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   g84~g107:IDCT data                           */
1103b705cfSriastradh//mov (2) g31.0<1>UD g82.12<2,2,1>UW {align1};          //restore x and y
1203b705cfSriastradhmov (2) g31.0<1>UD g82.0<2,2,1>UD {align1};          //restore x and y
1303b705cfSriastradh
1403b705cfSriastradhand.nz (1) null g82.30<1,1,1>UB 0x1UW {align1};       //dct_type
1503b705cfSriastradh(f0) jmpi field_dct;
1603b705cfSriastradh
1703b705cfSriastradhadd.sat (32) g58.0<2>UB g84.0<16,16,1>W g58.0<16,16,2>UB {align1 compr};
1803b705cfSriastradhadd.sat (32) g60.0<2>UB g86.0<16,16,1>W g60.0<16,16,2>UB {align1 compr};
1903b705cfSriastradhadd.sat (32) g62.0<2>UB g88.0<16,16,1>W g62.0<16,16,2>UB {align1 compr};
2003b705cfSriastradhadd.sat (32) g64.0<2>UB g90.0<16,16,1>W g64.0<16,16,2>UB {align1 compr};
2103b705cfSriastradhadd.sat (32) g66.0<2>UB g92.0<16,16,1>W g66.0<16,16,2>UB {align1 compr};
2203b705cfSriastradhadd.sat (32) g68.0<2>UB g94.0<16,16,1>W g68.0<16,16,2>UB {align1 compr};
2303b705cfSriastradhadd.sat (32) g70.0<2>UB g96.0<16,16,1>W g70.0<16,16,2>UB {align1 compr};
2403b705cfSriastradhadd.sat (32) g72.0<2>UB g98.0<16,16,1>W g72.0<16,16,2>UB {align1 compr};
2503b705cfSriastradh
2603b705cfSriastradhadd.sat (32) g74.0<2>UB g100.0<16,16,1>W g74.0<16,16,2>UB {align1 compr};
2703b705cfSriastradhadd.sat (32) g76.0<2>UB g102.0<16,16,1>W g76.0<16,16,2>UB {align1 compr};
2803b705cfSriastradhadd.sat (32) g78.0<2>UB g104.0<16,16,1>W g78.0<16,16,2>UB {align1 compr};
2903b705cfSriastradhadd.sat (32) g80.0<2>UB g106.0<16,16,1>W g80.0<16,16,2>UB {align1 compr};
3003b705cfSriastradhjmpi write_back;
3103b705cfSriastradh
3203b705cfSriastradhfield_dct:
3303b705cfSriastradhadd.sat (16) g58.0<2>UB g84.0<16,16,1>W g58.0<16,16,2>UB {align1};
3403b705cfSriastradhadd.sat (16) g59.0<2>UB g92.0<16,16,1>W g59.0<16,16,2>UB {align1};
3503b705cfSriastradhadd.sat (16) g60.0<2>UB g85.0<16,16,1>W g60.0<16,16,2>UB {align1};
3603b705cfSriastradhadd.sat (16) g61.0<2>UB g93.0<16,16,1>W g61.0<16,16,2>UB {align1};
3703b705cfSriastradhadd.sat (16) g62.0<2>UB g86.0<16,16,1>W g62.0<16,16,2>UB {align1};
3803b705cfSriastradhadd.sat (16) g63.0<2>UB g94.0<16,16,1>W g63.0<16,16,2>UB {align1};
3903b705cfSriastradhadd.sat (16) g64.0<2>UB g87.0<16,16,1>W g64.0<16,16,2>UB {align1};
4003b705cfSriastradhadd.sat (16) g65.0<2>UB g95.0<16,16,1>W g65.0<16,16,2>UB {align1};
4103b705cfSriastradhadd.sat (16) g66.0<2>UB g88.0<16,16,1>W g66.0<16,16,2>UB {align1};
4203b705cfSriastradhadd.sat (16) g67.0<2>UB g96.0<16,16,1>W g67.0<16,16,2>UB {align1};
4303b705cfSriastradhadd.sat (16) g68.0<2>UB g89.0<16,16,1>W g68.0<16,16,2>UB {align1};
4403b705cfSriastradhadd.sat (16) g69.0<2>UB g97.0<16,16,1>W g69.0<16,16,2>UB {align1};
4503b705cfSriastradhadd.sat (16) g70.0<2>UB g90.0<16,16,1>W g70.0<16,16,2>UB {align1};
4603b705cfSriastradhadd.sat (16) g71.0<2>UB g98.0<16,16,1>W g71.0<16,16,2>UB {align1};
4703b705cfSriastradhadd.sat (16) g72.0<2>UB g91.0<16,16,1>W g72.0<16,16,2>UB {align1};
4803b705cfSriastradhadd.sat (16) g73.0<2>UB g99.0<16,16,1>W g73.0<16,16,2>UB {align1};
4903b705cfSriastradh/*
5003b705cfSriastradhadd.sat (16) g74.0<2>UB g100.0<16,16,1>W g74.0<16,16,2>UB {align1};
5103b705cfSriastradhadd.sat (16) g75.0<2>UB g101.0<16,16,1>W g75.0<16,16,2>UB {align1};
5203b705cfSriastradhadd.sat (16) g76.0<2>UB g102.0<16,16,1>W g76.0<16,16,2>UB {align1};
5303b705cfSriastradhadd.sat (16) g77.0<2>UB g103.0<16,16,1>W g77.0<16,16,2>UB {align1};
5403b705cfSriastradhadd.sat (16) g78.0<2>UB g104.0<16,16,1>W g78.0<16,16,2>UB {align1};
5503b705cfSriastradhadd.sat (16) g79.0<2>UB g105.0<16,16,1>W g79.0<16,16,2>UB {align1};
5603b705cfSriastradhadd.sat (16) g80.0<2>UB g106.0<16,16,1>W g80.0<16,16,2>UB {align1};
5703b705cfSriastradhadd.sat (16) g81.0<2>UB g107.0<16,16,1>W g81.0<16,16,2>UB {align1};
5803b705cfSriastradh*/
5903b705cfSriastradhadd (16) g100.0<1>W g100.0<16,16,1>W g74.0<16,16,2>UB {align1};
6003b705cfSriastradhadd (16) g101.0<1>W g101.0<16,16,1>W g75.0<16,16,2>UB {align1};
6103b705cfSriastradhadd (16) g102.0<1>W g102.0<16,16,1>W g76.0<16,16,2>UB {align1};
6203b705cfSriastradhadd (16) g103.0<1>W g103.0<16,16,1>W g77.0<16,16,2>UB {align1};
6303b705cfSriastradhadd (16) g104.0<1>W g104.0<16,16,1>W g78.0<16,16,2>UB {align1};
6403b705cfSriastradhadd (16) g105.0<1>W g105.0<16,16,1>W g79.0<16,16,2>UB {align1};
6503b705cfSriastradhadd (16) g106.0<1>W g106.0<16,16,1>W g80.0<16,16,2>UB {align1};
6603b705cfSriastradhadd (16) g107.0<1>W g107.0<16,16,1>W g81.0<16,16,2>UB {align1};
6703b705cfSriastradh
6803b705cfSriastradhmov.sat (16) g74.0<2>UB g100.0<16,16,1>W {align1};
6903b705cfSriastradhmov.sat (16) g75.0<2>UB g101.0<16,16,1>W {align1};
7003b705cfSriastradhmov.sat (16) g76.0<2>UB g102.0<16,16,1>W {align1};
7103b705cfSriastradhmov.sat (16) g77.0<2>UB g103.0<16,16,1>W {align1};
7203b705cfSriastradhmov.sat (16) g78.0<2>UB g104.0<16,16,1>W {align1};
7303b705cfSriastradhmov.sat (16) g79.0<2>UB g105.0<16,16,1>W {align1};
7403b705cfSriastradhmov.sat (16) g80.0<2>UB g106.0<16,16,1>W {align1};
7503b705cfSriastradhmov.sat (16) g81.0<2>UB g107.0<16,16,1>W {align1};
7603b705cfSriastradh
7703b705cfSriastradh
7803b705cfSriastradhwrite_back:
7903b705cfSriastradhmov (1) g31.8<1>UD 0x00F000FUD {align1};
8003b705cfSriastradh
8103b705cfSriastradhmov (16) m1.0<1>UB  g58.0<16,16,2>UB {align1};
8203b705cfSriastradhmov (16) m1.16<1>UB g59.0<16,16,2>UB {align1};
8303b705cfSriastradhmov (16) m2.0<1>UB  g60.0<16,16,2>UB {align1};
8403b705cfSriastradhmov (16) m2.16<1>UB g61.0<16,16,2>UB {align1};
8503b705cfSriastradhmov (16) m3.0<1>UB  g62.0<16,16,2>UB {align1};
8603b705cfSriastradhmov (16) m3.16<1>UB g63.0<16,16,2>UB {align1};
8703b705cfSriastradhmov (16) m4.0<1>UB  g64.0<16,16,2>UB {align1};
8803b705cfSriastradhmov (16) m4.16<1>UB g65.0<16,16,2>UB {align1};
8903b705cfSriastradhmov (16) m5.0<1>UB  g66.0<16,16,2>UB {align1};
9003b705cfSriastradhmov (16) m5.16<1>UB g67.0<16,16,2>UB {align1};
9103b705cfSriastradhmov (16) m6.0<1>UB  g68.0<16,16,2>UB {align1};
9203b705cfSriastradhmov (16) m6.16<1>UB g69.0<16,16,2>UB {align1};
9303b705cfSriastradhmov (16) m7.0<1>UB  g70.0<16,16,2>UB {align1};
9403b705cfSriastradhmov (16) m7.16<1>UB g71.0<16,16,2>UB {align1};
9503b705cfSriastradhmov (16) m8.0<1>UB  g72.0<16,16,2>UB {align1};
9603b705cfSriastradhmov (16) m8.16<1>UB g73.0<16,16,2>UB {align1};
9703b705cfSriastradhsend (16) 0 acc0<1>UW g31<8,8,1>UW write(0,0,2,0) mlen 9 rlen 0 {align1};
9803b705cfSriastradh
9903b705cfSriastradh//U
10003b705cfSriastradhmov (1) g31.8<1>UD 0x0070007UD  { align1 };
10103b705cfSriastradhshr (2) g31.0<1>UD g31.0<2,2,1>UD 1D {align1};
10203b705cfSriastradh
10303b705cfSriastradhmov (16) m1.0<1>UB g74.0<16,16,2>UB {align1};
10403b705cfSriastradhmov (16) m1.16<1>UB g75.0<16,16,2>UB {align1};
10503b705cfSriastradhmov (16) m2.0<1>UB g76.0<16,16,2>UB {align1};
10603b705cfSriastradhmov (16) m2.16<1>UB g77.0<16,16,2>UB {align1};
10703b705cfSriastradhsend (16) 0 acc0<1>UW g31<8,8,1>UW write(1, 0, 2, 0) mlen 3 rlen 0 { align1 };
10803b705cfSriastradh
10903b705cfSriastradh//V
11003b705cfSriastradhmov (16) m1.0<1>UB g78.0<16,16,2>UB {align1};
11103b705cfSriastradhmov (16) m1.16<1>UB g79.0<16,16,2>UB {align1};
11203b705cfSriastradhmov (16) m2.0<1>UB g80.0<16,16,2>UB {align1};
11303b705cfSriastradhmov (16) m2.16<1>UB g81.0<16,16,2>UB {align1};
11403b705cfSriastradhsend (16) 0 acc0<1>UW g31<8,8,1>UW write(2, 0, 2, 0) mlen 3 rlen 0 { align1 };
11503b705cfSriastradh
11603b705cfSriastradhsend (16) 0 acc0<1>UW g0<8,8,1>UW
11703b705cfSriastradh        thread_spawner(0, 0, 0) mlen 1 rlen 0 { align1 EOT};
118