103b705cfSriastradh/*
203b705cfSriastradh * Copyright © 2009 Intel Corporation
303b705cfSriastradh *
403b705cfSriastradh * Permission is hereby granted, free of charge, to any person obtaining a
503b705cfSriastradh * copy of this software and associated documentation files (the "Software"),
603b705cfSriastradh * to deal in the Software without restriction, including without limitation
703b705cfSriastradh * the rights to use, copy, modify, merge, publish, distribute, sublicense,
803b705cfSriastradh * and/or sell copies of the Software, and to permit persons to whom the
903b705cfSriastradh * Software is furnished to do so, subject to the following conditions:
1003b705cfSriastradh *
1103b705cfSriastradh * The above copyright notice and this permission notice (including the next
1203b705cfSriastradh * paragraph) shall be included in all copies or substantial portions of the
1303b705cfSriastradh * Software.
1403b705cfSriastradh *
1503b705cfSriastradh * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
1603b705cfSriastradh * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
1703b705cfSriastradh * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
1803b705cfSriastradh * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
1903b705cfSriastradh * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
2003b705cfSriastradh * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
2103b705cfSriastradh * SOFTWARE.
2203b705cfSriastradh *
2303b705cfSriastradh * Author:
2403b705cfSriastradh *    Zou Nan hai <nanhai.zou@intel.com>
2503b705cfSriastradh *    Yan Li <li.l.yan@intel.com>
2603b705cfSriastradh *    Liu Xi bin<xibin.liu@intel.com>
2703b705cfSriastradh */
2803b705cfSriastradh/* GRF allocation:
2903b705cfSriastradh   g1~g30: constant buffer
3003b705cfSriastradh           g1~g2:intra IQ matrix in UB format
3103b705cfSriastradh           g3~g4:non intra IQ matrix in UB format
3203b705cfSriastradh           g5~g20:IDCT table
3303b705cfSriastradh   g32~g55:DCT data before IQ
3403b705cfSriastradh   g56~g79:DCT data after IQ
3503b705cfSriastradh   g82:    thread payload backup
3603b705cfSriastradh   g109:   q_scale_code
3703b705cfSriastradh   g110:   q_scale_code
3803b705cfSriastradh   g112~g115: non intra IQ matrix in UW format (in order to use instruction compress), copys from g3~g4
3903b705cfSriastradh   g125:   ip before jump
4003b705cfSriastradh*/
4103b705cfSriastradhand (1) g109.0<1>UW g82.8<1,1,1>UW 0x1fUW {align1}; //q_scale_code
4203b705cfSriastradh
4303b705cfSriastradhand.z (1) null g82.8<1,1,1>UW 0x20UW {align1};   //if(q_scale_type==0) q_scale=q_scale_code*2;
4403b705cfSriastradh(f0) jmpi Q_SCALE_TYPE_0;
4503b705cfSriastradh
4603b705cfSriastradhcmp.l (1) null g109.0<1,1,1>UW 9UW {align1};      //if(q_scale_type!=0) calculate q_scale
4703b705cfSriastradh(f0) jmpi DO_IQ;
4803b705cfSriastradhcmp.l (1) null g109.0<1,1,1>UW 17UW {align1};
4903b705cfSriastradh(f0) jmpi RANG_9_16;
5003b705cfSriastradhcmp.l (1) null g109.0<1,1,1>UW 25UW {align1};
5103b705cfSriastradh(f0) jmpi RANG_17_24;
5203b705cfSriastradh
5303b705cfSriastradhRANG_25_31:
5403b705cfSriastradhadd (1) g109.0<1>UW g109.0<1,1,1>UW -25W {align1};
5503b705cfSriastradhshl (1) g109.0<1>UW g109.0<1,1,1>UW 3UW {align1};
5603b705cfSriastradhadd (1) g109.0<1>UW g109.0<1,1,1>UW 64UW {align1};
5703b705cfSriastradhjmpi DO_IQ;
5803b705cfSriastradh
5903b705cfSriastradhRANG_9_16:
6003b705cfSriastradhadd (1) g109.0<1>UW g109.0<1,1,1>UW -9W {align1};
6103b705cfSriastradhshl (1) g109.0<1>UW g109.0<1,1,1>UW 1UW {align1};
6203b705cfSriastradhadd (1) g109.0<1>UW g109.0<1,1,1>UW 10UW {align1};
6303b705cfSriastradhjmpi DO_IQ;
6403b705cfSriastradh
6503b705cfSriastradhRANG_17_24:
6603b705cfSriastradhadd (1) g109.0<1>UW g109.0<1,1,1>UW -17W {align1};
6703b705cfSriastradhshl (1) g109.0<1>UW g109.0<1,1,1>UW 2UW {align1};
6803b705cfSriastradhadd (1) g109.0<1>UW g109.0<1,1,1>UW 28UW {align1};
6903b705cfSriastradhjmpi DO_IQ;
7003b705cfSriastradh
7103b705cfSriastradhQ_SCALE_TYPE_0:
7203b705cfSriastradhshl (1) g109.0<1>UW g109.0<1,1,1>UW 1UD {align1};
7303b705cfSriastradh
7403b705cfSriastradhDO_IQ:
7503b705cfSriastradhmov (1) g110.0<1>UW g109.0<1,1,1>UW {align1};
7603b705cfSriastradh
7703b705cfSriastradhmov (16) g112.0<1>UW g3.0<16,16,1>UB {align1};
7803b705cfSriastradhmov (16) g113.0<1>UW g3.16<16,16,1>UB {align1};
7903b705cfSriastradhmov (16) g114.0<1>UW g4.0<16,16,1>UB {align1};
8003b705cfSriastradhmov (16) g115.0<1>UW g4.16<16,16,1>UB {align1};
8103b705cfSriastradh
8203b705cfSriastradhmov (1) a0.0<1>UD 0x03F003E0UD {align1};
8303b705cfSriastradh
8403b705cfSriastradh//Y0
8503b705cfSriastradhiq_non_intra_y0:
8603b705cfSriastradhand.z (1) null g82.8<1,1,1>UW 0x800UW {align1};
8703b705cfSriastradh(f0) jmpi iq_non_intra_y1;
8803b705cfSriastradhmov (1) g125.0<1>UD ip {align1};
8903b705cfSriastradhjmpi DO_IQ_NON_INTRA;
9003b705cfSriastradhmov (16) g56.0<1>W g116.0<16,8,2>W {align1};
9103b705cfSriastradhmov (16) g57.0<1>W g118.0<16,8,2>W {align1};
9203b705cfSriastradhmov (16) g58.0<1>W g120.0<16,8,2>W {align1};
9303b705cfSriastradhmov (16) g59.0<1>W g122.0<16,8,2>W {align1};
9403b705cfSriastradh
9503b705cfSriastradh//Y1
9603b705cfSriastradhiq_non_intra_y1:
9703b705cfSriastradhand.z (1) null g82.8<1,1,1>UW 0x400UW {align1};
9803b705cfSriastradh(f0) jmpi iq_non_intra_y2;
9903b705cfSriastradhmov (1) g125.0<1>UD ip {align1};
10003b705cfSriastradhjmpi DO_IQ_NON_INTRA;
10103b705cfSriastradhmov (16) g60.0<1>W g116.0<16,8,2>W {align1};
10203b705cfSriastradhmov (16) g61.0<1>W g118.0<16,8,2>W {align1};
10303b705cfSriastradhmov (16) g62.0<1>W g120.0<16,8,2>W {align1};
10403b705cfSriastradhmov (16) g63.0<1>W g122.0<16,8,2>W {align1};
10503b705cfSriastradh
10603b705cfSriastradh//Y2
10703b705cfSriastradhiq_non_intra_y2:
10803b705cfSriastradhand.z (1) null g82.8<1,1,1>UW 0x200UW {align1};
10903b705cfSriastradh(f0) jmpi iq_non_intra_y3;
11003b705cfSriastradhmov (1) g125.0<1>UD ip {align1};
11103b705cfSriastradhjmpi DO_IQ_NON_INTRA;
11203b705cfSriastradhmov (16) g64.0<1>W g116.0<16,8,2>W {align1};
11303b705cfSriastradhmov (16) g65.0<1>W g118.0<16,8,2>W {align1};
11403b705cfSriastradhmov (16) g66.0<1>W g120.0<16,8,2>W {align1};
11503b705cfSriastradhmov (16) g67.0<1>W g122.0<16,8,2>W {align1};
11603b705cfSriastradh
11703b705cfSriastradh//Y3
11803b705cfSriastradhiq_non_intra_y3:
11903b705cfSriastradhand.z (1) null g82.8<1,1,1>UW 0x100UW {align1};
12003b705cfSriastradh(f0) jmpi iq_non_intra_u;
12103b705cfSriastradhmov (1) g125.0<1>UD ip {align1};
12203b705cfSriastradhjmpi DO_IQ_NON_INTRA;
12303b705cfSriastradhmov (16) g68.0<1>W g116.0<16,8,2>W {align1};
12403b705cfSriastradhmov (16) g69.0<1>W g118.0<16,8,2>W {align1};
12503b705cfSriastradhmov (16) g70.0<1>W g120.0<16,8,2>W {align1};
12603b705cfSriastradhmov (16) g71.0<1>W g122.0<16,8,2>W {align1};
12703b705cfSriastradh
12803b705cfSriastradh//U
12903b705cfSriastradhiq_non_intra_u:
13003b705cfSriastradhand.z (1) null g82.8<1,1,1>UW 0x80UW {align1};
13103b705cfSriastradh(f0) jmpi iq_non_intra_v;
13203b705cfSriastradhmov (1) g125.0<1>UD ip {align1};
13303b705cfSriastradhjmpi DO_IQ_NON_INTRA;
13403b705cfSriastradhmov (16) g72.0<1>W g116.0<16,8,2>W {align1};
13503b705cfSriastradhmov (16) g73.0<1>W g118.0<16,8,2>W {align1};
13603b705cfSriastradhmov (16) g74.0<1>W g120.0<16,8,2>W {align1};
13703b705cfSriastradhmov (16) g75.0<1>W g122.0<16,8,2>W {align1};
13803b705cfSriastradh
13903b705cfSriastradh//V
14003b705cfSriastradhiq_non_intra_v:
14103b705cfSriastradhand.z (1) null g82.8<1,1,1>UW 0x40UW {align1};
14203b705cfSriastradh(f0) jmpi iq_non_intra_end;
14303b705cfSriastradhmov (1) g125.0<1>UD ip {align1};
14403b705cfSriastradhjmpi DO_IQ_NON_INTRA;
14503b705cfSriastradhmov (16) g76.0<1>W g116.0<16,8,2>W {align1};
14603b705cfSriastradhmov (16) g77.0<1>W g118.0<16,8,2>W {align1};
14703b705cfSriastradhmov (16) g78.0<1>W g120.0<16,8,2>W {align1};
14803b705cfSriastradhmov (16) g79.0<1>W g122.0<16,8,2>W {align1};
14903b705cfSriastradh
15003b705cfSriastradhiq_non_intra_end:
151