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   g83~g106: IDCT data after idct
36fa225cbcSrjs   g82:    thread payload backup
37fa225cbcSrjs   g125:   ip before jump
38fa225cbcSrjs*/
39fa225cbcSrjsinclude(`iq_non_intra.g4i')
40fa225cbcSrjs
41fa225cbcSrjsdefine(`ROW_SHIFT', `11UD') //define for idct
42fa225cbcSrjsdefine(`ROW_ADD', `0x400UD')            
43fa225cbcSrjsdefine(`COL_SHIFT', `20UD') 
44fa225cbcSrjsdefine(`COL_ADD', `0x80000UD')           
45fa225cbcSrjs
46fa225cbcSrjsmov (1) a0.0<1>UD 0x06F006E0UD {align1};//0x06F006E0UD+0x00200020UD=0x07100700UD (g56.0 and g56.16,the start of DCT data)
47fa225cbcSrjs
48fa225cbcSrjs//Y0
49fa225cbcSrjsand.nz (1) null g82.8<1,1,1>UW 0x800UW {align1};
50fa225cbcSrjs(f0) jmpi do_idct_y0;
51fa225cbcSrjs
52fa225cbcSrjsmov (8) g83.0<1>UW 0UW {align1}; 
53fa225cbcSrjsmov (8) g84.0<1>UW 0UW {align1}; 
54fa225cbcSrjsmov (8) g85.0<1>UW 0UW {align1}; 
55fa225cbcSrjsmov (8) g86.0<1>UW 0UW {align1}; 
56fa225cbcSrjsmov (8) g87.0<1>UW 0UW {align1}; 
57fa225cbcSrjsmov (8) g88.0<1>UW 0UW {align1}; 
58fa225cbcSrjsmov (8) g89.0<1>UW 0UW {align1}; 
59fa225cbcSrjsmov (8) g90.0<1>UW 0UW {align1}; 
60fa225cbcSrjsadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
61fa225cbcSrjsjmpi block_y1;
62fa225cbcSrjsdo_idct_y0:
63fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
64fa225cbcSrjsjmpi IDCT_START;
65fa225cbcSrjsmov (8) g83.0<1>W g32.0<16,8,2>W {align1};
66fa225cbcSrjsmov (8) g84.0<1>W g33.0<16,8,2>W {align1};
67fa225cbcSrjsmov (8) g85.0<1>W g34.0<16,8,2>W {align1};
68fa225cbcSrjsmov (8) g86.0<1>W g35.0<16,8,2>W {align1};
69fa225cbcSrjsmov (8) g87.0<1>W g36.0<16,8,2>W {align1};
70fa225cbcSrjsmov (8) g88.0<1>W g37.0<16,8,2>W {align1};
71fa225cbcSrjsmov (8) g89.0<1>W g38.0<16,8,2>W {align1};
72fa225cbcSrjsmov (8) g90.0<1>W g39.0<16,8,2>W {align1};
73fa225cbcSrjs
74fa225cbcSrjs//Y1
75fa225cbcSrjsblock_y1:
76fa225cbcSrjsand.nz (1) null g82.8<1,1,1>UW 0x400UW {align1};
77fa225cbcSrjs(f0) jmpi do_idct_y1;
78fa225cbcSrjsmov (8) g83.16<1>UW 0UW {align1}; 
79fa225cbcSrjsmov (8) g84.16<1>UW 0UW {align1}; 
80fa225cbcSrjsmov (8) g85.16<1>UW 0UW {align1}; 
81fa225cbcSrjsmov (8) g86.16<1>UW 0UW {align1}; 
82fa225cbcSrjsmov (8) g87.16<1>UW 0UW {align1}; 
83fa225cbcSrjsmov (8) g88.16<1>UW 0UW {align1}; 
84fa225cbcSrjsmov (8) g89.16<1>UW 0UW {align1}; 
85fa225cbcSrjsmov (8) g90.16<1>UW 0UW {align1}; 
86fa225cbcSrjsadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
87fa225cbcSrjsjmpi block_y2;
88fa225cbcSrjsdo_idct_y1:
89fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
90fa225cbcSrjsjmpi IDCT_START;
91fa225cbcSrjsmov (8) g83.16<1>W g32.0<16,8,2>W {align1};
92fa225cbcSrjsmov (8) g84.16<1>W g33.0<16,8,2>W {align1};
93fa225cbcSrjsmov (8) g85.16<1>W g34.0<16,8,2>W {align1};
94fa225cbcSrjsmov (8) g86.16<1>W g35.0<16,8,2>W {align1};
95fa225cbcSrjsmov (8) g87.16<1>W g36.0<16,8,2>W {align1};
96fa225cbcSrjsmov (8) g88.16<1>W g37.0<16,8,2>W {align1};
97fa225cbcSrjsmov (8) g89.16<1>W g38.0<16,8,2>W {align1};
98fa225cbcSrjsmov (8) g90.16<1>W g39.0<16,8,2>W {align1};
99fa225cbcSrjs
100fa225cbcSrjs//Y2
101fa225cbcSrjsblock_y2:
102fa225cbcSrjsand.nz (1) null g82.8<1,1,1>UW 0x200UW {align1};
103fa225cbcSrjs(f0) jmpi do_idct_y2;
104fa225cbcSrjsmov (8) g91.0<1>UW 0UW {align1}; 
105fa225cbcSrjsmov (8) g92.0<1>UW 0UW {align1}; 
106fa225cbcSrjsmov (8) g93.0<1>UW 0UW {align1}; 
107fa225cbcSrjsmov (8) g94.0<1>UW 0UW {align1}; 
108fa225cbcSrjsmov (8) g95.0<1>UW 0UW {align1}; 
109fa225cbcSrjsmov (8) g96.0<1>UW 0UW {align1}; 
110fa225cbcSrjsmov (8) g97.0<1>UW 0UW {align1}; 
111fa225cbcSrjsmov (8) g98.0<1>UW 0UW {align1}; 
112fa225cbcSrjsadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
113fa225cbcSrjsjmpi block_y3;
114fa225cbcSrjsdo_idct_y2:
115fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
116fa225cbcSrjsjmpi IDCT_START;
117fa225cbcSrjsmov (8) g91.0<1>W g32.0<16,8,2>W {align1};
118fa225cbcSrjsmov (8) g92.0<1>W g33.0<16,8,2>W {align1};
119fa225cbcSrjsmov (8) g93.0<1>W g34.0<16,8,2>W {align1};
120fa225cbcSrjsmov (8) g94.0<1>W g35.0<16,8,2>W {align1};
121fa225cbcSrjsmov (8) g95.0<1>W g36.0<16,8,2>W {align1};
122fa225cbcSrjsmov (8) g96.0<1>W g37.0<16,8,2>W {align1};
123fa225cbcSrjsmov (8) g97.0<1>W g38.0<16,8,2>W {align1};
124fa225cbcSrjsmov (8) g98.0<1>W g39.0<16,8,2>W {align1};
125fa225cbcSrjs
126fa225cbcSrjs//Y3
127fa225cbcSrjsblock_y3:
128fa225cbcSrjsand.nz (1) null g82.8<1,1,1>UW 0x100UW {align1};
129fa225cbcSrjs(f0) jmpi do_idct_y3;
130fa225cbcSrjsmov (8) g91.16<1>UW 0UW {align1}; 
131fa225cbcSrjsmov (8) g92.16<1>UW 0UW {align1}; 
132fa225cbcSrjsmov (8) g93.16<1>UW 0UW {align1}; 
133fa225cbcSrjsmov (8) g94.16<1>UW 0UW {align1}; 
134fa225cbcSrjsmov (8) g95.16<1>UW 0UW {align1}; 
135fa225cbcSrjsmov (8) g96.16<1>UW 0UW {align1}; 
136fa225cbcSrjsmov (8) g97.16<1>UW 0UW {align1}; 
137fa225cbcSrjsmov (8) g98.16<1>UW 0UW {align1}; 
138fa225cbcSrjsadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
139fa225cbcSrjsjmpi block_u;
140fa225cbcSrjsdo_idct_y3:
141fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
142fa225cbcSrjsjmpi IDCT_START;
143fa225cbcSrjsmov (8) g91.16<1>W g32.0<16,8,2>W {align1};
144fa225cbcSrjsmov (8) g92.16<1>W g33.0<16,8,2>W {align1};
145fa225cbcSrjsmov (8) g93.16<1>W g34.0<16,8,2>W {align1};
146fa225cbcSrjsmov (8) g94.16<1>W g35.0<16,8,2>W {align1};
147fa225cbcSrjsmov (8) g95.16<1>W g36.0<16,8,2>W {align1};
148fa225cbcSrjsmov (8) g96.16<1>W g37.0<16,8,2>W {align1};
149fa225cbcSrjsmov (8) g97.16<1>W g38.0<16,8,2>W {align1};
150fa225cbcSrjsmov (8) g98.16<1>W g39.0<16,8,2>W {align1};
151fa225cbcSrjs
152fa225cbcSrjs//U
153fa225cbcSrjsblock_u:
154fa225cbcSrjsand.nz (1) null g82.8<1,1,1>UW 0x80UW {align1};
155fa225cbcSrjs(f0) jmpi do_idct_u;
156fa225cbcSrjsmov (16) g99.0<1>UW 0UW {align1}; 
157fa225cbcSrjsmov (16) g100.0<1>UW 0UW {align1}; 
158fa225cbcSrjsmov (16) g101.0<1>UW 0UW {align1}; 
159fa225cbcSrjsmov (16) g102.0<1>UW 0UW {align1}; 
160fa225cbcSrjsadd (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
161fa225cbcSrjsjmpi block_v;
162fa225cbcSrjsdo_idct_u:
163fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
164fa225cbcSrjsjmpi IDCT_START;
165fa225cbcSrjsmov (16) g99.0<1>W g32.0<16,8,2>W {align1};
166fa225cbcSrjsmov (16) g100.0<1>W g34.0<16,8,2>W {align1};
167fa225cbcSrjsmov (16) g101.0<1>W g36.0<16,8,2>W {align1};
168fa225cbcSrjsmov (16) g102.0<1>W g38.0<16,8,2>W {align1};
169fa225cbcSrjs
170fa225cbcSrjs//V
171fa225cbcSrjsblock_v:
172fa225cbcSrjsand.nz (1) null g82.8<1,1,1>UW 0x40UW {align1};
173fa225cbcSrjs(f0) jmpi do_idct_v;
174fa225cbcSrjsmov (16) g103.0<1>UW 0UW {align1}; 
175fa225cbcSrjsmov (16) g104.0<1>UW 0UW {align1}; 
176fa225cbcSrjsmov (16) g105.0<1>UW 0UW {align1}; 
177fa225cbcSrjsmov (16) g106.0<1>UW 0UW {align1};  
178fa225cbcSrjsjmpi block_end;
179fa225cbcSrjsdo_idct_v:
180fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
181fa225cbcSrjsjmpi IDCT_START;
182fa225cbcSrjsmov (16) g103.0<1>W g32.0<16,8,2>W {align1};
183fa225cbcSrjsmov (16) g104.0<1>W g34.0<16,8,2>W {align1};
184fa225cbcSrjsmov (16) g105.0<1>W g36.0<16,8,2>W {align1};
185fa225cbcSrjsmov (16) g106.0<1>W g38.0<16,8,2>W {align1};
186fa225cbcSrjsblock_end:
187fa225cbcSrjs
188fa225cbcSrjsadd (1) ip g126.8<1,1,1>UD 0x20UD {align1};            //jump back
189fa225cbcSrjsinclude(`do_iq_non_intra.g4i')
190fa225cbcSrjsinclude(`idct.g4i')
191