1*87aa9c9eSJonas Devlieghere #include <cstdint>
2*87aa9c9eSJonas Devlieghere
3*87aa9c9eSJonas Devlieghere struct alignas(64) zmm_t {
4*87aa9c9eSJonas Devlieghere uint64_t a, b, c, d, e, f, g, h;
5*87aa9c9eSJonas Devlieghere };
6*87aa9c9eSJonas Devlieghere
main()7*87aa9c9eSJonas Devlieghere int main() {
8*87aa9c9eSJonas Devlieghere constexpr zmm_t zmm[] = {
9*87aa9c9eSJonas Devlieghere { 0x0706050403020100, 0x0F0E0D0C0B0A0908,
10*87aa9c9eSJonas Devlieghere 0x1716151413121110, 0x1F1E1D1C1B1A1918,
11*87aa9c9eSJonas Devlieghere 0x2726252423222120, 0x2F2E2D2C2B2A2928,
12*87aa9c9eSJonas Devlieghere 0x3736353433323130, 0x3F3E3D3C3B3A3938, },
13*87aa9c9eSJonas Devlieghere { 0x0807060504030201, 0x100F0E0D0C0B0A09,
14*87aa9c9eSJonas Devlieghere 0x1817161514131211, 0x201F1E1D1C1B1A19,
15*87aa9c9eSJonas Devlieghere 0x2827262524232221, 0x302F2E2D2C2B2A29,
16*87aa9c9eSJonas Devlieghere 0x3837363534333231, 0x403F3E3D3C3B3A39, },
17*87aa9c9eSJonas Devlieghere { 0x0908070605040302, 0x11100F0E0D0C0B0A,
18*87aa9c9eSJonas Devlieghere 0x1918171615141312, 0x21201F1E1D1C1B1A,
19*87aa9c9eSJonas Devlieghere 0x2928272625242322, 0x31302F2E2D2C2B2A,
20*87aa9c9eSJonas Devlieghere 0x3938373635343332, 0x41403F3E3D3C3B3A, },
21*87aa9c9eSJonas Devlieghere { 0x0A09080706050403, 0x1211100F0E0D0C0B,
22*87aa9c9eSJonas Devlieghere 0x1A19181716151413, 0x2221201F1E1D1C1B,
23*87aa9c9eSJonas Devlieghere 0x2A29282726252423, 0x3231302F2E2D2C2B,
24*87aa9c9eSJonas Devlieghere 0x3A39383736353433, 0x4241403F3E3D3C3B, },
25*87aa9c9eSJonas Devlieghere { 0x0B0A090807060504, 0x131211100F0E0D0C,
26*87aa9c9eSJonas Devlieghere 0x1B1A191817161514, 0x232221201F1E1D1C,
27*87aa9c9eSJonas Devlieghere 0x2B2A292827262524, 0x333231302F2E2D2C,
28*87aa9c9eSJonas Devlieghere 0x3B3A393837363534, 0x434241403F3E3D3C, },
29*87aa9c9eSJonas Devlieghere { 0x0C0B0A0908070605, 0x14131211100F0E0D,
30*87aa9c9eSJonas Devlieghere 0x1C1B1A1918171615, 0x24232221201F1E1D,
31*87aa9c9eSJonas Devlieghere 0x2C2B2A2928272625, 0x34333231302F2E2D,
32*87aa9c9eSJonas Devlieghere 0x3C3B3A3938373635, 0x44434241403F3E3D, },
33*87aa9c9eSJonas Devlieghere { 0x0D0C0B0A09080706, 0x1514131211100F0E,
34*87aa9c9eSJonas Devlieghere 0x1D1C1B1A19181716, 0x2524232221201F1E,
35*87aa9c9eSJonas Devlieghere 0x2D2C2B2A29282726, 0x3534333231302F2E,
36*87aa9c9eSJonas Devlieghere 0x3D3C3B3A39383736, 0x4544434241403F3E, },
37*87aa9c9eSJonas Devlieghere { 0x0E0D0C0B0A090807, 0x161514131211100F,
38*87aa9c9eSJonas Devlieghere 0x1E1D1C1B1A191817, 0x262524232221201F,
39*87aa9c9eSJonas Devlieghere 0x2E2D2C2B2A292827, 0x363534333231302F,
40*87aa9c9eSJonas Devlieghere 0x3E3D3C3B3A393837, 0x464544434241403F, },
41*87aa9c9eSJonas Devlieghere #if defined(__x86_64__) || defined(_M_X64)
42*87aa9c9eSJonas Devlieghere { 0x0F0E0D0C0B0A0908, 0x1716151413121110,
43*87aa9c9eSJonas Devlieghere 0x1F1E1D1C1B1A1918, 0x2726252423222120,
44*87aa9c9eSJonas Devlieghere 0x2F2E2D2C2B2A2928, 0x3736353433323130,
45*87aa9c9eSJonas Devlieghere 0x3F3E3D3C3B3A3938, 0x4746454443424140, },
46*87aa9c9eSJonas Devlieghere { 0x100F0E0D0C0B0A09, 0x1817161514131211,
47*87aa9c9eSJonas Devlieghere 0x201F1E1D1C1B1A19, 0x2827262524232221,
48*87aa9c9eSJonas Devlieghere 0x302F2E2D2C2B2A29, 0x3837363534333231,
49*87aa9c9eSJonas Devlieghere 0x403F3E3D3C3B3A39, 0x4847464544434241, },
50*87aa9c9eSJonas Devlieghere { 0x11100F0E0D0C0B0A, 0x1918171615141312,
51*87aa9c9eSJonas Devlieghere 0x21201F1E1D1C1B1A, 0x2928272625242322,
52*87aa9c9eSJonas Devlieghere 0x31302F2E2D2C2B2A, 0x3938373635343332,
53*87aa9c9eSJonas Devlieghere 0x41403F3E3D3C3B3A, 0x4948474645444342, },
54*87aa9c9eSJonas Devlieghere { 0x1211100F0E0D0C0B, 0x1A19181716151413,
55*87aa9c9eSJonas Devlieghere 0x2221201F1E1D1C1B, 0x2A29282726252423,
56*87aa9c9eSJonas Devlieghere 0x3231302F2E2D2C2B, 0x3A39383736353433,
57*87aa9c9eSJonas Devlieghere 0x4241403F3E3D3C3B, 0x4A49484746454443, },
58*87aa9c9eSJonas Devlieghere { 0x131211100F0E0D0C, 0x1B1A191817161514,
59*87aa9c9eSJonas Devlieghere 0x232221201F1E1D1C, 0x2B2A292827262524,
60*87aa9c9eSJonas Devlieghere 0x333231302F2E2D2C, 0x3B3A393837363534,
61*87aa9c9eSJonas Devlieghere 0x434241403F3E3D3C, 0x4B4A494847464544, },
62*87aa9c9eSJonas Devlieghere { 0x14131211100F0E0D, 0x1C1B1A1918171615,
63*87aa9c9eSJonas Devlieghere 0x24232221201F1E1D, 0x2C2B2A2928272625,
64*87aa9c9eSJonas Devlieghere 0x34333231302F2E2D, 0x3C3B3A3938373635,
65*87aa9c9eSJonas Devlieghere 0x44434241403F3E3D, 0x4C4B4A4948474645, },
66*87aa9c9eSJonas Devlieghere { 0x1514131211100F0E, 0x1D1C1B1A19181716,
67*87aa9c9eSJonas Devlieghere 0x2524232221201F1E, 0x2D2C2B2A29282726,
68*87aa9c9eSJonas Devlieghere 0x3534333231302F2E, 0x3D3C3B3A39383736,
69*87aa9c9eSJonas Devlieghere 0x4544434241403F3E, 0x4D4C4B4A49484746, },
70*87aa9c9eSJonas Devlieghere { 0x161514131211100F, 0x1E1D1C1B1A191817,
71*87aa9c9eSJonas Devlieghere 0x262524232221201F, 0x2E2D2C2B2A292827,
72*87aa9c9eSJonas Devlieghere 0x363534333231302F, 0x3E3D3C3B3A393837,
73*87aa9c9eSJonas Devlieghere 0x464544434241403F, 0x4E4D4C4B4A494847, },
74*87aa9c9eSJonas Devlieghere { 0x1716151413121110, 0x1F1E1D1C1B1A1918,
75*87aa9c9eSJonas Devlieghere 0x2726252423222120, 0x2F2E2D2C2B2A2928,
76*87aa9c9eSJonas Devlieghere 0x3736353433323130, 0x3F3E3D3C3B3A3938,
77*87aa9c9eSJonas Devlieghere 0x4746454443424140, 0x4F4E4D4C4B4A4948, },
78*87aa9c9eSJonas Devlieghere { 0x1817161514131211, 0x201F1E1D1C1B1A19,
79*87aa9c9eSJonas Devlieghere 0x2827262524232221, 0x302F2E2D2C2B2A29,
80*87aa9c9eSJonas Devlieghere 0x3837363534333231, 0x403F3E3D3C3B3A39,
81*87aa9c9eSJonas Devlieghere 0x4847464544434241, 0x504F4E4D4C4B4A49, },
82*87aa9c9eSJonas Devlieghere { 0x1918171615141312, 0x21201F1E1D1C1B1A,
83*87aa9c9eSJonas Devlieghere 0x2928272625242322, 0x31302F2E2D2C2B2A,
84*87aa9c9eSJonas Devlieghere 0x3938373635343332, 0x41403F3E3D3C3B3A,
85*87aa9c9eSJonas Devlieghere 0x4948474645444342, 0x51504F4E4D4C4B4A, },
86*87aa9c9eSJonas Devlieghere { 0x1A19181716151413, 0x2221201F1E1D1C1B,
87*87aa9c9eSJonas Devlieghere 0x2A29282726252423, 0x3231302F2E2D2C2B,
88*87aa9c9eSJonas Devlieghere 0x3A39383736353433, 0x4241403F3E3D3C3B,
89*87aa9c9eSJonas Devlieghere 0x4A49484746454443, 0x5251504F4E4D4C4B, },
90*87aa9c9eSJonas Devlieghere { 0x1B1A191817161514, 0x232221201F1E1D1C,
91*87aa9c9eSJonas Devlieghere 0x2B2A292827262524, 0x333231302F2E2D2C,
92*87aa9c9eSJonas Devlieghere 0x3B3A393837363534, 0x434241403F3E3D3C,
93*87aa9c9eSJonas Devlieghere 0x4B4A494847464544, 0x535251504F4E4D4C, },
94*87aa9c9eSJonas Devlieghere { 0x1C1B1A1918171615, 0x24232221201F1E1D,
95*87aa9c9eSJonas Devlieghere 0x2C2B2A2928272625, 0x34333231302F2E2D,
96*87aa9c9eSJonas Devlieghere 0x3C3B3A3938373635, 0x44434241403F3E3D,
97*87aa9c9eSJonas Devlieghere 0x4C4B4A4948474645, 0x54535251504F4E4D, },
98*87aa9c9eSJonas Devlieghere { 0x1D1C1B1A19181716, 0x2524232221201F1E,
99*87aa9c9eSJonas Devlieghere 0x2D2C2B2A29282726, 0x3534333231302F2E,
100*87aa9c9eSJonas Devlieghere 0x3D3C3B3A39383736, 0x4544434241403F3E,
101*87aa9c9eSJonas Devlieghere 0x4D4C4B4A49484746, 0x5554535251504F4E, },
102*87aa9c9eSJonas Devlieghere { 0x1E1D1C1B1A191817, 0x262524232221201F,
103*87aa9c9eSJonas Devlieghere 0x2E2D2C2B2A292827, 0x363534333231302F,
104*87aa9c9eSJonas Devlieghere 0x3E3D3C3B3A393837, 0x464544434241403F,
105*87aa9c9eSJonas Devlieghere 0x4E4D4C4B4A494847, 0x565554535251504F, },
106*87aa9c9eSJonas Devlieghere { 0x1F1E1D1C1B1A1918, 0x2726252423222120,
107*87aa9c9eSJonas Devlieghere 0x2F2E2D2C2B2A2928, 0x3736353433323130,
108*87aa9c9eSJonas Devlieghere 0x3F3E3D3C3B3A3938, 0x4746454443424140,
109*87aa9c9eSJonas Devlieghere 0x4F4E4D4C4B4A4948, 0x5756555453525150, },
110*87aa9c9eSJonas Devlieghere { 0x201F1E1D1C1B1A19, 0x2827262524232221,
111*87aa9c9eSJonas Devlieghere 0x302F2E2D2C2B2A29, 0x3837363534333231,
112*87aa9c9eSJonas Devlieghere 0x403F3E3D3C3B3A39, 0x4847464544434241,
113*87aa9c9eSJonas Devlieghere 0x504F4E4D4C4B4A49, 0x5857565554535251, },
114*87aa9c9eSJonas Devlieghere { 0x21201F1E1D1C1B1A, 0x2928272625242322,
115*87aa9c9eSJonas Devlieghere 0x31302F2E2D2C2B2A, 0x3938373635343332,
116*87aa9c9eSJonas Devlieghere 0x41403F3E3D3C3B3A, 0x4948474645444342,
117*87aa9c9eSJonas Devlieghere 0x51504F4E4D4C4B4A, 0x5958575655545352, },
118*87aa9c9eSJonas Devlieghere { 0x2221201F1E1D1C1B, 0x2A29282726252423,
119*87aa9c9eSJonas Devlieghere 0x3231302F2E2D2C2B, 0x3A39383736353433,
120*87aa9c9eSJonas Devlieghere 0x4241403F3E3D3C3B, 0x4A49484746454443,
121*87aa9c9eSJonas Devlieghere 0x5251504F4E4D4C4B, 0x5A59585756555453, },
122*87aa9c9eSJonas Devlieghere { 0x232221201F1E1D1C, 0x2B2A292827262524,
123*87aa9c9eSJonas Devlieghere 0x333231302F2E2D2C, 0x3B3A393837363534,
124*87aa9c9eSJonas Devlieghere 0x434241403F3E3D3C, 0x4B4A494847464544,
125*87aa9c9eSJonas Devlieghere 0x535251504F4E4D4C, 0x5B5A595857565554, },
126*87aa9c9eSJonas Devlieghere { 0x24232221201F1E1D, 0x2C2B2A2928272625,
127*87aa9c9eSJonas Devlieghere 0x34333231302F2E2D, 0x3C3B3A3938373635,
128*87aa9c9eSJonas Devlieghere 0x44434241403F3E3D, 0x4C4B4A4948474645,
129*87aa9c9eSJonas Devlieghere 0x54535251504F4E4D, 0x5C5B5A5958575655, },
130*87aa9c9eSJonas Devlieghere { 0x2524232221201F1E, 0x2D2C2B2A29282726,
131*87aa9c9eSJonas Devlieghere 0x3534333231302F2E, 0x3D3C3B3A39383736,
132*87aa9c9eSJonas Devlieghere 0x4544434241403F3E, 0x4D4C4B4A49484746,
133*87aa9c9eSJonas Devlieghere 0x5554535251504F4E, 0x5D5C5B5A59585756, },
134*87aa9c9eSJonas Devlieghere { 0x262524232221201F, 0x2E2D2C2B2A292827,
135*87aa9c9eSJonas Devlieghere 0x363534333231302F, 0x3E3D3C3B3A393837,
136*87aa9c9eSJonas Devlieghere 0x464544434241403F, 0x4E4D4C4B4A494847,
137*87aa9c9eSJonas Devlieghere 0x565554535251504F, 0x5E5D5C5B5A595857, },
138*87aa9c9eSJonas Devlieghere #endif
139*87aa9c9eSJonas Devlieghere };
140*87aa9c9eSJonas Devlieghere
141*87aa9c9eSJonas Devlieghere asm volatile(
142*87aa9c9eSJonas Devlieghere "vmovaps 0x000(%0), %%zmm0\n\t"
143*87aa9c9eSJonas Devlieghere "vmovaps 0x040(%0), %%zmm1\n\t"
144*87aa9c9eSJonas Devlieghere "vmovaps 0x080(%0), %%zmm2\n\t"
145*87aa9c9eSJonas Devlieghere "vmovaps 0x0C0(%0), %%zmm3\n\t"
146*87aa9c9eSJonas Devlieghere "vmovaps 0x100(%0), %%zmm4\n\t"
147*87aa9c9eSJonas Devlieghere "vmovaps 0x140(%0), %%zmm5\n\t"
148*87aa9c9eSJonas Devlieghere "vmovaps 0x180(%0), %%zmm6\n\t"
149*87aa9c9eSJonas Devlieghere "vmovaps 0x1C0(%0), %%zmm7\n\t"
150*87aa9c9eSJonas Devlieghere #if defined(__x86_64__) || defined(_M_X64)
151*87aa9c9eSJonas Devlieghere "vmovaps 0x200(%0), %%zmm8\n\t"
152*87aa9c9eSJonas Devlieghere "vmovaps 0x240(%0), %%zmm9\n\t"
153*87aa9c9eSJonas Devlieghere "vmovaps 0x280(%0), %%zmm10\n\t"
154*87aa9c9eSJonas Devlieghere "vmovaps 0x2C0(%0), %%zmm11\n\t"
155*87aa9c9eSJonas Devlieghere "vmovaps 0x300(%0), %%zmm12\n\t"
156*87aa9c9eSJonas Devlieghere "vmovaps 0x340(%0), %%zmm13\n\t"
157*87aa9c9eSJonas Devlieghere "vmovaps 0x380(%0), %%zmm14\n\t"
158*87aa9c9eSJonas Devlieghere "vmovaps 0x3C0(%0), %%zmm15\n\t"
159*87aa9c9eSJonas Devlieghere "vmovaps 0x400(%0), %%zmm16\n\t"
160*87aa9c9eSJonas Devlieghere "vmovaps 0x440(%0), %%zmm17\n\t"
161*87aa9c9eSJonas Devlieghere "vmovaps 0x480(%0), %%zmm18\n\t"
162*87aa9c9eSJonas Devlieghere "vmovaps 0x4C0(%0), %%zmm19\n\t"
163*87aa9c9eSJonas Devlieghere "vmovaps 0x500(%0), %%zmm20\n\t"
164*87aa9c9eSJonas Devlieghere "vmovaps 0x540(%0), %%zmm21\n\t"
165*87aa9c9eSJonas Devlieghere "vmovaps 0x580(%0), %%zmm22\n\t"
166*87aa9c9eSJonas Devlieghere "vmovaps 0x5C0(%0), %%zmm23\n\t"
167*87aa9c9eSJonas Devlieghere "vmovaps 0x600(%0), %%zmm24\n\t"
168*87aa9c9eSJonas Devlieghere "vmovaps 0x640(%0), %%zmm25\n\t"
169*87aa9c9eSJonas Devlieghere "vmovaps 0x680(%0), %%zmm26\n\t"
170*87aa9c9eSJonas Devlieghere "vmovaps 0x6C0(%0), %%zmm27\n\t"
171*87aa9c9eSJonas Devlieghere "vmovaps 0x700(%0), %%zmm28\n\t"
172*87aa9c9eSJonas Devlieghere "vmovaps 0x740(%0), %%zmm29\n\t"
173*87aa9c9eSJonas Devlieghere "vmovaps 0x780(%0), %%zmm30\n\t"
174*87aa9c9eSJonas Devlieghere "vmovaps 0x7C0(%0), %%zmm31\n\t"
175*87aa9c9eSJonas Devlieghere #endif
176*87aa9c9eSJonas Devlieghere "\n\t"
177*87aa9c9eSJonas Devlieghere "int3\n\t"
178*87aa9c9eSJonas Devlieghere :
179*87aa9c9eSJonas Devlieghere : "b"(zmm)
180*87aa9c9eSJonas Devlieghere : "%zmm0", "%zmm1", "%zmm2", "%zmm3", "%zmm4", "%zmm5", "%zmm6", "%zmm7"
181*87aa9c9eSJonas Devlieghere #if defined(__x86_64__) || defined(_M_X64)
182*87aa9c9eSJonas Devlieghere , "%zmm8", "%zmm9", "%zmm10", "%zmm11", "%zmm12", "%zmm13", "%zmm14",
183*87aa9c9eSJonas Devlieghere "%zmm15", "%zmm16", "%zmm17", "%zmm18", "%zmm19", "%zmm20", "%zmm21",
184*87aa9c9eSJonas Devlieghere "%zmm22", "%zmm23", "%zmm24", "%zmm25", "%zmm26", "%zmm27", "%zmm28",
185*87aa9c9eSJonas Devlieghere "%zmm29", "%zmm30", "%zmm31"
186*87aa9c9eSJonas Devlieghere #endif
187*87aa9c9eSJonas Devlieghere );
188*87aa9c9eSJonas Devlieghere
189*87aa9c9eSJonas Devlieghere return 0;
190*87aa9c9eSJonas Devlieghere }
191