mt/am_matmul/matmul2.c

   1 /*size_t i;
   2   size_t j;
   3   size_t max_dim = lda*lda;
   4   if (coreid==0){
   5     for (i=0; i<max_dim/(ncores*2); i+=8){
   6       data_t elementA1 = A[i];
   7       data_t elementA12 = A[i+1];
   8       data_t elementA13 = A[i+2];
   9       data_t elementA14 = A[i+3];
  10       data_t elementA15 = A[i+4];
  11       data_t elementA16 = A[i+5];
  12       data_t elementA17 = A[i+6];
  13       data_t elementA18 = A[i+7];
  14       data_t elementA2 = A[i+32*8];
  15       data_t elementA21 = A[i+32*8+1];
  16       data_t elementA22 = A[i+32*8+2];
  17       data_t elementA23 = A[i+32*8+3];
  18       data_t elementA24 = A[i+32*8+4];
  19       data_t elementA25 = A[i+32*8+5];
  20       data_t elementA26 = A[i+32*8+6];
  21       data_t elementA27 = A[i+32*8+7];
  22       int row= (int)(i/32)*32;
  23       int row2 = row+8*32;
  24       int column1 = i%32*32;
  25       int column12 = (i+1)%32*32;
  26       int column13 = (i+2)%32*32;
  27       int column14 = (i+3)%32*32;
  28       int column15 = (i+4)%32*32;
  29       int column16 = (i+5)%32*32;
  30       int column17 = (i+6)%32*32;
  31       int column18 = (i+7)%32*32;
  32
  33       for (j=0; j<lda; j++){
  34         C[row+j]+=elementA1*B[column1+j]+elementA12*B[column12+j]+elementA13*B[column13+j]+elementA14*B[column14+j]+elementA15*B[column15+j]+elementA16*B[column16+j]+elementA17*B[column17+j]+elementA18*B[column18+j]
  35
  36         C[row2+j]+=elementA2*B[column1+j]+elementA21*B[column12+j]+elementA22*B[column13+j]+elementA23*B[column14+j]+elementA24*B[column15+j]+elementA25*B[column16+j]+elementA26*B[column17+j]+elementA27*B[column18+j];
  37       }
  38   }}else{
  39     for (i=max_dim/2; i<(max_dim/(ncores*2)+max_dim/2); i+=8){
  40       data_t elementA1 = A[i];
  41       data_t elementA12 = A[i+1];
  42       data_t elementA13 = A[i+2];
  43       data_t elementA14 = A[i+3];
  44       data_t elementA15 = A[i+4];
  45       data_t elementA16 = A[i+5];
  46       data_t elementA17 = A[i+6];
  47       data_t elementA18 = A[i+7];
  48       data_t elementA2 = A[i+32*8];
  49       data_t elementA21 = A[i+32*8+1];
  50       data_t elementA22 = A[i+32*8+2];
  51       data_t elementA23 = A[i+32*8+3];
  52       data_t elementA24 = A[i+32*8+4];
  53       data_t elementA25 = A[i+32*8+5];
  54       data_t elementA26 = A[i+32*8+6];
  55       data_t elementA27 = A[i+32*8+7];
  56       int row= (int)(i/32)*32;
  57       int row2 = row+8*32;
  58       int column1 = i%32*32;
  59       int column12 = (i+1)%32*32;
  60       int column13 = (i+2)%32*32;
  61       int column14 = (i+3)%32*32;
  62       int column15 = (i+4)%32*32;
  63       int column16 = (i+5)%32*32;
  64       int column17 = (i+6)%32*32;
  65       int column18 = (i+7)%32*32;
  66
  67       for (j=0; j<lda; j++){
  68         C[row+j]+=elementA1*B[column1+j]+elementA12*B[column12+j]+elementA13*B[column13+j]+elementA14*B[column14+j]+elementA15*B[column15+j]+elementA16*B[column16+j]+elementA17*B[column17+j]+elementA18*B[column18+j];
  69         C[row2+j]+=elementA2*B[column1+j]+elementA21*B[column12+j]+elementA22*B[column13+j]+elementA23*B[column14+j]+elementA24*B[column15+j]+elementA25*B[column16+j]+elementA26*B[column17+j]+elementA27*B[column18+j];
  70
  71       }
  72     }
  73     }*/