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};
12and.nz (1) null g115.16<1,1,1>UW 1UW {align1};
13(f0) jmpi read_backward;
14
15send (16) 0 g40.0<1>UW g115<8,8,1>UW read(5, 2, 0, 2) mlen 1 rlen 4 {align1};//U
16send (16) 0 g45.0<1>UW g115<8,8,1>UW read(6, 2, 0, 2) mlen 1 rlen 4 {align1};//V
17mov (1) g115.8<1>UD 0xFUD {align1};
18add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
19send (16) 0 g44.0<1>UW g115<8,8,1>UW read(5, 2, 0, 2) mlen 1 rlen 1 {align1};//U
20send (16) 0 g49.0<1>UW g115<8,8,1>UW read(6, 2, 0, 2) mlen 1 rlen 1 {align1};//V
21jmpi put_data;
22
23read_backward:
24send (16) 0 g40.0<1>UW g115<8,8,1>UW read(8, 2, 0, 2) mlen 1 rlen 4 {align1};//U
25send (16) 0 g45.0<1>UW g115<8,8,1>UW read(9, 2, 0, 2) mlen 1 rlen 4 {align1};//V
26mov (1) g115.8<1>UD 0xFUD {align1};
27add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
28send (16) 0 g44.0<1>UW g115<8,8,1>UW read(8, 2, 0, 2) mlen 1 rlen 1 {align1};//U
29send (16) 0 g49.0<1>UW g115<8,8,1>UW read(9, 2, 0, 2) mlen 1 rlen 1 {align1};//V
30
31put_data:
32//U
33add (8) g32.0<1>UW g40.0<8,8,1>UB g41.0<8,8,1>UB {align1};
34add (8) g33.0<1>UW g41.0<8,8,1>UB g42.0<8,8,1>UB {align1};
35add (8) g34.0<1>UW g42.0<8,8,1>UB g43.0<8,8,1>UB {align1};
36add (8) g35.0<1>UW g43.0<8,8,1>UB g44.0<8,8,1>UB {align1};
37
38add (8) g32.0<1>UW g32.0<8,8,1>UW g40.1<8,8,1>UB {align1};
39add (8) g33.0<1>UW g33.0<8,8,1>UW g41.1<8,8,1>UB {align1};
40add (8) g34.0<1>UW g34.0<8,8,1>UW g42.1<8,8,1>UB {align1};
41add (8) g35.0<1>UW g35.0<8,8,1>UW g43.1<8,8,1>UB {align1};
42
43add (8) g32.0<1>UW g32.0<8,8,1>UW g41.1<8,8,1>UB {align1};
44add (8) g33.0<1>UW g33.0<8,8,1>UW g42.1<8,8,1>UB {align1};
45add (8) g34.0<1>UW g34.0<8,8,1>UW g43.1<8,8,1>UB {align1};
46add (8) g35.0<1>UW g35.0<8,8,1>UW g44.1<8,8,1>UB {align1};
47//V
48add (8) g36.0<1>UW g45.0<8,8,1>UB g46.0<8,8,1>UB {align1};
49add (8) g37.0<1>UW g46.0<8,8,1>UB g47.0<8,8,1>UB {align1};
50add (8) g38.0<1>UW g47.0<8,8,1>UB g48.0<8,8,1>UB {align1};
51add (8) g39.0<1>UW g48.0<8,8,1>UB g49.0<8,8,1>UB {align1};
52
53add (8) g36.0<1>UW g36.0<8,8,1>UW g45.1<8,8,1>UB {align1};
54add (8) g37.0<1>UW g37.0<8,8,1>UW g46.1<8,8,1>UB {align1};
55add (8) g38.0<1>UW g38.0<8,8,1>UW g47.1<8,8,1>UB {align1};
56add (8) g39.0<1>UW g39.0<8,8,1>UW g48.1<8,8,1>UB {align1};
57
58add (8) g36.0<1>UW g36.0<8,8,1>UW g46.1<8,8,1>UB {align1};
59add (8) g37.0<1>UW g37.0<8,8,1>UW g47.1<8,8,1>UB {align1};
60add (8) g38.0<1>UW g38.0<8,8,1>UW g48.1<8,8,1>UB {align1};
61add (8) g39.0<1>UW g39.0<8,8,1>UW g49.1<8,8,1>UB {align1};
62
63shr (32) g32.0<1>UW g32.0<16,16,1>UW 2UW {align1 compr};
64shr (32) g34.0<1>UW g34.0<16,16,1>UW 2UW {align1 compr};
65shr (32) g36.0<1>UW g36.0<16,16,1>UW 2UW {align1 compr};
66shr (32) g38.0<1>UW g38.0<16,16,1>UW 2UW {align1 compr};
67