/* GRF allocation: g1~g30: constant buffer g1~g2:intra IQ matrix g3~g4:non intra IQ matrix g5~g20:IDCT table g31: thread payload g32: message descriptor for reading reference data g58~g81:reference data g82: thread payload backup g83~g106:IDCT data */ mov (1) g32.8<1>UD 0x007001FUD {align1}; and.nz (1) null g32.16<1,1,1>UW 1UW {align1}; (f0) jmpi read_backward; send (16) 0 g38.0<1>UW g32<8,8,1>UW read(4, 2, 0, 2) mlen 1 rlen 8 {align1}; add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1}; send (16) 0 g46.0<1>UW g32<8,8,1>UW read(4, 2, 0, 2) mlen 1 rlen 8 {align1}; add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1}; mov (1) g32.8<1>UD 0x1FUD {align1}; send (16) 0 g54.0<1>UW g32<8,8,1>UW read(4, 2, 0, 2) mlen 1 rlen 1 {align1}; jmpi put_data; read_backward: send (16) 0 g38.0<1>UW g32<8,8,1>UW read(7, 2, 0, 2) mlen 1 rlen 8 {align1}; add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1}; send (16) 0 g46.0<1>UW g32<8,8,1>UW read(7, 2, 0, 2) mlen 1 rlen 8 {align1}; add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1}; mov (1) g32.8<1>UD 0x1FUD {align1}; send (16) 0 g54.0<1>UW g32<8,8,1>UW read(7, 2, 0, 2) mlen 1 rlen 1 {align1}; put_data: add (16) g58.0<1>UW g38.0<16,16,1>UB g38.1<16,16,1>UB {align1}; add (16) g59.0<1>UW g39.0<16,16,1>UB g39.1<16,16,1>UB {align1}; add (16) g60.0<1>UW g40.0<16,16,1>UB g40.1<16,16,1>UB {align1}; add (16) g61.0<1>UW g41.0<16,16,1>UB g41.1<16,16,1>UB {align1}; add (16) g62.0<1>UW g42.0<16,16,1>UB g42.1<16,16,1>UB {align1}; add (16) g63.0<1>UW g43.0<16,16,1>UB g43.1<16,16,1>UB {align1}; add (16) g64.0<1>UW g44.0<16,16,1>UB g44.1<16,16,1>UB {align1}; add (16) g65.0<1>UW g45.0<16,16,1>UB g45.1<16,16,1>UB {align1}; add (16) g66.0<1>UW g46.0<16,16,1>UB g46.1<16,16,1>UB {align1}; add (16) g67.0<1>UW g47.0<16,16,1>UB g47.1<16,16,1>UB {align1}; add (16) g68.0<1>UW g48.0<16,16,1>UB g48.1<16,16,1>UB {align1}; add (16) g69.0<1>UW g49.0<16,16,1>UB g49.1<16,16,1>UB {align1}; add (16) g70.0<1>UW g50.0<16,16,1>UB g50.1<16,16,1>UB {align1}; add (16) g71.0<1>UW g51.0<16,16,1>UB g51.1<16,16,1>UB {align1}; add (16) g72.0<1>UW g52.0<16,16,1>UB g52.1<16,16,1>UB {align1}; add (16) g73.0<1>UW g53.0<16,16,1>UB g53.1<16,16,1>UB {align1}; add (16) g58.0<1>UW g58.0<16,16,1>UW g39.0<16,16,1>UB {align1}; add (16) g59.0<1>UW g59.0<16,16,1>UW g40.0<16,16,1>UB {align1}; add (16) g60.0<1>UW g60.0<16,16,1>UW g41.0<16,16,1>UB {align1}; add (16) g61.0<1>UW g61.0<16,16,1>UW g42.0<16,16,1>UB {align1}; add (16) g62.0<1>UW g62.0<16,16,1>UW g43.0<16,16,1>UB {align1}; add (16) g63.0<1>UW g63.0<16,16,1>UW g44.0<16,16,1>UB {align1}; add (16) g64.0<1>UW g64.0<16,16,1>UW g45.0<16,16,1>UB {align1}; add (16) g65.0<1>UW g65.0<16,16,1>UW g46.0<16,16,1>UB {align1}; add (16) g66.0<1>UW g66.0<16,16,1>UW g47.0<16,16,1>UB {align1}; add (16) g67.0<1>UW g67.0<16,16,1>UW g48.0<16,16,1>UB {align1}; add (16) g68.0<1>UW g68.0<16,16,1>UW g49.0<16,16,1>UB {align1}; add (16) g69.0<1>UW g69.0<16,16,1>UW g50.0<16,16,1>UB {align1}; add (16) g70.0<1>UW g70.0<16,16,1>UW g51.0<16,16,1>UB {align1}; add (16) g71.0<1>UW g71.0<16,16,1>UW g52.0<16,16,1>UB {align1}; add (16) g72.0<1>UW g72.0<16,16,1>UW g53.0<16,16,1>UB {align1}; add (16) g73.0<1>UW g73.0<16,16,1>UW g54.0<16,16,1>UB {align1}; add (16) g58.0<1>UW g58.0<16,16,1>UW g39.1<16,16,1>UB {align1}; add (16) g59.0<1>UW g59.0<16,16,1>UW g40.1<16,16,1>UB {align1}; add (16) g60.0<1>UW g60.0<16,16,1>UW g41.1<16,16,1>UB {align1}; add (16) g61.0<1>UW g61.0<16,16,1>UW g42.1<16,16,1>UB {align1}; add (16) g62.0<1>UW g62.0<16,16,1>UW g43.1<16,16,1>UB {align1}; add (16) g63.0<1>UW g63.0<16,16,1>UW g44.1<16,16,1>UB {align1}; add (16) g64.0<1>UW g64.0<16,16,1>UW g45.1<16,16,1>UB {align1}; add (16) g65.0<1>UW g65.0<16,16,1>UW g46.1<16,16,1>UB {align1}; add (16) g66.0<1>UW g66.0<16,16,1>UW g47.1<16,16,1>UB {align1}; add (16) g67.0<1>UW g67.0<16,16,1>UW g48.1<16,16,1>UB {align1}; add (16) g68.0<1>UW g68.0<16,16,1>UW g49.1<16,16,1>UB {align1}; add (16) g69.0<1>UW g69.0<16,16,1>UW g50.1<16,16,1>UB {align1}; add (16) g70.0<1>UW g70.0<16,16,1>UW g51.1<16,16,1>UB {align1}; add (16) g71.0<1>UW g71.0<16,16,1>UW g52.1<16,16,1>UB {align1}; add (16) g72.0<1>UW g72.0<16,16,1>UW g53.1<16,16,1>UB {align1}; add (16) g73.0<1>UW g73.0<16,16,1>UW g54.1<16,16,1>UB {align1}; shr (32) g58.0<1>UW g58.0<16,16,1>UW 2UW {align1 compr}; shr (32) g60.0<1>UW g60.0<16,16,1>UW 2UW {align1 compr}; shr (32) g62.0<1>UW g62.0<16,16,1>UW 2UW {align1 compr}; shr (32) g64.0<1>UW g64.0<16,16,1>UW 2UW {align1 compr}; shr (32) g66.0<1>UW g66.0<16,16,1>UW 2UW {align1 compr}; shr (32) g68.0<1>UW g68.0<16,16,1>UW 2UW {align1 compr}; shr (32) g70.0<1>UW g70.0<16,16,1>UW 2UW {align1 compr}; shr (32) g72.0<1>UW g72.0<16,16,1>UW 2UW {align1 compr};