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   */
11mov (1) g115.8<1>UD 0x07000FUD {align1};
12send (16) 0 g40.0<1>UW g115<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 4 {align1};//U
13send (16) 0 g45.0<1>UW g115<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 4 {align1};//V
14mov (1) g115.8<1>UD 0x01000FUD {align1};
15add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
16send (16) 0 g44.0<1>UW g115<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 1 {align1};//U
17send (16) 0 g49.0<1>UW g115<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 1 {align1};//V
18//U
19add (16) g32.0<1>UW g40.0<16,8,1>UB g41.0<16,8,1>UB {align1};
20add (16) g33.0<1>UW g41.0<16,8,1>UB g42.0<16,8,1>UB {align1};
21add (16) g34.0<1>UW g42.0<16,8,1>UB g43.0<16,8,1>UB {align1};
22add (16) g35.0<1>UW g43.0<16,8,1>UB g44.0<16,8,1>UB {align1};
23
24add (16) g32.0<1>UW g32.0<16,8,1>UW g40.1<16,8,1>UB {align1};
25add (16) g33.0<1>UW g33.0<16,8,1>UW g41.1<16,8,1>UB {align1};
26add (16) g34.0<1>UW g34.0<16,8,1>UW g42.1<16,8,1>UB {align1};
27add (16) g35.0<1>UW g35.0<16,8,1>UW g43.1<16,8,1>UB {align1};
28
29add (16) g32.0<1>UW g32.0<16,8,1>UW g41.1<16,8,1>UB {align1};
30add (16) g33.0<1>UW g33.0<16,8,1>UW g42.1<16,8,1>UB {align1};
31add (16) g34.0<1>UW g34.0<16,8,1>UW g43.1<16,8,1>UB {align1};
32add (16) g35.0<1>UW g35.0<16,8,1>UW g44.1<16,8,1>UB {align1};
33//V
34add (16) g36.0<1>UW g45.0<16,8,1>UB g46.0<16,8,1>UB {align1};
35add (16) g37.0<1>UW g46.0<16,8,1>UB g47.0<16,8,1>UB {align1};
36add (16) g38.0<1>UW g47.0<16,8,1>UB g48.0<16,8,1>UB {align1};
37add (16) g39.0<1>UW g48.0<16,8,1>UB g49.0<16,8,1>UB {align1};
38
39add (16) g36.0<1>UW g36.0<16,8,1>UW g45.1<16,8,1>UB {align1};
40add (16) g37.0<1>UW g37.0<16,8,1>UW g46.1<16,8,1>UB {align1};
41add (16) g38.0<1>UW g38.0<16,8,1>UW g47.1<16,8,1>UB {align1};
42add (16) g39.0<1>UW g39.0<16,8,1>UW g48.1<16,8,1>UB {align1};
43
44add (16) g36.0<1>UW g36.0<16,8,1>UW g46.1<16,8,1>UB {align1};
45add (16) g37.0<1>UW g37.0<16,8,1>UW g47.1<16,8,1>UB {align1};
46add (16) g38.0<1>UW g38.0<16,8,1>UW g48.1<16,8,1>UB {align1};
47add (16) g39.0<1>UW g39.0<16,8,1>UW g49.1<16,8,1>UB {align1};
48
49shr (32) g32.0<1>UW g32.0<16,16,1>UW 2UW {align1 compr};
50shr (32) g34.0<1>UW g34.0<16,16,1>UW 2UW {align1 compr};
51shr (32) g36.0<1>UW g36.0<16,16,1>UW 2UW {align1 compr};
52shr (32) g38.0<1>UW g38.0<16,16,1>UW 2UW {align1 compr};
53
54