idxgen_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(idxGen, KERNELS_DIR "idxgen_kc.uc");

// *******************
// idxGen_kc : generate indices for loading source images
// *******************

kernel idxGen(ostream<uint>   indices,    // indices into src image
              uc<int>&        uc_size,    // ls16bits: image width
                                          // ms16bits: image height
              uc<int>&        uc_params)  // ls16bits: num macroblocks
                                          // ms16bits: edges: 0 - neither
                                          //                  1 - right
                                          //                  2 - bottom
                                          //                  3 - bottom right
{
  // process parameters from uc regs
  byte4 unpack = 0x88883120;
  int size = commclperm(ucid(), 0, uc_size);
  double<int> size2 = shuffled(size, unpack);
  int width = lo(size2);
  int height = hi(size2);
  int parms = commclperm(ucid(), 0, uc_params);
  double<int> params2 = shuffled(parms, unpack);
  int cnt = lo(params2);
  int type = hi(params2);
  uc<int> uc_cnt;
  int dummy = commclperm(ucid(), cnt, uc_cnt, 0, uc_cnt);

  // calculate x values for each cluster for last mb (others are cid(), cid8)
  int base_loc = 0;
  int right_side = ((type == 1) | (type == 3));
  int right_margin = (width & 0xf);
  right_margin = select(itocc((right_margin>0) & right_side), right_margin,16);
  right_margin = right_margin - 1;
  int cid8 = cid() + 8;
  expand<int> x(2);
  x[0] = select(itocc(cid() <= right_margin), cid(), right_margin);
  x[1] = select(itocc(cid8 <= right_margin), cid8, right_margin);

  // calculate start value of each row
  int bottom_row = ((type == 2) | (type == 3));
  int bottom_margin = (height & 0xf);
  bottom_margin = select(itocc((bottom_margin>0)&bottom_row),bottom_margin,16);
  expand<int> y(15);
  y[0]  = select(itocc(1 < bottom_margin), width, 0);
  y[1]  = select(itocc(2 < bottom_margin), width, 0);
  y[2]  = select(itocc(3 < bottom_margin), width, 0);
  y[3]  = select(itocc(4 < bottom_margin), width, 0);
  y[4]  = select(itocc(5 < bottom_margin), width, 0);
  y[5]  = select(itocc(6 < bottom_margin), width, 0);
  y[6]  = select(itocc(7 < bottom_margin), width, 0);
  y[7]  = select(itocc(8 < bottom_margin), width, 0);
  y[8]  = select(itocc(9 < bottom_margin), width, 0);
  y[9]  = select(itocc(10 < bottom_margin), width, 0);
  y[10]  = select(itocc(11 < bottom_margin), width, 0);
  y[11]  = select(itocc(12 < bottom_margin), width, 0);
  y[12]  = select(itocc(13 < bottom_margin), width, 0);
  y[13]  = select(itocc(14 < bottom_margin), width, 0);
  y[14]  = select(itocc(15 < bottom_margin), width, 0);

  loop_count(uc_cnt) {
    // is this the last (rightmost) macroblock?
    cc last_iter = itocc(cnt == 1);
    int first = select(last_iter, x[0], cid());
    int second = select(last_iter, x[1], cid8);

    // output values
    int loc = base_loc;
    indices << loc + first << loc + second; loc = loc + y[0];
    indices << loc + first << loc + second; loc = loc + y[1];
    indices << loc + first << loc + second; loc = loc + y[2];
    indices << loc + first << loc + second; loc = loc + y[3];
    indices << loc + first << loc + second; loc = loc + y[4];
    indices << loc + first << loc + second; loc = loc + y[5];
    indices << loc + first << loc + second; loc = loc + y[6];
    indices << loc + first << loc + second; loc = loc + y[7];
    indices << loc + first << loc + second; loc = loc + y[8];
    indices << loc + first << loc + second; loc = loc + y[9];
    indices << loc + first << loc + second; loc = loc + y[10];
    indices << loc + first << loc + second; loc = loc + y[11];
    indices << loc + first << loc + second; loc = loc + y[12];
    indices << loc + first << loc + second; loc = loc + y[13];
    indices << loc + first << loc + second; loc = loc + y[14];
    indices << loc + first << loc + second;

    // update vars
    cnt = cnt - 1;
    base_loc = base_loc + 16;
  }
}