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   g32:    message descriptor for reading reference data
8   g58~g81:reference data
9   g82:    thread payload backup
10   g83~g106:IDCT data                           */
11
12mov (1) g32.8<1>UD 0x007001FUD {align1};
13and.nz (1) null g32.16<1,1,1>UW 1UW {align1};
14(f0) jmpi read_backward;
15
16send (16) 0 g38.0<1>UW g32<8,8,1>UW read(4, 2, 0, 2) mlen 1 rlen 8 {align1};
17add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
18send (16) 0 g46.0<1>UW g32<8,8,1>UW read(4, 2, 0, 2) mlen 1 rlen 8 {align1};
19add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
20mov (1) g32.8<1>UD 0x1FUD {align1};
21send (16) 0 g54.0<1>UW g32<8,8,1>UW read(4, 2, 0, 2) mlen 1 rlen 1 {align1};
22jmpi put_data;
23
24read_backward:
25send (16) 0 g38.0<1>UW g32<8,8,1>UW read(7, 2, 0, 2) mlen 1 rlen 8 {align1};
26add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
27send (16) 0 g46.0<1>UW g32<8,8,1>UW read(7, 2, 0, 2) mlen 1 rlen 8 {align1};
28add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
29mov (1) g32.8<1>UD 0x1FUD {align1};
30send (16) 0 g54.0<1>UW g32<8,8,1>UW read(7, 2, 0, 2) mlen 1 rlen 1 {align1};
31
32put_data:
33add (16) g58.0<1>UW g38.0<16,16,1>UB g38.1<16,16,1>UB {align1};
34add (16) g59.0<1>UW g39.0<16,16,1>UB g39.1<16,16,1>UB {align1};
35add (16) g60.0<1>UW g40.0<16,16,1>UB g40.1<16,16,1>UB {align1};
36add (16) g61.0<1>UW g41.0<16,16,1>UB g41.1<16,16,1>UB {align1};
37add (16) g62.0<1>UW g42.0<16,16,1>UB g42.1<16,16,1>UB {align1};
38add (16) g63.0<1>UW g43.0<16,16,1>UB g43.1<16,16,1>UB {align1};
39add (16) g64.0<1>UW g44.0<16,16,1>UB g44.1<16,16,1>UB {align1};
40add (16) g65.0<1>UW g45.0<16,16,1>UB g45.1<16,16,1>UB {align1};
41add (16) g66.0<1>UW g46.0<16,16,1>UB g46.1<16,16,1>UB {align1};
42add (16) g67.0<1>UW g47.0<16,16,1>UB g47.1<16,16,1>UB {align1};
43add (16) g68.0<1>UW g48.0<16,16,1>UB g48.1<16,16,1>UB {align1};
44add (16) g69.0<1>UW g49.0<16,16,1>UB g49.1<16,16,1>UB {align1};
45add (16) g70.0<1>UW g50.0<16,16,1>UB g50.1<16,16,1>UB {align1};
46add (16) g71.0<1>UW g51.0<16,16,1>UB g51.1<16,16,1>UB {align1};
47add (16) g72.0<1>UW g52.0<16,16,1>UB g52.1<16,16,1>UB {align1};
48add (16) g73.0<1>UW g53.0<16,16,1>UB g53.1<16,16,1>UB {align1};
49
50add (16) g58.0<1>UW g58.0<16,16,1>UW g39.0<16,16,1>UB {align1};
51add (16) g59.0<1>UW g59.0<16,16,1>UW g40.0<16,16,1>UB {align1};
52add (16) g60.0<1>UW g60.0<16,16,1>UW g41.0<16,16,1>UB {align1};
53add (16) g61.0<1>UW g61.0<16,16,1>UW g42.0<16,16,1>UB {align1};
54add (16) g62.0<1>UW g62.0<16,16,1>UW g43.0<16,16,1>UB {align1};
55add (16) g63.0<1>UW g63.0<16,16,1>UW g44.0<16,16,1>UB {align1};
56add (16) g64.0<1>UW g64.0<16,16,1>UW g45.0<16,16,1>UB {align1};
57add (16) g65.0<1>UW g65.0<16,16,1>UW g46.0<16,16,1>UB {align1};
58add (16) g66.0<1>UW g66.0<16,16,1>UW g47.0<16,16,1>UB {align1};
59add (16) g67.0<1>UW g67.0<16,16,1>UW g48.0<16,16,1>UB {align1};
60add (16) g68.0<1>UW g68.0<16,16,1>UW g49.0<16,16,1>UB {align1};
61add (16) g69.0<1>UW g69.0<16,16,1>UW g50.0<16,16,1>UB {align1};
62add (16) g70.0<1>UW g70.0<16,16,1>UW g51.0<16,16,1>UB {align1};
63add (16) g71.0<1>UW g71.0<16,16,1>UW g52.0<16,16,1>UB {align1};
64add (16) g72.0<1>UW g72.0<16,16,1>UW g53.0<16,16,1>UB {align1};
65add (16) g73.0<1>UW g73.0<16,16,1>UW g54.0<16,16,1>UB {align1};
66
67add (16) g58.0<1>UW g58.0<16,16,1>UW g39.1<16,16,1>UB {align1};
68add (16) g59.0<1>UW g59.0<16,16,1>UW g40.1<16,16,1>UB {align1};
69add (16) g60.0<1>UW g60.0<16,16,1>UW g41.1<16,16,1>UB {align1};
70add (16) g61.0<1>UW g61.0<16,16,1>UW g42.1<16,16,1>UB {align1};
71add (16) g62.0<1>UW g62.0<16,16,1>UW g43.1<16,16,1>UB {align1};
72add (16) g63.0<1>UW g63.0<16,16,1>UW g44.1<16,16,1>UB {align1};
73add (16) g64.0<1>UW g64.0<16,16,1>UW g45.1<16,16,1>UB {align1};
74add (16) g65.0<1>UW g65.0<16,16,1>UW g46.1<16,16,1>UB {align1};
75add (16) g66.0<1>UW g66.0<16,16,1>UW g47.1<16,16,1>UB {align1};
76add (16) g67.0<1>UW g67.0<16,16,1>UW g48.1<16,16,1>UB {align1};
77add (16) g68.0<1>UW g68.0<16,16,1>UW g49.1<16,16,1>UB {align1};
78add (16) g69.0<1>UW g69.0<16,16,1>UW g50.1<16,16,1>UB {align1};
79add (16) g70.0<1>UW g70.0<16,16,1>UW g51.1<16,16,1>UB {align1};
80add (16) g71.0<1>UW g71.0<16,16,1>UW g52.1<16,16,1>UB {align1};
81add (16) g72.0<1>UW g72.0<16,16,1>UW g53.1<16,16,1>UB {align1};
82add (16) g73.0<1>UW g73.0<16,16,1>UW g54.1<16,16,1>UB {align1};
83
84shr (32) g58.0<1>UW g58.0<16,16,1>UW 2UW {align1 compr};
85shr (32) g60.0<1>UW g60.0<16,16,1>UW 2UW {align1 compr};
86shr (32) g62.0<1>UW g62.0<16,16,1>UW 2UW {align1 compr};
87shr (32) g64.0<1>UW g64.0<16,16,1>UW 2UW {align1 compr};
88shr (32) g66.0<1>UW g66.0<16,16,1>UW 2UW {align1 compr};
89shr (32) g68.0<1>UW g68.0<16,16,1>UW 2UW {align1 compr};
90shr (32) g70.0<1>UW g70.0<16,16,1>UW 2UW {align1 compr};
91shr (32) g72.0<1>UW g72.0<16,16,1>UW 2UW {align1 compr};
92
93