Question

我正在将MMX SSE转换为等效C代码。我几乎已经转换了它，但我得到的图像质量不合适，或者我可以看到图像中出现了一些噪音。我正在调试过去5天的代码，但我没有理由为什么会发生这种情况。如果你们调查这个问题并帮助我，我将非常高兴。

原始SSE代码：

void unpack_8bit_to_16bit( __m128i *a, __m128i* b0, __m128i* b1 ) 
{
    __m128i zero = _mm_setzero_si128();
    b0 = _mm_unpacklo_epi8( a, zero );
    b1 = _mm_unpackhi_epi8( a, zero );
}

void convolve_cols_3x3( const unsigned char* in, int16_t* out_v, int16_t* out_h, int w, int h )
{
    using namespace std;
    assert( w % 16 == 0 && "width must be multiple of 16!" );
    const int w_chunk  = w/16;
    __m128i*    i0       = (__m128i*)( in );
    __m128i*    i1       = (__m128i*)( in ) + w_chunk*1;
    __m128i*    i2       = (__m128i*)( in ) + w_chunk*2;
    __m128i* result_h  = (__m128i*)( out_h ) + 2*w_chunk;
    __m128i* result_v  = (__m128i*)( out_v ) + 2*w_chunk;
    __m128i* end_input = (__m128i*)( in ) + w_chunk*h;

    for( ; i2 != end_input; i0++, i1++, i2++, result_v+=2, result_h+=2 ) 
    {
        *result_h     = _mm_setzero_si128();
        *(result_h+1) = _mm_setzero_si128();
        *result_v     = _mm_setzero_si128();
        *(result_v+1) = _mm_setzero_si128();
        __m128i ilo, ihi;
        unpack_8bit_to_16bit( *i0, ihi, ilo ); 
        *result_h     = _mm_add_epi16( ihi, *result_h );
        *(result_h+1) = _mm_add_epi16( ilo, *(result_h+1) );
        *result_v     = _mm_add_epi16( *result_v, ihi );
        *(result_v+1) = _mm_add_epi16( *(result_v+1), ilo );
        unpack_8bit_to_16bit( *i1, ihi, ilo );
        *result_v     = _mm_add_epi16( *result_v, ihi );
        *(result_v+1) = _mm_add_epi16( *(result_v+1), ilo );
        *result_v     = _mm_add_epi16( *result_v, ihi );
        *(result_v+1) = _mm_add_epi16( *(result_v+1), ilo );
        unpack_8bit_to_16bit( *i2, ihi, ilo );
        *result_h     = _mm_sub_epi16( *result_h, ihi );
        *(result_h+1) = _mm_sub_epi16( *(result_h+1), ilo );
        *result_v     = _mm_add_epi16( *result_v, ihi );
        *(result_v+1) = _mm_add_epi16( *(result_v+1), ilo );
    }
}

我转换的代码如下所示

void convolve_cols_3x3( const unsigned char* in, int16_t* out_v, int16_t* out_h, int w, int h )
{
    using namespace std;
    assert( w % 16 == 0 && "width must be multiple of 16!" );
    const int w_chunk  = w/16;

    uint8_t*    i0       = (uint8_t*)( in );
    uint8_t*    i1       = (uint8_t*)( in ) + w_chunk*1*16;
    uint8_t*    i2       = (uint8_t*)( in ) + w_chunk*2*16;
    int16_t* result_h  = (int16_t*)( out_h ) + 2*w_chunk*16;
    int16_t* result_v  = (int16_t*)( out_v ) + 2*w_chunk*16;
    uint8_t* end_input = (uint8_t*)( in ) + w_chunk*h*16;
    for( ; i2 != end_input; i0+= 16, i1+= 16, i2+= 16, result_v+= 16, result_h+= 16 ) 
    {
        for (int i=0; i<8;i++)
        {
            result_h[i]     = 0;
            result_h[i + 8] = 0;
            result_v[i]        = 0;
            result_v[i + 8] = 0;
            result_h[i]     = (int16_t)(i0[i]) + result_h[i] ;
            result_h[i + 8] = (int16_t)(i0[i + 8]) + result_h[i + 8] ;
            result_v[i]     = (int16_t)(i0[i]) + result_v[i] ;
            result_v[i + 8] = (int16_t)(i0[i + 8]) + result_v[i + 8] ;
            result_v[i]     = (int16_t)(i1[i]) + result_v[i] ;
            result_v[i + 8] = (int16_t)(i1[i + 8]) + result_v[i + 8] ;
            result_v[i]     = (int16_t)(i1[i]) + result_v[i] ;
            result_v[i + 8] = (int16_t)(i1[i + 8]) + result_v[i + 8] ;
            result_h[i]     = result_h[i] - (int16_t)(i2[i]);
            result_h[i + 8] = result_h[i + 8] - (int16_t)(i2[i + 8]);
            result_v[i]     = (int16_t)(i2[i]) + result_v[i] ;
            result_v[i + 8] = (int16_t)(i2[i + 8]) + result_v[i + 8] ;
        }
    }
}

很抱歉，如果代码不可读。 w和h代表宽度和高度。 out_h和out_v是两个参数，稍后用于其他目的。

Answer 1

错误似乎出现在指针数学和源数据的读取中。指针变量i0，i1，i2是unsigned char。代码中的这样的行：

 result_h[i + 8] = (int16_t)(i0[i + 8]) + result_h[i + 8] ;

应该是这样的：

result_h[i + 8] = (int16_t)(i0[i*2 + 16]) + result_h[i + 8] ;

强制转换为int16_t不会影响i0方括号内的偏移量。您正在使用16字节结构（__m128i），但是以8字节偏移量访问它们。您也只使用i0和i1指向的整数的低8位。在原始SSE代码中，您正在读取16位整数。如果您需要在添加之前读取16位整数，则最终更正的代码可能如下所示：

result_h[i + 8] = *(int16_t *)(&i0[i*2 + 16]) + result_h[i + 8];

MMX SSE到C代码转换时图像质量下降

1 个答案: