SIMD: Transpose

/**

 * @[Description] : Transpose

 * @ [Param] : pResidual

 * @ [Param] : buffer for result of transpose.

 * @ [Param] : Transform size

 */

void Transpose(short* pResi, short *pBuf, int iTrSize)

{

  short*pSrc;

  short*pDst;

  int iWidthBlk;

  int iHeightBlk;

  int i, j;

  __m128i Line0, Line1, Line2, Line3, Line4, Line5, Line6, Line7;

  __m128i L0, L1, L2, L3, H0, H1, H2, H3;

  iHeightBlk = iWidthBlk = (iTrSize) >> 3;

  for (i = 0; i < iHeightBlk; i++)

  {

    for (j = 0; j < iWidthBlk; j++)

    {

      pSrc = pResi + (j * 8) + (i * iTrSize * 8);     // 

      pDst = pBuf + (i * 8) + (j * iTrSize * 8);      // 

      Line0 = _mm_loadu_si128((__m128i *)(pSrc)); pSrc += iTrSize;

      Line1 = _mm_loadu_si128((__m128i *)(pSrc)); pSrc += iTrSize;

      Line2 = _mm_loadu_si128((__m128i *)(pSrc)); pSrc += iTrSize;

      Line3 = _mm_loadu_si128((__m128i *)(pSrc)); pSrc += iTrSize;

      Line4 = _mm_loadu_si128((__m128i *)(pSrc)); pSrc += iTrSize;

      Line5 = _mm_loadu_si128((__m128i *)(pSrc)); pSrc += iTrSize;

      Line6 = _mm_loadu_si128((__m128i *)(pSrc)); pSrc += iTrSize;

      Line7 = _mm_loadu_si128((__m128i *)(pSrc));

      L0 = _mm_unpacklo_epi16(Line0, Line1);

      L1 = _mm_unpacklo_epi16(Line2, Line3);

      L2 = _mm_unpacklo_epi16(Line4, Line5);

      L3 = _mm_unpacklo_epi16(Line6, Line7);

      H0 = _mm_unpackhi_epi16(Line0, Line1);

      H1 = _mm_unpackhi_epi16(Line2, Line3);

      H2 = _mm_unpackhi_epi16(Line4, Line5);

      H3 = _mm_unpackhi_epi16(Line6, Line7);

      Line0 = _mm_unpacklo_epi32(L0, L1);

      Line1 = _mm_unpackhi_epi32(L0, L1);

      Line2 = _mm_unpacklo_epi32(H0, H1);

      Line3 = _mm_unpackhi_epi32(H0, H1);

      Line4 = _mm_unpacklo_epi32(L2, L3);

      Line5 = _mm_unpackhi_epi32(L2, L3);

      Line6 = _mm_unpacklo_epi32(H2, H3);

      Line7 = _mm_unpackhi_epi32(H2, H3);

      L0 = _mm_unpacklo_epi64(Line0, Line4);

      H0 = _mm_unpackhi_epi64(Line0, Line4);

      L2 = _mm_unpacklo_epi64(Line1, Line5);

      H2 = _mm_unpackhi_epi64(Line1, Line5);

      L1 = _mm_unpacklo_epi64(Line2, Line6);

      H1 = _mm_unpackhi_epi64(Line2, Line6);

      L3 = _mm_unpacklo_epi64(Line3, Line7);

      H3 = _mm_unpackhi_epi64(Line3, Line7);

      _mm_store_si128((__m128i *)(pDst), L0); pDst += iTrSize;

      _mm_store_si128((__m128i *)(pDst), H0); pDst += iTrSize;

      _mm_store_si128((__m128i *)(pDst), L2); pDst += iTrSize;

      _mm_store_si128((__m128i *)(pDst), H2); pDst += iTrSize;

      _mm_store_si128((__m128i *)(pDst), L1); pDst += iTrSize;

      _mm_store_si128((__m128i *)(pDst), H1); pDst += iTrSize;

      _mm_store_si128((__m128i *)(pDst), L3); pDst += iTrSize;

      _mm_store_si128((__m128i *)(pDst), H3);

    }

  }

}