标签：c x86-64 optimization sse simd
分析表明这个函数对我的应用来说是一个真正的瓶颈：
static inline int countEqualChars(const char* string1, const char* string2, int size) {
int r = 0;
for (int j = 0; j < size; ++j) {
if (string1[j] == string2[j]) {
++r;
}
}
return r;
}
即使使用-O3和-march = native,G 4.7.2也不会对此函数进行矢量化(我检查了汇编程序输出).现在,我不是SSE和朋友的专家,但我认为同时比较多个角色应该更快.关于如何加快速度的任何想法？目标架构是x86-64.
解决方法:
矢量化的编译器标志：
-ftree,矢量化
-ftree-vectorize -march =< your_architecture> (使用计算机上可用的所有指令集扩展,而不仅仅是像SSE2 for x86-64这样的基线).使用-march = native来优化运行编译器的机器.) – march =< foo>还设置-mtune =< foo>,这也是一件好事.
使用SSEx内在函数：
> Padd并将缓冲区对齐到16个字节(根据您实际要使用的向量大小)
>用_mm_set1_epi8(0)创建累加器countU8
>对于所有n / 16个输入(子)向量,请执行以下操作：
>使用_mm_load_si128或_mm_loadu_si128从两个字符串加载16个字符(对于未对齐的载荷)
> _mm_cmpeq_epi8
并行比较八位字节.每个匹配产生0xFF(-1),否则产生0x00.
>使用_mm_sub_epi8(减-1 – > 1)从countU8中减去上述结果向量
>始终在255个周期后,必须将16个8位计数器提取为更大的整数类型以防止溢出.请参阅解压缩和水平添加这个很好的答案,如何做到这一点：https://stackoverflow.com/a/10930706/1175253
码：
#include <iostream>
#include <vector>
#include <cassert>
#include <cstdint>
#include <climits>
#include <cstring>
#include <emmintrin.h>
#ifdef __SSE2__
#if !defined(UINTPTR_MAX) || !defined(UINT64_MAX) || !defined(UINT32_MAX)
# error "Limit macros are not defined"
#endif
#if UINTPTR_MAX == UINT64_MAX
#define PTR_64
#elif UINTPTR_MAX == UINT32_MAX
#define PTR_32
#else
# error "Current UINTPTR_MAX is not supported"
#endif
template<typename T>
void print_vector(std::ostream& out,const __m128i& vec)
{
static_assert(sizeof(vec) % sizeof(T) == 0,"Invalid element size");
std::cout << '{';
const T* const end = reinterpret_cast<const T*>(&vec)-1;
const T* const upper = end+(sizeof(vec)/sizeof(T));
for(const T* elem = upper;
elem != end;
--elem
)
{
if(elem != upper)
std::cout << ',';
std::cout << +(*elem);
}
std::cout << '}' << std::endl;
}
#define PRINT_VECTOR(_TYPE,_VEC) do{ std::cout << #_VEC << " : "; print_vector<_TYPE>(std::cout,_VEC); } while(0)
///@note SSE2 required (macro: __SSE2__)
///@warning Not tested!
size_t counteq_epi8(const __m128i* a_in,const __m128i* b_in,size_t count)
{
assert(a_in != nullptr && (uintptr_t(a_in) % 16) == 0);
assert(b_in != nullptr && (uintptr_t(b_in) % 16) == 0);
//assert(count > 0);
/*
//maybe not so good with all that branching and additional loop variables
__m128i accumulatorU8 = _mm_set1_epi8(0);
__m128i sum2xU64 = _mm_set1_epi8(0);
for(size_t i = 0;i < count;++i)
{
//this operation could also be unrolled, where multiple result registers would be accumulated
accumulatorU8 = _mm_sub_epi8(accumulatorU8,_mm_cmpeq_epi8(*a_in++,*b_in++));
if(i % 255 == 0)
{
//before overflow of uint8, the counter will be extracted
__m128i sum2xU16 = _mm_sad_epu8(accumulatorU8,_mm_set1_epi8(0));
sum2xU64 = _mm_add_epi64(sum2xU64,sum2xU16);
//reset accumulatorU8
accumulatorU8 = _mm_set1_epi8(0);
}
}
//blindly accumulate remaining values
__m128i sum2xU16 = _mm_sad_epu8(accumulatorU8,_mm_set1_epi8(0));
sum2xU64 = _mm_add_epi64(sum2xU64,sum2xU16);
//do a horizontal addition of the two counter values
sum2xU64 = _mm_add_epi64(sum2xU64,_mm_srli_si128(sum2xU64,64/8));
#if defined PTR_64
return _mm_cvtsi128_si64(sum2xU64);
#elif defined PTR_32
return _mm_cvtsi128_si32(sum2xU64);
#else
# error "macro PTR_(32|64) is not set"
#endif
*/
__m128i sum2xU64 = _mm_set1_epi32(0);
while(count--)
{
__m128i matches = _mm_sub_epi8(_mm_set1_epi32(0),_mm_cmpeq_epi8(*a_in++,*b_in++));
__m128i sum2xU16 = _mm_sad_epu8(matches,_mm_set1_epi32(0));
sum2xU64 = _mm_add_epi64(sum2xU64,sum2xU16);
#ifndef NDEBUG
PRINT_VECTOR(uint16_t,sum2xU64);
#endif
}
//do a horizontal addition of the two counter values
sum2xU64 = _mm_add_epi64(sum2xU64,_mm_srli_si128(sum2xU64,64/8));
#ifndef NDEBUG
std::cout << "----------------------------------------" << std::endl;
PRINT_VECTOR(uint16_t,sum2xU64);
#endif
#if !defined(UINTPTR_MAX) || !defined(UINT64_MAX) || !defined(UINT32_MAX)
# error "Limit macros are not defined"
#endif
#if defined PTR_64
return _mm_cvtsi128_si64(sum2xU64);
#elif defined PTR_32
return _mm_cvtsi128_si32(sum2xU64);
#else
# error "macro PTR_(32|64) is not set"
#endif
}
#endif
int main(int argc, char* argv[])
{
std::vector<__m128i> a(64); // * 16 bytes
std::vector<__m128i> b(a.size());
const size_t nBytes = a.size() * sizeof(std::vector<__m128i>::value_type);
char* const a_out = reinterpret_cast<char*>(a.data());
char* const b_out = reinterpret_cast<char*>(b.data());
memset(a_out,0,nBytes);
memset(b_out,0,nBytes);
a_out[1023] = 1;
b_out[1023] = 1;
size_t equalBytes = counteq_epi8(a.data(),b.data(),a.size());
std::cout << "equalBytes = " << equalBytes << std::endl;
return 0;
}
我为大型和小型阵列实现的最快的SSE实现：
size_t counteq_epi8(const __m128i* a_in,const __m128i* b_in,size_t count)
{
assert((count > 0 ? a_in != nullptr : true) && (uintptr_t(a_in) % sizeof(__m128i)) == 0);
assert((count > 0 ? b_in != nullptr : true) && (uintptr_t(b_in) % sizeof(__m128i)) == 0);
//assert(count > 0);
const size_t maxInnerLoops = 255;
const size_t nNestedLoops = count / maxInnerLoops;
const size_t nRemainderLoops = count % maxInnerLoops;
const __m128i zero = _mm_setzero_si128();
__m128i sum16xU8 = zero;
__m128i sum2xU64 = zero;
for(size_t i = 0;i < nNestedLoops;++i)
{
for(size_t j = 0;j < maxInnerLoops;++j)
{
sum16xU8 = _mm_sub_epi8(sum16xU8,_mm_cmpeq_epi8(*a_in++,*b_in++));
}
sum2xU64 = _mm_add_epi64(sum2xU64,_mm_sad_epu8(sum16xU8,zero));
sum16xU8 = zero;
}
for(size_t j = 0;j < nRemainderLoops;++j)
{
sum16xU8 = _mm_sub_epi8(sum16xU8,_mm_cmpeq_epi8(*a_in++,*b_in++));
}
sum2xU64 = _mm_add_epi64(sum2xU64,_mm_sad_epu8(sum16xU8,zero));
sum2xU64 = _mm_add_epi64(sum2xU64,_mm_srli_si128(sum2xU64,64/8));
#if UINTPTR_MAX == UINT64_MAX
return _mm_cvtsi128_si64(sum2xU64);
#elif UINTPTR_MAX == UINT32_MAX
return _mm_cvtsi128_si32(sum2xU64);
#else
# error "macro PTR_(32|64) is not set"
#endif
}
标签：c,x86-64,optimization,sse,simd 来源： https://codeday.me/bug/20191006/1861301.html
本站声明： 1. iCode9 技术分享网（下文简称本站）提供的所有内容，仅供技术学习、探讨和分享； 2. 关于本站的所有留言、评论、转载及引用，纯属内容发起人的个人观点，与本站观点和立场无关； 3. 关于本站的所有言论和文字，纯属内容发起人的个人观点，与本站观点和立场无关； 4. 本站文章均是网友提供，不完全保证技术分享内容的完整性、准确性、时效性、风险性和版权归属；如您发现该文章侵犯了您的权益，可联系我们第一时间进行删除； 5. 本站为非盈利性的个人网站，所有内容不会用来进行牟利，也不会利用任何形式的广告来间接获益，纯粹是为了广大技术爱好者提供技术内容和技术思想的分享性交流网站。