1fa225cbcSrjs/*
2fa225cbcSrjs * Copyright © 2009 Intel Corporation
3fa225cbcSrjs *
4fa225cbcSrjs * Permission is hereby granted, free of charge, to any person obtaining a
5fa225cbcSrjs * copy of this software and associated documentation files (the "Software"),
6fa225cbcSrjs * to deal in the Software without restriction, including without limitation
7fa225cbcSrjs * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8fa225cbcSrjs * and/or sell copies of the Software, and to permit persons to whom the
9fa225cbcSrjs * Software is furnished to do so, subject to the following conditions:
10fa225cbcSrjs *
11fa225cbcSrjs * The above copyright notice and this permission notice (including the next
12fa225cbcSrjs * paragraph) shall be included in all copies or substantial portions of the
13fa225cbcSrjs * Software.
14fa225cbcSrjs *
15fa225cbcSrjs * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16fa225cbcSrjs * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17fa225cbcSrjs * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18fa225cbcSrjs * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19fa225cbcSrjs * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20fa225cbcSrjs * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21fa225cbcSrjs * SOFTWARE.
22fa225cbcSrjs *
23fa225cbcSrjs * Author:
24fa225cbcSrjs *    Zou Nan hai <nanhai.zou@intel.com>
25fa225cbcSrjs *    Yan Li <li.l.yan@intel.com>
26fa225cbcSrjs *    Liu Xi bin<xibin.liu@intel.com>
27fa225cbcSrjs */
28fa225cbcSrjs/* GRF allocation:
29fa225cbcSrjs   g1~g30: constant buffer
30fa225cbcSrjs           g1~g2:intra IQ matrix in UB format
31fa225cbcSrjs           g3~g4:non intra IQ matrix in UB format
32fa225cbcSrjs           g5~g20:IDCT table
33fa225cbcSrjs   g32~g55:DCT data before IQ
34fa225cbcSrjs   g56~g79:DCT data after IQ
35fa225cbcSrjs   g82:    thread payload backup
36fa225cbcSrjs   g109:   g109.0:q_scale_code, g109.4:intra_dc_mult, 
37fa225cbcSrjs   g110:   q_scale_code
38fa225cbcSrjs   g111:   intra DC coefficient 
39fa225cbcSrjs   g112~g115: intra IQ matrix in UW format (in order to use instruction compress), copys from g1~g2
40fa225cbcSrjs   g125:   ip before jump
41fa225cbcSrjs*/
42fa225cbcSrjsand (1) g109.0<1>UW g82.8<1,1,1>UW 0x1fUW {align1};    //q_scale_code
43fa225cbcSrjs
44fa225cbcSrjsand (1) g109.4<1>UW g82.8<1,1,1>UW 0x6000UW {align1};   //intra_dc_presion
45fa225cbcSrjsshr (1) g109.4<1>UW g109.4<1,1,1>UW 13UW {align1};
46fa225cbcSrjsmov (1) g109.6<1>UW 0x8UW {align1};
47fa225cbcSrjsshr (1) g109.4<1>UW g109.6<1,1,1>UW g109.4<1,1,1>UW {align1};  //intra_dc_mult
48fa225cbcSrjs 
49fa225cbcSrjsand.z (1) null g82.8<1,1,1>UW 0x20UW {align1};   //if(q_scale_type==0) q_scale=q_scale_code*2;
50fa225cbcSrjs(f0) jmpi Q_SCALE_TYPE_0;
51fa225cbcSrjs
52fa225cbcSrjscmp.l (1) null g109.0<1,1,1>UW 9UW {align1};      //if(q_scale_type!=0) calculate q_scale
53fa225cbcSrjs(f0) jmpi DO_IQ;
54fa225cbcSrjscmp.l (1) null g109.0<1,1,1>UW 17UW {align1};
55fa225cbcSrjs(f0) jmpi RANG_9_16;
56fa225cbcSrjscmp.l (1) null g109.0<1,1,1>UW 25UW {align1};
57fa225cbcSrjs(f0) jmpi RANG_17_24;
58fa225cbcSrjs
59fa225cbcSrjsRANG_25_31:
60fa225cbcSrjsadd (1) g109.0<1>UW g109.0<1,1,1>UW -25W {align1};
61fa225cbcSrjsshl (1) g109.0<1>UW g109.0<1,1,1>UW 3UW {align1};
62fa225cbcSrjsadd (1) g109.0<1>UW g109.0<1,1,1>UW 64UW {align1};
63fa225cbcSrjsjmpi DO_IQ;
64fa225cbcSrjs
65fa225cbcSrjsRANG_9_16:
66fa225cbcSrjsadd (1) g109.0<1>UW g109.0<1,1,1>UW -9W {align1};
67fa225cbcSrjsshl (1) g109.0<1>UW g109.0<1,1,1>UW 1UW {align1};
68fa225cbcSrjsadd (1) g109.0<1>UW g109.0<1,1,1>UW 10UW {align1};
69fa225cbcSrjsjmpi DO_IQ;
70fa225cbcSrjs
71fa225cbcSrjsRANG_17_24:
72fa225cbcSrjsadd (1) g109.0<1>UW g109.0<1,1,1>UW -17W {align1};
73fa225cbcSrjsshl (1) g109.0<1>UW g109.0<1,1,1>UW 2UW {align1};
74fa225cbcSrjsadd (1) g109.0<1>UW g109.0<1,1,1>UW 28UW {align1};
75fa225cbcSrjsjmpi DO_IQ;
76fa225cbcSrjs
77fa225cbcSrjsQ_SCALE_TYPE_0:
78fa225cbcSrjsshl (1) g109.0<1>UW g109.0<1,1,1>UW 1UD {align1};
79fa225cbcSrjs
80fa225cbcSrjsDO_IQ:
81fa225cbcSrjsmov (1) g110.0<1>UW g109.0<1,1,1>UW {align1};
82fa225cbcSrjs
83fa225cbcSrjsmov (16) g112.0<1>UW g1.0<16,16,1>UB {align1};
84fa225cbcSrjsmov (16) g113.0<1>UW g1.16<16,16,1>UB {align1};
85fa225cbcSrjsmov (16) g114.0<1>UW g2.0<16,16,1>UB {align1};
86fa225cbcSrjsmov (16) g115.0<1>UW g2.16<16,16,1>UB {align1};
87fa225cbcSrjs
88fa225cbcSrjsmov (1) a0.0<1>UD 0x03F003E0UD {align1};
89fa225cbcSrjs
90fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
91fa225cbcSrjsjmpi DO_IQ_INTRA;
92fa225cbcSrjsmov (16) g56.0<1>W g116.0<16,8,2>W {align1};
93fa225cbcSrjsmov (16) g57.0<1>W g118.0<16,8,2>W {align1};
94fa225cbcSrjsmov (16) g58.0<1>W g120.0<16,8,2>W {align1};
95fa225cbcSrjsmov (16) g59.0<1>W g122.0<16,8,2>W {align1};
96fa225cbcSrjs
97fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
98fa225cbcSrjsjmpi DO_IQ_INTRA;
99fa225cbcSrjsmov (16) g60.0<1>W g116.0<16,8,2>W {align1};
100fa225cbcSrjsmov (16) g61.0<1>W g118.0<16,8,2>W {align1};
101fa225cbcSrjsmov (16) g62.0<1>W g120.0<16,8,2>W {align1};
102fa225cbcSrjsmov (16) g63.0<1>W g122.0<16,8,2>W {align1};
103fa225cbcSrjs
104fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
105fa225cbcSrjsjmpi DO_IQ_INTRA;
106fa225cbcSrjsmov (16) g64.0<1>W g116.0<16,8,2>W {align1};
107fa225cbcSrjsmov (16) g65.0<1>W g118.0<16,8,2>W {align1};
108fa225cbcSrjsmov (16) g66.0<1>W g120.0<16,8,2>W {align1};
109fa225cbcSrjsmov (16) g67.0<1>W g122.0<16,8,2>W {align1};
110fa225cbcSrjs
111fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
112fa225cbcSrjsjmpi DO_IQ_INTRA;
113fa225cbcSrjsmov (16) g68.0<1>W g116.0<16,8,2>W {align1};
114fa225cbcSrjsmov (16) g69.0<1>W g118.0<16,8,2>W {align1};
115fa225cbcSrjsmov (16) g70.0<1>W g120.0<16,8,2>W {align1};
116fa225cbcSrjsmov (16) g71.0<1>W g122.0<16,8,2>W {align1};
117fa225cbcSrjs
118fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
119fa225cbcSrjsjmpi DO_IQ_INTRA;
120fa225cbcSrjsmov (16) g72.0<1>W g116.0<16,8,2>W {align1};
121fa225cbcSrjsmov (16) g73.0<1>W g118.0<16,8,2>W {align1};
122fa225cbcSrjsmov (16) g74.0<1>W g120.0<16,8,2>W {align1};
123fa225cbcSrjsmov (16) g75.0<1>W g122.0<16,8,2>W {align1};
124fa225cbcSrjs
125fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
126fa225cbcSrjsjmpi DO_IQ_INTRA;
127fa225cbcSrjsmov (16) g76.0<1>W g116.0<16,8,2>W {align1};
128fa225cbcSrjsmov (16) g77.0<1>W g118.0<16,8,2>W {align1};
129fa225cbcSrjsmov (16) g78.0<1>W g120.0<16,8,2>W {align1};
130fa225cbcSrjsmov (16) g79.0<1>W g122.0<16,8,2>W {align1};
131fa225cbcSrjs
132