Geek Answers Handbook

Fastest 50% scaling (A) of RGB32 images using sse intrinsics

I want to scale images as fast as I can in C ++. This article describes how effectively average 32-bit rgb images are reduced by 50%. It is fast and looks good.

I tried modifying this approach using sse intrinsics. The code below works with or without SSE enabled. Surprisingly, acceleration is negligible.

Can anyone see a way to improve the SSE code. The two lines creating vars shuffle1 and shuffle2 seem to be two candidates (using some clever shifts or the like).

/*
 * Calculates the average of two rgb32 pixels.
 */
inline static uint32_t avg(uint32_t a, uint32_t b)
{
    return (((a^b) & 0xfefefefeUL) >> 1) + (a&b);
}

/*
 * Calculates the average of four rgb32 pixels.
 */
inline static uint32_t avg(const uint32_t a[2], const uint32_t b[2])
{
    return avg(avg(a[0], a[1]), avg(b[0], b[1]));
}

/*
 * Calculates the average of two rows of rgb32 pixels.
 */
void average2Rows(const uint32_t* src_row1, const uint32_t* src_row2, uint32_t* dst_row, int w)
{
#if !defined(__SSE)
        for (int x = w; x; --x, dst_row++, src_row1 += 2, src_row2 += 2)
            * dst_row = avg(src_row1, src_row2);
#else
        for (int x = w; x; x-=4, dst_row+=4, src_row1 += 8, src_row2 += 8)
        {
            __m128i left  = _mm_avg_epu8(_mm_load_si128((__m128i const*)src_row1), _mm_load_si128((__m128i const*)src_row2));
            __m128i right = _mm_avg_epu8(_mm_load_si128((__m128i const*)(src_row1+4)), _mm_load_si128((__m128i const*)(src_row2+4)));

            __m128i shuffle1 = _mm_set_epi32( right.m128i_u32[2], right.m128i_u32[0], left.m128i_u32[2], left.m128i_u32[0]);
            __m128i shuffle2 = _mm_set_epi32( right.m128i_u32[3], right.m128i_u32[1], left.m128i_u32[3], left.m128i_u32[1]);

            _mm_store_si128((__m128i *)dst_row, _mm_avg_epu8(shuffle1, shuffle2));
        }
#endif
}
+5
source share
3 answers

SSE , , :

__m128i shuffle1 = _mm_set_epi32( right.m128i_u32[2], right.m128i_u32[0], left.m128i_u32[2], left.m128i_u32[0]);
__m128i shuffle2 = _mm_set_epi32( right.m128i_u32[3], right.m128i_u32[1], left.m128i_u32[3], left.m128i_u32[1]);

SSE .

, :

__m128i t0 = _mm_unpacklo_epi32( left, right ); // right.m128i_u32[1] left.m128i_u32[1] right.m128i_u32[0] left.m128i_u32[0]
__m128i t1 = _mm_unpackhi_epi32( left, right ); // right.m128i_u32[3] left.m128i_u32[3] right.m128i_u32[2] left.m128i_u32[2]
__m128i shuffle1 = _mm_unpacklo_epi32( t0, t1 );    // right.m128i_u32[2] right.m128i_u32[0] left.m128i_u32[2] left.m128i_u32[0]
__m128i shuffle2 = _mm_unpackhi_epi32( t0, t1 );    // right.m128i_u32[3] right.m128i_u32[1] left.m128i_u32[3] left.m128i_u32[1]
+8

SSE , , , .

(_mm_set_epi32 - , ) .

/ , . , , , , , SSE.

. , , , / . , , , , .

+4

_mm_set_epi32 - SSE, SSE - , , SSE-. , SSE .

The second problem is that you do very few calculations regarding the number of downloads and storages. This will cause the code to be limited in bandwidth rather than being calculated, and you will not see a significant performance improvement even with perfect SSE code. Look at combining more operations in your loop so that you do more with your data while in the cache.

+4
source

More articles:


All Articles