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 0x07001FUD {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(4,2,0,2) mlen 1 rlen 8 {align1};
18add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
19send (16) 0 g48.0<1>UW g115<8,8,1>UW read(4,2,0,2) mlen 1 rlen 8 {align1};
20add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
21mov (1) g115.8<1>UD 0x1FUD {align1};
22send (16) 0 g56.0<1>UW g115<8,8,1>UW read(4,2,0,2) mlen 1 rlen 1 {align1};
23jmpi put_data;
24
25read_backward:
26send (16) 0 g40.0<1>UW g115<8,8,1>UW read(7,2,0,2) mlen 1 rlen 8 {align1};
27add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
28send (16) 0 g48.0<1>UW g115<8,8,1>UW read(7,2,0,2) mlen 1 rlen 8 {align1};
29add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
30mov (1) g115.8<1>UD 0x1FUD {align1};
31send (16) 0 g56.0<1>UW g115<8,8,1>UW read(7,2,0,2) mlen 1 rlen 1 {align1};
32
33put_data:
34add (16) g32.0<1>UW g40.0<16,16,1>UB g42.0<16,16,1>UB {align1};
35add (16) g33.0<1>UW g42.0<16,16,1>UB g44.0<16,16,1>UB {align1};
36add (16) g34.0<1>UW g44.0<16,16,1>UB g46.0<16,16,1>UB {align1};
37add (16) g35.0<1>UW g46.0<16,16,1>UB g48.0<16,16,1>UB {align1};
38add (16) g36.0<1>UW g48.0<16,16,1>UB g50.0<16,16,1>UB {align1};
39add (16) g37.0<1>UW g50.0<16,16,1>UB g52.0<16,16,1>UB {align1};
40add (16) g38.0<1>UW g52.0<16,16,1>UB g54.0<16,16,1>UB {align1};
41add (16) g39.0<1>UW g54.0<16,16,1>UB g56.0<16,16,1>UB {align1};
42
43add (16) g32.0<1>UW g32.0<16,16,1>UW g40.1<16,16,1>UB {align1};
44add (16) g33.0<1>UW g33.0<16,16,1>UW g42.1<16,16,1>UB {align1};
45add (16) g34.0<1>UW g34.0<16,16,1>UW g44.1<16,16,1>UB {align1};
46add (16) g35.0<1>UW g35.0<16,16,1>UW g46.1<16,16,1>UB {align1};
47add (16) g36.0<1>UW g36.0<16,16,1>UW g48.1<16,16,1>UB {align1};
48add (16) g37.0<1>UW g37.0<16,16,1>UW g50.1<16,16,1>UB {align1};
49add (16) g38.0<1>UW g38.0<16,16,1>UW g52.1<16,16,1>UB {align1};
50add (16) g39.0<1>UW g39.0<16,16,1>UW g54.1<16,16,1>UB {align1};
51
52add (16) g32.0<1>UW g32.0<16,16,1>UW g42.1<16,16,1>UB {align1};
53add (16) g33.0<1>UW g33.0<16,16,1>UW g44.1<16,16,1>UB {align1};
54add (16) g34.0<1>UW g34.0<16,16,1>UW g46.1<16,16,1>UB {align1};
55add (16) g35.0<1>UW g35.0<16,16,1>UW g48.1<16,16,1>UB {align1};
56add (16) g36.0<1>UW g36.0<16,16,1>UW g50.1<16,16,1>UB {align1};
57add (16) g37.0<1>UW g37.0<16,16,1>UW g52.1<16,16,1>UB {align1};
58add (16) g38.0<1>UW g38.0<16,16,1>UW g54.1<16,16,1>UB {align1};
59add (16) g39.0<1>UW g39.0<16,16,1>UW g56.1<16,16,1>UB {align1};
60
61shr (32) g32.0<1>UW g32.0<16,16,1>UW 2UW {align1 compr};
62shr (32) g34.0<1>UW g34.0<16,16,1>UW 2UW {align1 compr};
63shr (32) g36.0<1>UW g36.0<16,16,1>UW 2UW {align1 compr};
64shr (32) g38.0<1>UW g38.0<16,16,1>UW 2UW {align1 compr};
65