xenocara/driver/xf86-video-intel/xvmc/shader/mc/lib_igd.g4a

134 lines
3.8 KiB
Text
Raw Permalink Normal View History

/* GRF allocation:
g1~g30: constant buffer
g1~g2:intra IQ matrix in UB format
g3~g4:non intra IQ matrix in UB format
g5~g20:IDCT table
g32~g55:DCT data before IQ
g56~g79:DCT data after IQ
g84~g107: IDCT data after idct
g82: thread payload backup
g126.8: ip before jump to the lib
*/
jmpi MOTION_FRAME_Y;
jmpi MOTION_FRAME_UV;
jmpi MOTION_FIELD_Y;
jmpi MOTION_FIELD_UV;
jmpi ADD_IDCT;
mov (16) g84.0<1>W g33.0<16,16,1>W {align1};
mov (16) g85.0<1>W g34.0<16,16,1>W {align1};
mov (16) g86.0<1>W g35.0<16,16,1>W {align1};
mov (16) g87.0<1>W g36.0<16,16,1>W {align1};
mov (16) g88.0<1>W g37.0<16,16,1>W {align1};
mov (16) g89.0<1>W g38.0<16,16,1>W {align1};
mov (16) g90.0<1>W g39.0<16,16,1>W {align1};
mov (16) g91.0<1>W g40.0<16,16,1>W {align1};
mov (16) g92.0<1>W g41.0<16,16,1>W {align1};
mov (16) g93.0<1>W g42.0<16,16,1>W {align1};
mov (16) g94.0<1>W g43.0<16,16,1>W {align1};
mov (16) g95.0<1>W g44.0<16,16,1>W {align1};
mov (16) g96.0<1>W g45.0<16,16,1>W {align1};
mov (16) g97.0<1>W g46.0<16,16,1>W {align1};
mov (16) g98.0<1>W g47.0<16,16,1>W {align1};
mov (16) g99.0<1>W g48.0<16,16,1>W {align1};
mov (16) g100.0<1>W g49.0<16,16,1>W {align1};
mov (16) g101.0<1>W g50.0<16,16,1>W {align1};
mov (16) g102.0<1>W g51.0<16,16,1>W {align1};
mov (16) g103.0<1>W g52.0<16,16,1>W {align1};
mov (16) g104.0<1>W g53.0<16,16,1>W {align1};
mov (16) g105.0<1>W g54.0<16,16,1>W {align1};
mov (16) g106.0<1>W g55.0<16,16,1>W {align1};
mov (16) g107.0<1>W g56.0<16,16,1>W {align1};
//Y0
and.nz (1) null g82.12<1,1,1>UW 0x20UW {align1};
(f0) jmpi block_y1;
mov (8) g84.0<1>W 0W {align1};
mov (8) g85.0<1>W 0W {align1};
mov (8) g86.0<1>W 0W {align1};
mov (8) g87.0<1>W 0W {align1};
mov (8) g88.0<1>W 0W {align1};
mov (8) g89.0<1>W 0W {align1};
mov (8) g90.0<1>W 0W {align1};
mov (8) g91.0<1>W 0W {align1};
//Y1
block_y1:
and.nz (1) null g82.12<1,1,1>UW 0x10UW {align1};
(f0) jmpi block_y2;
mov (8) g84.16<1>W 0W {align1};
mov (8) g85.16<1>W 0W {align1};
mov (8) g86.16<1>W 0W {align1};
mov (8) g87.16<1>W 0W {align1};
mov (8) g88.16<1>W 0W {align1};
mov (8) g89.16<1>W 0W {align1};
mov (8) g90.16<1>W 0W {align1};
mov (8) g91.16<1>W 0W {align1};
//Y2
block_y2:
and.nz (1) null g82.12<1,1,1>UW 0x08UW {align1};
(f0) jmpi block_y3;
mov (8) g92.0<1>W 0W {align1};
mov (8) g93.0<1>W 0W {align1};
mov (8) g94.0<1>W 0W {align1};
mov (8) g95.0<1>W 0W {align1};
mov (8) g96.0<1>W 0W {align1};
mov (8) g97.0<1>W 0W {align1};
mov (8) g98.0<1>W 0W {align1};
mov (8) g99.0<1>W 0W {align1};
//Y3
block_y3:
and.nz (1) null g82.12<1,1,1>UW 0x04UW {align1};
(f0) jmpi block_u;
mov (8) g92.16<1>W 0W {align1};
mov (8) g93.16<1>W 0W {align1};
mov (8) g94.16<1>W 0W {align1};
mov (8) g95.16<1>W 0W {align1};
mov (8) g96.16<1>W 0W {align1};
mov (8) g97.16<1>W 0W {align1};
mov (8) g98.16<1>W 0W {align1};
mov (8) g99.16<1>W 0W {align1};
//U
block_u:
and.nz (1) null g82.12<1,1,1>UW 0x02UW {align1};
(f0) jmpi block_v;
mov (16) g100.0<1>W 0W {align1};
mov (16) g101.0<1>W 0W {align1};
mov (16) g102.0<1>W 0W {align1};
mov (16) g103.0<1>W 0W {align1};
//V
block_v:
and.nz (1) null g82.12<1,1,1>UW 0x01UW {align1};
(f0) jmpi out;
mov (16) g104.0<1>W 0W {align1};
mov (16) g105.0<1>W 0W {align1};
mov (16) g106.0<1>W 0W {align1};
mov (16) g107.0<1>W 0W {align1};
out:
add (1) ip g126.8<1,1,1>UD 0x20UD {align1}; //jump back
MOTION_FRAME_Y:
include(`motion_frame_y_igd.g4i')
add (1) ip g126.8<1,1,1>UD 0x20UD {align1}; //jump back
MOTION_FRAME_UV:
include(`motion_frame_uv_igd.g4i')
add (1) ip g126.8<1,1,1>UD 0x20UD {align1}; //jump back
MOTION_FIELD_Y:
include(`motion_field_y_igd.g4i')
add (1) ip g126.8<1,1,1>UD 0x20UD {align1}; //jump back
MOTION_FIELD_UV:
include(`motion_field_uv_igd.g4i')
add (1) ip g126.8<1,1,1>UD 0x20UD {align1}; //jump back
ADD_IDCT:
include(`addidct_igd.g4i')