1/*
2 */
3mov (1) g32.8<1>UD 0x007001FUD {align1};
4send (16) 0 g34.0<1>UW g32<8,8,1>UW read(input_surface1, 2, 0, 2) mlen 1 rlen 8 {align1}; //U
5send (16) 0 g44.0<1>UW g32<8,8,1>UW read(input_surface2, 2, 0, 2) mlen 1 rlen 8 {align1}; //V
6add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
7mov (1) g32.8<1>UD 0x1FUD {align1};
8send (16) 0 g42.0<1>UW g32<8,8,1>UW read(input_surface1, 2, 0, 2) mlen 1 rlen 1 {align1}; //U
9send (16) 0 g52.0<1>UW g32<8,8,1>UW read(input_surface2, 2, 0, 2) mlen 1 rlen 1 {align1}; //V
10
11//U
12add (8) g74.0<1>UW  g34.0<8,8,1>UB g34.1<8,8,1>UB {align1};
13add (8) g74.16<1>UW g35.0<8,8,1>UB g35.1<8,8,1>UB {align1};
14add (8) g75.0<1>UW  g36.0<8,8,1>UB g36.1<8,8,1>UB {align1};
15add (8) g75.16<1>UW g37.0<8,8,1>UB g37.1<8,8,1>UB {align1};
16add (8) g76.0<1>UW  g38.0<8,8,1>UB g38.1<8,8,1>UB {align1};
17add (8) g76.16<1>UW g39.0<8,8,1>UB g39.1<8,8,1>UB {align1};
18add (8) g77.0<1>UW  g40.0<8,8,1>UB g40.1<8,8,1>UB {align1};
19add (8) g77.16<1>UW g41.0<8,8,1>UB g41.1<8,8,1>UB {align1};
20
21add (8) g74.0<1>UW  g74.0<8,8,1>UW  g35.0<8,8,1>UB {align1};
22add (8) g74.16<1>UW g74.16<8,8,1>UW g36.0<8,8,1>UB {align1};
23add (8) g75.0<1>UW  g75.0<8,8,1>UW  g37.0<8,8,1>UB {align1};
24add (8) g75.16<1>UW g75.16<8,8,1>UW g38.0<8,8,1>UB {align1};
25add (8) g76.0<1>UW  g76.0<8,8,1>UW  g39.0<8,8,1>UB {align1};
26add (8) g76.16<1>UW g76.16<8,8,1>UW g40.0<8,8,1>UB {align1};
27add (8) g77.0<1>UW  g77.0<8,8,1>UW  g41.0<8,8,1>UB {align1};
28add (8) g77.16<1>UW g77.16<8,8,1>UW g42.0<8,8,1>UB {align1};
29
30add (8) g74.0<1>UW  g74.0<8,8,1>UW  g35.1<8,8,1>UB {align1};
31add (8) g74.16<1>UW g74.16<8,8,1>UW g36.1<8,8,1>UB {align1};
32add (8) g75.0<1>UW  g75.0<8,8,1>UW  g37.1<8,8,1>UB {align1};
33add (8) g75.16<1>UW g75.16<8,8,1>UW g38.1<8,8,1>UB {align1};
34add (8) g76.0<1>UW  g76.0<8,8,1>UW  g39.1<8,8,1>UB {align1};
35add (8) g76.16<1>UW g76.16<8,8,1>UW g40.1<8,8,1>UB {align1};
36add (8) g77.0<1>UW  g77.0<8,8,1>UW  g41.1<8,8,1>UB {align1};
37add (8) g77.16<1>UW g77.16<8,8,1>UW g42.1<8,8,1>UB {align1};
38
39//V
40add (8) g78.0<1>UW  g44.0<8,8,1>UB g44.1<8,8,1>UB {align1};
41add (8) g78.16<1>UW g45.0<8,8,1>UB g45.1<8,8,1>UB {align1};
42add (8) g79.0<1>UW  g46.0<8,8,1>UB g46.1<8,8,1>UB {align1};
43add (8) g79.16<1>UW g47.0<8,8,1>UB g47.1<8,8,1>UB {align1};
44add (8) g80.0<1>UW  g48.0<8,8,1>UB g48.1<8,8,1>UB {align1};
45add (8) g80.16<1>UW g49.0<8,8,1>UB g49.1<8,8,1>UB {align1};
46add (8) g81.0<1>UW  g50.0<8,8,1>UB g50.1<8,8,1>UB {align1};
47add (8) g81.16<1>UW g51.0<8,8,1>UB g51.1<8,8,1>UB {align1};
48
49add (8) g78.0<1>UW  g78.0<8,8,1>UW  g45.0<8,8,1>UB {align1};
50add (8) g78.16<1>UW g78.16<8,8,1>UW g46.0<8,8,1>UB {align1};
51add (8) g79.0<1>UW  g79.0<8,8,1>UW  g47.0<8,8,1>UB {align1};
52add (8) g79.16<1>UW g79.16<8,8,1>UW g48.0<8,8,1>UB {align1};
53add (8) g80.0<1>UW  g80.0<8,8,1>UW  g49.0<8,8,1>UB {align1};
54add (8) g80.16<1>UW g80.16<8,8,1>UW g50.0<8,8,1>UB {align1};
55add (8) g81.0<1>UW  g81.0<8,8,1>UW  g51.0<8,8,1>UB {align1};
56add (8) g81.16<1>UW g81.16<8,8,1>UW g52.0<8,8,1>UB {align1};
57
58add (8) g78.0<1>UW  g78.0<8,8,1>UW  g45.1<8,8,1>UB {align1};
59add (8) g78.16<1>UW g78.16<8,8,1>UW g46.1<8,8,1>UB {align1};
60add (8) g79.0<1>UW  g79.0<8,8,1>UW  g47.1<8,8,1>UB {align1};
61add (8) g79.16<1>UW g79.16<8,8,1>UW g48.1<8,8,1>UB {align1};
62add (8) g80.0<1>UW  g80.0<8,8,1>UW  g49.1<8,8,1>UB {align1};
63add (8) g80.16<1>UW g80.16<8,8,1>UW g50.1<8,8,1>UB {align1};
64add (8) g81.0<1>UW  g81.0<8,8,1>UW  g51.1<8,8,1>UB {align1};
65add (8) g81.16<1>UW g81.16<8,8,1>UW g52.1<8,8,1>UB {align1};
66
67shr (16) g74.0<1>UW g74.0<16,16,1>UW 2UW {align1};
68shr (16) g75.0<1>UW g75.0<16,16,1>UW 2UW {align1};
69shr (16) g76.0<1>UW g76.0<16,16,1>UW 2UW {align1};
70shr (16) g77.0<1>UW g77.0<16,16,1>UW 2UW {align1};
71shr (16) g78.0<1>UW g78.0<16,16,1>UW 2UW {align1};
72shr (16) g79.0<1>UW g79.0<16,16,1>UW 2UW {align1};
73shr (16) g80.0<1>UW g80.0<16,16,1>UW 2UW {align1};
74shr (16) g81.0<1>UW g81.0<16,16,1>UW 2UW {align1};
75