Accélérer certaines intrinsions SSE2 pour une conversion de couleur

Question

J'essaie d'effectuer une conversion de couleur d'image de YCBCR en BGRA (ne pas poser de questions sur le peu, un tel mal de tête).

Quoi qu'il en soit, cela doit fonctionner aussi vite que possible, alors j'ai écrit de l'utiliser à l'aide de l'intrigue de compilateur pour profiter de SSE2. C'est ma première entreprise à Simd Terre, je suis essentiellement un débutant et je suis donc sûr qu'il y a beaucoup de choses que je fais de manière inefficace.

Mon code arithmétique pour faire la conversion de couleur réelle s'avère être particulièrement lent et que Intel's Vtutune le montre comme un goulot d'étranglement important.

Donc, de quelque manière que je peux accélérer le code suivant? Cela se fait dans 32 bits, 4 pixels à la fois. J'ai essayé à l'origine de le faire dans 8 bits, 16 pixels à la fois (comme dans la boucle supérieure), mais les calculs provoquent un débordement entier et une conversion cassée. Tout ce processus, y compris le décode Intel Jpeg, prend environ 14 ms pour un seul champ de Full HD. Ce serait génial si je pouvais la descendre à au moins 12 ms, idéalement 10 ms.

Toute aide ou conseils appréciés avec gratitude. Merci!

const __m128i s128_8    = _mm_set1_epi8((char)128);


const int nNumPixels = roi.width * roi.height;

for (int i=0; i<nNumPixels; i+=32)
{
    // Go ahead and prefetch our packed UV Data.
    // As long as the load remains directly next, this saves us time.
    _mm_prefetch((const char*)&pSrc8u[2][i],_MM_HINT_T0); 

    // We need to fetch and blit out our k before we write over it with UV data.
    __m128i sK1 = _mm_load_si128((__m128i*)&pSrc8u[2][i]);
    __m128i sK2 = _mm_load_si128((__m128i*)&pSrc8u[2][i+16]);

    // Using the destination buffer temporarily here so we don't need to waste time doing a memory allocation.
    _mm_store_si128 ((__m128i*)&m_pKBuffer[i],      sK1);
    _mm_store_si128 ((__m128i*)&m_pKBuffer[i+16],   sK2);

    // In theory, this prefetch needs to be some cycles ahead of the first read. It isn't, yet it does appear to save us time. Worth investigating.
    _mm_prefetch((const char*)&pSrc8u[1][i],_MM_HINT_T0); 

    __m128i sUVI1 = _mm_load_si128((__m128i*)&pSrc8u[1][i]);
    __m128i sUVI2 = _mm_load_si128((__m128i*)&pSrc8u[1][i+16]);  

    // Subtract the 128 here ahead of our YCbCr -> BGRA conversion so we can go 16 pixels at a time rather than 4.
    sUVI1 = _mm_sub_epi8(sUVI1, s128_8);
    sUVI2 = _mm_sub_epi8(sUVI2, s128_8);

    // Swizzle and double up UV data from interleaved 8x1 byte blocks into planar
    __m128i sU1 = _mm_unpacklo_epi8(sUVI1, sUVI1);
    __m128i sV1 = _mm_unpackhi_epi8(sUVI1, sUVI1);

    __m128i sU2 = _mm_unpacklo_epi8(sUVI2, sUVI2);  
    __m128i sV2 = _mm_unpackhi_epi8(sUVI2, sUVI2);  

    _mm_store_si128((__m128i*)&pSrc8u[1][i],        sU1);
    _mm_store_si128((__m128i*)&pSrc8u[1][i+16],     sU2); 

    _mm_store_si128((__m128i*)&pSrc8u[2][i],        sV1);
    _mm_store_si128((__m128i*)&pSrc8u[2][i+16],     sV2); 
}

const __m128i s16   = _mm_set1_epi32(16);
const __m128i s299  = _mm_set1_epi32(299);
const __m128i s410  = _mm_set1_epi32(410);
const __m128i s518  = _mm_set1_epi32(518);
const __m128i s101  = _mm_set1_epi32(101);
const __m128i s209  = _mm_set1_epi32(209);

Ipp8u* pDstP = pDst8u;
for (int i=0; i<nNumPixels; i+=4, pDstP+=16)
{
    __m128i sK = _mm_set_epi32(m_pKBuffer[i],           m_pKBuffer[i+1],            m_pKBuffer[i+2],            m_pKBuffer[i+3]);

    __m128i sY = _mm_set_epi32(pSrc8u[0][i],            pSrc8u[0][i+1],             pSrc8u[0][i+2],             pSrc8u[0][i+3]);
    __m128i sU = _mm_set_epi32((char)pSrc8u[1][i],      (char)pSrc8u[1][i+1],       (char)pSrc8u[1][i+2],       (char)pSrc8u[1][i+3]);
    __m128i sV = _mm_set_epi32((char)pSrc8u[2][i],      (char)pSrc8u[2][i+1],       (char)pSrc8u[2][i+2],       (char)pSrc8u[2][i+3]);

    // N.b. - Attempted to do the sub 16 in 8 bits similar to the sub 128 for U and V - however doing it here is quicker
    // as the time saved on the arithmetic is less than the time taken by the additional loads/stores needed in the swizzle loop
    sY = _mm_mullo_epi32(_mm_sub_epi32(sY, s16), s299);

    __m128i sR  = _mm_srli_epi32(_mm_add_epi32(sY,_mm_mullo_epi32(s410, sV)), 8);
    __m128i sG  = _mm_srli_epi32(_mm_sub_epi32(_mm_sub_epi32(sY, _mm_mullo_epi32(s101, sU)),_mm_mullo_epi32(s209, sV)), 8);
    __m128i sB  = _mm_srli_epi32(_mm_add_epi32(sY, _mm_mullo_epi32(s518, sU)), 8);

    //Microsoft's YUV Conversion
    //__m128i sC = _mm_sub_epi32(sY, s16);
    //__m128i sD = _mm_sub_epi32(sU, s128);
    //__m128i sE = _mm_sub_epi32(sV, s128);
    //
    //__m128i sR =  _mm_srli_epi32(_mm_add_epi32(_mm_add_epi32(_mm_mullo_epi32(s298, sC), _mm_mullo_epi32(s409, sE)), s128), 8);
    //__m128i sG    = _mm_srli_epi32(_mm_add_epi32(_mm_sub_epi32(_mm_mullo_epi32(s298, sC), _mm_sub_epi32(_mm_mullo_epi32(s100, sD), _mm_mullo_epi32(s208, sE))), s128), 8);
    //__m128i sB    = _mm_srli_epi32(_mm_add_epi32(_mm_add_epi32(_mm_mullo_epi32(s298, sC), _mm_mullo_epi32(s516, sD)), s128), 8);

    __m128i sKGl = _mm_unpacklo_epi32(sK, sG);
    __m128i sKGh = _mm_unpackhi_epi32(sK, sG);

    __m128i sRBl = _mm_unpacklo_epi32(sR, sB);
    __m128i sRBh = _mm_unpackhi_epi32(sR, sB);

    __m128i sKRGB1 = _mm_unpackhi_epi32(sKGh,sRBh);
    __m128i sKRGB2 = _mm_unpacklo_epi32(sKGh,sRBh);
    __m128i sKRGB3 = _mm_unpackhi_epi32(sKGl,sRBl);
    __m128i sKRGB4 = _mm_unpacklo_epi32(sKGl,sRBl);

    __m128i p1 = _mm_packus_epi16(sKRGB1, sKRGB2);
    __m128i p2 = _mm_packus_epi16(sKRGB3, sKRGB4);

    __m128i po = _mm_packus_epi16(p1, p2);

    _mm_store_si128((__m128i*)pDstP, po);
}

Answer 1

You may be bandwidth limited here, as there is very little computation relative to the number of loads and stores.

One suggestion: get rid of the _mm_prefetch intrinsics - they are almost certainly not helping and may even hinder operation on more recent CPUs (which already do a pretty good job with automatic prefetching).

Another area to look at:

__m128i sK = _mm_set_epi32(m_pKBuffer[i],           m_pKBuffer[i+1],            m_pKBuffer[i+2],            m_pKBuffer[i+3]);
__m128i sY = _mm_set_epi32(pSrc8u[0][i],            pSrc8u[0][i+1],             pSrc8u[0][i+2],             pSrc8u[0][i+3]);
__m128i sU = _mm_set_epi32((char)pSrc8u[1][i],      (char)pSrc8u[1][i+1],       (char)pSrc8u[1][i+2],       (char)pSrc8u[1][i+3]);
__m128i sV = _mm_set_epi32((char)pSrc8u[2][i],      (char)pSrc8u[2][i+1],       (char)pSrc8u[2][i+2],       (char)pSrc8u[2][i+3]);

This is generating a lot of unnecessary instructions - you should be using _mm_load_xxx and _mm_unpackxx_xxx here. It will look like more code, but it will be a lot more efficient. And you should probably be processing 16 pixels per iteration of the loop, rather than 4 - that way you load a vector of 8 bit values once, and unpack to get each set of 4 values as a vector of 32 bit ints as needed.