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*/
12
13mov (1) g115.8<1>UD 0x07000FUD {align1}; 
14and.nz (1) null g115.16<1,1,1>UW 1UW {align1};
15(f0) jmpi read_backward;
16
17send (16) 0 g40.0<1>UW g115<8,8,1>UW read(5, 2, 0, 2) mlen 1 rlen 4 {align1};//U
18send (16) 0 g45.0<1>UW g115<8,8,1>UW read(6, 2, 0, 2) mlen 1 rlen 4 {align1};//V
19mov (1) g115.8<1>UD 0xFUD {align1};
20add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
21send (16) 0 g44.0<1>UW g115<8,8,1>UW read(5, 2, 0, 2) mlen 1 rlen 1 {align1};//U
22send (16) 0 g49.0<1>UW g115<8,8,1>UW read(6, 2, 0, 2) mlen 1 rlen 1 {align1};//V
23jmpi put_data;
24
25read_backward:
26send (16) 0 g40.0<1>UW g115<8,8,1>UW read(8, 2, 0, 2) mlen 1 rlen 4 {align1};//U
27send (16) 0 g45.0<1>UW g115<8,8,1>UW read(9, 2, 0, 2) mlen 1 rlen 4 {align1};//V
28mov (1) g115.8<1>UD 0xFUD {align1};
29add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
30send (16) 0 g44.0<1>UW g115<8,8,1>UW read(8, 2, 0, 2) mlen 1 rlen 1 {align1};//U
31send (16) 0 g49.0<1>UW g115<8,8,1>UW read(9, 2, 0, 2) mlen 1 rlen 1 {align1};//V
32
33put_data:
34avg (8) g32.0<1>UW g40.0<8,8,1>UB g41.0<8,8,1>UB {align1};
35avg (8) g33.0<1>UW g41.0<8,8,1>UB g42.0<8,8,1>UB {align1};
36avg (8) g34.0<1>UW g42.0<8,8,1>UB g43.0<8,8,1>UB {align1};
37avg (8) g35.0<1>UW g43.0<8,8,1>UB g44.0<8,8,1>UB {align1};
38
39avg (8) g36.0<1>UW g45.0<8,8,1>UB g46.0<8,8,1>UB {align1};
40avg (8) g37.0<1>UW g46.0<8,8,1>UB g47.0<8,8,1>UB {align1};
41avg (8) g38.0<1>UW g47.0<8,8,1>UB g48.0<8,8,1>UB {align1};
42avg (8) g39.0<1>UW g48.0<8,8,1>UB g49.0<8,8,1>UB {align1};
43