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/*
29fa225cbcSrjs   GRF allocation:
30fa225cbcSrjs   g1~g30: constant buffer
31fa225cbcSrjs           g1~g2:intra IQ matrix
32fa225cbcSrjs           g3~g4:non intra IQ matrix
33fa225cbcSrjs           g5~g20:IDCT tab
34fa225cbcSrjs   g31:    read and write message descriptor
35fa225cbcSrjs   g32~g55:DCT data
36fa225cbcSrjs   g58~g81:reference data
37fa225cbcSrjs   g82:    thread payload 
38fa225cbcSrjs   g83~g106:IDCT data 
39fa225cbcSrjs*/
40fa225cbcSrjsmov (8) g82.0<1>UD g31.0<8,8,1>UD {align1};
41fa225cbcSrjsmov (2) g31.0<1>UD g82.12<2,2,1>UW {align1};
42fa225cbcSrjs
43fa225cbcSrjsinclude(`iq_intra.g4i')
44fa225cbcSrjs
45fa225cbcSrjs//defined for idct
46fa225cbcSrjsdefine(`ROW_SHIFT', `11UD') 
47fa225cbcSrjsdefine(`ROW_ADD', `0x400UD')          
48fa225cbcSrjsdefine(`COL_SHIFT', `20UD') 
49fa225cbcSrjsdefine(`COL_ADD', `0x80000UD')          
50fa225cbcSrjs
51fa225cbcSrjsmov (1) a0.0<1>UD 0x06F006E0UD {align1};      //0x06F006E0UD+0x00200020UD=0x07100700UD (g56.0 and g56.16)
52fa225cbcSrjs
53fa225cbcSrjs//Y0
54fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
55fa225cbcSrjsjmpi IDCT_START;
56fa225cbcSrjsadd (8) g83.0<1>W g32.0<16,8,2>W 128UW {align1};
57fa225cbcSrjsadd (8) g84.0<1>W g33.0<16,8,2>W 128UW {align1};
58fa225cbcSrjsadd (8) g85.0<1>W g34.0<16,8,2>W 128UW {align1};
59fa225cbcSrjsadd (8) g86.0<1>W g35.0<16,8,2>W 128UW {align1};
60fa225cbcSrjsadd (8) g87.0<1>W g36.0<16,8,2>W 128UW {align1};
61fa225cbcSrjsadd (8) g88.0<1>W g37.0<16,8,2>W 128UW {align1};
62fa225cbcSrjsadd (8) g89.0<1>W g38.0<16,8,2>W 128UW {align1};
63fa225cbcSrjsadd (8) g90.0<1>W g39.0<16,8,2>W 128UW {align1};
64fa225cbcSrjs
65fa225cbcSrjs//Y1
66fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
67fa225cbcSrjsjmpi IDCT_START;
68fa225cbcSrjsadd (8) g83.16<1>W g32.0<16,8,2>W 128UW {align1};
69fa225cbcSrjsadd (8) g84.16<1>W g33.0<16,8,2>W 128UW {align1};
70fa225cbcSrjsadd (8) g85.16<1>W g34.0<16,8,2>W 128UW {align1};
71fa225cbcSrjsadd (8) g86.16<1>W g35.0<16,8,2>W 128UW {align1};
72fa225cbcSrjsadd (8) g87.16<1>W g36.0<16,8,2>W 128UW {align1};
73fa225cbcSrjsadd (8) g88.16<1>W g37.0<16,8,2>W 128UW {align1};
74fa225cbcSrjsadd (8) g89.16<1>W g38.0<16,8,2>W 128UW {align1};
75fa225cbcSrjsadd (8) g90.16<1>W g39.0<16,8,2>W 128UW {align1};
76fa225cbcSrjs
77fa225cbcSrjs//Y2
78fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
79fa225cbcSrjsjmpi IDCT_START;
80fa225cbcSrjsadd (8) g91.0<1>W g32.0<16,8,2>W 128UW {align1};
81fa225cbcSrjsadd (8) g92.0<1>W g33.0<16,8,2>W 128UW {align1};
82fa225cbcSrjsadd (8) g93.0<1>W g34.0<16,8,2>W 128UW {align1};
83fa225cbcSrjsadd (8) g94.0<1>W g35.0<16,8,2>W 128UW {align1};
84fa225cbcSrjsadd (8) g95.0<1>W g36.0<16,8,2>W 128UW {align1};
85fa225cbcSrjsadd (8) g96.0<1>W g37.0<16,8,2>W 128UW {align1};
86fa225cbcSrjsadd (8) g97.0<1>W g38.0<16,8,2>W 128UW {align1};
87fa225cbcSrjsadd (8) g98.0<1>W g39.0<16,8,2>W 128UW {align1};
88fa225cbcSrjs
89fa225cbcSrjs//Y3
90fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
91fa225cbcSrjsjmpi IDCT_START;
92fa225cbcSrjsadd (8) g91.16<1>W g32.0<16,8,2>W 128UW {align1};
93fa225cbcSrjsadd (8) g92.16<1>W g33.0<16,8,2>W 128UW {align1};
94fa225cbcSrjsadd (8) g93.16<1>W g34.0<16,8,2>W 128UW {align1};
95fa225cbcSrjsadd (8) g94.16<1>W g35.0<16,8,2>W 128UW {align1};
96fa225cbcSrjsadd (8) g95.16<1>W g36.0<16,8,2>W 128UW {align1};
97fa225cbcSrjsadd (8) g96.16<1>W g37.0<16,8,2>W 128UW {align1};
98fa225cbcSrjsadd (8) g97.16<1>W g38.0<16,8,2>W 128UW {align1};
99fa225cbcSrjsadd (8) g98.16<1>W g39.0<16,8,2>W 128UW {align1};
100fa225cbcSrjs
101fa225cbcSrjs//U
102fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
103fa225cbcSrjsjmpi IDCT_START;
104fa225cbcSrjsadd (16) g99.0<1>W g32.0<16,8,2>W 128UW {align1};
105fa225cbcSrjsadd (16) g100.0<1>W g34.0<16,8,2>W 128UW {align1};
106fa225cbcSrjsadd (16) g101.0<1>W g36.0<16,8,2>W 128UW {align1};
107fa225cbcSrjsadd (16) g102.0<1>W g38.0<16,8,2>W 128UW {align1};
108fa225cbcSrjs
109fa225cbcSrjs//V
110fa225cbcSrjsmov (1) g125.0<1>UD ip {align1};
111fa225cbcSrjsjmpi IDCT_START;
112fa225cbcSrjsadd (16) g103.0<1>W g32.0<16,8,2>W 128UW {align1};
113fa225cbcSrjsadd (16) g104.0<1>W g34.0<16,8,2>W 128UW {align1};
114fa225cbcSrjsadd (16) g105.0<1>W g36.0<16,8,2>W 128UW {align1};
115fa225cbcSrjsadd (16) g106.0<1>W g38.0<16,8,2>W 128UW {align1};
116fa225cbcSrjs
117fa225cbcSrjs/******************* MC ************************/
118fa225cbcSrjsmov (1) g31.8<1>UD 0x00F000FUD {align1};
119fa225cbcSrjsmov.sat (16) g83.0<2>UB g83.0<16,16,1>W {align1};
120fa225cbcSrjsmov.sat (16) g84.0<2>UB g84.0<16,16,1>W {align1};
121fa225cbcSrjsmov.sat (16) g85.0<2>UB g85.0<16,16,1>W {align1};
122fa225cbcSrjsmov.sat (16) g86.0<2>UB g86.0<16,16,1>W {align1};
123fa225cbcSrjsmov.sat (16) g87.0<2>UB g87.0<16,16,1>W {align1};
124fa225cbcSrjsmov.sat (16) g88.0<2>UB g88.0<16,16,1>W {align1};
125fa225cbcSrjsmov.sat (16) g89.0<2>UB g89.0<16,16,1>W {align1};
126fa225cbcSrjsmov.sat (16) g90.0<2>UB g90.0<16,16,1>W {align1};
127fa225cbcSrjsmov.sat (16) g91.0<2>UB g91.0<16,16,1>W {align1};
128fa225cbcSrjsmov.sat (16) g92.0<2>UB g92.0<16,16,1>W {align1};
129fa225cbcSrjsmov.sat (16) g93.0<2>UB g93.0<16,16,1>W {align1};
130fa225cbcSrjsmov.sat (16) g94.0<2>UB g94.0<16,16,1>W {align1};
131fa225cbcSrjsmov.sat (16) g95.0<2>UB g95.0<16,16,1>W {align1};
132fa225cbcSrjsmov.sat (16) g96.0<2>UB g96.0<16,16,1>W {align1};
133fa225cbcSrjsmov.sat (16) g97.0<2>UB g97.0<16,16,1>W {align1};
134fa225cbcSrjsmov.sat (16) g98.0<2>UB g98.0<16,16,1>W {align1};
135fa225cbcSrjs
136fa225cbcSrjsand.nz (1) null g82.2<1,1,1>UW 0x20UW{align1};
137fa225cbcSrjs(f0) jmpi field_dct;
138fa225cbcSrjs
139fa225cbcSrjsmov (16) m1.0<1>UB  g83.0<16,16,2>UB {align1};
140fa225cbcSrjsmov (16) m1.16<1>UB g84.0<16,16,2>UB {align1};
141fa225cbcSrjsmov (16) m2.0<1>UB  g85.0<16,16,2>UB {align1};
142fa225cbcSrjsmov (16) m2.16<1>UB g86.0<16,16,2>UB {align1};
143fa225cbcSrjsmov (16) m3.0<1>UB  g87.0<16,16,2>UB {align1};
144fa225cbcSrjsmov (16) m3.16<1>UB g88.0<16,16,2>UB {align1};
145fa225cbcSrjsmov (16) m4.0<1>UB  g89.0<16,16,2>UB {align1};
146fa225cbcSrjsmov (16) m4.16<1>UB g90.0<16,16,2>UB {align1};
147fa225cbcSrjsmov (16) m5.0<1>UB  g91.0<16,16,2>UB {align1};
148fa225cbcSrjsmov (16) m5.16<1>UB g92.0<16,16,2>UB {align1};
149fa225cbcSrjsmov (16) m6.0<1>UB  g93.0<16,16,2>UB {align1};
150fa225cbcSrjsmov (16) m6.16<1>UB g94.0<16,16,2>UB {align1};
151fa225cbcSrjsmov (16) m7.0<1>UB  g95.0<16,16,2>UB {align1};
152fa225cbcSrjsmov (16) m7.16<1>UB g96.0<16,16,2>UB {align1};
153fa225cbcSrjsmov (16) m8.0<1>UB  g97.0<16,16,2>UB {align1};
154fa225cbcSrjsmov (16) m8.16<1>UB g98.0<16,16,2>UB {align1};
155fa225cbcSrjsjmpi write_back;
156fa225cbcSrjs
157fa225cbcSrjsfield_dct:
158fa225cbcSrjsmov (16) m1.0<1>UB  g83.0<16,16,2>UB {align1};
159fa225cbcSrjsmov (16) m1.16<1>UB g91.0<16,16,2>UB {align1};
160fa225cbcSrjsmov (16) m2.0<1>UB  g84.0<16,16,2>UB {align1};
161fa225cbcSrjsmov (16) m2.16<1>UB g92.0<16,16,2>UB {align1};
162fa225cbcSrjsmov (16) m3.0<1>UB  g85.0<16,16,2>UB {align1};
163fa225cbcSrjsmov (16) m3.16<1>UB g93.0<16,16,2>UB {align1};
164fa225cbcSrjsmov (16) m4.0<1>UB  g86.0<16,16,2>UB {align1};
165fa225cbcSrjsmov (16) m4.16<1>UB g94.0<16,16,2>UB {align1};
166fa225cbcSrjsmov (16) m5.0<1>UB  g87.0<16,16,2>UB {align1};
167fa225cbcSrjsmov (16) m5.16<1>UB g95.0<16,16,2>UB {align1};
168fa225cbcSrjsmov (16) m6.0<1>UB  g88.0<16,16,2>UB {align1};
169fa225cbcSrjsmov (16) m6.16<1>UB g96.0<16,16,2>UB {align1};
170fa225cbcSrjsmov (16) m7.0<1>UB  g89.0<16,16,2>UB {align1};
171fa225cbcSrjsmov (16) m7.16<1>UB g97.0<16,16,2>UB {align1};
172fa225cbcSrjsmov (16) m8.0<1>UB  g90.0<16,16,2>UB {align1};
173fa225cbcSrjsmov (16) m8.16<1>UB g98.0<16,16,2>UB {align1};
174fa225cbcSrjs
175fa225cbcSrjswrite_back:
176fa225cbcSrjssend (16) 0 acc0<1>UW g31<8,8,1>UW write(0,0,2,0) mlen 9 rlen 0 {align1};
177fa225cbcSrjs
178fa225cbcSrjs//U
179fa225cbcSrjsmov (1) g31.8<1>UD 0x0070007UD  { align1 };
180fa225cbcSrjsshr (2) g31.0<1>UD g82.12<2,2,1>UW 1D {align1};
181fa225cbcSrjsmov.sat (16) g99.0<2>UB g99.0<16,16,1>W {align1};
182fa225cbcSrjsmov.sat (16) g100.0<2>UB g100.0<16,16,1>W {align1};
183fa225cbcSrjsmov.sat (16) g101.0<2>UB g101.0<16,16,1>W {align1};
184fa225cbcSrjsmov.sat (16) g102.0<2>UB g102.0<16,16,1>W {align1};
185fa225cbcSrjs
186fa225cbcSrjsmov (16) m1.0<1>UB  g99.0<16,16,2>UB {align1};
187fa225cbcSrjsmov (16) m1.16<1>UB g100.0<16,16,2>UB {align1};
188fa225cbcSrjsmov (16) m2.0<1>UB  g101.0<16,16,2>UB {align1};
189fa225cbcSrjsmov (16) m2.16<1>UB g102.0<16,16,2>UB {align1};
190fa225cbcSrjssend (16) 0 acc0<1>UW g31<8,8,1>UW write(1, 0, 2, 0) mlen 3 rlen 0 { align1 };
191fa225cbcSrjs
192fa225cbcSrjs//V
193fa225cbcSrjsmov.sat (16) g103.0<2>UB g103.0<16,16,1>W {align1};
194fa225cbcSrjsmov.sat (16) g104.0<2>UB g104.0<16,16,1>W {align1};
195fa225cbcSrjsmov.sat (16) g105.0<2>UB g105.0<16,16,1>W {align1};
196fa225cbcSrjsmov.sat (16) g106.0<2>UB g106.0<16,16,1>W {align1};
197fa225cbcSrjs
198fa225cbcSrjsmov (16) m1.0<1>UB  g103.0<16,16,2>UB {align1};
199fa225cbcSrjsmov (16) m1.16<1>UB g104.0<16,16,2>UB {align1};
200fa225cbcSrjsmov (16) m2.0<1>UB  g105.0<16,16,2>UB {align1};
201fa225cbcSrjsmov (16) m2.16<1>UB g106.0<16,16,2>UB {align1};
202fa225cbcSrjssend (16) 0 acc0<1>UW g31<8,8,1>UW write(2, 0, 2, 0) mlen 3 rlen 0 { align1 };
203fa225cbcSrjs
204fa225cbcSrjsOUT:
205fa225cbcSrjssend (16) 0 acc0<1>UW g0<8,8,1>UW
206fa225cbcSrjs        thread_spawner(0, 0, 0) mlen 1 rlen 0 { align1 EOT};
207fa225cbcSrjs
208fa225cbcSrjsinclude(`do_iq_intra.g4i')
209fa225cbcSrjsinclude(`idct.g4i')
210