我正在IntelXeon®Phi®上实现超快速弹出计数，因为它是各种生物信息学软件的性能热点。
我已经实现了五段代码，
#if defined(__MIC__)
#include <zmmintrin.h>
__attribute__((align(64))) static const uint32_t POPCOUNT_4bit[16] = {0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4};
__attribute__((align(64))) static const uint32_t MASK_4bit[16] = {0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF, 0xF};
inline uint64_t vpu_popcount1(uint64_t* buf, size_t n) {
register size_t result = 0;
size_t i;
register const __m512i popcnt = _mm512_load_epi32((void*)POPCOUNT_4bit);
register const __m512i mask = _mm512_load_epi32((void*)MASK_4bit);
register __m512i total;
register __m512i shuf;
#pragma unroll(8)
for (i = 0; i < n; i+=8) {
shuf = _mm512_load_epi32(&buf[i]);
_mm_prefetch((const char *)&buf[i+256], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+64], _MM_HINT_T0); // vprefetch0
total = _mm512_setzero_epi32();
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(shuf, mask), popcnt), total);
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(_mm512_srli_epi32(shuf, 4), mask), popcnt), total);
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(_mm512_srli_epi32(shuf, 8), mask), popcnt), total);
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(_mm512_srli_epi32(shuf, 12), mask), popcnt), total);
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(_mm512_srli_epi32(shuf, 16), mask), popcnt), total);
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(_mm512_srli_epi32(shuf, 20), mask), popcnt), total);
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(_mm512_srli_epi32(shuf, 24), mask), popcnt), total);
total = _mm512_add_epi32(_mm512_permutevar_epi32(_mm512_and_epi32(_mm512_srli_epi32(shuf, 28), mask), popcnt), total);
/* Reduce add, which is analogous to SSSE3's PSADBW instruction,
is not implementated as a single instruction in VPUv1, thus
emulated by multiple instructions*/
result += _mm512_reduce_add_epi32(total);
}
return result;
}
__attribute__((align(64))) static const unsigned magic[] = {\
0x55555555, 0x55555555, 0x55555555, 0x55555555,\
0x55555555, 0x55555555, 0x55555555, 0x55555555,\
0x55555555, 0x55555555, 0x55555555, 0x55555555,\
0x55555555, 0x55555555, 0x55555555, 0x55555555,\
0x33333333, 0x33333333, 0x33333333, 0x33333333,\
0x33333333, 0x33333333, 0x33333333, 0x33333333,\
0x33333333, 0x33333333, 0x33333333, 0x33333333,\
0x33333333, 0x33333333, 0x33333333, 0x33333333,\
0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F,\
0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F,\
0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F,\
0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F,\
0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF,\
0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF,\
0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF,\
0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF,\
0x0000FFFF, 0x0000FFFF, 0x0000FFFF, 0x0000FFFF,\
0x0000FFFF, 0x0000FFFF, 0x0000FFFF, 0x0000FFFF,\
0x0000FFFF, 0x0000FFFF, 0x0000FFFF, 0x0000FFFF,\
0x0000FFFF, 0x0000FFFF, 0x0000FFFF, 0x0000FFFF,\
0x000000FF, 0x000000FF, 0x000000FF, 0x000000FF,\
0x000000FF, 0x000000FF, 0x000000FF, 0x000000FF,\
0x000000FF, 0x000000FF, 0x000000FF, 0x000000FF,\
0x000000FF, 0x000000FF, 0x000000FF, 0x000000FF
};
inline uint64_t vpu_popcount2(uint64_t* buf, size_t n) {
register size_t result = 0;
size_t i;
register const __m512i B0 = _mm512_load_epi32((void*)(magic+0));
register const __m512i B1 = _mm512_load_epi32((void*)(magic+16));
register const __m512i B2 = _mm512_load_epi32((void*)(magic+32));
register const __m512i B3 = _mm512_load_epi32((void*)(magic+48));
register const __m512i B4 = _mm512_load_epi32((void*)(magic+64));
register __m512i total;
register __m512i shuf;
#pragma unroll(8)
for (i = 0; i < n; i+=8) {
shuf = _mm512_load_epi32(&buf[i]);
_mm_prefetch((const char *)&buf[i+512], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+64], _MM_HINT_T0); // vprefetch0
total = _mm512_sub_epi32(shuf, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf,1)));
total = _mm512_add_epi32(_mm512_and_epi32(B1, total), _mm512_and_epi32(B1,_mm512_srli_epi32(total,2)));
total = _mm512_and_epi32(B2, _mm512_add_epi32(total, _mm512_srli_epi32(total,4)));
total = _mm512_and_epi32(B3, _mm512_add_epi32(total, _mm512_srli_epi32(total,8)));
total = _mm512_and_epi32(B4, _mm512_add_epi32(total, _mm512_srli_epi32(total,16)));
/* Reduce add, which is analogous to SSSE3's PSADBW instruction,
is not implementated as a single instruction in VPUv1, thus
emulated by multiple instructions*/
result += _mm512_reduce_add_epi32(total);
}
return result;
}
inline uint64_t vpu_popcount3(uint64_t* buf, size_t n) {
register size_t result = 0;
size_t i;
register const __m512i B0 = _mm512_load_epi32((void*)(magic+0));
register const __m512i B1 = _mm512_load_epi32((void*)(magic+16));
register const __m512i B2 = _mm512_load_epi32((void*)(magic+32));
register const __m512i B3 = _mm512_load_epi32((void*)(magic+48));
register const __m512i B4 = _mm512_load_epi32((void*)(magic+64));
register __m512i total;
register __m512i shuf;
#pragma unroll(4)
for (i = 0; i < n; i+=16) {
shuf = _mm512_load_epi32(&buf[i]);
result += _mm_countbits_64(buf[i+8]);
_mm_prefetch((const char *)&buf[i+512], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+576], _MM_HINT_T1); // vprefetch1
result += _mm_countbits_64(buf[i+9]);
_mm_prefetch((const char *)&buf[i+64], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+128], _MM_HINT_T0); // vprefetch0
total = _mm512_sub_epi32(shuf, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf,1)));
result += _mm_countbits_64(buf[i+10]);
total = _mm512_add_epi32(_mm512_and_epi32(B1, total), _mm512_and_epi32(B1,_mm512_srli_epi32(total,2)));
result += _mm_countbits_64(buf[i+11]);
total = _mm512_and_epi32(B2, _mm512_add_epi32(total, _mm512_srli_epi32(total,4)));
result += _mm_countbits_64(buf[i+12]);
total = _mm512_and_epi32(B3, _mm512_add_epi32(total, _mm512_srli_epi32(total,8)));
result += _mm_countbits_64(buf[i+13]);
total = _mm512_and_epi32(B4, _mm512_add_epi32(total, _mm512_srli_epi32(total,16)));
result += _mm_countbits_64(buf[i+14]);
/* Reduce add, which is analogous to SSSE3's PSADBW instruction,
is not implementated as a single instruction in VPUv1, thus
emulated by multiple instructions*/
result += _mm512_reduce_add_epi32(total);
result += _mm_countbits_64(buf[i+15]);
}
return result;
}
/* Using VPU or SSE's machine intrinsic, CPUs not supporting SIMD
* will use compiler's implementation, the speed of which depends */
static inline size_t scalar_popcountu(unsigned *buf, size_t n) {
register size_t cnt = 0;
size_t i;
#pragma vector always
#pragma unroll(8)
for (i = 0; i < n; i++) {
cnt += _mm_countbits_32(buf[i]);
_mm_prefetch((const char *)&buf[i+512], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+64], _MM_HINT_T0); // vprefetch0
}
return cnt;
}
static inline size_t scalar_popcountlu(uint64_t *buf, size_t n) {
register size_t cnt = 0;
size_t i;
#pragma vector always
#pragma unroll(8)
for (i = 0; i < n; i++) {
cnt += _mm_countbits_64(buf[i]);
_mm_prefetch((const char *)&buf[i+512], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+64], _MM_HINT_T0); // vprefetch0
}
return cnt;
}
#endif
可以从https://www.dropbox.com/sh/b3sfqps19wa2oi4/iFQ9wQ1NTg下载支持OpenMP的代码总结
该代码是使用Intel C/C++编译器XE 13使用以下命令编译的:
icc -debug inline-debug-info -O3 -mmic -fno-alias -ansi-alias -opt-streaming-stores always -ipo popcnt-mmic.cpp -o popcnt-mmic -vec-report=2 -openmp
该代码在 native (61个内核)上以“122线程”和线程亲和性通过导出以“平衡”的方式在 native 上运行:
export OMP_NUM_THREADS=122;export KMP_AFFINITY=balanced
我正在使用至强融核SE10p，B1步进，CentOS6.4
测试28兆字节的垃圾邮件(由rand()填充)并进行10000次迭代，其性能如下:
Buffer allocated at: 0x7f456b000000
OpenMP scalar_popcountu 4310169 us; cnt = 28439328
OpenMP scalar_popcountlu 1421139 us; cnt = 28439328
OpenMP vpu_popcount 1489992 us; cnt = 28439328
OpenMP vpu_popcount2 1109530 us; cnt = 28439328
OpenMP vpu_popcount3 951122 us; cnt = 28439328
“scalar_popcountu”和“scalar_popcountlu”分别使用“_mm_countbits_32”和“_mm_countbits_64”内在函数，它们使用了标量“popcnt”指令。设置“#pragma vector always”会要求编译器将负载和总和向量化为16个无符号整数或8个无符号长整型时间，尽管popcount本身仍是标量指令。
vpu_popcount1的实现类似于SSSE3 popcount的实现http://wm.ite.pl/articles/sse-popcount.html。但是，1)Xeon Phi不支持对整数的压缩字节操作(最小为双字，也就是32位)，并且2)它不实现“绝对差的压缩总和”指令(如SSSE3中的_mm_sad_epu8)，因此通过四组“vpermf32x4”，“vpaddd”和“movslq”的组合执行还原添加。因此，与原始SSSE3版本相比，该实现产生了更多的指令。
vpu_popcount2的实现类似于SSE2 popcount的实现(可以引用“Hacker's Delight”)。与vpu_popcount1相比，该实现生成的指令更少，并且速度提高了约30％。但是，繁琐的“减少添加”仍然无法避免。
vpu_popcount3的实现是至强融核所特有的。 vector 和标量运算的混合使用，比vpu_popcount2快15％(在我的实现中， vector 运算中散布了标量运算，我可以根据编译器生成的汇编代码重新排列标量运算，但是可以预期得到改进)就我而言是有限的)。该改进基于以下观察:1)Xeon Phi是按顺序调度的； 2)每个时钟周期可以发出两个标量指令或“1 vector + 1标量”指令。我已将展开次数从8减少到4，以避免寄存器文件饱和。
在每个函数中，从内存到L2 8个循环的显式预取以及从L2到L1 1循环的显式预取已将L1命中率从0.38提高到了0.994。
展开确实会使性能提高约15％。由于Xeon Phi是按顺序安排的，因此这很不直观。但是展开可以使icc编译器进行尽可能多的编译时间调度。
我们是否还有更多技术可以提高性能？
布赖恩·尼克森(Brian Nickerson)的两段更快的代码，
OpenMP vpu_popcount2 1110737 us; cnt = 28439328
OpenMP vpu_popcount3 951459 us; cnt = 28439328
OpenMP vpu_popcount3_r 815126 us; cnt = 28439328
OpenMP vpu_popcount5 746852 us; cnt = 28439328
vpu_popcount3_revised:
inline uint64_t vpu_popcount3_revised(uint64_t* buf, size_t n) {
_mm_prefetch((const char *)&buf[0], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[8], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[16], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[24], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[32], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[40], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[48], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[56], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[64], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[72], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[80], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[88], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[96], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[104], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[112], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[120], _MM_HINT_T1); // vprefetch1
register size_t result;
size_t i;
register const __m512i B0 = _mm512_load_epi32((void*)(magic+0));
register const __m512i B1 = _mm512_load_epi32((void*)(magic+16));
register const __m512i B2 = _mm512_load_epi32((void*)(magic+32));
register const __m512i B3 = _mm512_load_epi32((void*)(magic+48));
register const __m512i B4 = _mm512_load_epi32((void*)(magic+64));
register __m512i total0;
register __m512i total1;
register __m512i shuf0;
register __m512i shuf1;
register __m512i result0;
register __m512i result1;
result0 = _mm512_setzero_epi32();
result1 = _mm512_setzero_epi32();
for (i = 0; i < n; i+=16) {
shuf0 = _mm512_load_epi32(&buf[i ]);
shuf1 = _mm512_load_epi32(&buf[i+8]);
_mm_prefetch((const char *)&buf[i+128], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+136], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+16], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+24], _MM_HINT_T0); // vprefetch0
total0 = _mm512_sub_epi32(shuf0, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf0,1)));
total1 = _mm512_sub_epi32(shuf1, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf1,1)));
total0 = _mm512_add_epi32(_mm512_and_epi32(B1, total0), _mm512_and_epi32(B1,_mm512_srli_epi32(total0,2)));
total1 = _mm512_add_epi32(_mm512_and_epi32(B1, total1), _mm512_and_epi32(B1,_mm512_srli_epi32(total1,2)));
total0 = _mm512_and_epi32(B2, _mm512_add_epi32(total0, _mm512_srli_epi32(total0,4)));
total1 = _mm512_and_epi32(B2, _mm512_add_epi32(total1, _mm512_srli_epi32(total1,4)));
total0 = _mm512_and_epi32(B3, _mm512_add_epi32(total0, _mm512_srli_epi32(total0,8)));
total1 = _mm512_and_epi32(B3, _mm512_add_epi32(total1, _mm512_srli_epi32(total1,8)));
total0 = _mm512_and_epi32(B4, _mm512_add_epi32(total0, _mm512_srli_epi32(total0,16)));
total1 = _mm512_and_epi32(B4, _mm512_add_epi32(total1, _mm512_srli_epi32(total1,16)));
result0 = _mm512_add_epi32(result0,total0);
result1 = _mm512_add_epi32(result1,total1);
}
result0 = _mm512_add_epi32(result0,result1);
result = _mm512_reduce_add_epi32(result0);
return result;
}
vpu_popcount5:
inline uint64_t vpu_popcount5(uint64_t* buf, size_t n) {
_mm_prefetch((const char *)&buf[0], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[8], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[16], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[24], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[32], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[40], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[48], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[56], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[64], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[72], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[80], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[88], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[96], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[104], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[112], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[120], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[128], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[136], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[144], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[152], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[160], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[168], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[176], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[184], _MM_HINT_T1); // vprefetch1
register size_t result;
size_t i;
register const __m512i B0 = _mm512_load_epi32((void*)(magic+0));
register const __m512i B1 = _mm512_load_epi32((void*)(magic+16));
register const __m512i B2 = _mm512_load_epi32((void*)(magic+32));
register const __m512i B3 = _mm512_load_epi32((void*)(magic+48));
register const __m512i B4 = _mm512_load_epi32((void*)(magic+64));
register const __m512i B6 = _mm512_load_epi32((void*)(magic+80));
register __m512i total0;
register __m512i total1;
register __m512i total2;
register __m512i total3;
register __m512i shuf0;
register __m512i shuf1;
register __m512i shuf2;
register __m512i shuf3;
register __m512i result0;
register __m512i result1;
result0 = _mm512_setzero_epi32();
result1 = _mm512_setzero_epi32();
for (i = 0; i < n; i+=32) {
shuf0 = _mm512_load_epi32(&buf[i ]);
shuf1 = _mm512_load_epi32(&buf[i+ 8]);
shuf2 = _mm512_load_epi32(&buf[i+16]);
shuf3 = _mm512_load_epi32(&buf[i+24]);
_mm_prefetch((const char *)&buf[i+192], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+200], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+208], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+216], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+32], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+40], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+48], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+56], _MM_HINT_T0); // vprefetch0
total0 = _mm512_sub_epi32(shuf0, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf0,1))); // max value in nn is 10
total1 = _mm512_sub_epi32(shuf1, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf1,1)));
total2 = _mm512_sub_epi32(shuf2, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf2,1)));
total3 = _mm512_sub_epi32(shuf3, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf3,1)));
total0 = _mm512_add_epi32(_mm512_and_epi32(B1, total0), _mm512_and_epi32(B1,_mm512_srli_epi32(total0,2))); // max value in nnnn is 0100
total1 = _mm512_add_epi32(_mm512_and_epi32(B1, total1), _mm512_and_epi32(B1,_mm512_srli_epi32(total1,2)));
total2 = _mm512_add_epi32(_mm512_and_epi32(B1, total2), _mm512_and_epi32(B1,_mm512_srli_epi32(total2,2)));
total3 = _mm512_add_epi32(_mm512_and_epi32(B1, total3), _mm512_and_epi32(B1,_mm512_srli_epi32(total3,2)));
total0 = _mm512_and_epi32(B2, _mm512_add_epi32(total0, _mm512_srli_epi32(total0,4))); // max value in 0000nnnn is 00001000
total1 = _mm512_and_epi32(B2, _mm512_add_epi32(total1, _mm512_srli_epi32(total1,4)));
total2 = _mm512_and_epi32(B2, _mm512_add_epi32(total2, _mm512_srli_epi32(total2,4)));
total3 = _mm512_and_epi32(B2, _mm512_add_epi32(total3, _mm512_srli_epi32(total3,4)));
total0 = _mm512_add_epi32(total0, total1); // max value in 000nnnnn is 00010000
total1 = _mm512_add_epi32(total2, total3);
total0 = _mm512_add_epi32(total0, _mm512_srli_epi32(total0,8)); // max value in xxxxxxxx00nnnnnn is 00100000
total1 = _mm512_add_epi32(total1, _mm512_srli_epi32(total1,8));
total0 = _mm512_and_epi32(B6, _mm512_add_epi32(total0, _mm512_srli_epi32(total0,16))); // max value in each element is 01000000, i.e. 64
total1 = _mm512_and_epi32(B6, _mm512_add_epi32(total1, _mm512_srli_epi32(total1,16)));
result0 = _mm512_add_epi32(result0,total0);
result1 = _mm512_add_epi32(result1,total1);
}
result0 = _mm512_add_epi32(result0,result1);
result = _mm512_reduce_add_epi32(result0);
return result;
}
最佳答案
自昨天发布以来，我已经能够在自己的卡上运行您的代码和我的建议。由于硬件的步进，可能与我的编译器的版本有关，所以我得到的计时与您不完全相同。但是这种趋势持续了下来，我的建议似乎使性能提高了约15％。
如下面的代码所示，我还做了一些小的改进，提高了5％到10％，并进行了一些调整。请注意，在以下代码段中，B6的每个元素均设置为0x000000FF。在这一点上，我认为该算法可能会逼近从GDDR到L2缓存的最大可持续带宽。
(已添加注释:此断言的一个证明是，如果我用重复十次的for循环包装popcount5函数的主体，请注意，这是输入数据的“chunk_size”的十次快速重复，因此在L2中，有9次会变得很热-测试的总时间仅增加了大约5而不是10倍。我提出这一点是因为我认为您的目标是调整位计数的速度逻辑，但也许您希望在其中部署的应用程序实际上具有较小的和/或较热的工作集；如果是这样，则由DRAM-> L2带宽引入的节流使情况变得模糊不清。测试输入的数据，使其倾向于在L2中保持较高的温度，似乎会导致其他开销(可能是openmp开销)变得相对更重要。)
inline uint64_t vpu_popcount5(uint64_t* buf, size_t n) {
_mm_prefetch((const char *)&buf[0], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[8], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[16], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[24], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[32], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[40], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[48], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[56], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[64], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[72], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[80], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[88], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[96], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[104], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[112], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[120], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[128], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[136], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[144], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[152], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[160], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[168], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[176], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[184], _MM_HINT_T1); // vprefetch1
register size_t result;
size_t i;
register const __m512i B0 = _mm512_load_epi32((void*)(magic+0));
register const __m512i B1 = _mm512_load_epi32((void*)(magic+16));
register const __m512i B2 = _mm512_load_epi32((void*)(magic+32));
register const __m512i B6 = _mm512_load_epi32((void*)(magic+80));
register __m512i total0;
register __m512i total1;
register __m512i total2;
register __m512i total3;
register __m512i shuf0;
register __m512i shuf1;
register __m512i shuf2;
register __m512i shuf3;
register __m512i result0;
register __m512i result1;
result0 = _mm512_setzero_epi32();
result1 = _mm512_setzero_epi32();
for (i = 0; i < n; i+=32) {
shuf0 = _mm512_load_epi32(&buf[i ]);
shuf1 = _mm512_load_epi32(&buf[i+ 8]);
shuf2 = _mm512_load_epi32(&buf[i+16]);
shuf3 = _mm512_load_epi32(&buf[i+24]);
_mm_prefetch((const char *)&buf[i+192], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+200], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+208], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+216], _MM_HINT_T1); // vprefetch1
_mm_prefetch((const char *)&buf[i+32], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+40], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+48], _MM_HINT_T0); // vprefetch0
_mm_prefetch((const char *)&buf[i+56], _MM_HINT_T0); // vprefetch0
total0 = _mm512_sub_epi32(shuf0, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf0,1))); // max value in nn is 10
total1 = _mm512_sub_epi32(shuf1, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf1,1)));
total2 = _mm512_sub_epi32(shuf2, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf2,1)));
total3 = _mm512_sub_epi32(shuf3, _mm512_and_epi32(B0, _mm512_srli_epi32(shuf3,1)));
total0 = _mm512_add_epi32(_mm512_and_epi32(B1, total0), _mm512_and_epi32(B1,_mm512_srli_epi32(total0,2))); // max value in nnnn is 0100
total1 = _mm512_add_epi32(_mm512_and_epi32(B1, total1), _mm512_and_epi32(B1,_mm512_srli_epi32(total1,2)));
total2 = _mm512_add_epi32(_mm512_and_epi32(B1, total2), _mm512_and_epi32(B1,_mm512_srli_epi32(total2,2)));
total3 = _mm512_add_epi32(_mm512_and_epi32(B1, total3), _mm512_and_epi32(B1,_mm512_srli_epi32(total3,2)));
total0 = _mm512_and_epi32(B2, _mm512_add_epi32(total0, _mm512_srli_epi32(total0,4))); // max value in 0000nnnn is 00001000
total1 = _mm512_and_epi32(B2, _mm512_add_epi32(total1, _mm512_srli_epi32(total1,4)));
total2 = _mm512_and_epi32(B2, _mm512_add_epi32(total2, _mm512_srli_epi32(total2,4)));
total3 = _mm512_and_epi32(B2, _mm512_add_epi32(total3, _mm512_srli_epi32(total3,4)));
total0 = _mm512_add_epi32(total0, total1); // max value in 000nnnnn is 00010000
total1 = _mm512_add_epi32(total2, total3);
total0 = _mm512_add_epi32(total0, _mm512_srli_epi32(total0,8)); // max value in xxxxxxxx00nnnnnn is 00100000
total1 = _mm512_add_epi32(total1, _mm512_srli_epi32(total1,8));
total0 = _mm512_and_epi32(B6, _mm512_add_epi32(total0, _mm512_srli_epi32(total0,16))); // max value in each element is 01000000, i.e. 64
total1 = _mm512_and_epi32(B6, _mm512_add_epi32(total1, _mm512_srli_epi32(total1,16)));
result0 = _mm512_add_epi32(result0,total0);
result1 = _mm512_add_epi32(result1,total1);
/* Reduce add, which is analogous to SSSE3's PSADBW instruction,
is not implementated as a single instruction in VPUv1, thus
emulated by multiple instructions*/
}
result0 = _mm512_add_epi32(result0,result1);
result = _mm512_reduce_add_epi32(result0);
return result;
}
关于c - 英特尔至强融核上的快速弹出计数，我们在Stack Overflow上找到一个类似的问题：https://stackoverflow.com/questions/16164507/