1428d7b3dSmrg/*
2428d7b3dSmrg   GRF allocation:
3428d7b3dSmrg   g1~g30: constant buffer
4428d7b3dSmrg           g1~g2:intra IQ matrix
5428d7b3dSmrg           g3~g4:non intra IQ matrix
6428d7b3dSmrg           g5~g20:IDCT tab
7428d7b3dSmrg   g31:    read and write message descriptor
8428d7b3dSmrg   g32~g55:DCT data
9428d7b3dSmrg   g58~g81:reference data
10428d7b3dSmrg   g82:    thread payload 
11428d7b3dSmrg   g83~g106:IDCT data 
12428d7b3dSmrg*/
13428d7b3dSmrgmov (8) g82.0<1>UD g31.0<8,8,1>UD {align1};
14428d7b3dSmrg
15428d7b3dSmrgmov.sat (16) g33.0<2>UB g33.0<16,16,1>W {align1};
16428d7b3dSmrgmov.sat (16) g34.0<2>UB g34.0<16,16,1>W {align1};
17428d7b3dSmrgmov.sat (16) g35.0<2>UB g35.0<16,16,1>W {align1};
18428d7b3dSmrgmov.sat (16) g36.0<2>UB g36.0<16,16,1>W {align1};
19428d7b3dSmrgmov.sat (16) g37.0<2>UB g37.0<16,16,1>W {align1};
20428d7b3dSmrgmov.sat (16) g38.0<2>UB g38.0<16,16,1>W {align1};
21428d7b3dSmrgmov.sat (16) g39.0<2>UB g39.0<16,16,1>W {align1};
22428d7b3dSmrgmov.sat (16) g40.0<2>UB g40.0<16,16,1>W {align1};
23428d7b3dSmrgmov.sat (16) g41.0<2>UB g41.0<16,16,1>W {align1};
24428d7b3dSmrgmov.sat (16) g42.0<2>UB g42.0<16,16,1>W {align1};
25428d7b3dSmrgmov.sat (16) g43.0<2>UB g43.0<16,16,1>W {align1};
26428d7b3dSmrgmov.sat (16) g44.0<2>UB g44.0<16,16,1>W {align1};
27428d7b3dSmrgmov.sat (16) g45.0<2>UB g45.0<16,16,1>W {align1};
28428d7b3dSmrgmov.sat (16) g46.0<2>UB g46.0<16,16,1>W {align1};
29428d7b3dSmrgmov.sat (16) g47.0<2>UB g47.0<16,16,1>W {align1};
30428d7b3dSmrgmov.sat (16) g48.0<2>UB g48.0<16,16,1>W {align1};
31428d7b3dSmrg
32428d7b3dSmrgmov.sat (16) g49.0<2>UB g49.0<16,16,1>W {align1};
33428d7b3dSmrgmov.sat (16) g50.0<2>UB g50.0<16,16,1>W {align1};
34428d7b3dSmrgmov.sat (16) g51.0<2>UB g51.0<16,16,1>W {align1};
35428d7b3dSmrgmov.sat (16) g52.0<2>UB g52.0<16,16,1>W {align1};
36428d7b3dSmrgmov.sat (16) g53.0<2>UB g53.0<16,16,1>W {align1};
37428d7b3dSmrgmov.sat (16) g54.0<2>UB g54.0<16,16,1>W {align1};
38428d7b3dSmrgmov.sat (16) g55.0<2>UB g55.0<16,16,1>W {align1};
39428d7b3dSmrgmov.sat (16) g56.0<2>UB g56.0<16,16,1>W {align1};
40428d7b3dSmrg
41428d7b3dSmrgmov (1) g31.8<1>UD 0x00F000FUD {align1};
42428d7b3dSmrg
43428d7b3dSmrgand.nz (1) null g82.30<1,1,1>UB 0x1UW{align1};
44428d7b3dSmrg(f0) jmpi field_dct_y;
45428d7b3dSmrg
46428d7b3dSmrgmov (16) m1.0<1>UB  g33.0<16,16,2>UB {align1};
47428d7b3dSmrgmov (16) m1.16<1>UB g34.0<16,16,2>UB {align1};
48428d7b3dSmrgmov (16) m2.0<1>UB  g35.0<16,16,2>UB {align1};
49428d7b3dSmrgmov (16) m2.16<1>UB g36.0<16,16,2>UB {align1};
50428d7b3dSmrgmov (16) m3.0<1>UB  g37.0<16,16,2>UB {align1};
51428d7b3dSmrgmov (16) m3.16<1>UB g38.0<16,16,2>UB {align1};
52428d7b3dSmrgmov (16) m4.0<1>UB  g39.0<16,16,2>UB {align1};
53428d7b3dSmrgmov (16) m4.16<1>UB g40.0<16,16,2>UB {align1};
54428d7b3dSmrgmov (16) m5.0<1>UB  g41.0<16,16,2>UB {align1};
55428d7b3dSmrgmov (16) m5.16<1>UB g42.0<16,16,2>UB {align1};
56428d7b3dSmrgmov (16) m6.0<1>UB  g43.0<16,16,2>UB {align1};
57428d7b3dSmrgmov (16) m6.16<1>UB g44.0<16,16,2>UB {align1};
58428d7b3dSmrgmov (16) m7.0<1>UB  g45.0<16,16,2>UB {align1};
59428d7b3dSmrgmov (16) m7.16<1>UB g46.0<16,16,2>UB {align1};
60428d7b3dSmrgmov (16) m8.0<1>UB  g47.0<16,16,2>UB {align1};
61428d7b3dSmrgmov (16) m8.16<1>UB g48.0<16,16,2>UB {align1};
62428d7b3dSmrgjmpi write_back_y;
63428d7b3dSmrg
64428d7b3dSmrgfield_dct_y:
65428d7b3dSmrgmov (16) m1.0<1>UB  g33.0<16,16,2>UB {align1};
66428d7b3dSmrgmov (16) m1.16<1>UB g41.0<16,16,2>UB {align1};
67428d7b3dSmrgmov (16) m2.0<1>UB  g34.0<16,16,2>UB {align1};
68428d7b3dSmrgmov (16) m2.16<1>UB g42.0<16,16,2>UB {align1};
69428d7b3dSmrgmov (16) m3.0<1>UB  g35.0<16,16,2>UB {align1};
70428d7b3dSmrgmov (16) m3.16<1>UB g43.0<16,16,2>UB {align1};
71428d7b3dSmrgmov (16) m4.0<1>UB  g36.0<16,16,2>UB {align1};
72428d7b3dSmrgmov (16) m4.16<1>UB g44.0<16,16,2>UB {align1};
73428d7b3dSmrgmov (16) m5.0<1>UB  g37.0<16,16,2>UB {align1};
74428d7b3dSmrgmov (16) m5.16<1>UB g45.0<16,16,2>UB {align1};
75428d7b3dSmrgmov (16) m6.0<1>UB  g38.0<16,16,2>UB {align1};
76428d7b3dSmrgmov (16) m6.16<1>UB g46.0<16,16,2>UB {align1};
77428d7b3dSmrgmov (16) m7.0<1>UB  g39.0<16,16,2>UB {align1};
78428d7b3dSmrgmov (16) m7.16<1>UB g47.0<16,16,2>UB {align1};
79428d7b3dSmrgmov (16) m8.0<1>UB  g40.0<16,16,2>UB {align1};
80428d7b3dSmrgmov (16) m8.16<1>UB g48.0<16,16,2>UB {align1};
81428d7b3dSmrg
82428d7b3dSmrgwrite_back_y:
83428d7b3dSmrgsend (16) 0 acc0<1>UW g31<8,8,1>UW write(0,0,2,0) mlen 9 rlen 0 {align1};
84428d7b3dSmrg
85428d7b3dSmrg//U
86428d7b3dSmrgmov (1) g31.8<1>UD 0x0070007UD  { align1 };
87428d7b3dSmrgshr (2) g31.0<1>UD g82.0<2,2,1>UD 1D {align1};
88428d7b3dSmrg
89428d7b3dSmrgmov (16) m1.0<1>UB  g49.0<16,16,2>UB {align1};
90428d7b3dSmrgmov (16) m1.16<1>UB g50.0<16,16,2>UB {align1};
91428d7b3dSmrgmov (16) m2.0<1>UB  g51.0<16,16,2>UB {align1};
92428d7b3dSmrgmov (16) m2.16<1>UB g52.0<16,16,2>UB {align1};
93428d7b3dSmrgsend (16) 0 acc0<1>UW g31<8,8,1>UW write(1, 0, 2, 0) mlen 3 rlen 0 { align1 };
94428d7b3dSmrg
95428d7b3dSmrg//V
96428d7b3dSmrgmov (16) m1.0<1>UB  g53.0<16,16,2>UB {align1};
97428d7b3dSmrgmov (16) m1.16<1>UB g54.0<16,16,2>UB {align1};
98428d7b3dSmrgmov (16) m2.0<1>UB  g55.0<16,16,2>UB {align1};
99428d7b3dSmrgmov (16) m2.16<1>UB g56.0<16,16,2>UB {align1};
100428d7b3dSmrgsend (16) 0 acc0<1>UW g31<8,8,1>UW write(2, 0, 2, 0) mlen 3 rlen 0 { align1 };
101428d7b3dSmrg
102428d7b3dSmrgOUT:
103428d7b3dSmrgsend (16) 0 acc0<1>UW g0<8,8,1>UW
104428d7b3dSmrg        thread_spawner(0, 0, 0) mlen 1 rlen 0 { align1 EOT};
105428d7b3dSmrg
106