f6224eae74
Tested by jsg@, kettenis@ and myself on a wide range of intel cards.
29 lines
1.3 KiB
Plaintext
29 lines
1.3 KiB
Plaintext
/* GRF allocation:
|
|
g1~g30: constant buffer
|
|
g1~g2:intra IQ matrix
|
|
g3~g4:non intra IQ matrix
|
|
g5~g20:IDCT table
|
|
g31: thread payload
|
|
g58~g81:reference data
|
|
g82: thread payload backup
|
|
g83~g106:IDCT data
|
|
g115: message descriptor for reading reference data */
|
|
|
|
mov (1) g115.8<1>UD 0x07000FUD {align1}; // 8*16/32=4
|
|
send (16) 0 g40.0<1>UW g115<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 4 {align1};//U
|
|
send (16) 0 g45.0<1>UW g115<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 4 {align1};//V
|
|
mov (1) g115.8<1>UD 0xFUD {align1};
|
|
add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1};
|
|
send (16) 0 g44.0<1>UW g115<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 1 {align1};//U
|
|
send (16) 0 g49.0<1>UW g115<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 1 {align1};//V
|
|
|
|
avg (16) g32.0<1>UW g40.0<16,8,1>UB g41.0<16,8,1>UB {align1};
|
|
avg (16) g33.0<1>UW g41.0<16,8,1>UB g42.0<16,8,1>UB {align1};
|
|
avg (16) g34.0<1>UW g42.0<16,8,1>UB g43.0<16,8,1>UB {align1};
|
|
avg (16) g35.0<1>UW g43.0<16,8,1>UB g44.0<16,8,1>UB {align1};
|
|
|
|
avg (16) g36.0<1>UW g45.0<16,8,1>UB g46.0<16,8,1>UB {align1};
|
|
avg (16) g37.0<1>UW g46.0<16,8,1>UB g47.0<16,8,1>UB {align1};
|
|
avg (16) g38.0<1>UW g47.0<16,8,1>UB g48.0<16,8,1>UB {align1};
|
|
avg (16) g39.0<1>UW g48.0<16,8,1>UB g49.0<16,8,1>UB {align1};
|