1/*
2 * Copyright © 2008 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Author:
24 *    Zou Nan hai <nanhai.zou@intel.com>
25 *    Zhang Hua jun <huajun.zhang@intel.com>
26 *    Xing Dong sheng <dongsheng.xing@intel.com>
27 *
28 */
29mov (1) g2.8<1>UD 0x007001FUD {align1};
30send (16) 0 g98.0<1>UW g2<8,8,1>UW read(input_surface, 2, 0, 2) mlen 1 rlen 8 {align1};
31
32and (1) g2.24<1>UD g2.0<1,1,1>UD 3UD {align1};
33mul(1) g2.24<1>UD g2.24<1,1,1>UD 49UD {align1};
34add (1) g2.4<1>UD g2.4<1,1,1>UD 8UD {align1};
35send (16) 0 g106.0<1>UW g2<8,8,1>UW read(input_surface, 2, 0, 2) mlen 1 rlen 8 {align1};
36add (1) g2.4<1>UD g2.4<1,1,1>UD 8UD {align1};
37mov (1) g2.8<1>UD 0x1FUD {align1};
38send (16) 0 g120.0<1>UW g2<8,8,1>UW read(input_surface, 2, 0, 2) mlen 1 rlen 1 {align1};
39
40jmpi g2.24<1,1,1>D;
41
42add (16) g28.0<1>UW g98.0<16,16,1>UB g98.1<16,16,1>UB {align1};
43add (16) g29.0<1>UW g99.0<16,16,1>UB g99.1<16,16,1>UB {align1};
44add (16) g30.0<1>UW g100.0<16,16,1>UB g100.1<16,16,1>UB {align1};
45add (16) g31.0<1>UW g101.0<16,16,1>UB g101.1<16,16,1>UB {align1};
46add (16) g32.0<1>UW g102.0<16,16,1>UB g102.1<16,16,1>UB {align1};
47add (16) g33.0<1>UW g103.0<16,16,1>UB g103.1<16,16,1>UB {align1};
48add (16) g34.0<1>UW g104.0<16,16,1>UB g104.1<16,16,1>UB {align1};
49add (16) g35.0<1>UW g105.0<16,16,1>UB g105.1<16,16,1>UB {align1};
50add (16) g36.0<1>UW g106.0<16,16,1>UB g106.1<16,16,1>UB {align1};
51add (16) g37.0<1>UW g107.0<16,16,1>UB g107.1<16,16,1>UB {align1};
52add (16) g38.0<1>UW g108.0<16,16,1>UB g108.1<16,16,1>UB {align1};
53add (16) g39.0<1>UW g109.0<16,16,1>UB g109.1<16,16,1>UB {align1};
54add (16) g40.0<1>UW g110.0<16,16,1>UB g110.1<16,16,1>UB {align1};
55add (16) g41.0<1>UW g111.0<16,16,1>UB g111.1<16,16,1>UB {align1};
56add (16) g42.0<1>UW g112.0<16,16,1>UB g112.1<16,16,1>UB {align1};
57add (16) g43.0<1>UW g113.0<16,16,1>UB g113.1<16,16,1>UB {align1};
58
59add (16) g28.0<1>UW g28.0<16,16,1>UW g99.0<16,16,1>UB {align1};
60add (16) g29.0<1>UW g29.0<16,16,1>UW g100.0<16,16,1>UB {align1};
61add (16) g30.0<1>UW g30.0<16,16,1>UW g101.0<16,16,1>UB {align1};
62add (16) g31.0<1>UW g31.0<16,16,1>UW g102.0<16,16,1>UB {align1};
63add (16) g32.0<1>UW g32.0<16,16,1>UW g103.0<16,16,1>UB {align1};
64add (16) g33.0<1>UW g33.0<16,16,1>UW g104.0<16,16,1>UB {align1};
65add (16) g34.0<1>UW g34.0<16,16,1>UW g105.0<16,16,1>UB {align1};
66add (16) g35.0<1>UW g35.0<16,16,1>UW g106.0<16,16,1>UB {align1};
67add (16) g36.0<1>UW g36.0<16,16,1>UW g107.0<16,16,1>UB {align1};
68add (16) g37.0<1>UW g37.0<16,16,1>UW g108.0<16,16,1>UB {align1};
69add (16) g38.0<1>UW g38.0<16,16,1>UW g109.0<16,16,1>UB {align1};
70add (16) g39.0<1>UW g39.0<16,16,1>UW g110.0<16,16,1>UB {align1};
71add (16) g40.0<1>UW g40.0<16,16,1>UW g111.0<16,16,1>UB {align1};
72add (16) g41.0<1>UW g41.0<16,16,1>UW g112.0<16,16,1>UB {align1};
73add (16) g42.0<1>UW g42.0<16,16,1>UW g113.0<16,16,1>UB {align1};
74add (16) g43.0<1>UW g43.0<16,16,1>UW g120.0<16,16,1>UB {align1};
75
76add (16) g28.0<1>UW g28.0<16,16,1>UW g99.1<16,16,1>UB {align1};
77add (16) g29.0<1>UW g29.0<16,16,1>UW g100.1<16,16,1>UB {align1};
78add (16) g30.0<1>UW g30.0<16,16,1>UW g101.1<16,16,1>UB {align1};
79add (16) g31.0<1>UW g31.0<16,16,1>UW g102.1<16,16,1>UB {align1};
80add (16) g32.0<1>UW g32.0<16,16,1>UW g103.1<16,16,1>UB {align1};
81add (16) g33.0<1>UW g33.0<16,16,1>UW g104.1<16,16,1>UB {align1};
82add (16) g34.0<1>UW g34.0<16,16,1>UW g105.1<16,16,1>UB {align1};
83add (16) g35.0<1>UW g35.0<16,16,1>UW g106.1<16,16,1>UB {align1};
84add (16) g36.0<1>UW g36.0<16,16,1>UW g107.1<16,16,1>UB {align1};
85add (16) g37.0<1>UW g37.0<16,16,1>UW g108.1<16,16,1>UB {align1};
86add (16) g38.0<1>UW g38.0<16,16,1>UW g109.1<16,16,1>UB {align1};
87add (16) g39.0<1>UW g39.0<16,16,1>UW g110.1<16,16,1>UB {align1};
88add (16) g40.0<1>UW g40.0<16,16,1>UW g111.1<16,16,1>UB {align1};
89add (16) g41.0<1>UW g41.0<16,16,1>UW g112.1<16,16,1>UB {align1};
90add (16) g42.0<1>UW g42.0<16,16,1>UW g113.1<16,16,1>UB {align1};
91add (16) g43.0<1>UW g43.0<16,16,1>UW g120.1<16,16,1>UB {align1};
92jmpi out;
93
94add (16) g28.0<1>UW g98.1<16,16,1>UB g98.2<16,16,1>UB {align1};
95add (16) g29.0<1>UW g99.1<16,16,1>UB g99.2<16,16,1>UB {align1};
96add (16) g30.0<1>UW g100.1<16,16,1>UB g100.2<16,16,1>UB {align1};
97add (16) g31.0<1>UW g101.1<16,16,1>UB g101.2<16,16,1>UB {align1};
98add (16) g32.0<1>UW g102.1<16,16,1>UB g102.2<16,16,1>UB {align1};
99add (16) g33.0<1>UW g103.1<16,16,1>UB g103.2<16,16,1>UB {align1};
100add (16) g34.0<1>UW g104.1<16,16,1>UB g104.2<16,16,1>UB {align1};
101add (16) g35.0<1>UW g105.1<16,16,1>UB g105.2<16,16,1>UB {align1};
102add (16) g36.0<1>UW g106.1<16,16,1>UB g106.2<16,16,1>UB {align1};
103add (16) g37.0<1>UW g107.1<16,16,1>UB g107.2<16,16,1>UB {align1};
104add (16) g38.0<1>UW g108.1<16,16,1>UB g108.2<16,16,1>UB {align1};
105add (16) g39.0<1>UW g109.1<16,16,1>UB g109.2<16,16,1>UB {align1};
106add (16) g40.0<1>UW g110.1<16,16,1>UB g110.2<16,16,1>UB {align1};
107add (16) g41.0<1>UW g111.1<16,16,1>UB g111.2<16,16,1>UB {align1};
108add (16) g42.0<1>UW g112.1<16,16,1>UB g112.2<16,16,1>UB {align1};
109add (16) g43.0<1>UW g113.1<16,16,1>UB g113.2<16,16,1>UB {align1};
110
111add (16) g28.0<1>UW g28.0<16,16,1>UW g99.1<16,16,1>UB {align1};
112add (16) g29.0<1>UW g29.0<16,16,1>UW g100.1<16,16,1>UB {align1};
113add (16) g30.0<1>UW g30.0<16,16,1>UW g101.1<16,16,1>UB {align1};
114add (16) g31.0<1>UW g31.0<16,16,1>UW g102.1<16,16,1>UB {align1};
115add (16) g32.0<1>UW g32.0<16,16,1>UW g103.1<16,16,1>UB {align1};
116add (16) g33.0<1>UW g33.0<16,16,1>UW g104.1<16,16,1>UB {align1};
117add (16) g34.0<1>UW g34.0<16,16,1>UW g105.1<16,16,1>UB {align1};
118add (16) g35.0<1>UW g35.0<16,16,1>UW g106.1<16,16,1>UB {align1};
119add (16) g36.0<1>UW g36.0<16,16,1>UW g107.1<16,16,1>UB {align1};
120add (16) g37.0<1>UW g37.0<16,16,1>UW g108.1<16,16,1>UB {align1};
121add (16) g38.0<1>UW g38.0<16,16,1>UW g109.1<16,16,1>UB {align1};
122add (16) g39.0<1>UW g39.0<16,16,1>UW g110.1<16,16,1>UB {align1};
123add (16) g40.0<1>UW g40.0<16,16,1>UW g111.1<16,16,1>UB {align1};
124add (16) g41.0<1>UW g41.0<16,16,1>UW g112.1<16,16,1>UB {align1};
125add (16) g42.0<1>UW g42.0<16,16,1>UW g113.1<16,16,1>UB {align1};
126add (16) g43.0<1>UW g43.0<16,16,1>UW g120.1<16,16,1>UB {align1};
127
128add (16) g28.0<1>UW g28.0<16,16,1>UW g99.2<16,16,1>UB {align1};
129add (16) g29.0<1>UW g29.0<16,16,1>UW g100.2<16,16,1>UB {align1};
130add (16) g30.0<1>UW g30.0<16,16,1>UW g101.2<16,16,1>UB {align1};
131add (16) g31.0<1>UW g31.0<16,16,1>UW g102.2<16,16,1>UB {align1};
132add (16) g32.0<1>UW g32.0<16,16,1>UW g103.2<16,16,1>UB {align1};
133add (16) g33.0<1>UW g33.0<16,16,1>UW g104.2<16,16,1>UB {align1};
134add (16) g34.0<1>UW g34.0<16,16,1>UW g105.2<16,16,1>UB {align1};
135add (16) g35.0<1>UW g35.0<16,16,1>UW g106.2<16,16,1>UB {align1};
136add (16) g36.0<1>UW g36.0<16,16,1>UW g107.2<16,16,1>UB {align1};
137add (16) g37.0<1>UW g37.0<16,16,1>UW g108.2<16,16,1>UB {align1};
138add (16) g38.0<1>UW g38.0<16,16,1>UW g109.2<16,16,1>UB {align1};
139add (16) g39.0<1>UW g39.0<16,16,1>UW g110.2<16,16,1>UB {align1};
140add (16) g40.0<1>UW g40.0<16,16,1>UW g111.2<16,16,1>UB {align1};
141add (16) g41.0<1>UW g41.0<16,16,1>UW g112.2<16,16,1>UB {align1};
142add (16) g42.0<1>UW g42.0<16,16,1>UW g113.2<16,16,1>UB {align1};
143add (16) g43.0<1>UW g43.0<16,16,1>UW g120.2<16,16,1>UB {align1};
144jmpi out;
145
146add (16) g28.0<1>UW g98.2<16,16,1>UB g98.3<16,16,1>UB {align1};
147add (16) g29.0<1>UW g99.2<16,16,1>UB g99.3<16,16,1>UB {align1};
148add (16) g30.0<1>UW g100.2<16,16,1>UB g100.3<16,16,1>UB {align1};
149add (16) g31.0<1>UW g101.2<16,16,1>UB g101.3<16,16,1>UB {align1};
150add (16) g32.0<1>UW g102.2<16,16,1>UB g102.3<16,16,1>UB {align1};
151add (16) g33.0<1>UW g103.2<16,16,1>UB g103.3<16,16,1>UB {align1};
152add (16) g34.0<1>UW g104.2<16,16,1>UB g104.3<16,16,1>UB {align1};
153add (16) g35.0<1>UW g105.2<16,16,1>UB g105.3<16,16,1>UB {align1};
154add (16) g36.0<1>UW g106.2<16,16,1>UB g106.3<16,16,1>UB {align1};
155add (16) g37.0<1>UW g107.2<16,16,1>UB g107.3<16,16,1>UB {align1};
156add (16) g38.0<1>UW g108.2<16,16,1>UB g108.3<16,16,1>UB {align1};
157add (16) g39.0<1>UW g109.2<16,16,1>UB g109.3<16,16,1>UB {align1};
158add (16) g40.0<1>UW g110.2<16,16,1>UB g110.3<16,16,1>UB {align1};
159add (16) g41.0<1>UW g111.2<16,16,1>UB g111.3<16,16,1>UB {align1};
160add (16) g42.0<1>UW g112.2<16,16,1>UB g112.3<16,16,1>UB {align1};
161add (16) g43.0<1>UW g113.2<16,16,1>UB g113.3<16,16,1>UB {align1};
162
163add (16) g28.0<1>UW g28.0<16,16,1>UW g99.2<16,16,1>UB {align1};
164add (16) g29.0<1>UW g29.0<16,16,1>UW g100.2<16,16,1>UB {align1};
165add (16) g30.0<1>UW g30.0<16,16,1>UW g101.2<16,16,1>UB {align1};
166add (16) g31.0<1>UW g31.0<16,16,1>UW g102.2<16,16,1>UB {align1};
167add (16) g32.0<1>UW g32.0<16,16,1>UW g103.2<16,16,1>UB {align1};
168add (16) g33.0<1>UW g33.0<16,16,1>UW g104.2<16,16,1>UB {align1};
169add (16) g34.0<1>UW g34.0<16,16,1>UW g105.2<16,16,1>UB {align1};
170add (16) g35.0<1>UW g35.0<16,16,1>UW g106.2<16,16,1>UB {align1};
171add (16) g36.0<1>UW g36.0<16,16,1>UW g107.2<16,16,1>UB {align1};
172add (16) g37.0<1>UW g37.0<16,16,1>UW g108.2<16,16,1>UB {align1};
173add (16) g38.0<1>UW g38.0<16,16,1>UW g109.2<16,16,1>UB {align1};
174add (16) g39.0<1>UW g39.0<16,16,1>UW g110.2<16,16,1>UB {align1};
175add (16) g40.0<1>UW g40.0<16,16,1>UW g111.2<16,16,1>UB {align1};
176add (16) g41.0<1>UW g41.0<16,16,1>UW g112.2<16,16,1>UB {align1};
177add (16) g42.0<1>UW g42.0<16,16,1>UW g113.2<16,16,1>UB {align1};
178add (16) g43.0<1>UW g43.0<16,16,1>UW g120.2<16,16,1>UB {align1};
179
180add (16) g28.0<1>UW g28.0<16,16,1>UW g99.3<16,16,1>UB {align1};
181add (16) g29.0<1>UW g29.0<16,16,1>UW g100.3<16,16,1>UB {align1};
182add (16) g30.0<1>UW g30.0<16,16,1>UW g101.3<16,16,1>UB {align1};
183add (16) g31.0<1>UW g31.0<16,16,1>UW g102.3<16,16,1>UB {align1};
184add (16) g32.0<1>UW g32.0<16,16,1>UW g103.3<16,16,1>UB {align1};
185add (16) g33.0<1>UW g33.0<16,16,1>UW g104.3<16,16,1>UB {align1};
186add (16) g34.0<1>UW g34.0<16,16,1>UW g105.3<16,16,1>UB {align1};
187add (16) g35.0<1>UW g35.0<16,16,1>UW g106.3<16,16,1>UB {align1};
188add (16) g36.0<1>UW g36.0<16,16,1>UW g107.3<16,16,1>UB {align1};
189add (16) g37.0<1>UW g37.0<16,16,1>UW g108.3<16,16,1>UB {align1};
190add (16) g38.0<1>UW g38.0<16,16,1>UW g109.3<16,16,1>UB {align1};
191add (16) g39.0<1>UW g39.0<16,16,1>UW g110.3<16,16,1>UB {align1};
192add (16) g40.0<1>UW g40.0<16,16,1>UW g111.3<16,16,1>UB {align1};
193add (16) g41.0<1>UW g41.0<16,16,1>UW g112.3<16,16,1>UB {align1};
194add (16) g42.0<1>UW g42.0<16,16,1>UW g113.3<16,16,1>UB {align1};
195add (16) g43.0<1>UW g43.0<16,16,1>UW g120.3<16,16,1>UB {align1};
196jmpi out;
197add (16) g28.0<1>UW g98.3<16,16,1>UB g98.4<16,16,1>UB {align1};
198add (16) g29.0<1>UW g99.3<16,16,1>UB g99.4<16,16,1>UB {align1};
199add (16) g30.0<1>UW g100.3<16,16,1>UB g100.4<16,16,1>UB {align1};
200add (16) g31.0<1>UW g101.3<16,16,1>UB g101.4<16,16,1>UB {align1};
201add (16) g32.0<1>UW g102.3<16,16,1>UB g102.4<16,16,1>UB {align1};
202add (16) g33.0<1>UW g103.3<16,16,1>UB g103.4<16,16,1>UB {align1};
203add (16) g34.0<1>UW g104.3<16,16,1>UB g104.4<16,16,1>UB {align1};
204add (16) g35.0<1>UW g105.3<16,16,1>UB g105.4<16,16,1>UB {align1};
205add (16) g36.0<1>UW g106.3<16,16,1>UB g106.4<16,16,1>UB {align1};
206add (16) g37.0<1>UW g107.3<16,16,1>UB g107.4<16,16,1>UB {align1};
207add (16) g38.0<1>UW g108.3<16,16,1>UB g108.4<16,16,1>UB {align1};
208add (16) g39.0<1>UW g109.3<16,16,1>UB g109.4<16,16,1>UB {align1};
209add (16) g40.0<1>UW g110.3<16,16,1>UB g110.4<16,16,1>UB {align1};
210add (16) g41.0<1>UW g111.3<16,16,1>UB g111.4<16,16,1>UB {align1};
211add (16) g42.0<1>UW g112.3<16,16,1>UB g112.4<16,16,1>UB {align1};
212add (16) g43.0<1>UW g113.3<16,16,1>UB g113.4<16,16,1>UB {align1};
213
214add (16) g28.0<1>UW g28.0<16,16,1>UW g99.3<16,16,1>UB {align1};
215add (16) g29.0<1>UW g29.0<16,16,1>UW g100.3<16,16,1>UB {align1};
216add (16) g30.0<1>UW g30.0<16,16,1>UW g101.3<16,16,1>UB {align1};
217add (16) g31.0<1>UW g31.0<16,16,1>UW g102.3<16,16,1>UB {align1};
218add (16) g32.0<1>UW g32.0<16,16,1>UW g103.3<16,16,1>UB {align1};
219add (16) g33.0<1>UW g33.0<16,16,1>UW g104.3<16,16,1>UB {align1};
220add (16) g34.0<1>UW g34.0<16,16,1>UW g105.3<16,16,1>UB {align1};
221add (16) g35.0<1>UW g35.0<16,16,1>UW g106.3<16,16,1>UB {align1};
222add (16) g36.0<1>UW g36.0<16,16,1>UW g107.3<16,16,1>UB {align1};
223add (16) g37.0<1>UW g37.0<16,16,1>UW g108.3<16,16,1>UB {align1};
224add (16) g38.0<1>UW g38.0<16,16,1>UW g109.3<16,16,1>UB {align1};
225add (16) g39.0<1>UW g39.0<16,16,1>UW g110.3<16,16,1>UB {align1};
226add (16) g40.0<1>UW g40.0<16,16,1>UW g111.3<16,16,1>UB {align1};
227add (16) g41.0<1>UW g41.0<16,16,1>UW g112.3<16,16,1>UB {align1};
228add (16) g42.0<1>UW g42.0<16,16,1>UW g113.3<16,16,1>UB {align1};
229add (16) g43.0<1>UW g43.0<16,16,1>UW g120.3<16,16,1>UB {align1};
230
231add (16) g28.0<1>UW g28.0<16,16,1>UW g99.4<16,16,1>UB {align1};
232add (16) g29.0<1>UW g29.0<16,16,1>UW g100.4<16,16,1>UB {align1};
233add (16) g30.0<1>UW g30.0<16,16,1>UW g101.4<16,16,1>UB {align1};
234add (16) g31.0<1>UW g31.0<16,16,1>UW g102.4<16,16,1>UB {align1};
235add (16) g32.0<1>UW g32.0<16,16,1>UW g103.4<16,16,1>UB {align1};
236add (16) g33.0<1>UW g33.0<16,16,1>UW g104.4<16,16,1>UB {align1};
237add (16) g34.0<1>UW g34.0<16,16,1>UW g105.4<16,16,1>UB {align1};
238add (16) g35.0<1>UW g35.0<16,16,1>UW g106.4<16,16,1>UB {align1};
239add (16) g36.0<1>UW g36.0<16,16,1>UW g107.4<16,16,1>UB {align1};
240add (16) g37.0<1>UW g37.0<16,16,1>UW g108.4<16,16,1>UB {align1};
241add (16) g38.0<1>UW g38.0<16,16,1>UW g109.4<16,16,1>UB {align1};
242add (16) g39.0<1>UW g39.0<16,16,1>UW g110.4<16,16,1>UB {align1};
243add (16) g40.0<1>UW g40.0<16,16,1>UW g111.4<16,16,1>UB {align1};
244add (16) g41.0<1>UW g41.0<16,16,1>UW g112.4<16,16,1>UB {align1};
245add (16) g42.0<1>UW g42.0<16,16,1>UW g113.4<16,16,1>UB {align1};
246add (16) g43.0<1>UW g43.0<16,16,1>UW g120.4<16,16,1>UB {align1};
247
248out:
249shr.sat (16) g28.0<1>UW g28.0<16,16,1>UW 2UW {align1};
250shr.sat (16) g29.0<1>UW g29.0<16,16,1>UW 2UW {align1};
251shr.sat (16) g30.0<1>UW g30.0<16,16,1>UW 2UW {align1};
252shr.sat (16) g31.0<1>UW g31.0<16,16,1>UW 2UW {align1};
253shr.sat (16) g32.0<1>UW g32.0<16,16,1>UW 2UW {align1};
254shr.sat (16) g33.0<1>UW g33.0<16,16,1>UW 2UW {align1};
255shr.sat (16) g34.0<1>UW g34.0<16,16,1>UW 2UW {align1};
256shr.sat (16) g35.0<1>UW g35.0<16,16,1>UW 2UW {align1};
257shr.sat (16) g36.0<1>UW g36.0<16,16,1>UW 2UW {align1};
258shr.sat (16) g37.0<1>UW g37.0<16,16,1>UW 2UW {align1};
259shr.sat (16) g38.0<1>UW g38.0<16,16,1>UW 2UW {align1};
260shr.sat (16) g39.0<1>UW g39.0<16,16,1>UW 2UW {align1};
261shr.sat (16) g40.0<1>UW g40.0<16,16,1>UW 2UW {align1};
262shr.sat (16) g41.0<1>UW g41.0<16,16,1>UW 2UW {align1};
263shr.sat (16) g42.0<1>UW g42.0<16,16,1>UW 2UW {align1};
264shr.sat (16) g43.0<1>UW g43.0<16,16,1>UW 2UW {align1};
265