mv2idx_kc.cpp

H.264完整的C语言代码和DCT的代码

#include "idb_kernelc.hpp"   
#include "mpeg.hpp"
#include "idb_kernelc2.hpp"   

KERNELDEF(MV2idx, KERNELS_DIR "mv2idx_kc.uc");

// MV2idx.i    -- generate load indices from a row of motion vectors

kernel MV2idx(cistream<half2> motion,      // Motion Vectors
              ostream<uint>   yindices,    // indices for 4 Y blocks
              ostream<uint>   crcbindices, // indices for 2 C blocks
              uc<int>&        uc_offsets,  // x, y offset of current block
              uc<int>&        uc_mblks,    // number of motion vectors
              uc<int>&        uc_mb_width) // number of macroblocks in row
{
  uint fetchaddress;
  uint fetchx, fetchy;
  int yrefx, yrefy;
  int crefx, crefy;

  uint yout0, yout1, yout2, yout3, yout4, yout5, yout6, yout7;
  uint cout0, cout1, cout2, cout3, cout4, cout5, cout6, cout7;

  int offsets    = commclperm(0x8, 0, uc_offsets);
  int xoffset    = offsets & 0xffff;
  int yoffset    = shift(offsets, -16);
  uint mb_width  = uint(commclperm(0x8, 0, uc_mb_width));
  uint rowlen    = shift(mb_width, 2);
  uint crowlen   = shift(mb_width, 6);

  yrefx = xoffset;
  yrefy = yoffset + cid();

  crefx = shift(xoffset, -1) + cid();
  crefy = shift(yoffset, -1);


  // Shuffle control words
  byte4 unpackmv = 0x57571302; // lo: sign | sign |  3rd |  2nd
                               // hi: sign | sign |  1st |  0th

  cc ccin = itocc (cid() == 0);
  cc ccend = itocc(0);

  // positive shift value means left shift

  half2 dxdy, bestval;
  loop_count(uc_mblks) pipeline(1) {
    // get motion vectors
    double<half2> dxdy2;
    int dx, dy;

    motion(ccin, ccend) >> dxdy >> bestval;

    // Change 1/2 pixel MV's to full pixel MV's
    dxdy = shifta(dxdy, -1);

    // Broadcast and unpack MV's
    dxdy = commclperm(0, dxdy);
    dxdy2 = shuffled(dxdy, unpackmv);
    dx = int(hi(dxdy2));
    dy = int(lo(dxdy2));

    // luminance
    fetchx = uint(yrefx + dx);
    fetchy = uint(yrefy + dy);

    expand<uint> fetchxoffsets(4), fetchyoffsets(8);

    fetchyoffsets[0] = shift(lo(shift(fetchy, -4) * rowlen), 4)
      + (fetchy & 0xf);
    fetchy = fetchy + 8;
    fetchyoffsets[1] = shift(lo(shift(fetchy, -4) * rowlen), 4)
      + (fetchy & 0xf);

    fetchxoffsets[0] = shift(shift(fetchx, -4), 6)
      + shift(shift(fetchx & 0xf, -2), 4);
    fetchx = fetchx + 4;
    fetchxoffsets[1] = shift(shift(fetchx, -4), 6)
      + shift(shift(fetchx & 0xf, -2), 4);
    fetchx = fetchx + 4;
    fetchxoffsets[2] = shift(shift(fetchx, -4), 6)
      + shift(shift(fetchx & 0xf, -2), 4);
    fetchx = fetchx + 4;
    fetchxoffsets[3] = shift(shift(fetchx, -4), 6)
      + shift(shift(fetchx & 0xf, -2), 4);

    yout0 = fetchyoffsets[0] + fetchxoffsets[0];
    yout1 = fetchyoffsets[1] + fetchxoffsets[0];
    yout2 = fetchyoffsets[0] + fetchxoffsets[1];
    yout3 = fetchyoffsets[1] + fetchxoffsets[1];
    yout4 = fetchyoffsets[0] + fetchxoffsets[2];
    yout5 = fetchyoffsets[1] + fetchxoffsets[2];
    yout6 = fetchyoffsets[0] + fetchxoffsets[3];
    yout7 = fetchyoffsets[1] + fetchxoffsets[3];

    yindices << yout0 << yout1 << yout2 << yout3;
    yindices << yout4 << yout5 << yout6 << yout7;

    // chromanance
    // add one if deltas are negative to get correct biasing
    //  when using shift by 1 in place of divide by 2
    dx = dx + select(itocc(dx < 0), 1, 0);
    dy = dy + select(itocc(dy < 0), 1, 0);
    fetchx = uint(crefx + shifta(int(dx), -1));
    fetchy = uint(crefy + shifta(int(dy), -1));

    uint fetchxoffset = shift(shift(fetchx, -3), 6) + (fetchx & 0x7);

    fetchyoffsets[0] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);
    fetchy = fetchy + 1;
    fetchyoffsets[1] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);
    fetchy = fetchy + 1;
    fetchyoffsets[2] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);
    fetchy = fetchy + 1;
    fetchyoffsets[3] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);
    fetchy = fetchy + 1;
    fetchyoffsets[4] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);
    fetchy = fetchy + 1;
    fetchyoffsets[5] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);
    fetchy = fetchy + 1;
    fetchyoffsets[6] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);
    fetchy = fetchy + 1;
    fetchyoffsets[7] = lo(shift(fetchy, -3) * crowlen)
      + shift(fetchy & 0x7, 3);

    cout0 = fetchxoffset + fetchyoffsets[0];
    cout1 = fetchxoffset + fetchyoffsets[1];
    cout2 = fetchxoffset + fetchyoffsets[2];
    cout3 = fetchxoffset + fetchyoffsets[3];
    cout4 = fetchxoffset + fetchyoffsets[4];
    cout5 = fetchxoffset + fetchyoffsets[5];
    cout6 = fetchxoffset + fetchyoffsets[6];
    cout7 = fetchxoffset + fetchyoffsets[7];

    crcbindices << cout0 << cout1 << cout2 << cout3;
    crcbindices << cout4 << cout5 << cout6 << cout7;

    yrefx = yrefx + 16;
    crefx = crefx + 8;
  }

  cc always = itocc( 0xffffffff );

  loop_until_any(ccend) {
    motion( always, ccend) >> dxdy >> bestval;
  }
}