1 // Code generated by avx512test. DO NOT EDIT.
2
3 #include "../../../../../../runtime/textflag.h"
4
5 TEXT asmtest_avx512_vnni(SB), NOSPLIT, $0
6 VPDPBUSD X15, X16, K2, X6 // 62d27d0250f7
7 VPDPBUSD X11, X16, K2, X6 // 62d27d0250f3
8 VPDPBUSD X1, X16, K2, X6 // 62f27d0250f1
9 VPDPBUSD -15(R14)(R15*1), X16, K2, X6 // 62927d0250b43ef1ffffff
10 VPDPBUSD -15(BX), X16, K2, X6 // 62f27d0250b3f1ffffff
11 VPDPBUSD X15, X28, K2, X6 // 62d21d0250f7
12 VPDPBUSD X11, X28, K2, X6 // 62d21d0250f3
13 VPDPBUSD X1, X28, K2, X6 // 62f21d0250f1
14 VPDPBUSD -15(R14)(R15*1), X28, K2, X6 // 62921d0250b43ef1ffffff
15 VPDPBUSD -15(BX), X28, K2, X6 // 62f21d0250b3f1ffffff
16 VPDPBUSD X15, X8, K2, X6 // 62d23d0a50f7
17 VPDPBUSD X11, X8, K2, X6 // 62d23d0a50f3
18 VPDPBUSD X1, X8, K2, X6 // 62f23d0a50f1
19 VPDPBUSD -15(R14)(R15*1), X8, K2, X6 // 62923d0a50b43ef1ffffff
20 VPDPBUSD -15(BX), X8, K2, X6 // 62f23d0a50b3f1ffffff
21 VPDPBUSD X15, X16, K2, X22 // 62c27d0250f7
22 VPDPBUSD X11, X16, K2, X22 // 62c27d0250f3
23 VPDPBUSD X1, X16, K2, X22 // 62e27d0250f1
24 VPDPBUSD -15(R14)(R15*1), X16, K2, X22 // 62827d0250b43ef1ffffff
25 VPDPBUSD -15(BX), X16, K2, X22 // 62e27d0250b3f1ffffff
26 VPDPBUSD X15, X28, K2, X22 // 62c21d0250f7
27 VPDPBUSD X11, X28, K2, X22 // 62c21d0250f3
28 VPDPBUSD X1, X28, K2, X22 // 62e21d0250f1
29 VPDPBUSD -15(R14)(R15*1), X28, K2, X22 // 62821d0250b43ef1ffffff
30 VPDPBUSD -15(BX), X28, K2, X22 // 62e21d0250b3f1ffffff
31 VPDPBUSD X15, X8, K2, X22 // 62c23d0a50f7
32 VPDPBUSD X11, X8, K2, X22 // 62c23d0a50f3
33 VPDPBUSD X1, X8, K2, X22 // 62e23d0a50f1
34 VPDPBUSD -15(R14)(R15*1), X8, K2, X22 // 62823d0a50b43ef1ffffff
35 VPDPBUSD -15(BX), X8, K2, X22 // 62e23d0a50b3f1ffffff
36 VPDPBUSD X15, X16, K2, X12 // 62527d0250e7
37 VPDPBUSD X11, X16, K2, X12 // 62527d0250e3
38 VPDPBUSD X1, X16, K2, X12 // 62727d0250e1
39 VPDPBUSD -15(R14)(R15*1), X16, K2, X12 // 62127d0250a43ef1ffffff
40 VPDPBUSD -15(BX), X16, K2, X12 // 62727d0250a3f1ffffff
41 VPDPBUSD X15, X28, K2, X12 // 62521d0250e7
42 VPDPBUSD X11, X28, K2, X12 // 62521d0250e3
43 VPDPBUSD X1, X28, K2, X12 // 62721d0250e1
44 VPDPBUSD -15(R14)(R15*1), X28, K2, X12 // 62121d0250a43ef1ffffff
45 VPDPBUSD -15(BX), X28, K2, X12 // 62721d0250a3f1ffffff
46 VPDPBUSD X15, X8, K2, X12 // 62523d0a50e7
47 VPDPBUSD X11, X8, K2, X12 // 62523d0a50e3
48 VPDPBUSD X1, X8, K2, X12 // 62723d0a50e1
49 VPDPBUSD -15(R14)(R15*1), X8, K2, X12 // 62123d0a50a43ef1ffffff
50 VPDPBUSD -15(BX), X8, K2, X12 // 62723d0a50a3f1ffffff
51 VPDPBUSD Y11, Y28, K5, Y20 // 62c21d2550e3
52 VPDPBUSD Y27, Y28, K5, Y20 // 62821d2550e3
53 VPDPBUSD Y17, Y28, K5, Y20 // 62a21d2550e1
54 VPDPBUSD (AX), Y28, K5, Y20 // 62e21d255020
55 VPDPBUSD 7(SI), Y28, K5, Y20 // 62e21d2550a607000000
56 VPDPBUSD Y11, Y1, K5, Y20 // 62c2752d50e3
57 VPDPBUSD Y27, Y1, K5, Y20 // 6282752d50e3
58 VPDPBUSD Y17, Y1, K5, Y20 // 62a2752d50e1
59 VPDPBUSD (AX), Y1, K5, Y20 // 62e2752d5020
60 VPDPBUSD 7(SI), Y1, K5, Y20 // 62e2752d50a607000000
61 VPDPBUSD Y11, Y8, K5, Y20 // 62c23d2d50e3
62 VPDPBUSD Y27, Y8, K5, Y20 // 62823d2d50e3
63 VPDPBUSD Y17, Y8, K5, Y20 // 62a23d2d50e1
64 VPDPBUSD (AX), Y8, K5, Y20 // 62e23d2d5020
65 VPDPBUSD 7(SI), Y8, K5, Y20 // 62e23d2d50a607000000
66 VPDPBUSD Y11, Y28, K5, Y9 // 62521d2550cb
67 VPDPBUSD Y27, Y28, K5, Y9 // 62121d2550cb
68 VPDPBUSD Y17, Y28, K5, Y9 // 62321d2550c9
69 VPDPBUSD (AX), Y28, K5, Y9 // 62721d255008
70 VPDPBUSD 7(SI), Y28, K5, Y9 // 62721d25508e07000000
71 VPDPBUSD Y11, Y1, K5, Y9 // 6252752d50cb
72 VPDPBUSD Y27, Y1, K5, Y9 // 6212752d50cb
73 VPDPBUSD Y17, Y1, K5, Y9 // 6232752d50c9
74 VPDPBUSD (AX), Y1, K5, Y9 // 6272752d5008
75 VPDPBUSD 7(SI), Y1, K5, Y9 // 6272752d508e07000000
76 VPDPBUSD Y11, Y8, K5, Y9 // 62523d2d50cb
77 VPDPBUSD Y27, Y8, K5, Y9 // 62123d2d50cb
78 VPDPBUSD Y17, Y8, K5, Y9 // 62323d2d50c9
79 VPDPBUSD (AX), Y8, K5, Y9 // 62723d2d5008
80 VPDPBUSD 7(SI), Y8, K5, Y9 // 62723d2d508e07000000
81 VPDPBUSD Y11, Y28, K5, Y28 // 62421d2550e3
82 VPDPBUSD Y27, Y28, K5, Y28 // 62021d2550e3
83 VPDPBUSD Y17, Y28, K5, Y28 // 62221d2550e1
84 VPDPBUSD (AX), Y28, K5, Y28 // 62621d255020
85 VPDPBUSD 7(SI), Y28, K5, Y28 // 62621d2550a607000000
86 VPDPBUSD Y11, Y1, K5, Y28 // 6242752d50e3
87 VPDPBUSD Y27, Y1, K5, Y28 // 6202752d50e3
88 VPDPBUSD Y17, Y1, K5, Y28 // 6222752d50e1
89 VPDPBUSD (AX), Y1, K5, Y28 // 6262752d5020
90 VPDPBUSD 7(SI), Y1, K5, Y28 // 6262752d50a607000000
91 VPDPBUSD Y11, Y8, K5, Y28 // 62423d2d50e3
92 VPDPBUSD Y27, Y8, K5, Y28 // 62023d2d50e3
93 VPDPBUSD Y17, Y8, K5, Y28 // 62223d2d50e1
94 VPDPBUSD (AX), Y8, K5, Y28 // 62623d2d5020
95 VPDPBUSD 7(SI), Y8, K5, Y28 // 62623d2d50a607000000
96 VPDPBUSD Z8, Z23, K3, Z23 // 62c2454350f8
97 VPDPBUSD Z28, Z23, K3, Z23 // 6282454350fc
98 VPDPBUSD (SI), Z23, K3, Z23 // 62e24543503e
99 VPDPBUSD 7(SI)(DI*2), Z23, K3, Z23 // 62e2454350bc7e07000000
100 VPDPBUSD Z8, Z6, K3, Z23 // 62c24d4b50f8
101 VPDPBUSD Z28, Z6, K3, Z23 // 62824d4b50fc
102 VPDPBUSD (SI), Z6, K3, Z23 // 62e24d4b503e
103 VPDPBUSD 7(SI)(DI*2), Z6, K3, Z23 // 62e24d4b50bc7e07000000
104 VPDPBUSD Z8, Z23, K3, Z5 // 62d2454350e8
105 VPDPBUSD Z28, Z23, K3, Z5 // 6292454350ec
106 VPDPBUSD (SI), Z23, K3, Z5 // 62f24543502e
107 VPDPBUSD 7(SI)(DI*2), Z23, K3, Z5 // 62f2454350ac7e07000000
108 VPDPBUSD Z8, Z6, K3, Z5 // 62d24d4b50e8
109 VPDPBUSD Z28, Z6, K3, Z5 // 62924d4b50ec
110 VPDPBUSD (SI), Z6, K3, Z5 // 62f24d4b502e
111 VPDPBUSD 7(SI)(DI*2), Z6, K3, Z5 // 62f24d4b50ac7e07000000
112 VPDPBUSDS X25, X14, K4, X19 // 62820d0c51d9
113 VPDPBUSDS X11, X14, K4, X19 // 62c20d0c51db
114 VPDPBUSDS X17, X14, K4, X19 // 62a20d0c51d9
115 VPDPBUSDS 7(AX)(CX*4), X14, K4, X19 // 62e20d0c519c8807000000
116 VPDPBUSDS 7(AX)(CX*1), X14, K4, X19 // 62e20d0c519c0807000000
117 VPDPBUSDS X25, X0, K4, X19 // 62827d0c51d9
118 VPDPBUSDS X11, X0, K4, X19 // 62c27d0c51db
119 VPDPBUSDS X17, X0, K4, X19 // 62a27d0c51d9
120 VPDPBUSDS 7(AX)(CX*4), X0, K4, X19 // 62e27d0c519c8807000000
121 VPDPBUSDS 7(AX)(CX*1), X0, K4, X19 // 62e27d0c519c0807000000
122 VPDPBUSDS X25, X14, K4, X13 // 62120d0c51e9
123 VPDPBUSDS X11, X14, K4, X13 // 62520d0c51eb
124 VPDPBUSDS X17, X14, K4, X13 // 62320d0c51e9
125 VPDPBUSDS 7(AX)(CX*4), X14, K4, X13 // 62720d0c51ac8807000000
126 VPDPBUSDS 7(AX)(CX*1), X14, K4, X13 // 62720d0c51ac0807000000
127 VPDPBUSDS X25, X0, K4, X13 // 62127d0c51e9
128 VPDPBUSDS X11, X0, K4, X13 // 62527d0c51eb
129 VPDPBUSDS X17, X0, K4, X13 // 62327d0c51e9
130 VPDPBUSDS 7(AX)(CX*4), X0, K4, X13 // 62727d0c51ac8807000000
131 VPDPBUSDS 7(AX)(CX*1), X0, K4, X13 // 62727d0c51ac0807000000
132 VPDPBUSDS X25, X14, K4, X2 // 62920d0c51d1
133 VPDPBUSDS X11, X14, K4, X2 // 62d20d0c51d3
134 VPDPBUSDS X17, X14, K4, X2 // 62b20d0c51d1
135 VPDPBUSDS 7(AX)(CX*4), X14, K4, X2 // 62f20d0c51948807000000
136 VPDPBUSDS 7(AX)(CX*1), X14, K4, X2 // 62f20d0c51940807000000
137 VPDPBUSDS X25, X0, K4, X2 // 62927d0c51d1
138 VPDPBUSDS X11, X0, K4, X2 // 62d27d0c51d3
139 VPDPBUSDS X17, X0, K4, X2 // 62b27d0c51d1
140 VPDPBUSDS 7(AX)(CX*4), X0, K4, X2 // 62f27d0c51948807000000
141 VPDPBUSDS 7(AX)(CX*1), X0, K4, X2 // 62f27d0c51940807000000
142 VPDPBUSDS Y28, Y26, K2, Y16 // 62822d2251c4
143 VPDPBUSDS Y1, Y26, K2, Y16 // 62e22d2251c1
144 VPDPBUSDS Y23, Y26, K2, Y16 // 62a22d2251c7
145 VPDPBUSDS (BX), Y26, K2, Y16 // 62e22d225103
146 VPDPBUSDS -17(BP)(SI*1), Y26, K2, Y16 // 62e22d22518435efffffff
147 VPDPBUSDS Y28, Y3, K2, Y16 // 6282652a51c4
148 VPDPBUSDS Y1, Y3, K2, Y16 // 62e2652a51c1
149 VPDPBUSDS Y23, Y3, K2, Y16 // 62a2652a51c7
150 VPDPBUSDS (BX), Y3, K2, Y16 // 62e2652a5103
151 VPDPBUSDS -17(BP)(SI*1), Y3, K2, Y16 // 62e2652a518435efffffff
152 VPDPBUSDS Y28, Y8, K2, Y16 // 62823d2a51c4
153 VPDPBUSDS Y1, Y8, K2, Y16 // 62e23d2a51c1
154 VPDPBUSDS Y23, Y8, K2, Y16 // 62a23d2a51c7
155 VPDPBUSDS (BX), Y8, K2, Y16 // 62e23d2a5103
156 VPDPBUSDS -17(BP)(SI*1), Y8, K2, Y16 // 62e23d2a518435efffffff
157 VPDPBUSDS Y28, Y26, K2, Y12 // 62122d2251e4
158 VPDPBUSDS Y1, Y26, K2, Y12 // 62722d2251e1
159 VPDPBUSDS Y23, Y26, K2, Y12 // 62322d2251e7
160 VPDPBUSDS (BX), Y26, K2, Y12 // 62722d225123
161 VPDPBUSDS -17(BP)(SI*1), Y26, K2, Y12 // 62722d2251a435efffffff
162 VPDPBUSDS Y28, Y3, K2, Y12 // 6212652a51e4
163 VPDPBUSDS Y1, Y3, K2, Y12 // 6272652a51e1
164 VPDPBUSDS Y23, Y3, K2, Y12 // 6232652a51e7
165 VPDPBUSDS (BX), Y3, K2, Y12 // 6272652a5123
166 VPDPBUSDS -17(BP)(SI*1), Y3, K2, Y12 // 6272652a51a435efffffff
167 VPDPBUSDS Y28, Y8, K2, Y12 // 62123d2a51e4
168 VPDPBUSDS Y1, Y8, K2, Y12 // 62723d2a51e1
169 VPDPBUSDS Y23, Y8, K2, Y12 // 62323d2a51e7
170 VPDPBUSDS (BX), Y8, K2, Y12 // 62723d2a5123
171 VPDPBUSDS -17(BP)(SI*1), Y8, K2, Y12 // 62723d2a51a435efffffff
172 VPDPBUSDS Y28, Y26, K2, Y6 // 62922d2251f4
173 VPDPBUSDS Y1, Y26, K2, Y6 // 62f22d2251f1
174 VPDPBUSDS Y23, Y26, K2, Y6 // 62b22d2251f7
175 VPDPBUSDS (BX), Y26, K2, Y6 // 62f22d225133
176 VPDPBUSDS -17(BP)(SI*1), Y26, K2, Y6 // 62f22d2251b435efffffff
177 VPDPBUSDS Y28, Y3, K2, Y6 // 6292652a51f4
178 VPDPBUSDS Y1, Y3, K2, Y6 // 62f2652a51f1
179 VPDPBUSDS Y23, Y3, K2, Y6 // 62b2652a51f7
180 VPDPBUSDS (BX), Y3, K2, Y6 // 62f2652a5133
181 VPDPBUSDS -17(BP)(SI*1), Y3, K2, Y6 // 62f2652a51b435efffffff
182 VPDPBUSDS Y28, Y8, K2, Y6 // 62923d2a51f4
183 VPDPBUSDS Y1, Y8, K2, Y6 // 62f23d2a51f1
184 VPDPBUSDS Y23, Y8, K2, Y6 // 62b23d2a51f7
185 VPDPBUSDS (BX), Y8, K2, Y6 // 62f23d2a5133
186 VPDPBUSDS -17(BP)(SI*1), Y8, K2, Y6 // 62f23d2a51b435efffffff
187 VPDPBUSDS Z12, Z16, K2, Z21 // 62c27d4251ec
188 VPDPBUSDS Z27, Z16, K2, Z21 // 62827d4251eb
189 VPDPBUSDS 17(SP)(BP*8), Z16, K2, Z21 // 62e27d4251acec11000000
190 VPDPBUSDS 17(SP)(BP*4), Z16, K2, Z21 // 62e27d4251acac11000000
191 VPDPBUSDS Z12, Z13, K2, Z21 // 62c2154a51ec
192 VPDPBUSDS Z27, Z13, K2, Z21 // 6282154a51eb
193 VPDPBUSDS 17(SP)(BP*8), Z13, K2, Z21 // 62e2154a51acec11000000
194 VPDPBUSDS 17(SP)(BP*4), Z13, K2, Z21 // 62e2154a51acac11000000
195 VPDPBUSDS Z12, Z16, K2, Z5 // 62d27d4251ec
196 VPDPBUSDS Z27, Z16, K2, Z5 // 62927d4251eb
197 VPDPBUSDS 17(SP)(BP*8), Z16, K2, Z5 // 62f27d4251acec11000000
198 VPDPBUSDS 17(SP)(BP*4), Z16, K2, Z5 // 62f27d4251acac11000000
199 VPDPBUSDS Z12, Z13, K2, Z5 // 62d2154a51ec
200 VPDPBUSDS Z27, Z13, K2, Z5 // 6292154a51eb
201 VPDPBUSDS 17(SP)(BP*8), Z13, K2, Z5 // 62f2154a51acec11000000
202 VPDPBUSDS 17(SP)(BP*4), Z13, K2, Z5 // 62f2154a51acac11000000
203 VPDPWSSD X2, X2, K3, X18 // 62e26d0b52d2
204 VPDPWSSD X27, X2, K3, X18 // 62826d0b52d3
205 VPDPWSSD X26, X2, K3, X18 // 62826d0b52d2
206 VPDPWSSD (SI), X2, K3, X18 // 62e26d0b5216
207 VPDPWSSD 7(SI)(DI*2), X2, K3, X18 // 62e26d0b52947e07000000
208 VPDPWSSD X2, X24, K3, X18 // 62e23d0352d2
209 VPDPWSSD X27, X24, K3, X18 // 62823d0352d3
210 VPDPWSSD X26, X24, K3, X18 // 62823d0352d2
211 VPDPWSSD (SI), X24, K3, X18 // 62e23d035216
212 VPDPWSSD 7(SI)(DI*2), X24, K3, X18 // 62e23d0352947e07000000
213 VPDPWSSD X2, X2, K3, X11 // 62726d0b52da
214 VPDPWSSD X27, X2, K3, X11 // 62126d0b52db
215 VPDPWSSD X26, X2, K3, X11 // 62126d0b52da
216 VPDPWSSD (SI), X2, K3, X11 // 62726d0b521e
217 VPDPWSSD 7(SI)(DI*2), X2, K3, X11 // 62726d0b529c7e07000000
218 VPDPWSSD X2, X24, K3, X11 // 62723d0352da
219 VPDPWSSD X27, X24, K3, X11 // 62123d0352db
220 VPDPWSSD X26, X24, K3, X11 // 62123d0352da
221 VPDPWSSD (SI), X24, K3, X11 // 62723d03521e
222 VPDPWSSD 7(SI)(DI*2), X24, K3, X11 // 62723d03529c7e07000000
223 VPDPWSSD X2, X2, K3, X9 // 62726d0b52ca
224 VPDPWSSD X27, X2, K3, X9 // 62126d0b52cb
225 VPDPWSSD X26, X2, K3, X9 // 62126d0b52ca
226 VPDPWSSD (SI), X2, K3, X9 // 62726d0b520e
227 VPDPWSSD 7(SI)(DI*2), X2, K3, X9 // 62726d0b528c7e07000000
228 VPDPWSSD X2, X24, K3, X9 // 62723d0352ca
229 VPDPWSSD X27, X24, K3, X9 // 62123d0352cb
230 VPDPWSSD X26, X24, K3, X9 // 62123d0352ca
231 VPDPWSSD (SI), X24, K3, X9 // 62723d03520e
232 VPDPWSSD 7(SI)(DI*2), X24, K3, X9 // 62723d03528c7e07000000
233 VPDPWSSD Y8, Y2, K3, Y14 // 62526d2b52f0
234 VPDPWSSD Y9, Y2, K3, Y14 // 62526d2b52f1
235 VPDPWSSD Y22, Y2, K3, Y14 // 62326d2b52f6
236 VPDPWSSD 15(R8)(R14*4), Y2, K3, Y14 // 62126d2b52b4b00f000000
237 VPDPWSSD -7(CX)(DX*4), Y2, K3, Y14 // 62726d2b52b491f9ffffff
238 VPDPWSSD Y8, Y22, K3, Y14 // 62524d2352f0
239 VPDPWSSD Y9, Y22, K3, Y14 // 62524d2352f1
240 VPDPWSSD Y22, Y22, K3, Y14 // 62324d2352f6
241 VPDPWSSD 15(R8)(R14*4), Y22, K3, Y14 // 62124d2352b4b00f000000
242 VPDPWSSD -7(CX)(DX*4), Y22, K3, Y14 // 62724d2352b491f9ffffff
243 VPDPWSSD Y8, Y27, K3, Y14 // 6252252352f0
244 VPDPWSSD Y9, Y27, K3, Y14 // 6252252352f1
245 VPDPWSSD Y22, Y27, K3, Y14 // 6232252352f6
246 VPDPWSSD 15(R8)(R14*4), Y27, K3, Y14 // 6212252352b4b00f000000
247 VPDPWSSD -7(CX)(DX*4), Y27, K3, Y14 // 6272252352b491f9ffffff
248 VPDPWSSD Y8, Y2, K3, Y31 // 62426d2b52f8
249 VPDPWSSD Y9, Y2, K3, Y31 // 62426d2b52f9
250 VPDPWSSD Y22, Y2, K3, Y31 // 62226d2b52fe
251 VPDPWSSD 15(R8)(R14*4), Y2, K3, Y31 // 62026d2b52bcb00f000000
252 VPDPWSSD -7(CX)(DX*4), Y2, K3, Y31 // 62626d2b52bc91f9ffffff
253 VPDPWSSD Y8, Y22, K3, Y31 // 62424d2352f8
254 VPDPWSSD Y9, Y22, K3, Y31 // 62424d2352f9
255 VPDPWSSD Y22, Y22, K3, Y31 // 62224d2352fe
256 VPDPWSSD 15(R8)(R14*4), Y22, K3, Y31 // 62024d2352bcb00f000000
257 VPDPWSSD -7(CX)(DX*4), Y22, K3, Y31 // 62624d2352bc91f9ffffff
258 VPDPWSSD Y8, Y27, K3, Y31 // 6242252352f8
259 VPDPWSSD Y9, Y27, K3, Y31 // 6242252352f9
260 VPDPWSSD Y22, Y27, K3, Y31 // 6222252352fe
261 VPDPWSSD 15(R8)(R14*4), Y27, K3, Y31 // 6202252352bcb00f000000
262 VPDPWSSD -7(CX)(DX*4), Y27, K3, Y31 // 6262252352bc91f9ffffff
263 VPDPWSSD Y8, Y2, K3, Y25 // 62426d2b52c8
264 VPDPWSSD Y9, Y2, K3, Y25 // 62426d2b52c9
265 VPDPWSSD Y22, Y2, K3, Y25 // 62226d2b52ce
266 VPDPWSSD 15(R8)(R14*4), Y2, K3, Y25 // 62026d2b528cb00f000000
267 VPDPWSSD -7(CX)(DX*4), Y2, K3, Y25 // 62626d2b528c91f9ffffff
268 VPDPWSSD Y8, Y22, K3, Y25 // 62424d2352c8
269 VPDPWSSD Y9, Y22, K3, Y25 // 62424d2352c9
270 VPDPWSSD Y22, Y22, K3, Y25 // 62224d2352ce
271 VPDPWSSD 15(R8)(R14*4), Y22, K3, Y25 // 62024d23528cb00f000000
272 VPDPWSSD -7(CX)(DX*4), Y22, K3, Y25 // 62624d23528c91f9ffffff
273 VPDPWSSD Y8, Y27, K3, Y25 // 6242252352c8
274 VPDPWSSD Y9, Y27, K3, Y25 // 6242252352c9
275 VPDPWSSD Y22, Y27, K3, Y25 // 6222252352ce
276 VPDPWSSD 15(R8)(R14*4), Y27, K3, Y25 // 62022523528cb00f000000
277 VPDPWSSD -7(CX)(DX*4), Y27, K3, Y25 // 62622523528c91f9ffffff
278 VPDPWSSD Z25, Z6, K3, Z22 // 62824d4b52f1
279 VPDPWSSD Z12, Z6, K3, Z22 // 62c24d4b52f4
280 VPDPWSSD 7(SI)(DI*4), Z6, K3, Z22 // 62e24d4b52b4be07000000
281 VPDPWSSD -7(DI)(R8*2), Z6, K3, Z22 // 62a24d4b52b447f9ffffff
282 VPDPWSSD Z25, Z8, K3, Z22 // 62823d4b52f1
283 VPDPWSSD Z12, Z8, K3, Z22 // 62c23d4b52f4
284 VPDPWSSD 7(SI)(DI*4), Z8, K3, Z22 // 62e23d4b52b4be07000000
285 VPDPWSSD -7(DI)(R8*2), Z8, K3, Z22 // 62a23d4b52b447f9ffffff
286 VPDPWSSD Z25, Z6, K3, Z11 // 62124d4b52d9
287 VPDPWSSD Z12, Z6, K3, Z11 // 62524d4b52dc
288 VPDPWSSD 7(SI)(DI*4), Z6, K3, Z11 // 62724d4b529cbe07000000
289 VPDPWSSD -7(DI)(R8*2), Z6, K3, Z11 // 62324d4b529c47f9ffffff
290 VPDPWSSD Z25, Z8, K3, Z11 // 62123d4b52d9
291 VPDPWSSD Z12, Z8, K3, Z11 // 62523d4b52dc
292 VPDPWSSD 7(SI)(DI*4), Z8, K3, Z11 // 62723d4b529cbe07000000
293 VPDPWSSD -7(DI)(R8*2), Z8, K3, Z11 // 62323d4b529c47f9ffffff
294 VPDPWSSDS X13, X11, K2, X22 // 62c2250a53f5
295 VPDPWSSDS X6, X11, K2, X22 // 62e2250a53f6
296 VPDPWSSDS X12, X11, K2, X22 // 62c2250a53f4
297 VPDPWSSDS 17(SP)(BP*8), X11, K2, X22 // 62e2250a53b4ec11000000
298 VPDPWSSDS 17(SP)(BP*4), X11, K2, X22 // 62e2250a53b4ac11000000
299 VPDPWSSDS X13, X15, K2, X22 // 62c2050a53f5
300 VPDPWSSDS X6, X15, K2, X22 // 62e2050a53f6
301 VPDPWSSDS X12, X15, K2, X22 // 62c2050a53f4
302 VPDPWSSDS 17(SP)(BP*8), X15, K2, X22 // 62e2050a53b4ec11000000
303 VPDPWSSDS 17(SP)(BP*4), X15, K2, X22 // 62e2050a53b4ac11000000
304 VPDPWSSDS X13, X30, K2, X22 // 62c20d0253f5
305 VPDPWSSDS X6, X30, K2, X22 // 62e20d0253f6
306 VPDPWSSDS X12, X30, K2, X22 // 62c20d0253f4
307 VPDPWSSDS 17(SP)(BP*8), X30, K2, X22 // 62e20d0253b4ec11000000
308 VPDPWSSDS 17(SP)(BP*4), X30, K2, X22 // 62e20d0253b4ac11000000
309 VPDPWSSDS X13, X11, K2, X30 // 6242250a53f5
310 VPDPWSSDS X6, X11, K2, X30 // 6262250a53f6
311 VPDPWSSDS X12, X11, K2, X30 // 6242250a53f4
312 VPDPWSSDS 17(SP)(BP*8), X11, K2, X30 // 6262250a53b4ec11000000
313 VPDPWSSDS 17(SP)(BP*4), X11, K2, X30 // 6262250a53b4ac11000000
314 VPDPWSSDS X13, X15, K2, X30 // 6242050a53f5
315 VPDPWSSDS X6, X15, K2, X30 // 6262050a53f6
316 VPDPWSSDS X12, X15, K2, X30 // 6242050a53f4
317 VPDPWSSDS 17(SP)(BP*8), X15, K2, X30 // 6262050a53b4ec11000000
318 VPDPWSSDS 17(SP)(BP*4), X15, K2, X30 // 6262050a53b4ac11000000
319 VPDPWSSDS X13, X30, K2, X30 // 62420d0253f5
320 VPDPWSSDS X6, X30, K2, X30 // 62620d0253f6
321 VPDPWSSDS X12, X30, K2, X30 // 62420d0253f4
322 VPDPWSSDS 17(SP)(BP*8), X30, K2, X30 // 62620d0253b4ec11000000
323 VPDPWSSDS 17(SP)(BP*4), X30, K2, X30 // 62620d0253b4ac11000000
324 VPDPWSSDS X13, X11, K2, X3 // 62d2250a53dd
325 VPDPWSSDS X6, X11, K2, X3 // 62f2250a53de
326 VPDPWSSDS X12, X11, K2, X3 // 62d2250a53dc
327 VPDPWSSDS 17(SP)(BP*8), X11, K2, X3 // 62f2250a539cec11000000
328 VPDPWSSDS 17(SP)(BP*4), X11, K2, X3 // 62f2250a539cac11000000
329 VPDPWSSDS X13, X15, K2, X3 // 62d2050a53dd
330 VPDPWSSDS X6, X15, K2, X3 // 62f2050a53de
331 VPDPWSSDS X12, X15, K2, X3 // 62d2050a53dc
332 VPDPWSSDS 17(SP)(BP*8), X15, K2, X3 // 62f2050a539cec11000000
333 VPDPWSSDS 17(SP)(BP*4), X15, K2, X3 // 62f2050a539cac11000000
334 VPDPWSSDS X13, X30, K2, X3 // 62d20d0253dd
335 VPDPWSSDS X6, X30, K2, X3 // 62f20d0253de
336 VPDPWSSDS X12, X30, K2, X3 // 62d20d0253dc
337 VPDPWSSDS 17(SP)(BP*8), X30, K2, X3 // 62f20d02539cec11000000
338 VPDPWSSDS 17(SP)(BP*4), X30, K2, X3 // 62f20d02539cac11000000
339 VPDPWSSDS Y0, Y6, K1, Y9 // 62724d2953c8
340 VPDPWSSDS Y19, Y6, K1, Y9 // 62324d2953cb
341 VPDPWSSDS Y31, Y6, K1, Y9 // 62124d2953cf
342 VPDPWSSDS (R8), Y6, K1, Y9 // 62524d295308
343 VPDPWSSDS 15(DX)(BX*2), Y6, K1, Y9 // 62724d29538c5a0f000000
344 VPDPWSSDS Y0, Y1, K1, Y9 // 6272752953c8
345 VPDPWSSDS Y19, Y1, K1, Y9 // 6232752953cb
346 VPDPWSSDS Y31, Y1, K1, Y9 // 6212752953cf
347 VPDPWSSDS (R8), Y1, K1, Y9 // 625275295308
348 VPDPWSSDS 15(DX)(BX*2), Y1, K1, Y9 // 62727529538c5a0f000000
349 VPDPWSSDS Y0, Y9, K1, Y9 // 6272352953c8
350 VPDPWSSDS Y19, Y9, K1, Y9 // 6232352953cb
351 VPDPWSSDS Y31, Y9, K1, Y9 // 6212352953cf
352 VPDPWSSDS (R8), Y9, K1, Y9 // 625235295308
353 VPDPWSSDS 15(DX)(BX*2), Y9, K1, Y9 // 62723529538c5a0f000000
354 VPDPWSSDS Y0, Y6, K1, Y14 // 62724d2953f0
355 VPDPWSSDS Y19, Y6, K1, Y14 // 62324d2953f3
356 VPDPWSSDS Y31, Y6, K1, Y14 // 62124d2953f7
357 VPDPWSSDS (R8), Y6, K1, Y14 // 62524d295330
358 VPDPWSSDS 15(DX)(BX*2), Y6, K1, Y14 // 62724d2953b45a0f000000
359 VPDPWSSDS Y0, Y1, K1, Y14 // 6272752953f0
360 VPDPWSSDS Y19, Y1, K1, Y14 // 6232752953f3
361 VPDPWSSDS Y31, Y1, K1, Y14 // 6212752953f7
362 VPDPWSSDS (R8), Y1, K1, Y14 // 625275295330
363 VPDPWSSDS 15(DX)(BX*2), Y1, K1, Y14 // 6272752953b45a0f000000
364 VPDPWSSDS Y0, Y9, K1, Y14 // 6272352953f0
365 VPDPWSSDS Y19, Y9, K1, Y14 // 6232352953f3
366 VPDPWSSDS Y31, Y9, K1, Y14 // 6212352953f7
367 VPDPWSSDS (R8), Y9, K1, Y14 // 625235295330
368 VPDPWSSDS 15(DX)(BX*2), Y9, K1, Y14 // 6272352953b45a0f000000
369 VPDPWSSDS Y0, Y6, K1, Y1 // 62f24d2953c8
370 VPDPWSSDS Y19, Y6, K1, Y1 // 62b24d2953cb
371 VPDPWSSDS Y31, Y6, K1, Y1 // 62924d2953cf
372 VPDPWSSDS (R8), Y6, K1, Y1 // 62d24d295308
373 VPDPWSSDS 15(DX)(BX*2), Y6, K1, Y1 // 62f24d29538c5a0f000000
374 VPDPWSSDS Y0, Y1, K1, Y1 // 62f2752953c8
375 VPDPWSSDS Y19, Y1, K1, Y1 // 62b2752953cb
376 VPDPWSSDS Y31, Y1, K1, Y1 // 6292752953cf
377 VPDPWSSDS (R8), Y1, K1, Y1 // 62d275295308
378 VPDPWSSDS 15(DX)(BX*2), Y1, K1, Y1 // 62f27529538c5a0f000000
379 VPDPWSSDS Y0, Y9, K1, Y1 // 62f2352953c8
380 VPDPWSSDS Y19, Y9, K1, Y1 // 62b2352953cb
381 VPDPWSSDS Y31, Y9, K1, Y1 // 6292352953cf
382 VPDPWSSDS (R8), Y9, K1, Y1 // 62d235295308
383 VPDPWSSDS 15(DX)(BX*2), Y9, K1, Y1 // 62f23529538c5a0f000000
384 VPDPWSSDS Z6, Z9, K2, Z12 // 6272354a53e6
385 VPDPWSSDS Z25, Z9, K2, Z12 // 6212354a53e1
386 VPDPWSSDS 17(SP), Z9, K2, Z12 // 6272354a53a42411000000
387 VPDPWSSDS -17(BP)(SI*4), Z9, K2, Z12 // 6272354a53a4b5efffffff
388 VPDPWSSDS Z6, Z12, K2, Z12 // 62721d4a53e6
389 VPDPWSSDS Z25, Z12, K2, Z12 // 62121d4a53e1
390 VPDPWSSDS 17(SP), Z12, K2, Z12 // 62721d4a53a42411000000
391 VPDPWSSDS -17(BP)(SI*4), Z12, K2, Z12 // 62721d4a53a4b5efffffff
392 VPDPWSSDS Z6, Z9, K2, Z17 // 62e2354a53ce
393 VPDPWSSDS Z25, Z9, K2, Z17 // 6282354a53c9
394 VPDPWSSDS 17(SP), Z9, K2, Z17 // 62e2354a538c2411000000
395 VPDPWSSDS -17(BP)(SI*4), Z9, K2, Z17 // 62e2354a538cb5efffffff
396 VPDPWSSDS Z6, Z12, K2, Z17 // 62e21d4a53ce
397 VPDPWSSDS Z25, Z12, K2, Z17 // 62821d4a53c9
398 VPDPWSSDS 17(SP), Z12, K2, Z17 // 62e21d4a538c2411000000
399 VPDPWSSDS -17(BP)(SI*4), Z12, K2, Z17 // 62e21d4a538cb5efffffff
400 RET
401
View as plain text