我必须计算汉明重量以获得相当快速的64位数据流，并使用popcnt汇编指令引起了我的英特尔酷睿i7-4650U的例外。
我检查了我的圣经黑客的喜悦，并在网上搜索了各种算法（自从他们开始解决这个问题＆＃39;计算机诞生之后）
我在周末玩了一些我自己的想法并提出了这些算法，我几乎可以将数据移入和移出CPU。
//64-bit popcnt using BMI2
_popcnt_bmi2:
mov (%rdi),%r11
pext %r11,%r11,%r11
not %r11
tzcnt %r11,%r11
mov %r11,(%rdx)
add $8h,%rdi
add $8h,%rdx
dec %rsi
jnz _popcnt_bmi2
ret
在上面的代码中，我使用pext（BMI2），其中传入的数据使用自身作为掩码。然后，所有存在的位将从结果寄存器中的最低有效位（本身再次）开始崩溃。然后我需要计算折叠位的数量，以便反转所有位，然后使用tzcnt来计算现在为零的数量。我认为这是个不错的主意。
然后我也尝试了AVX2方法：
//64-bit popcnt using AVX2
_popcnt_avx2:
vmovdqa (%rcx),%ymm2
add $20h,%rcx
vmovdqa (%rcx),%ymm3
add $20h,%rcx
vmovdqa (%rcx),%ymm4
popcnt_avx2_loop:
vmovdqa (%rdi),%ymm0
vpand %ymm0, %ymm2, %ymm1
vpandn %ymm0, %ymm2, %ymm0
vpsrld $4h,%ymm0, %ymm0
vpshufb %ymm1, %ymm3, %ymm1
vpshufb %ymm0, %ymm3, %ymm0
vpaddb %ymm1,%ymm0,%ymm0 //popcnt (8-bits)
vpsadbw %ymm0,%ymm4,%ymm0 //popcnt (64-bits)
vmovdqa %ymm0,(%rdx)
add $20h,%rdi
add $20h,%rdx
dec %rsi
jnz popcnt_avx2_loop
在AVX2的情况下，我读取了32个字节，然后屏蔽了半字节（ymm2），然后我使用ymm3作为查找表来对半字节进行位计数。然后我将结果添加到8位，然后我使用超级压缩vpsadbw将8个字节添加到64位值（ymm4 = 0）。
任何人都能更快地穿上他们的衣服？
修改
失败的POPCNT是由于我在代码中犯了一个错误，该功能与我的英特尔酷睿i7-4650U有关。请参阅下面的帖子，显示工作台成绩。
答案 0 :(得分：1)
好的结论是，我不知道是不是想要变得聪明，而是为了我的目标：
内置的内在popcount：_mm_popcnt_u64
bmi2：__tzcnt_u64(~_pext_u64(data[i],data[i]));
针对三个汇编函数
popcnt，bmi2和avx2。
它们都以你可以将记忆移入和移出我的速度运行：
cat /proc/cpuinfo
-Intel（R）Xeon（R）CPU E3-1275 v3 @ 3.50GHz
供参考：
main.c中：
// Hamming weight bench
#include <stdio.h>
#include <string.h>
#include <stdint.h>
#include <stdlib.h>
#include <math.h>
#include <sys/time.h>
#include <smmintrin.h>
#include <immintrin.h>
#include <x86intrin.h>
#include <math.h>
#define DISPLAY_HEIGHT 4
#define DISPLAY_WIDTH 32
#define NUM_DATA_OBJECTS 40000000
#define ITTERATIONS 20
// The source data (+32 to avoid the quantization out of memory problem)
__attribute__ ((aligned(32))) static long long unsigned data[NUM_DATA_OBJECTS+32]={};
__attribute__ ((aligned(32))) static long long unsigned data_out[NUM_DATA_OBJECTS+32]={};
__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,
0x00,0x01,0x01,0x02,0x01,0x02,0x02,0x03,0x01,0x02,0x02,0x03,0x02,0x03,0x03,0x04,0x00,0x01,0x01,0x02,0x01,0x02,0x02,0x03,0x01,0x02,0x02,0x03,0x02,0x03,0x03,0x04,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00
};
extern "C" {
void popcnt_popcnt(long long unsigned[],unsigned int,long long unsigned[]);
void popcnt_bmi2(long long unsigned[],unsigned int,long long unsigned[]);
void popcnt_avx2(long long unsigned[],unsigned int,long long unsigned[],unsigned char[]);
}
void populate_data()
{
for(unsigned int i = 0; i < NUM_DATA_OBJECTS; i++)
{
data[i] = rand();
}
}
void display_source_data()
{
printf ("\r\nData in(start):\r\n");
for (unsigned int j = 0; j < DISPLAY_HEIGHT; j++)
{
for (unsigned int i = 0; i < DISPLAY_WIDTH; i++)
{
printf ("0x%02llux,",data[i+(j*DISPLAY_WIDTH)]);
}
printf ("\r\n");
}
}
void bench_popcnt()
{
for(unsigned int i = 0; i < NUM_DATA_OBJECTS; i++)
{
data_out[i] = _mm_popcnt_u64(data[i]);
}
}
void bench_move_data_memcpy()
{
memcpy(data_out,data,NUM_DATA_OBJECTS*8);
}
// __tzcnt64 ??
void bench_bmi2()
{
for(unsigned int i = 0; i < NUM_DATA_OBJECTS; i++)
{
data_out[i]=__tzcnt_u64(~_pext_u64(data[i],data[i]));
}
}
void display_dest_data()
{
printf ("\r\nData out:\r\n");
for (unsigned int j = 0; j < DISPLAY_HEIGHT; j++)
{
for (unsigned int i = 0; i < DISPLAY_WIDTH; i++)
{
printf ("0x%02llux,",data_out[i+(j*DISPLAY_WIDTH)]);
}
printf ("\r\n");
}
}
int main() {
struct timeval t0;
struct timeval t1;
long elapsed[ITTERATIONS]={0};
long avrg=0;
for (unsigned int i = 0; i < ITTERATIONS; i++)
{
populate_data();
// display_source_data();
gettimeofday(&t0, 0);
bench_move_data_memcpy();
gettimeofday(&t1, 0);
elapsed[i]= (((t1.tv_sec-t0.tv_sec)*1000000 + t1.tv_usec-t0.tv_usec)/1000);
printf ("Time_to_move_data_without_processing: %ld\n",elapsed[i]);
}
avrg=0;
for (unsigned int i = 1; i < ITTERATIONS; i++){
avrg+=elapsed[i];
}
printf ("Average time_to_move_data: %ld\n",avrg/(ITTERATIONS-1));
//display_dest_data();
for (unsigned int i = 0; i < ITTERATIONS; i++)
{
populate_data();
// display_source_data();
gettimeofday(&t0, 0);
bench_popcnt();
gettimeofday(&t1, 0);
elapsed[i] = ((t1.tv_sec-t0.tv_sec)*1000000 + t1.tv_usec-t0.tv_usec)/1000;
printf ("popcnt: %ld\n",elapsed[i]);
}
avrg=0;
for (unsigned int i = 1; i < ITTERATIONS; i++){
avrg+=elapsed[i];
}
printf ("Average popcnt: %ld\n",avrg/(ITTERATIONS-1));
//display_dest_data();
for (unsigned int i = 0; i < ITTERATIONS; i++)
{
populate_data();
// display_source_data();
gettimeofday(&t0, 0);
bench_bmi2();
gettimeofday(&t1, 0);
elapsed[i] = ((t1.tv_sec-t0.tv_sec)*1000000 + t1.tv_usec-t0.tv_usec)/1000;
printf ("bmi2: %ld\n",elapsed[i]);
}
avrg=0;
for (unsigned int i = 1; i < ITTERATIONS; i++){
avrg+=elapsed[i];
}
printf ("Average bmi2: %ld\n",avrg/(ITTERATIONS-1));
//display_dest_data();
printf ("Now test the assembler functions\n");
for (unsigned int i = 0; i < ITTERATIONS; i++)
{
populate_data();
// display_source_data();
gettimeofday(&t0, 0);
popcnt_popcnt(data,NUM_DATA_OBJECTS,data_out);
gettimeofday(&t1, 0);
elapsed[i] = ((t1.tv_sec-t0.tv_sec)*1000000 + t1.tv_usec-t0.tv_usec)/1000;
printf ("popcnt_asm: %ld\n",elapsed[i]);
}
avrg=0;
for (unsigned int i = 1; i < ITTERATIONS; i++){
avrg+=elapsed[i];
}
printf ("Average popcnt_asm: %ld\n",avrg/(ITTERATIONS-1));
//display_dest_data();
for (unsigned int i = 0; i < ITTERATIONS; i++)
{
populate_data();
// display_source_data();
gettimeofday(&t0, 0);
popcnt_bmi2(data,NUM_DATA_OBJECTS,data_out);
gettimeofday(&t1, 0);
elapsed[i] = ((t1.tv_sec-t0.tv_sec)*1000000 + t1.tv_usec-t0.tv_usec)/1000;
printf ("bmi2_asm: %ld\n",elapsed[i]);
}
avrg=0;
for (unsigned int i = 1; i < ITTERATIONS; i++){
avrg+=elapsed[i];
}
printf ("Average bmi2_asm: %ld\n",avrg/(ITTERATIONS-1));
//display_dest_data();
for (unsigned int i = 0; i < ITTERATIONS; i++)
{
populate_data();
// display_source_data();
gettimeofday(&t0, 0);
popcnt_avx2(data,(unsigned int)ceil((NUM_DATA_OBJECTS*8)/32.0),data_out,k1);
gettimeofday(&t1, 0);
elapsed[i] = ((t1.tv_sec-t0.tv_sec)*1000000 + t1.tv_usec-t0.tv_usec)/1000;
printf ("avx2_asm: %ld\n",elapsed[i]);
}
avrg=0;
for (unsigned int i = 1; i < ITTERATIONS; i++){
avrg+=elapsed[i];
}
printf ("Average avx2_asm: %ld\n",avrg/(ITTERATIONS-1));
//display_dest_data();
return 0;
}
引擎。
//
// avx2_bmi2_popcnt bench
//
.global popcnt_bmi2 , popcnt_avx2, popcnt_popcnt
.align 2
//64-bit popcnt using the built-in popcnt instruction
popcnt_popcnt:
popcntq (%rdi), %r11
mov %r11,(%rdx)
add $8,%rdi
add $8,%rdx
dec %rsi
jnz popcnt_popcnt
ret
//64-bit popcnt using BMI2
popcnt_bmi2:
mov (%rdi),%r11
pextq %r11,%r11,%r11
not %r11
tzcnt %r11,%r11
mov %r11,(%rdx)
add $8,%rdi
add $8,%rdx
dec %rsi
jnz popcnt_bmi2
ret
//64-bit popcnt using AVX2
popcnt_avx2:
vmovdqa (%rcx),%ymm2
add $0x20,%rcx
vmovdqa (%rcx),%ymm3
add $0x20,%rcx
vmovdqa (%rcx),%ymm4
popcnt_avx2_loop:
vmovdqa (%rdi),%ymm0
vpand %ymm0, %ymm2, %ymm1
vpandn %ymm0, %ymm2, %ymm0
vpsrld $4,%ymm0, %ymm0
vpshufb %ymm1, %ymm3, %ymm1
vpshufb %ymm0, %ymm3, %ymm0
vpaddb %ymm1,%ymm0,%ymm0
vpsadbw %ymm0,%ymm4,%ymm0
vmovdqa %ymm0,(%rdx)
add $0x20,%rdi
add $0x20,%rdx
dec %rsi
jnz popcnt_avx2_loop
ret
编译来源：
g++ -march=native -mavx -mpopcnt -O3 main.c engine.s
将CPU设置为性能：
cpufreq-set -g performance
跑板凳：
sudo chrt -r 10 ./a.out
结果：
平均时间_to_move_data：61
平均popcnt：61
平均值bmi2：61
现在测试汇编程序函数
平均popcnt_asm：61
平均bmi2_asm：61
平均avx2_asm：61
答案 1 :(得分：0)
您是否尝试过基于表格的方法，例如：
unsigned char bitcnt[256] = {0,1,1,2,1, ... ,7,8};
unsigned char* p = &the64bitWord;
nbits = bitcnt[p[0]]
+ bitcnt[p[1]]
+ bitcnt[p[2]]
...
+ bitcnt[p[7]];
或者也许你自己在asm中滚动。