103b705cfSriastradh/* GRF allocation:
203b705cfSriastradh   g1~g30: constant buffer
303b705cfSriastradh           g1~g2:intra IQ matrix
403b705cfSriastradh           g3~g4:non intra IQ matrix
503b705cfSriastradh           g5~g20:IDCT table
603b705cfSriastradh   g31:    thread payload 
703b705cfSriastradh   g32:    message descriptor for reading reference data
803b705cfSriastradh   g58~g81:reference data
903b705cfSriastradh   g82:    thread payload backup
1003b705cfSriastradh   g83~g106:IDCT data                           */
1103b705cfSriastradh/*
1203b705cfSriastradhmov (8) g82.0<1>UD g31.0<8,8,1>UD {align1};
1303b705cfSriastradhdefine(`UV_red',`0xffffffffUD')
1403b705cfSriastradhdefine(`UV_white',`0x7f7f7f7fUD')
1503b705cfSriastradhdefine(`UV_green',`0x00000000UD')
1603b705cfSriastradhmov(1) g31.8<1>UD 0x000f000fUD  { align1 };
1703b705cfSriastradhmov(16) m1<1>UD 0xFFFFFFFFUD {align1 compr};
1803b705cfSriastradhmov(16) m3<1>UD 0xFFFFFFFFUD {align1 compr};
1903b705cfSriastradhmov(16) m5<1>UD 0xFFFFFFFFUD {align1 compr};
2003b705cfSriastradhmov(16) m7<1>UD 0xFFFFFFFFUD {align1 compr};
2103b705cfSriastradhsend (16) 0 acc0<1>UW g31<8,8,1>UW write(0, 0, 2, 0) mlen 9 rlen 0 { align1 };
2203b705cfSriastradhshr (2) g31.0<1>UD g82.0<2,2,1>UD  1UW {align1};
2303b705cfSriastradhmov(1) g31.8<1>UD 0x00070007UD  { align1 };
2403b705cfSriastradhmov (16) m1<1>UD UV_green {align1 compr};
2503b705cfSriastradhsend (16) 0 acc0<1>UW g31<8,8,1>UW write(2, 0, 2, 0) mlen 3 rlen 0 { align1 };
2603b705cfSriastradhsend (16) 0 acc0<1>UW g31<8,8,1>UW write(1, 0, 2, 0) mlen 3 rlen 0 { align1 };
2703b705cfSriastradhsend (16) 0 acc0<1>UW g0<8,8,1>UW 
2803b705cfSriastradh	thread_spawner(0, 0, 0) mlen 1 rlen 0 { align1 EOT};
2903b705cfSriastradh*/
3003b705cfSriastradh
3103b705cfSriastradhmov (8) g82.0<1>UD g31.0<8,8,1>UD {align1};
3203b705cfSriastradhmov (1) g126.8<1>UD ip {align1};
3303b705cfSriastradhadd (1) ip g21.0<1,1,1>UD 0x50UD {align1};  //jump to the lib to do IDCT         
3403b705cfSriastradh 
3503b705cfSriastradh//Y, (x', y') = (x, y) + (motion_vector.x >> 1, motion_vector.y >> 1) 
3603b705cfSriastradhasr (2) g31.14<1>W g82.18<2,2,1>W 1W {align1};
3703b705cfSriastradhadd (2) g32.0<1>UD g31.0<2,2,1>UD g31.14<2,2,1>W {align1};
3803b705cfSriastradhmov (1) g32.16<1>UW 1UW {align1};           //0:forward 1:backward
3903b705cfSriastradhmov (1) a0.0<1>UD 0x0A52UD {align1};        //g82.18,motion vector
4003b705cfSriastradhmov (1) g126.8<1>UD ip {align1};
4103b705cfSriastradhadd (1) ip g21.0<1,1,1>UD 0x0UD {align1};   //jump to the lib to read reference data  
4203b705cfSriastradh
4303b705cfSriastradh//UV, (x', y') = (x >> 1, y >> 1) + (motion_vector.x >> 2, motion_vector.y >> 2)
4403b705cfSriastradhshr (2) g31.0<1>UD g31.0<2,2,1>UD 1UD {align1};
4503b705cfSriastradhasr (2) g31.14<1>W g82.18<2,2,1>W 2W {align1};
4603b705cfSriastradhadd (2) g32.0<1>UD g31.0<2,2,1>UD g31.14<2,2,1>W {align1};
4703b705cfSriastradhmov (1) g126.8<1>UD ip {align1};
4803b705cfSriastradhadd (1) ip g21.0<1,1,1>UD 0x10UD {align1};          //jump to the lib to read reference data  
4903b705cfSriastradh
5003b705cfSriastradhadd (1) ip g21.0<1,1,1>UD 0x40UD {align1};          //jump to the lib to add the reference and idct data
51