/* GRF allocation:
   g1~g30: constant buffer
           g1~g2:intra IQ matrix
           g3~g4:non intra IQ matrix
           g5~g20:IDCT table
   g31:    thread payload 
   g58~g81:reference data
   g82:    thread payload backup
   g83~g106:IDCT data
   g115:   message descriptor for reading reference data   */
mov (1) g32.8<1>UD 0x7000FUD {align1}; //8*16
send (16) 0 g40.0<1>UW g32<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 4 {align1};//U
send (16) 0 g44.0<1>UW g32<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 4 {align1};//V
//U
avg (8) g74.0<1>UW  g40.0<8,8,1>UB g40.1<8,8,1>UB {align1};
avg (8) g74.16<1>UW g41.0<8,8,1>UB g41.1<8,8,1>UB {align1};
avg (8) g75.0<1>UW  g42.0<8,8,1>UB g42.1<8,8,1>UB {align1};
avg (8) g75.16<1>UW g43.0<8,8,1>UB g43.1<8,8,1>UB {align1};
//V 
avg (8) g78.0<1>UW  g44.0<8,8,1>UB g44.1<8,8,1>UB {align1};
avg (8) g78.16<1>UW g45.0<8,8,1>UB g45.1<8,8,1>UB {align1};
avg (8) g79.0<1>UW  g46.0<8,8,1>UB g46.1<8,8,1>UB {align1};
avg (8) g79.16<1>UW g47.0<8,8,1>UB g47.1<8,8,1>UB {align1};

add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
send (16) 0 g40.0<1>UW g32<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 4 {align1};//U
send (16) 0 g44.0<1>UW g32<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 4 {align1};//V
//U
avg (8) g76.0<1>UW  g40.0<8,8,1>UB g40.1<8,8,1>UB {align1};
avg (8) g76.16<1>UW g41.0<8,8,1>UB g41.1<8,8,1>UB {align1};
avg (8) g77.0<1>UW  g42.0<8,8,1>UB g42.1<8,8,1>UB {align1};
avg (8) g77.16<1>UW g43.0<8,8,1>UB g43.1<8,8,1>UB {align1};
//V                                              
avg (8) g80.0<1>UW  g44.0<8,8,1>UB g44.1<8,8,1>UB {align1};
avg (8) g80.16<1>UW g45.0<8,8,1>UB g45.1<8,8,1>UB {align1};
avg (8) g81.0<1>UW  g46.0<8,8,1>UB g46.1<8,8,1>UB {align1};
avg (8) g81.16<1>UW g47.0<8,8,1>UB g47.1<8,8,1>UB {align1};