和字节的SSE代码.错误在哪里？

Question

我编写了SSE代码来总结字节值.(VS2005).

因为它很简单,它运行得很好(而且速度很快).只有一些大小的数组崩溃.它只在发布模式下崩溃 - 在调试中永远不会.也许有人看到了"明显的"错误？任何帮助赞赏.

__int64 Sum (const unsigned char* pData, const unsigned int& nLength)
{
    __int64 nSum (0);

    __m128i* pp = (__m128i*)pData;

    ATLASSERT( ( (DWORD)pp & 15 ) == 0 ); // pointer must point to address multiple of 16 (cache line)

    __m128i zero = _mm_setzero_si128(),
        a, b, c, d, tmp;

    unsigned int i (0);

    for ( ; i < nLength; i+=64) // 4-fach loop-unroll (x 16)
    {
        a = _mm_sad_epu8( *(pp++), zero);           
        b = _mm_sad_epu8( *(pp++), zero);  // It crashes here.
        c = _mm_sad_epu8( *(pp++), zero);
        d = _mm_sad_epu8( *(pp++), zero);

        // commenting the following line prevents the crash (???)
        tmp = _mm_add_epi64( _mm_add_epi64( _mm_add_epi64( a, b ), c ), d);

        a = _mm_srli_si128 ( tmp, 8 );

        nSum += _mm_cvtsi128_si32( a ) + _mm_cvtsi128_si32( tmp );
    }

    // ... the rest
    if (nLength % 64)
        for (i -= 64; i < nLength; i++)
            nSum += pData [i];

    return nSum;
}

该函数调用如下:

unsigned int nLength = 3571653;  // One of the values that causes crash
unsigned char *pData = (unsigned char*) _aligned_malloc(nLength, 16);
Sum (pData,  nLength);

Answer 1

您的for循环需要定义如下:

for ( ; i < (nLength - 63); i+=64)

基本上想象你传入一个nLength 120的数组.你在第一次运行时没问题.我现在等于64.我<120所以你做了另一个循环.不幸的是,在达到128之前传递数组的末尾并进入未定义的行为区域.这可能表现为访问冲突(0xC0000005),这会导致您崩溃.

现在以nLength = 128为例,它应该在我的建议修改的优化循环中运行完美.第一个循环我很好,i = 64.我小于65,所以执行另一个循环.我现在等于128并循环退出.外部循环也不会运行,因为i == nLength.任务完成 :)

Answer 2

根据要求，这就是当我说“循环中的 4 个累加器并在循环后对它们求和”时我想到的。

__int64 Sum (const unsigned char* pData, const int& nLength)
{
    __int64 nSum (0);

    __m128i* pp = (__m128i*)pData;


    __m128i zero = _mm_setzero_si128(),
        a = _mm_setzero_si128(),
        b = _mm_setzero_si128(),
        c = _mm_setzero_si128(),
        d = _mm_setzero_si128(), tmp;

    int i (0);

    for ( ; i < (nLength - 63); i+=64)
    {
        a = _mm_add_epi64( _mm_sad_epu8( *(pp++), zero ), a );
        b = _mm_add_epi64( _mm_sad_epu8( *(pp++), zero ), b );
        c = _mm_add_epi64( _mm_sad_epu8( *(pp++), zero ), c );
        d = _mm_add_epi64( _mm_sad_epu8( *(pp++), zero ), d );
    }

    tmp = _mm_add_epi64( _mm_add_epi64( a, b ), _mm_add_epi64( c, d ));
    tmp = _mm_add_epi64( _mm_srli_si128( tmp, 8 ), tmp );
    nSum = (_mm_cvtsi128_si32( tmp ) & 0xFFFFFFFFULL) + 
               (((__int64)_mm_cvtsi128_si32( _mm_srli_si128( tmp, 4 ) )) << 32);

    // ... the rest
    for (; i < nLength; i++)
        nSum += pData [i];

    return nSum;
}