191 lines · cpp
1#include <cstdint>2 3struct alignas(64) zmm_t {4 uint64_t a, b, c, d, e, f, g, h;5};6 7int main() {8 constexpr zmm_t zmm[] = {9 { 0x0706050403020100, 0x0F0E0D0C0B0A0908,10 0x1716151413121110, 0x1F1E1D1C1B1A1918,11 0x2726252423222120, 0x2F2E2D2C2B2A2928,12 0x3736353433323130, 0x3F3E3D3C3B3A3938, },13 { 0x0807060504030201, 0x100F0E0D0C0B0A09,14 0x1817161514131211, 0x201F1E1D1C1B1A19,15 0x2827262524232221, 0x302F2E2D2C2B2A29,16 0x3837363534333231, 0x403F3E3D3C3B3A39, },17 { 0x0908070605040302, 0x11100F0E0D0C0B0A,18 0x1918171615141312, 0x21201F1E1D1C1B1A,19 0x2928272625242322, 0x31302F2E2D2C2B2A,20 0x3938373635343332, 0x41403F3E3D3C3B3A, },21 { 0x0A09080706050403, 0x1211100F0E0D0C0B,22 0x1A19181716151413, 0x2221201F1E1D1C1B,23 0x2A29282726252423, 0x3231302F2E2D2C2B,24 0x3A39383736353433, 0x4241403F3E3D3C3B, },25 { 0x0B0A090807060504, 0x131211100F0E0D0C,26 0x1B1A191817161514, 0x232221201F1E1D1C,27 0x2B2A292827262524, 0x333231302F2E2D2C,28 0x3B3A393837363534, 0x434241403F3E3D3C, },29 { 0x0C0B0A0908070605, 0x14131211100F0E0D,30 0x1C1B1A1918171615, 0x24232221201F1E1D,31 0x2C2B2A2928272625, 0x34333231302F2E2D,32 0x3C3B3A3938373635, 0x44434241403F3E3D, },33 { 0x0D0C0B0A09080706, 0x1514131211100F0E,34 0x1D1C1B1A19181716, 0x2524232221201F1E,35 0x2D2C2B2A29282726, 0x3534333231302F2E,36 0x3D3C3B3A39383736, 0x4544434241403F3E, },37 { 0x0E0D0C0B0A090807, 0x161514131211100F,38 0x1E1D1C1B1A191817, 0x262524232221201F,39 0x2E2D2C2B2A292827, 0x363534333231302F,40 0x3E3D3C3B3A393837, 0x464544434241403F, },41#if defined(__x86_64__) || defined(_M_X64)42 { 0x0F0E0D0C0B0A0908, 0x1716151413121110,43 0x1F1E1D1C1B1A1918, 0x2726252423222120,44 0x2F2E2D2C2B2A2928, 0x3736353433323130,45 0x3F3E3D3C3B3A3938, 0x4746454443424140, },46 { 0x100F0E0D0C0B0A09, 0x1817161514131211,47 0x201F1E1D1C1B1A19, 0x2827262524232221,48 0x302F2E2D2C2B2A29, 0x3837363534333231,49 0x403F3E3D3C3B3A39, 0x4847464544434241, },50 { 0x11100F0E0D0C0B0A, 0x1918171615141312,51 0x21201F1E1D1C1B1A, 0x2928272625242322,52 0x31302F2E2D2C2B2A, 0x3938373635343332,53 0x41403F3E3D3C3B3A, 0x4948474645444342, },54 { 0x1211100F0E0D0C0B, 0x1A19181716151413,55 0x2221201F1E1D1C1B, 0x2A29282726252423,56 0x3231302F2E2D2C2B, 0x3A39383736353433,57 0x4241403F3E3D3C3B, 0x4A49484746454443, },58 { 0x131211100F0E0D0C, 0x1B1A191817161514,59 0x232221201F1E1D1C, 0x2B2A292827262524,60 0x333231302F2E2D2C, 0x3B3A393837363534,61 0x434241403F3E3D3C, 0x4B4A494847464544, },62 { 0x14131211100F0E0D, 0x1C1B1A1918171615,63 0x24232221201F1E1D, 0x2C2B2A2928272625,64 0x34333231302F2E2D, 0x3C3B3A3938373635,65 0x44434241403F3E3D, 0x4C4B4A4948474645, },66 { 0x1514131211100F0E, 0x1D1C1B1A19181716,67 0x2524232221201F1E, 0x2D2C2B2A29282726,68 0x3534333231302F2E, 0x3D3C3B3A39383736,69 0x4544434241403F3E, 0x4D4C4B4A49484746, },70 { 0x161514131211100F, 0x1E1D1C1B1A191817,71 0x262524232221201F, 0x2E2D2C2B2A292827,72 0x363534333231302F, 0x3E3D3C3B3A393837,73 0x464544434241403F, 0x4E4D4C4B4A494847, },74 { 0x1716151413121110, 0x1F1E1D1C1B1A1918,75 0x2726252423222120, 0x2F2E2D2C2B2A2928,76 0x3736353433323130, 0x3F3E3D3C3B3A3938,77 0x4746454443424140, 0x4F4E4D4C4B4A4948, },78 { 0x1817161514131211, 0x201F1E1D1C1B1A19,79 0x2827262524232221, 0x302F2E2D2C2B2A29,80 0x3837363534333231, 0x403F3E3D3C3B3A39,81 0x4847464544434241, 0x504F4E4D4C4B4A49, },82 { 0x1918171615141312, 0x21201F1E1D1C1B1A,83 0x2928272625242322, 0x31302F2E2D2C2B2A,84 0x3938373635343332, 0x41403F3E3D3C3B3A,85 0x4948474645444342, 0x51504F4E4D4C4B4A, },86 { 0x1A19181716151413, 0x2221201F1E1D1C1B,87 0x2A29282726252423, 0x3231302F2E2D2C2B,88 0x3A39383736353433, 0x4241403F3E3D3C3B,89 0x4A49484746454443, 0x5251504F4E4D4C4B, },90 { 0x1B1A191817161514, 0x232221201F1E1D1C,91 0x2B2A292827262524, 0x333231302F2E2D2C,92 0x3B3A393837363534, 0x434241403F3E3D3C,93 0x4B4A494847464544, 0x535251504F4E4D4C, },94 { 0x1C1B1A1918171615, 0x24232221201F1E1D,95 0x2C2B2A2928272625, 0x34333231302F2E2D,96 0x3C3B3A3938373635, 0x44434241403F3E3D,97 0x4C4B4A4948474645, 0x54535251504F4E4D, },98 { 0x1D1C1B1A19181716, 0x2524232221201F1E,99 0x2D2C2B2A29282726, 0x3534333231302F2E,100 0x3D3C3B3A39383736, 0x4544434241403F3E,101 0x4D4C4B4A49484746, 0x5554535251504F4E, },102 { 0x1E1D1C1B1A191817, 0x262524232221201F,103 0x2E2D2C2B2A292827, 0x363534333231302F,104 0x3E3D3C3B3A393837, 0x464544434241403F,105 0x4E4D4C4B4A494847, 0x565554535251504F, },106 { 0x1F1E1D1C1B1A1918, 0x2726252423222120,107 0x2F2E2D2C2B2A2928, 0x3736353433323130,108 0x3F3E3D3C3B3A3938, 0x4746454443424140,109 0x4F4E4D4C4B4A4948, 0x5756555453525150, },110 { 0x201F1E1D1C1B1A19, 0x2827262524232221,111 0x302F2E2D2C2B2A29, 0x3837363534333231,112 0x403F3E3D3C3B3A39, 0x4847464544434241,113 0x504F4E4D4C4B4A49, 0x5857565554535251, },114 { 0x21201F1E1D1C1B1A, 0x2928272625242322,115 0x31302F2E2D2C2B2A, 0x3938373635343332,116 0x41403F3E3D3C3B3A, 0x4948474645444342,117 0x51504F4E4D4C4B4A, 0x5958575655545352, },118 { 0x2221201F1E1D1C1B, 0x2A29282726252423,119 0x3231302F2E2D2C2B, 0x3A39383736353433,120 0x4241403F3E3D3C3B, 0x4A49484746454443,121 0x5251504F4E4D4C4B, 0x5A59585756555453, },122 { 0x232221201F1E1D1C, 0x2B2A292827262524,123 0x333231302F2E2D2C, 0x3B3A393837363534,124 0x434241403F3E3D3C, 0x4B4A494847464544,125 0x535251504F4E4D4C, 0x5B5A595857565554, },126 { 0x24232221201F1E1D, 0x2C2B2A2928272625,127 0x34333231302F2E2D, 0x3C3B3A3938373635,128 0x44434241403F3E3D, 0x4C4B4A4948474645,129 0x54535251504F4E4D, 0x5C5B5A5958575655, },130 { 0x2524232221201F1E, 0x2D2C2B2A29282726,131 0x3534333231302F2E, 0x3D3C3B3A39383736,132 0x4544434241403F3E, 0x4D4C4B4A49484746,133 0x5554535251504F4E, 0x5D5C5B5A59585756, },134 { 0x262524232221201F, 0x2E2D2C2B2A292827,135 0x363534333231302F, 0x3E3D3C3B3A393837,136 0x464544434241403F, 0x4E4D4C4B4A494847,137 0x565554535251504F, 0x5E5D5C5B5A595857, },138#endif139 };140 141 asm volatile(142 "vmovaps 0x000(%0), %%zmm0\n\t"143 "vmovaps 0x040(%0), %%zmm1\n\t"144 "vmovaps 0x080(%0), %%zmm2\n\t"145 "vmovaps 0x0C0(%0), %%zmm3\n\t"146 "vmovaps 0x100(%0), %%zmm4\n\t"147 "vmovaps 0x140(%0), %%zmm5\n\t"148 "vmovaps 0x180(%0), %%zmm6\n\t"149 "vmovaps 0x1C0(%0), %%zmm7\n\t"150#if defined(__x86_64__) || defined(_M_X64)151 "vmovaps 0x200(%0), %%zmm8\n\t"152 "vmovaps 0x240(%0), %%zmm9\n\t"153 "vmovaps 0x280(%0), %%zmm10\n\t"154 "vmovaps 0x2C0(%0), %%zmm11\n\t"155 "vmovaps 0x300(%0), %%zmm12\n\t"156 "vmovaps 0x340(%0), %%zmm13\n\t"157 "vmovaps 0x380(%0), %%zmm14\n\t"158 "vmovaps 0x3C0(%0), %%zmm15\n\t"159 "vmovaps 0x400(%0), %%zmm16\n\t"160 "vmovaps 0x440(%0), %%zmm17\n\t"161 "vmovaps 0x480(%0), %%zmm18\n\t"162 "vmovaps 0x4C0(%0), %%zmm19\n\t"163 "vmovaps 0x500(%0), %%zmm20\n\t"164 "vmovaps 0x540(%0), %%zmm21\n\t"165 "vmovaps 0x580(%0), %%zmm22\n\t"166 "vmovaps 0x5C0(%0), %%zmm23\n\t"167 "vmovaps 0x600(%0), %%zmm24\n\t"168 "vmovaps 0x640(%0), %%zmm25\n\t"169 "vmovaps 0x680(%0), %%zmm26\n\t"170 "vmovaps 0x6C0(%0), %%zmm27\n\t"171 "vmovaps 0x700(%0), %%zmm28\n\t"172 "vmovaps 0x740(%0), %%zmm29\n\t"173 "vmovaps 0x780(%0), %%zmm30\n\t"174 "vmovaps 0x7C0(%0), %%zmm31\n\t"175#endif176 "\n\t"177 "int3\n\t"178 :179 : "b"(zmm)180 : "%zmm0", "%zmm1", "%zmm2", "%zmm3", "%zmm4", "%zmm5", "%zmm6", "%zmm7"181#if defined(__x86_64__) || defined(_M_X64)182 , "%zmm8", "%zmm9", "%zmm10", "%zmm11", "%zmm12", "%zmm13", "%zmm14",183 "%zmm15", "%zmm16", "%zmm17", "%zmm18", "%zmm19", "%zmm20", "%zmm21",184 "%zmm22", "%zmm23", "%zmm24", "%zmm25", "%zmm26", "%zmm27", "%zmm28",185 "%zmm29", "%zmm30", "%zmm31"186#endif187 );188 189 return 0;190}191