1/* GRF allocation:
2   g1~g30: constant buffer
3           g1~g2:intra IQ matrix
4           g3~g4:non intra IQ matrix
5           g5~g20:IDCT table
6   g31:    thread payload 
7   g58~g81:reference data
8   g82:    thread payload backup
9   g83~g106:IDCT data
10   g115:   message descriptor for reading reference data   */
11
12mov (1) g115.8<1>UD 0x07000FUD {align1}; // 8*16/32=4
13send (16) 0 g40.0<1>UW g115<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 4 {align1};//U
14send (16) 0 g45.0<1>UW g115<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 4 {align1};//V
15mov (1) g115.8<1>UD 0xFUD {align1};
16add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
17send (16) 0 g44.0<1>UW g115<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 1 {align1};//U
18send (16) 0 g49.0<1>UW g115<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 1 {align1};//V
19
20avg (16) g32.0<1>UW g40.0<16,8,1>UB g41.0<16,8,1>UB {align1};
21avg (16) g33.0<1>UW g41.0<16,8,1>UB g42.0<16,8,1>UB {align1};
22avg (16) g34.0<1>UW g42.0<16,8,1>UB g43.0<16,8,1>UB {align1};
23avg (16) g35.0<1>UW g43.0<16,8,1>UB g44.0<16,8,1>UB {align1};
24
25avg (16) g36.0<1>UW g45.0<16,8,1>UB g46.0<16,8,1>UB {align1};
26avg (16) g37.0<1>UW g46.0<16,8,1>UB g47.0<16,8,1>UB {align1};
27avg (16) g38.0<1>UW g47.0<16,8,1>UB g48.0<16,8,1>UB {align1};
28avg (16) g39.0<1>UW g48.0<16,8,1>UB g49.0<16,8,1>UB {align1};
29