
}
*data = temp;
}
return;
}
// Purpose: to reverse bits in an unsigned short integer
// Input: an unsigned short integer whose bits are to be reversed
// Output: an unsigned short integer with the reversed bits of the input one
unsigned short ReverseBits( unsigned short a )
{
// declare and initialize number of bits in the unsigned short integer
const char num_bits = sizeof(a) * CHAR_BIT;
// declare and initialize bitset representation of integer a
bitset<num_bits> bitset_a(a);
// declare and initialize bitset representation of integer b (0000000000000000)
bitset<num_bits> bitset_b(0);
// declare and initialize bitset representation of mask (0000000000000001)
bitset<num_bits> mask(1);
for ( char i = 0; i < num_bits; ++i )
{
bitset_b = (bitset_b << 1) | bitset_a & mask;
bitset_a >>= 1;
}
return (unsigned short) bitset_b.to_ulong();
}
void PrintBits( unsigned short a )
{
// declare and initialize bitset representation of a
bitset<sizeof(a) * CHAR_BIT> bitset(a);
// print out bits
cout << bitset << endl;
}
// Testing the functionality of the code
int main ()
{
unsigned short a = 17, b;
cout << "Original: ";
PrintBits(a);
b = ReverseBits( a );
cout << "Reversed: ";
PrintBits(b);
}
// Output:
Original: 0000000000010001
Reversed: 1000100000000000
另一个基于循环的解决方案,当数量较少时会快速退出(在C ++中为多种类型)
template<class T>
T reverse_bits(T in) {
T bit = static_cast<T>(1) << (sizeof(T) * 8 - 1);
T out;
for (out = 0; bit && in; bit >>= 1, in >>= 1) {
if (in & 1) {
out |= bit;
}
}
return out;
}
或用C表示一个无符号的int
unsigned int reverse_bits(unsigned int in) {
unsigned int bit = 1u << (sizeof(T) * 8 - 1);
unsigned int out;
for (out = 0; bit && in; bit >>= 1, in >>= 1) {
if (in & 1)
out |= bit;
}
return out;
}
似乎还有许多其他帖子都在关注速度(即最佳=最快)。那么简单呢?考虑:
char ReverseBits(char character) {
char reversed_character = 0;
for (int i = 0; i < 8; i++) {
char ith_bit = (c >> i) & 1;
reversed_character |= (ith_bit << (sizeof(char) - 1 - i));
}
return reversed_character;
}
并希望聪明的编译器能够为您优化。
如果要反转更长的位列表(包含sizeof(char) * n
位),可以使用此函数获取:
void ReverseNumber(char* number, int bit_count_in_number) {
int bytes_occupied = bit_count_in_number / sizeof(char);
// first reverse bytes
for (int i = 0; i <= (bytes_occupied / 2); i++) {
swap(long_number[i], long_number[n - i]);
}
// then reverse bits of each individual byte
for (int i = 0; i < bytes_occupied; i++) {
long_number[i] = ReverseBits(long_number[i]);
}
}
这会将[10000000,10101010]反转为[01010101,00000001]。
ith_bit = (c >> i) & 1
。还可以通过移位reversed_char
而不是移位来保存SUB ,除非您希望它将在x86上编译为sub something
/ bts reg,reg
以将目标寄存器中的第n位设置为SUB 。
伪代码中的位反转
源->要反转的字节b00101100目标->反转,也需要是无符号类型,因此不会向下传播符号位
复制到临时文件中以使原始文件不受影响,也需要为无符号类型,以使符号位不会自动移入
bytecopy = b0010110
LOOP8://进行8次测试,如果bytecopy <0(负)
set bit8 (msb) of reversed = reversed | b10000000
else do not set bit8
shift bytecopy left 1 place
bytecopy = bytecopy << 1 = b0101100 result
shift result right 1 place
reversed = reversed >> 1 = b00000000
8 times no then up^ LOOP8
8 times yes then done.
我的简单解决方案
BitReverse(IN)
OUT = 0x00;
R = 1; // Right mask ...0000.0001
L = 0; // Left mask 1000.0000...
L = ~0;
L = ~(i >> 1);
int size = sizeof(IN) * 4; // bit size
while(size--){
if(IN & L) OUT = OUT | R; // start from MSB 1000.xxxx
if(IN & R) OUT = OUT | L; // start from LSB xxxx.0001
L = L >> 1;
R = R << 1;
}
return OUT;
i
啊 另外,那个魔法常数是* 4
什么?是CHAR_BIT / 2
吗
这是32位,如果考虑8位,则需要更改大小。
void bitReverse(int num)
{
int num_reverse = 0;
int size = (sizeof(int)*8) -1;
int i=0,j=0;
for(i=0,j=size;i<=size,j>=0;i++,j--)
{
if((num >> i)&1)
{
num_reverse = (num_reverse | (1<<j));
}
}
printf("\n rev num = %d\n",num_reverse);
}
以LSB-> MSB的顺序读取输入整数“ num”,并以MSB-> LSB的顺序存储在num_reverse中。