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/*
29428d7b3dSmrg   GRF allocation:
30428d7b3dSmrg   g1~g30: constant buffer
31428d7b3dSmrg           g1~g2:intra IQ matrix
32428d7b3dSmrg           g3~g4:non intra IQ matrix
33428d7b3dSmrg           g5~g20:IDCT tab
34428d7b3dSmrg   g31:    read and write message descriptor
35428d7b3dSmrg   g32~g55:DCT data
36428d7b3dSmrg   g58~g81:reference data
37428d7b3dSmrg   g82:    thread payload 
38428d7b3dSmrg   g83~g106:IDCT data 
39428d7b3dSmrg*/
40428d7b3dSmrgmov (8) g82.0<1>UD g31.0<8,8,1>UD {align1};
41428d7b3dSmrgmov (2) g31.0<1>UD g82.12<2,2,1>UW {align1};
42428d7b3dSmrg
43428d7b3dSmrginclude(`iq_intra.g4i')
44428d7b3dSmrg
45428d7b3dSmrg//defined for idct
46428d7b3dSmrgdefine(`ROW_SHIFT', `11UD') 
47428d7b3dSmrgdefine(`ROW_ADD', `0x400UD')          
48428d7b3dSmrgdefine(`COL_SHIFT', `20UD') 
49428d7b3dSmrgdefine(`COL_ADD', `0x80000UD')          
50428d7b3dSmrg
51428d7b3dSmrgmov (1) a0.0<1>UD 0x06F006E0UD {align1};      //0x06F006E0UD+0x00200020UD=0x07100700UD (g56.0 and g56.16)
52428d7b3dSmrg
53428d7b3dSmrg//Y0
54428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
55428d7b3dSmrgjmpi IDCT_START;
56428d7b3dSmrgadd (8) g83.0<1>W g32.0<16,8,2>W 128UW {align1};
57428d7b3dSmrgadd (8) g84.0<1>W g33.0<16,8,2>W 128UW {align1};
58428d7b3dSmrgadd (8) g85.0<1>W g34.0<16,8,2>W 128UW {align1};
59428d7b3dSmrgadd (8) g86.0<1>W g35.0<16,8,2>W 128UW {align1};
60428d7b3dSmrgadd (8) g87.0<1>W g36.0<16,8,2>W 128UW {align1};
61428d7b3dSmrgadd (8) g88.0<1>W g37.0<16,8,2>W 128UW {align1};
62428d7b3dSmrgadd (8) g89.0<1>W g38.0<16,8,2>W 128UW {align1};
63428d7b3dSmrgadd (8) g90.0<1>W g39.0<16,8,2>W 128UW {align1};
64428d7b3dSmrg
65428d7b3dSmrg//Y1
66428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
67428d7b3dSmrgjmpi IDCT_START;
68428d7b3dSmrgadd (8) g83.16<1>W g32.0<16,8,2>W 128UW {align1};
69428d7b3dSmrgadd (8) g84.16<1>W g33.0<16,8,2>W 128UW {align1};
70428d7b3dSmrgadd (8) g85.16<1>W g34.0<16,8,2>W 128UW {align1};
71428d7b3dSmrgadd (8) g86.16<1>W g35.0<16,8,2>W 128UW {align1};
72428d7b3dSmrgadd (8) g87.16<1>W g36.0<16,8,2>W 128UW {align1};
73428d7b3dSmrgadd (8) g88.16<1>W g37.0<16,8,2>W 128UW {align1};
74428d7b3dSmrgadd (8) g89.16<1>W g38.0<16,8,2>W 128UW {align1};
75428d7b3dSmrgadd (8) g90.16<1>W g39.0<16,8,2>W 128UW {align1};
76428d7b3dSmrg
77428d7b3dSmrg//Y2
78428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
79428d7b3dSmrgjmpi IDCT_START;
80428d7b3dSmrgadd (8) g91.0<1>W g32.0<16,8,2>W 128UW {align1};
81428d7b3dSmrgadd (8) g92.0<1>W g33.0<16,8,2>W 128UW {align1};
82428d7b3dSmrgadd (8) g93.0<1>W g34.0<16,8,2>W 128UW {align1};
83428d7b3dSmrgadd (8) g94.0<1>W g35.0<16,8,2>W 128UW {align1};
84428d7b3dSmrgadd (8) g95.0<1>W g36.0<16,8,2>W 128UW {align1};
85428d7b3dSmrgadd (8) g96.0<1>W g37.0<16,8,2>W 128UW {align1};
86428d7b3dSmrgadd (8) g97.0<1>W g38.0<16,8,2>W 128UW {align1};
87428d7b3dSmrgadd (8) g98.0<1>W g39.0<16,8,2>W 128UW {align1};
88428d7b3dSmrg
89428d7b3dSmrg//Y3
90428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
91428d7b3dSmrgjmpi IDCT_START;
92428d7b3dSmrgadd (8) g91.16<1>W g32.0<16,8,2>W 128UW {align1};
93428d7b3dSmrgadd (8) g92.16<1>W g33.0<16,8,2>W 128UW {align1};
94428d7b3dSmrgadd (8) g93.16<1>W g34.0<16,8,2>W 128UW {align1};
95428d7b3dSmrgadd (8) g94.16<1>W g35.0<16,8,2>W 128UW {align1};
96428d7b3dSmrgadd (8) g95.16<1>W g36.0<16,8,2>W 128UW {align1};
97428d7b3dSmrgadd (8) g96.16<1>W g37.0<16,8,2>W 128UW {align1};
98428d7b3dSmrgadd (8) g97.16<1>W g38.0<16,8,2>W 128UW {align1};
99428d7b3dSmrgadd (8) g98.16<1>W g39.0<16,8,2>W 128UW {align1};
100428d7b3dSmrg
101428d7b3dSmrg//U
102428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
103428d7b3dSmrgjmpi IDCT_START;
104428d7b3dSmrgadd (16) g99.0<1>W g32.0<16,8,2>W 128UW {align1};
105428d7b3dSmrgadd (16) g100.0<1>W g34.0<16,8,2>W 128UW {align1};
106428d7b3dSmrgadd (16) g101.0<1>W g36.0<16,8,2>W 128UW {align1};
107428d7b3dSmrgadd (16) g102.0<1>W g38.0<16,8,2>W 128UW {align1};
108428d7b3dSmrg
109428d7b3dSmrg//V
110428d7b3dSmrgmov (1) g125.0<1>UD ip {align1};
111428d7b3dSmrgjmpi IDCT_START;
112428d7b3dSmrgadd (16) g103.0<1>W g32.0<16,8,2>W 128UW {align1};
113428d7b3dSmrgadd (16) g104.0<1>W g34.0<16,8,2>W 128UW {align1};
114428d7b3dSmrgadd (16) g105.0<1>W g36.0<16,8,2>W 128UW {align1};
115428d7b3dSmrgadd (16) g106.0<1>W g38.0<16,8,2>W 128UW {align1};
116428d7b3dSmrg
117428d7b3dSmrg/******************* MC ************************/
118428d7b3dSmrgmov (1) g31.8<1>UD 0x00F000FUD {align1};
119428d7b3dSmrgmov.sat (16) g83.0<2>UB g83.0<16,16,1>W {align1};
120428d7b3dSmrgmov.sat (16) g84.0<2>UB g84.0<16,16,1>W {align1};
121428d7b3dSmrgmov.sat (16) g85.0<2>UB g85.0<16,16,1>W {align1};
122428d7b3dSmrgmov.sat (16) g86.0<2>UB g86.0<16,16,1>W {align1};
123428d7b3dSmrgmov.sat (16) g87.0<2>UB g87.0<16,16,1>W {align1};
124428d7b3dSmrgmov.sat (16) g88.0<2>UB g88.0<16,16,1>W {align1};
125428d7b3dSmrgmov.sat (16) g89.0<2>UB g89.0<16,16,1>W {align1};
126428d7b3dSmrgmov.sat (16) g90.0<2>UB g90.0<16,16,1>W {align1};
127428d7b3dSmrgmov.sat (16) g91.0<2>UB g91.0<16,16,1>W {align1};
128428d7b3dSmrgmov.sat (16) g92.0<2>UB g92.0<16,16,1>W {align1};
129428d7b3dSmrgmov.sat (16) g93.0<2>UB g93.0<16,16,1>W {align1};
130428d7b3dSmrgmov.sat (16) g94.0<2>UB g94.0<16,16,1>W {align1};
131428d7b3dSmrgmov.sat (16) g95.0<2>UB g95.0<16,16,1>W {align1};
132428d7b3dSmrgmov.sat (16) g96.0<2>UB g96.0<16,16,1>W {align1};
133428d7b3dSmrgmov.sat (16) g97.0<2>UB g97.0<16,16,1>W {align1};
134428d7b3dSmrgmov.sat (16) g98.0<2>UB g98.0<16,16,1>W {align1};
135428d7b3dSmrg
136428d7b3dSmrgand.nz (1) null g82.2<1,1,1>UW 0x20UW{align1};
137428d7b3dSmrg(f0) jmpi field_dct;
138428d7b3dSmrg
139428d7b3dSmrgmov (16) m1.0<1>UB  g83.0<16,16,2>UB {align1};
140428d7b3dSmrgmov (16) m1.16<1>UB g84.0<16,16,2>UB {align1};
141428d7b3dSmrgmov (16) m2.0<1>UB  g85.0<16,16,2>UB {align1};
142428d7b3dSmrgmov (16) m2.16<1>UB g86.0<16,16,2>UB {align1};
143428d7b3dSmrgmov (16) m3.0<1>UB  g87.0<16,16,2>UB {align1};
144428d7b3dSmrgmov (16) m3.16<1>UB g88.0<16,16,2>UB {align1};
145428d7b3dSmrgmov (16) m4.0<1>UB  g89.0<16,16,2>UB {align1};
146428d7b3dSmrgmov (16) m4.16<1>UB g90.0<16,16,2>UB {align1};
147428d7b3dSmrgmov (16) m5.0<1>UB  g91.0<16,16,2>UB {align1};
148428d7b3dSmrgmov (16) m5.16<1>UB g92.0<16,16,2>UB {align1};
149428d7b3dSmrgmov (16) m6.0<1>UB  g93.0<16,16,2>UB {align1};
150428d7b3dSmrgmov (16) m6.16<1>UB g94.0<16,16,2>UB {align1};
151428d7b3dSmrgmov (16) m7.0<1>UB  g95.0<16,16,2>UB {align1};
152428d7b3dSmrgmov (16) m7.16<1>UB g96.0<16,16,2>UB {align1};
153428d7b3dSmrgmov (16) m8.0<1>UB  g97.0<16,16,2>UB {align1};
154428d7b3dSmrgmov (16) m8.16<1>UB g98.0<16,16,2>UB {align1};
155428d7b3dSmrgjmpi write_back;
156428d7b3dSmrg
157428d7b3dSmrgfield_dct:
158428d7b3dSmrgmov (16) m1.0<1>UB  g83.0<16,16,2>UB {align1};
159428d7b3dSmrgmov (16) m1.16<1>UB g91.0<16,16,2>UB {align1};
160428d7b3dSmrgmov (16) m2.0<1>UB  g84.0<16,16,2>UB {align1};
161428d7b3dSmrgmov (16) m2.16<1>UB g92.0<16,16,2>UB {align1};
162428d7b3dSmrgmov (16) m3.0<1>UB  g85.0<16,16,2>UB {align1};
163428d7b3dSmrgmov (16) m3.16<1>UB g93.0<16,16,2>UB {align1};
164428d7b3dSmrgmov (16) m4.0<1>UB  g86.0<16,16,2>UB {align1};
165428d7b3dSmrgmov (16) m4.16<1>UB g94.0<16,16,2>UB {align1};
166428d7b3dSmrgmov (16) m5.0<1>UB  g87.0<16,16,2>UB {align1};
167428d7b3dSmrgmov (16) m5.16<1>UB g95.0<16,16,2>UB {align1};
168428d7b3dSmrgmov (16) m6.0<1>UB  g88.0<16,16,2>UB {align1};
169428d7b3dSmrgmov (16) m6.16<1>UB g96.0<16,16,2>UB {align1};
170428d7b3dSmrgmov (16) m7.0<1>UB  g89.0<16,16,2>UB {align1};
171428d7b3dSmrgmov (16) m7.16<1>UB g97.0<16,16,2>UB {align1};
172428d7b3dSmrgmov (16) m8.0<1>UB  g90.0<16,16,2>UB {align1};
173428d7b3dSmrgmov (16) m8.16<1>UB g98.0<16,16,2>UB {align1};
174428d7b3dSmrg
175428d7b3dSmrgwrite_back:
176428d7b3dSmrgsend (16) 0 acc0<1>UW g31<8,8,1>UW write(0,0,2,0) mlen 9 rlen 0 {align1};
177428d7b3dSmrg
178428d7b3dSmrg//U
179428d7b3dSmrgmov (1) g31.8<1>UD 0x0070007UD  { align1 };
180428d7b3dSmrgshr (2) g31.0<1>UD g82.12<2,2,1>UW 1D {align1};
181428d7b3dSmrgmov.sat (16) g99.0<2>UB g99.0<16,16,1>W {align1};
182428d7b3dSmrgmov.sat (16) g100.0<2>UB g100.0<16,16,1>W {align1};
183428d7b3dSmrgmov.sat (16) g101.0<2>UB g101.0<16,16,1>W {align1};
184428d7b3dSmrgmov.sat (16) g102.0<2>UB g102.0<16,16,1>W {align1};
185428d7b3dSmrg
186428d7b3dSmrgmov (16) m1.0<1>UB  g99.0<16,16,2>UB {align1};
187428d7b3dSmrgmov (16) m1.16<1>UB g100.0<16,16,2>UB {align1};
188428d7b3dSmrgmov (16) m2.0<1>UB  g101.0<16,16,2>UB {align1};
189428d7b3dSmrgmov (16) m2.16<1>UB g102.0<16,16,2>UB {align1};
190428d7b3dSmrgsend (16) 0 acc0<1>UW g31<8,8,1>UW write(1, 0, 2, 0) mlen 3 rlen 0 { align1 };
191428d7b3dSmrg
192428d7b3dSmrg//V
193428d7b3dSmrgmov.sat (16) g103.0<2>UB g103.0<16,16,1>W {align1};
194428d7b3dSmrgmov.sat (16) g104.0<2>UB g104.0<16,16,1>W {align1};
195428d7b3dSmrgmov.sat (16) g105.0<2>UB g105.0<16,16,1>W {align1};
196428d7b3dSmrgmov.sat (16) g106.0<2>UB g106.0<16,16,1>W {align1};
197428d7b3dSmrg
198428d7b3dSmrgmov (16) m1.0<1>UB  g103.0<16,16,2>UB {align1};
199428d7b3dSmrgmov (16) m1.16<1>UB g104.0<16,16,2>UB {align1};
200428d7b3dSmrgmov (16) m2.0<1>UB  g105.0<16,16,2>UB {align1};
201428d7b3dSmrgmov (16) m2.16<1>UB g106.0<16,16,2>UB {align1};
202428d7b3dSmrgsend (16) 0 acc0<1>UW g31<8,8,1>UW write(2, 0, 2, 0) mlen 3 rlen 0 { align1 };
203428d7b3dSmrg
204428d7b3dSmrgOUT:
205428d7b3dSmrgsend (16) 0 acc0<1>UW g0<8,8,1>UW
206428d7b3dSmrg        thread_spawner(0, 0, 0) mlen 1 rlen 0 { align1 EOT};
207428d7b3dSmrg
208428d7b3dSmrginclude(`do_iq_intra.g4i')
209428d7b3dSmrginclude(`idct.g4i')
210