diff options
Diffstat (limited to 'xvmc/shader/mc/read_field_x1y1_y_igd.g4i')
-rw-r--r-- | xvmc/shader/mc/read_field_x1y1_y_igd.g4i | 64 |
1 files changed, 64 insertions, 0 deletions
diff --git a/xvmc/shader/mc/read_field_x1y1_y_igd.g4i b/xvmc/shader/mc/read_field_x1y1_y_igd.g4i new file mode 100644 index 00000000..db635fe5 --- /dev/null +++ b/xvmc/shader/mc/read_field_x1y1_y_igd.g4i @@ -0,0 +1,64 @@ +/* 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 0x07001FUD {align1}; +and.nz (1) null g115.16<1,1,1>UW 1UW {align1}; +(f0) jmpi read_backward; + +send (16) 0 g40.0<1>UW g115<8,8,1>UW read(4,2,0,2) mlen 1 rlen 8 {align1}; +add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1}; +send (16) 0 g48.0<1>UW g115<8,8,1>UW read(4,2,0,2) mlen 1 rlen 8 {align1}; +add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1}; +mov (1) g115.8<1>UD 0x1FUD {align1}; +send (16) 0 g56.0<1>UW g115<8,8,1>UW read(4,2,0,2) mlen 1 rlen 1 {align1}; +jmpi put_data; + +read_backward: +send (16) 0 g40.0<1>UW g115<8,8,1>UW read(7,2,0,2) mlen 1 rlen 8 {align1}; +add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1}; +send (16) 0 g48.0<1>UW g115<8,8,1>UW read(7,2,0,2) mlen 1 rlen 8 {align1}; +add (1) g115.4<1>UD g115.4<1,1,1>UD 8UD {align1}; +mov (1) g115.8<1>UD 0x1FUD {align1}; +send (16) 0 g56.0<1>UW g115<8,8,1>UW read(7,2,0,2) mlen 1 rlen 1 {align1}; + +put_data: +add (16) g32.0<1>UW g40.0<16,16,1>UB g42.0<16,16,1>UB {align1}; +add (16) g33.0<1>UW g42.0<16,16,1>UB g44.0<16,16,1>UB {align1}; +add (16) g34.0<1>UW g44.0<16,16,1>UB g46.0<16,16,1>UB {align1}; +add (16) g35.0<1>UW g46.0<16,16,1>UB g48.0<16,16,1>UB {align1}; +add (16) g36.0<1>UW g48.0<16,16,1>UB g50.0<16,16,1>UB {align1}; +add (16) g37.0<1>UW g50.0<16,16,1>UB g52.0<16,16,1>UB {align1}; +add (16) g38.0<1>UW g52.0<16,16,1>UB g54.0<16,16,1>UB {align1}; +add (16) g39.0<1>UW g54.0<16,16,1>UB g56.0<16,16,1>UB {align1}; + +add (16) g32.0<1>UW g32.0<16,16,1>UW g40.1<16,16,1>UB {align1}; +add (16) g33.0<1>UW g33.0<16,16,1>UW g42.1<16,16,1>UB {align1}; +add (16) g34.0<1>UW g34.0<16,16,1>UW g44.1<16,16,1>UB {align1}; +add (16) g35.0<1>UW g35.0<16,16,1>UW g46.1<16,16,1>UB {align1}; +add (16) g36.0<1>UW g36.0<16,16,1>UW g48.1<16,16,1>UB {align1}; +add (16) g37.0<1>UW g37.0<16,16,1>UW g50.1<16,16,1>UB {align1}; +add (16) g38.0<1>UW g38.0<16,16,1>UW g52.1<16,16,1>UB {align1}; +add (16) g39.0<1>UW g39.0<16,16,1>UW g54.1<16,16,1>UB {align1}; + +add (16) g32.0<1>UW g32.0<16,16,1>UW g42.1<16,16,1>UB {align1}; +add (16) g33.0<1>UW g33.0<16,16,1>UW g44.1<16,16,1>UB {align1}; +add (16) g34.0<1>UW g34.0<16,16,1>UW g46.1<16,16,1>UB {align1}; +add (16) g35.0<1>UW g35.0<16,16,1>UW g48.1<16,16,1>UB {align1}; +add (16) g36.0<1>UW g36.0<16,16,1>UW g50.1<16,16,1>UB {align1}; +add (16) g37.0<1>UW g37.0<16,16,1>UW g52.1<16,16,1>UB {align1}; +add (16) g38.0<1>UW g38.0<16,16,1>UW g54.1<16,16,1>UB {align1}; +add (16) g39.0<1>UW g39.0<16,16,1>UW g56.1<16,16,1>UB {align1}; + +shr (32) g32.0<1>UW g32.0<16,16,1>UW 2UW {align1 compr}; +shr (32) g34.0<1>UW g34.0<16,16,1>UW 2UW {align1 compr}; +shr (32) g36.0<1>UW g36.0<16,16,1>UW 2UW {align1 compr}; +shr (32) g38.0<1>UW g38.0<16,16,1>UW 2UW {align1 compr}; |