c++ - 两个8位数组协方差的快速实现

Question

我需要比较大量相似的小尺寸图片(最大 200x200)。所以我尝试实现 SSIM(结构相似性参见 https://en.wikipedia.org/wiki/Structural_similarity)算法。SSIM 需要计算两个 8 位灰度图像的协方差。一个简单的实现如下所示:

float SigmaXY(const uint8_t * x, const uint8_t * y, size_t size, float averageX, float averageY)
{
    float sum = 0;
    for(size_t i = 0; i < size; ++i)
        sum += (x[i] - averageX) * (y[i] - averageY);
    return sum / size;
}

但是性能很差。所以我希望用SIMD或者CUDA来改进一下(听说可以)。不幸的是，我没有这样做的经验。它看起来如何？我要去哪里？

Answer 1

我有另一个不错的解决方案!

首先我想提一些数学公式:

averageX = Sum(x[i])/size;
averageY = Sum(y[i])/size;

因此:

Sum((x[i] - averageX)*(y[i] - averageY))/size = 

Sum(x[i]*y[i])/size - Sum(x[i]*averageY)/size - 
Sum(averageX*y[i])/size + Sum(averageX*averageY)/size = 

Sum(x[i]*y[i])/size - averageY*Sum(x[i])/size - 
averageX*Sum(y[i])/size + averageX*averageY*Sum(1)/size =   

Sum(x[i]*y[i])/size - averageY*averageX - 
averageX*averageY + averageX*averageY = 

Sum(x[i]*y[i])/size - averageY*averageX;     

它允许修改我们的算法:

float SigmaXY(const uint8_t * x, const uint8_t * y, size_t size, float averageX, float averageY)
{
    uint32_t sum = 0; // If images will have size greater then 256x256 than you have to use uint64_t.
    for(size_t i = 0; i < size; ++i)
        sum += x[i]*y[i];
    return sum / size - averageY*averageX;
} 

只有在那之后我们才能使用 SIMD(我使用 SSE2):

#include <emmintrin.h>

inline __m128i SigmaXY(__m128i x, __m128i y)
{
    __m128i lo = _mm_madd_epi16(_mm_unpacklo_epi8(x, _mm_setzero_si128()), _mm_unpacklo_epi8(y, _mm_setzero_si128()));
    __m128i hi = _mm_madd_epi16(_mm_unpackhi_epi8(y, _mm_setzero_si128()), _mm_unpackhi_epi8(y, _mm_setzero_si128()));
    return _mm_add_epi32(lo, hi);
}

float SigmaXY(const uint8_t * x, const uint8_t * y, size_t size, float averageX, float averageY)
{
    uint32_t sum = 0;
    size_t i = 0, alignedSize = size/16*16;
    if(size >= 16)
    {
        __m128i sums = _mm_setzero_si128();
        for(; i < alignedSize; i += 16)
        {
            __m128i _x = _mm_loadu_si128((__m128i*)(x + i));
            __m128i _y = _mm_loadu_si128((__m128i*)(y + i));
            sums = _mm_add_epi32(sums, SigmaXY(_x, _y));
        }
        uint32_t _sums[4];
        _mm_storeu_si128(_sums, sums);
        sum = _sums[0] + _sums[1] + _sums[2] + _sums[3];
    } 
    for(; i < size; ++i)
        sum += x[i]*y[i];
    return sum / size - averageY*averageX;
}