我有两个vector<bool>
A和B。
我想比较它们并计算相等的元素数:
例如:
A = {0,1,0,1}
B = {0,0,1,1}
结果将等于2。
我可以使用
_mm_cmpeq_epi8
,但是它只能比较16个元素(即我应该将0和1转换为char然后进行比较)。是否可以每次用SSE(或SIMD指令)比较128个元素?
最佳答案
如果可以假设vector<bool>
使用连续的字节大小的元素进行存储,或者可以考虑使用类似vector<uint8_t>
的存储,则此示例应该为您提供一个良好的起点:
static size_t count_equal(const vector<uint8_t> &vec1, const vector<uint8_t> &vec2)
{
assert(vec1.size() == vec2.size()); // vectors must be same size
const size_t n = vec1.size();
const size_t max_block_size = 255 * 16; // max block size before possible overflow
__m128i vcount = _mm_setzero_si128();
size_t i, count = 0;
for (i = 0; i + 16 <= n; ) // for each block
{
size_t m = std::min(n, i + max_block_size);
for ( ; i + 16 <= m; i += 16) // for each vector in block
{
__m128i v1 = _mm_loadu_si128((__m128i *)&vec1[i]);
__m128i v2 = _mm_loadu_si128((__m128i *)&vec2[i]);
__m128i vcmp = _mm_cmpeq_epi8(v1, v2);
vcount = _mm_sub_epi8(vcount, vcmp);
}
vcount = _mm_sad_epu8(vcount, _mm_setzero_si128());
count += _mm_extract_epi16(vcount, 0) + _mm_extract_epi16(vcount, 4);
vcount = _mm_setzero_si128(); // update count from current block
}
vcount = _mm_sad_epu8(vcount, _mm_setzero_si128());
count += _mm_extract_epi16(vcount, 0) + _mm_extract_epi16(vcount, 4);
for ( ; i < n; ++i) // deal with any remaining partial vector
{
count += (vec1[i] == vec2[i]);
}
return count;
}
请注意,这是使用
vector<uint8_t>
。如果确实需要使用vector<bool>
并可以保证元素始终是连续的并且是字节大小的,那么您只需要将vector<bool>
强制转换为const uint8_t *
或类似的方式即可。测试线束:
#include <cassert>
#include <cstdlib>
#include <ctime>
#include <iostream>
#include <vector>
#include <emmintrin.h> // SSE2
using std::vector;
static size_t count_equal_ref(const vector<uint8_t> &vec1, const vector<uint8_t> &vec2)
{
assert(vec1.size() == vec2.size());
const size_t n = vec1.size();
size_t i, count = 0;
for (i = 0 ; i < n; ++i)
{
count += (vec1[i] == vec2[i]);
}
return count;
}
static size_t count_equal(const vector<uint8_t> &vec1, const vector<uint8_t> &vec2)
{
assert(vec1.size() == vec2.size()); // vectors must be same size
const size_t n = vec1.size();
const size_t max_block_size = 255 * 16; // max block size before possible overflow
__m128i vcount = _mm_setzero_si128();
size_t i, count = 0;
for (i = 0; i + 16 <= n; ) // for each block
{
size_t m = std::min(n, i + max_block_size);
for ( ; i + 16 <= m; i += 16) // for each vector in block
{
__m128i v1 = _mm_loadu_si128((__m128i *)&vec1[i]);
__m128i v2 = _mm_loadu_si128((__m128i *)&vec2[i]);
__m128i vcmp = _mm_cmpeq_epi8(v1, v2);
vcount = _mm_sub_epi8(vcount, vcmp);
}
vcount = _mm_sad_epu8(vcount, _mm_setzero_si128());
count += _mm_extract_epi16(vcount, 0) + _mm_extract_epi16(vcount, 4);
vcount = _mm_setzero_si128(); // update count from current block
}
vcount = _mm_sad_epu8(vcount, _mm_setzero_si128());
count += _mm_extract_epi16(vcount, 0) + _mm_extract_epi16(vcount, 4);
for ( ; i < n; ++i) // deal with any remaining partial vector
{
count += (vec1[i] == vec2[i]);
}
return count;
}
int main(int argc, char * argv[])
{
size_t n = 100;
if (argc > 1)
{
n = atoi(argv[1]);
}
vector<uint8_t> vec1(n);
vector<uint8_t> vec2(n);
srand((unsigned int)time(NULL));
for (size_t i = 0; i < n; ++i)
{
vec1[i] = rand() & 1;
vec2[i] = rand() & 1;
}
size_t n_ref = count_equal_ref(vec1, vec2);
size_t n_test = count_equal(vec1, vec2);
if (n_ref == n_test)
{
std::cout << "PASS" << std::endl;
}
else
{
std::cout << "FAIL: n_ref = " << n_ref << ", n_test = " << n_test << std::endl;
}
return 0;
}
编译并运行:
$ g++ -Wall -msse3 -O3 test.cpp && ./a.out
PASS