可以使用SIMD优化两个字符串之间的字节匹配计数吗?

时间:2013-03-24 13:23:53

标签: c++ optimization x86-64 sse simd

分析表明这个功能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-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。

3 个答案:

答案 0 :(得分:9)

当然可以。

pcmpeqb比较两个16字节的向量,并产生一个带有零的向量,它们不同,而-1则匹配。使用它来一次比较16个字节,将结果添加到累加器向量(确保累加最多255个向量比较的结果以避免溢出)。完成后,累加器中有16个结果。求它们并否定得到相等元素的数量。

如果长度非常短,那么很难从这种方法中获得显着的加速。如果长度很长,那么值得追求。

答案 1 :(得分:7)

矢量化的编译器标志:

-ftree-vectorize

-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>也设置-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)

  • 使用_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:

    • 使用_mm_load_si128_mm_loadu_si128从两个字符串加载16个字符(对于未对齐的加载)

    • _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。有关如何执行此操作的详细解答,请参阅解压缩和水平添加:https://stackoverflow.com/a/10930706/1175253

代码:

#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
}

答案 2 :(得分:3)

当前gcc中的自动向量化有助于编译器理解可以很容易地对代码进行矢量化。在您的情况下:如果您删除条件并以更强制的方式重写代码,它将理解向量化请求:

    static inline int count(const char* string1, const char* string2, int size) {
            int r = 0;
            bool b;

            for (int j = 0; j < size; ++j) {
                    b = (string1[j] == string2[j]);
                    r += b;
            }

            return r;
    }

在这种情况下:

movdqa  16(%rsp), %xmm1
movl    $.LC2, %esi
pxor    %xmm2, %xmm2
movzbl  416(%rsp), %edx
movdqa  .LC1(%rip), %xmm3
pcmpeqb 224(%rsp), %xmm1
cmpb    %dl, 208(%rsp)
movzbl  417(%rsp), %eax
movl    $1, %edi
pand    %xmm3, %xmm1
movdqa  %xmm1, %xmm5
sete    %dl
movdqa  %xmm1, %xmm4
movzbl  %dl, %edx
punpcklbw   %xmm2, %xmm5
punpckhbw   %xmm2, %xmm4
pxor    %xmm1, %xmm1
movdqa  %xmm5, %xmm6
movdqa  %xmm5, %xmm0
movdqa  %xmm4, %xmm5
punpcklwd   %xmm1, %xmm6

(等)