1428d7b3dSmrg/*
2428d7b3dSmrg * Copyright © 2009 Intel Corporation
3428d7b3dSmrg *
4428d7b3dSmrg * Permission is hereby granted, free of charge, to any person obtaining a
5428d7b3dSmrg * copy of this software and associated documentation files (the "Software"),
6428d7b3dSmrg * to deal in the Software without restriction, including without limitation
7428d7b3dSmrg * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8428d7b3dSmrg * and/or sell copies of the Software, and to permit persons to whom the
9428d7b3dSmrg * Software is furnished to do so, subject to the following conditions:
10428d7b3dSmrg *
11428d7b3dSmrg * The above copyright notice and this permission notice (including the next
12428d7b3dSmrg * paragraph) shall be included in all copies or substantial portions of the
13428d7b3dSmrg * Software.
14428d7b3dSmrg *
15428d7b3dSmrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16428d7b3dSmrg * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17428d7b3dSmrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18428d7b3dSmrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19428d7b3dSmrg * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20428d7b3dSmrg * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21428d7b3dSmrg * SOFTWARE.
22428d7b3dSmrg *
23428d7b3dSmrg * Author:
24428d7b3dSmrg *    Zou Nan hai <nanhai.zou@intel.com>
25428d7b3dSmrg *    Yan Li <li.l.yan@intel.com>
26428d7b3dSmrg *    Liu Xi bin<xibin.liu@intel.com>
27428d7b3dSmrg */
28428d7b3dSmrg/* GRF allocation:
29428d7b3dSmrg   g1~g30: constant buffer
30428d7b3dSmrg           g1~g2:intra IQ matrix in UB format
31428d7b3dSmrg           g3~g4:non intra IQ matrix in UB format
32428d7b3dSmrg           g5~g20:IDCT table
33428d7b3dSmrg   g32~g55:DCT data before IQ
34428d7b3dSmrg   g56~g79:DCT data after IQ
35428d7b3dSmrg   g83~g106: IDCT data after idct
36428d7b3dSmrg   g82:    thread payload backup
37428d7b3dSmrg   g125:   ip before jump
38428d7b3dSmrg*/
39428d7b3dSmrginclude(`iq_non_intra.g4i')
40428d7b3dSmrg
41428d7b3dSmrgdefine(`ROW_SHIFT', `11UD') //define for idct
42428d7b3dSmrgdefine(`ROW_ADD', `0x400UD')            
43428d7b3dSmrgdefine(`COL_SHIFT', `20UD') 
44428d7b3dSmrgdefine(`COL_ADD', `0x80000UD')           
45428d7b3dSmrg
46428d7b3dSmrgmov (1) a0.0<1>UD 0x06F006E0UD {align1};//0x06F006E0UD+0x00200020UD=0x07100700UD (g56.0 and g56.16,the start of DCT data)
47428d7b3dSmrg
48428d7b3dSmrg//Y0
49428d7b3dSmrgand.nz (1) null g82.8<1,1,1>UW 0x800UW {align1};
50428d7b3dSmrg(f0) jmpi do_idct_y0;
51428d7b3dSmrg
52428d7b3dSmrgmov (8) g83.0<1>UW 0UW {align1}; 
53428d7b3dSmrgmov (8) g84.0<1>UW 0UW {align1}; 
54428d7b3dSmrgmov (8) g85.0<1>UW 0UW {align1}; 
55428d7b3dSmrgmov (8) g86.0<1>UW 0UW {align1}; 
56428d7b3dSmrgmov (8) g87.0<1>UW 0UW {align1}; 
57428d7b3dSmrgmov (8) g88.0<1>UW 0UW {align1}; 
58428d7b3dSmrgmov (8) g89.0<1>UW 0UW {align1}; 
59428d7b3dSmrgmov (8) g90.0<1>UW 0UW {align1}; 
60428d7b3dSmrgadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
61428d7b3dSmrgjmpi block_y1;
62428d7b3dSmrgdo_idct_y0:
63428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
64428d7b3dSmrgjmpi IDCT_START;
65428d7b3dSmrgmov (8) g83.0<1>W g32.0<16,8,2>W {align1};
66428d7b3dSmrgmov (8) g84.0<1>W g33.0<16,8,2>W {align1};
67428d7b3dSmrgmov (8) g85.0<1>W g34.0<16,8,2>W {align1};
68428d7b3dSmrgmov (8) g86.0<1>W g35.0<16,8,2>W {align1};
69428d7b3dSmrgmov (8) g87.0<1>W g36.0<16,8,2>W {align1};
70428d7b3dSmrgmov (8) g88.0<1>W g37.0<16,8,2>W {align1};
71428d7b3dSmrgmov (8) g89.0<1>W g38.0<16,8,2>W {align1};
72428d7b3dSmrgmov (8) g90.0<1>W g39.0<16,8,2>W {align1};
73428d7b3dSmrg
74428d7b3dSmrg//Y1
75428d7b3dSmrgblock_y1:
76428d7b3dSmrgand.nz (1) null g82.8<1,1,1>UW 0x400UW {align1};
77428d7b3dSmrg(f0) jmpi do_idct_y1;
78428d7b3dSmrgmov (8) g83.16<1>UW 0UW {align1}; 
79428d7b3dSmrgmov (8) g84.16<1>UW 0UW {align1}; 
80428d7b3dSmrgmov (8) g85.16<1>UW 0UW {align1}; 
81428d7b3dSmrgmov (8) g86.16<1>UW 0UW {align1}; 
82428d7b3dSmrgmov (8) g87.16<1>UW 0UW {align1}; 
83428d7b3dSmrgmov (8) g88.16<1>UW 0UW {align1}; 
84428d7b3dSmrgmov (8) g89.16<1>UW 0UW {align1}; 
85428d7b3dSmrgmov (8) g90.16<1>UW 0UW {align1}; 
86428d7b3dSmrgadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
87428d7b3dSmrgjmpi block_y2;
88428d7b3dSmrgdo_idct_y1:
89428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
90428d7b3dSmrgjmpi IDCT_START;
91428d7b3dSmrgmov (8) g83.16<1>W g32.0<16,8,2>W {align1};
92428d7b3dSmrgmov (8) g84.16<1>W g33.0<16,8,2>W {align1};
93428d7b3dSmrgmov (8) g85.16<1>W g34.0<16,8,2>W {align1};
94428d7b3dSmrgmov (8) g86.16<1>W g35.0<16,8,2>W {align1};
95428d7b3dSmrgmov (8) g87.16<1>W g36.0<16,8,2>W {align1};
96428d7b3dSmrgmov (8) g88.16<1>W g37.0<16,8,2>W {align1};
97428d7b3dSmrgmov (8) g89.16<1>W g38.0<16,8,2>W {align1};
98428d7b3dSmrgmov (8) g90.16<1>W g39.0<16,8,2>W {align1};
99428d7b3dSmrg
100428d7b3dSmrg//Y2
101428d7b3dSmrgblock_y2:
102428d7b3dSmrgand.nz (1) null g82.8<1,1,1>UW 0x200UW {align1};
103428d7b3dSmrg(f0) jmpi do_idct_y2;
104428d7b3dSmrgmov (8) g91.0<1>UW 0UW {align1}; 
105428d7b3dSmrgmov (8) g92.0<1>UW 0UW {align1}; 
106428d7b3dSmrgmov (8) g93.0<1>UW 0UW {align1}; 
107428d7b3dSmrgmov (8) g94.0<1>UW 0UW {align1}; 
108428d7b3dSmrgmov (8) g95.0<1>UW 0UW {align1}; 
109428d7b3dSmrgmov (8) g96.0<1>UW 0UW {align1}; 
110428d7b3dSmrgmov (8) g97.0<1>UW 0UW {align1}; 
111428d7b3dSmrgmov (8) g98.0<1>UW 0UW {align1}; 
112428d7b3dSmrgadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
113428d7b3dSmrgjmpi block_y3;
114428d7b3dSmrgdo_idct_y2:
115428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
116428d7b3dSmrgjmpi IDCT_START;
117428d7b3dSmrgmov (8) g91.0<1>W g32.0<16,8,2>W {align1};
118428d7b3dSmrgmov (8) g92.0<1>W g33.0<16,8,2>W {align1};
119428d7b3dSmrgmov (8) g93.0<1>W g34.0<16,8,2>W {align1};
120428d7b3dSmrgmov (8) g94.0<1>W g35.0<16,8,2>W {align1};
121428d7b3dSmrgmov (8) g95.0<1>W g36.0<16,8,2>W {align1};
122428d7b3dSmrgmov (8) g96.0<1>W g37.0<16,8,2>W {align1};
123428d7b3dSmrgmov (8) g97.0<1>W g38.0<16,8,2>W {align1};
124428d7b3dSmrgmov (8) g98.0<1>W g39.0<16,8,2>W {align1};
125428d7b3dSmrg
126428d7b3dSmrg//Y3
127428d7b3dSmrgblock_y3:
128428d7b3dSmrgand.nz (1) null g82.8<1,1,1>UW 0x100UW {align1};
129428d7b3dSmrg(f0) jmpi do_idct_y3;
130428d7b3dSmrgmov (8) g91.16<1>UW 0UW {align1}; 
131428d7b3dSmrgmov (8) g92.16<1>UW 0UW {align1}; 
132428d7b3dSmrgmov (8) g93.16<1>UW 0UW {align1}; 
133428d7b3dSmrgmov (8) g94.16<1>UW 0UW {align1}; 
134428d7b3dSmrgmov (8) g95.16<1>UW 0UW {align1}; 
135428d7b3dSmrgmov (8) g96.16<1>UW 0UW {align1}; 
136428d7b3dSmrgmov (8) g97.16<1>UW 0UW {align1}; 
137428d7b3dSmrgmov (8) g98.16<1>UW 0UW {align1}; 
138428d7b3dSmrgadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
139428d7b3dSmrgjmpi block_u;
140428d7b3dSmrgdo_idct_y3:
141428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
142428d7b3dSmrgjmpi IDCT_START;
143428d7b3dSmrgmov (8) g91.16<1>W g32.0<16,8,2>W {align1};
144428d7b3dSmrgmov (8) g92.16<1>W g33.0<16,8,2>W {align1};
145428d7b3dSmrgmov (8) g93.16<1>W g34.0<16,8,2>W {align1};
146428d7b3dSmrgmov (8) g94.16<1>W g35.0<16,8,2>W {align1};
147428d7b3dSmrgmov (8) g95.16<1>W g36.0<16,8,2>W {align1};
148428d7b3dSmrgmov (8) g96.16<1>W g37.0<16,8,2>W {align1};
149428d7b3dSmrgmov (8) g97.16<1>W g38.0<16,8,2>W {align1};
150428d7b3dSmrgmov (8) g98.16<1>W g39.0<16,8,2>W {align1};
151428d7b3dSmrg
152428d7b3dSmrg//U
153428d7b3dSmrgblock_u:
154428d7b3dSmrgand.nz (1) null g82.8<1,1,1>UW 0x80UW {align1};
155428d7b3dSmrg(f0) jmpi do_idct_u;
156428d7b3dSmrgmov (16) g99.0<1>UW 0UW {align1}; 
157428d7b3dSmrgmov (16) g100.0<1>UW 0UW {align1}; 
158428d7b3dSmrgmov (16) g101.0<1>UW 0UW {align1}; 
159428d7b3dSmrgmov (16) g102.0<1>UW 0UW {align1}; 
160428d7b3dSmrgadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
161428d7b3dSmrgjmpi block_v;
162428d7b3dSmrgdo_idct_u:
163428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
164428d7b3dSmrgjmpi IDCT_START;
165428d7b3dSmrgmov (16) g99.0<1>W g32.0<16,8,2>W {align1};
166428d7b3dSmrgmov (16) g100.0<1>W g34.0<16,8,2>W {align1};
167428d7b3dSmrgmov (16) g101.0<1>W g36.0<16,8,2>W {align1};
168428d7b3dSmrgmov (16) g102.0<1>W g38.0<16,8,2>W {align1};
169428d7b3dSmrg
170428d7b3dSmrg//V
171428d7b3dSmrgblock_v:
172428d7b3dSmrgand.nz (1) null g82.8<1,1,1>UW 0x40UW {align1};
173428d7b3dSmrg(f0) jmpi do_idct_v;
174428d7b3dSmrgmov (16) g103.0<1>UW 0UW {align1}; 
175428d7b3dSmrgmov (16) g104.0<1>UW 0UW {align1}; 
176428d7b3dSmrgmov (16) g105.0<1>UW 0UW {align1}; 
177428d7b3dSmrgmov (16) g106.0<1>UW 0UW {align1};  
178428d7b3dSmrgjmpi block_end;
179428d7b3dSmrgdo_idct_v:
180428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
181428d7b3dSmrgjmpi IDCT_START;
182428d7b3dSmrgmov (16) g103.0<1>W g32.0<16,8,2>W {align1};
183428d7b3dSmrgmov (16) g104.0<1>W g34.0<16,8,2>W {align1};
184428d7b3dSmrgmov (16) g105.0<1>W g36.0<16,8,2>W {align1};
185428d7b3dSmrgmov (16) g106.0<1>W g38.0<16,8,2>W {align1};
186428d7b3dSmrgblock_end:
187428d7b3dSmrg
188428d7b3dSmrgadd (1) ip g126.8<1,1,1>UD 0x20UD {align1};            //jump back
189428d7b3dSmrginclude(`do_iq_non_intra.g4i')
190428d7b3dSmrginclude(`idct.g4i')
191