diff options
Diffstat (limited to 'src/xvmc/shader/mc/addidct_igd.g4i')
-rw-r--r-- | src/xvmc/shader/mc/addidct_igd.g4i | 117 |
1 files changed, 117 insertions, 0 deletions
diff --git a/src/xvmc/shader/mc/addidct_igd.g4i b/src/xvmc/shader/mc/addidct_igd.g4i new file mode 100644 index 00000000..56fa2af3 --- /dev/null +++ b/src/xvmc/shader/mc/addidct_igd.g4i @@ -0,0 +1,117 @@ +/* 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 + g84~g107:IDCT data */ +//mov (2) g31.0<1>UD g82.12<2,2,1>UW {align1}; //restore x and y +mov (2) g31.0<1>UD g82.0<2,2,1>UD {align1}; //restore x and y + +and.nz (1) null g82.30<1,1,1>UB 0x1UW {align1}; //dct_type +(f0) jmpi field_dct; + +add.sat (32) g58.0<2>UB g84.0<16,16,1>W g58.0<16,16,2>UB {align1 compr}; +add.sat (32) g60.0<2>UB g86.0<16,16,1>W g60.0<16,16,2>UB {align1 compr}; +add.sat (32) g62.0<2>UB g88.0<16,16,1>W g62.0<16,16,2>UB {align1 compr}; +add.sat (32) g64.0<2>UB g90.0<16,16,1>W g64.0<16,16,2>UB {align1 compr}; +add.sat (32) g66.0<2>UB g92.0<16,16,1>W g66.0<16,16,2>UB {align1 compr}; +add.sat (32) g68.0<2>UB g94.0<16,16,1>W g68.0<16,16,2>UB {align1 compr}; +add.sat (32) g70.0<2>UB g96.0<16,16,1>W g70.0<16,16,2>UB {align1 compr}; +add.sat (32) g72.0<2>UB g98.0<16,16,1>W g72.0<16,16,2>UB {align1 compr}; + +add.sat (32) g74.0<2>UB g100.0<16,16,1>W g74.0<16,16,2>UB {align1 compr}; +add.sat (32) g76.0<2>UB g102.0<16,16,1>W g76.0<16,16,2>UB {align1 compr}; +add.sat (32) g78.0<2>UB g104.0<16,16,1>W g78.0<16,16,2>UB {align1 compr}; +add.sat (32) g80.0<2>UB g106.0<16,16,1>W g80.0<16,16,2>UB {align1 compr}; +jmpi write_back; + +field_dct: +add.sat (16) g58.0<2>UB g84.0<16,16,1>W g58.0<16,16,2>UB {align1}; +add.sat (16) g59.0<2>UB g92.0<16,16,1>W g59.0<16,16,2>UB {align1}; +add.sat (16) g60.0<2>UB g85.0<16,16,1>W g60.0<16,16,2>UB {align1}; +add.sat (16) g61.0<2>UB g93.0<16,16,1>W g61.0<16,16,2>UB {align1}; +add.sat (16) g62.0<2>UB g86.0<16,16,1>W g62.0<16,16,2>UB {align1}; +add.sat (16) g63.0<2>UB g94.0<16,16,1>W g63.0<16,16,2>UB {align1}; +add.sat (16) g64.0<2>UB g87.0<16,16,1>W g64.0<16,16,2>UB {align1}; +add.sat (16) g65.0<2>UB g95.0<16,16,1>W g65.0<16,16,2>UB {align1}; +add.sat (16) g66.0<2>UB g88.0<16,16,1>W g66.0<16,16,2>UB {align1}; +add.sat (16) g67.0<2>UB g96.0<16,16,1>W g67.0<16,16,2>UB {align1}; +add.sat (16) g68.0<2>UB g89.0<16,16,1>W g68.0<16,16,2>UB {align1}; +add.sat (16) g69.0<2>UB g97.0<16,16,1>W g69.0<16,16,2>UB {align1}; +add.sat (16) g70.0<2>UB g90.0<16,16,1>W g70.0<16,16,2>UB {align1}; +add.sat (16) g71.0<2>UB g98.0<16,16,1>W g71.0<16,16,2>UB {align1}; +add.sat (16) g72.0<2>UB g91.0<16,16,1>W g72.0<16,16,2>UB {align1}; +add.sat (16) g73.0<2>UB g99.0<16,16,1>W g73.0<16,16,2>UB {align1}; +/* +add.sat (16) g74.0<2>UB g100.0<16,16,1>W g74.0<16,16,2>UB {align1}; +add.sat (16) g75.0<2>UB g101.0<16,16,1>W g75.0<16,16,2>UB {align1}; +add.sat (16) g76.0<2>UB g102.0<16,16,1>W g76.0<16,16,2>UB {align1}; +add.sat (16) g77.0<2>UB g103.0<16,16,1>W g77.0<16,16,2>UB {align1}; +add.sat (16) g78.0<2>UB g104.0<16,16,1>W g78.0<16,16,2>UB {align1}; +add.sat (16) g79.0<2>UB g105.0<16,16,1>W g79.0<16,16,2>UB {align1}; +add.sat (16) g80.0<2>UB g106.0<16,16,1>W g80.0<16,16,2>UB {align1}; +add.sat (16) g81.0<2>UB g107.0<16,16,1>W g81.0<16,16,2>UB {align1}; +*/ +add (16) g100.0<1>W g100.0<16,16,1>W g74.0<16,16,2>UB {align1}; +add (16) g101.0<1>W g101.0<16,16,1>W g75.0<16,16,2>UB {align1}; +add (16) g102.0<1>W g102.0<16,16,1>W g76.0<16,16,2>UB {align1}; +add (16) g103.0<1>W g103.0<16,16,1>W g77.0<16,16,2>UB {align1}; +add (16) g104.0<1>W g104.0<16,16,1>W g78.0<16,16,2>UB {align1}; +add (16) g105.0<1>W g105.0<16,16,1>W g79.0<16,16,2>UB {align1}; +add (16) g106.0<1>W g106.0<16,16,1>W g80.0<16,16,2>UB {align1}; +add (16) g107.0<1>W g107.0<16,16,1>W g81.0<16,16,2>UB {align1}; + +mov.sat (16) g74.0<2>UB g100.0<16,16,1>W {align1}; +mov.sat (16) g75.0<2>UB g101.0<16,16,1>W {align1}; +mov.sat (16) g76.0<2>UB g102.0<16,16,1>W {align1}; +mov.sat (16) g77.0<2>UB g103.0<16,16,1>W {align1}; +mov.sat (16) g78.0<2>UB g104.0<16,16,1>W {align1}; +mov.sat (16) g79.0<2>UB g105.0<16,16,1>W {align1}; +mov.sat (16) g80.0<2>UB g106.0<16,16,1>W {align1}; +mov.sat (16) g81.0<2>UB g107.0<16,16,1>W {align1}; + + +write_back: +mov (1) g31.8<1>UD 0x00F000FUD {align1}; + +mov (16) m1.0<1>UB g58.0<16,16,2>UB {align1}; +mov (16) m1.16<1>UB g59.0<16,16,2>UB {align1}; +mov (16) m2.0<1>UB g60.0<16,16,2>UB {align1}; +mov (16) m2.16<1>UB g61.0<16,16,2>UB {align1}; +mov (16) m3.0<1>UB g62.0<16,16,2>UB {align1}; +mov (16) m3.16<1>UB g63.0<16,16,2>UB {align1}; +mov (16) m4.0<1>UB g64.0<16,16,2>UB {align1}; +mov (16) m4.16<1>UB g65.0<16,16,2>UB {align1}; +mov (16) m5.0<1>UB g66.0<16,16,2>UB {align1}; +mov (16) m5.16<1>UB g67.0<16,16,2>UB {align1}; +mov (16) m6.0<1>UB g68.0<16,16,2>UB {align1}; +mov (16) m6.16<1>UB g69.0<16,16,2>UB {align1}; +mov (16) m7.0<1>UB g70.0<16,16,2>UB {align1}; +mov (16) m7.16<1>UB g71.0<16,16,2>UB {align1}; +mov (16) m8.0<1>UB g72.0<16,16,2>UB {align1}; +mov (16) m8.16<1>UB g73.0<16,16,2>UB {align1}; +send (16) 0 acc0<1>UW g31<8,8,1>UW write(0,0,2,0) mlen 9 rlen 0 {align1}; + +//U +mov (1) g31.8<1>UD 0x0070007UD { align1 }; +shr (2) g31.0<1>UD g31.0<2,2,1>UD 1D {align1}; + +mov (16) m1.0<1>UB g74.0<16,16,2>UB {align1}; +mov (16) m1.16<1>UB g75.0<16,16,2>UB {align1}; +mov (16) m2.0<1>UB g76.0<16,16,2>UB {align1}; +mov (16) m2.16<1>UB g77.0<16,16,2>UB {align1}; +send (16) 0 acc0<1>UW g31<8,8,1>UW write(1, 0, 2, 0) mlen 3 rlen 0 { align1 }; + +//V +mov (16) m1.0<1>UB g78.0<16,16,2>UB {align1}; +mov (16) m1.16<1>UB g79.0<16,16,2>UB {align1}; +mov (16) m2.0<1>UB g80.0<16,16,2>UB {align1}; +mov (16) m2.16<1>UB g81.0<16,16,2>UB {align1}; +send (16) 0 acc0<1>UW g31<8,8,1>UW write(2, 0, 2, 0) mlen 3 rlen 0 { align1 }; + +send (16) 0 acc0<1>UW g0<8,8,1>UW + thread_spawner(0, 0, 0) mlen 1 rlen 0 { align1 EOT}; |