|
| 1 | +/*************************************************************************** |
| 2 | + * Copyright (c) Johan Mabille, Sylvain Corlay, Wolf Vollprecht and * |
| 3 | + * Martin Renou * |
| 4 | + * Copyright (c) QuantStack * |
| 5 | + * Copyright (c) Serge Guelton * |
| 6 | + * * |
| 7 | + * Distributed under the terms of the BSD 3-Clause License. * |
| 8 | + * * |
| 9 | + * The full license is in the file LICENSE, distributed with this software. * |
| 10 | + ***************************************************************************/ |
| 11 | + |
| 12 | +#ifndef XSIMD_GETAUXVAL_HPP |
| 13 | +#define XSIMD_GETAUXVAL_HPP |
| 14 | + |
| 15 | +#include "../utils/bits.hpp" |
| 16 | +#include "./xsimd_config.hpp" |
| 17 | + |
| 18 | +#if XSIMD_WITH_LINUX_GETAUXVAL |
| 19 | +#include <sys/auxv.h> |
| 20 | +#endif |
| 21 | + |
| 22 | +namespace xsimd |
| 23 | +{ |
| 24 | + namespace detail |
| 25 | + { |
| 26 | + using linux_getauxval_t = unsigned long; |
| 27 | + |
| 28 | + inline linux_getauxval_t linux_getauxval(linux_getauxval_t type) noexcept; |
| 29 | + |
| 30 | + /** |
| 31 | + * Base class for getauxval querying. |
| 32 | + */ |
| 33 | + template <linux_getauxval_t type, typename A> |
| 34 | + class linux_auxval : private utils::uint_bitset<A, linux_getauxval_t> |
| 35 | + { |
| 36 | + using bitset_t = utils::uint_bitset<A, linux_getauxval_t>; |
| 37 | + |
| 38 | + public: |
| 39 | + using aux = A; |
| 40 | + |
| 41 | + inline static linux_auxval read() |
| 42 | + { |
| 43 | + return bitset_t(linux_getauxval(type)); |
| 44 | + } |
| 45 | + |
| 46 | + /** Create a value which returns false to everything. */ |
| 47 | + constexpr linux_auxval() noexcept = default; |
| 48 | + |
| 49 | + using bitset_t::all_bits_set; |
| 50 | + }; |
| 51 | + |
| 52 | + template <typename Traits> |
| 53 | + using make_auxiliary_val_t = linux_auxval<Traits::type, typename Traits::aux>; |
| 54 | + } |
| 55 | + |
| 56 | + /* |
| 57 | + * Hardware Capabilities Register (HWCAP) for Linux. |
| 58 | + * |
| 59 | + * On Linux systems, the kernel exposes some CPU features through the |
| 60 | + * auxiliary vector, which can be queried via `getauxval(AT_HWCAP)`. |
| 61 | + * This utility parses such bit values. |
| 62 | + * |
| 63 | + * @see https://www.kernel.org/doc/Documentation/arm64/elf_hwcaps.txt |
| 64 | + */ |
| 65 | + struct linux_hwcap_traits |
| 66 | + { |
| 67 | +#if XSIMD_WITH_LINUX_GETAUXVAL |
| 68 | + static constexpr detail::linux_getauxval_t type = AT_HWCAP; |
| 69 | +#else |
| 70 | + static constexpr detail::linux_getauxval_t type = 0; |
| 71 | +#endif |
| 72 | + |
| 73 | + enum class aux |
| 74 | + { |
| 75 | +#if XSIMD_WITH_LINUX_GETAUXVAL |
| 76 | +#if XSIMD_TARGET_ARM64 |
| 77 | + /** Scalable Vector Extension. */ |
| 78 | + sve = HWCAP_SVE, |
| 79 | +#elif XSIMD_TARGET_ARM |
| 80 | + /** Neon vector extension. */ |
| 81 | + neon = HWCAP_NEON, |
| 82 | +#endif |
| 83 | +#endif |
| 84 | + }; |
| 85 | + }; |
| 86 | + |
| 87 | + using linux_hwcap = detail::make_auxiliary_val_t<linux_hwcap_traits>; |
| 88 | + |
| 89 | + /* |
| 90 | + * Extended Hardware Capabilities Register (HWCAP2) for Linux. |
| 91 | + * |
| 92 | + * On Linux systems, the kernel exposes some CPU additional features through the |
| 93 | + * auxiliary vector, which can be queried via `getauxval(AT_HWCAP2)`. |
| 94 | + * |
| 95 | + * @see https://www.kernel.org/doc/Documentation/arm64/elf_hwcaps.txt |
| 96 | + */ |
| 97 | + struct linux_hwcap2_traits |
| 98 | + { |
| 99 | +#if XSIMD_WITH_LINUX_GETAUXVAL |
| 100 | + static constexpr detail::linux_getauxval_t type = AT_HWCAP2; |
| 101 | +#else |
| 102 | + static constexpr detail::linux_getauxval_t type = 0; |
| 103 | +#endif |
| 104 | + |
| 105 | + enum class aux |
| 106 | + { |
| 107 | +#if XSIMD_WITH_LINUX_GETAUXVAL |
| 108 | +#if XSIMD_TARGET_ARM64 |
| 109 | +#ifndef HWCAP2_I8MM |
| 110 | +#define HWCAP2_I8MM (1 << 13) |
| 111 | +#endif |
| 112 | + /** 8 bits integer matrix multiplication. */ |
| 113 | + i8mm = HWCAP2_I8MM, |
| 114 | +#endif |
| 115 | +#endif |
| 116 | + }; |
| 117 | + }; |
| 118 | + |
| 119 | + using linux_hwcap2 = detail::make_auxiliary_val_t<linux_hwcap2_traits>; |
| 120 | + |
| 121 | + /******************** |
| 122 | + * Implementation * |
| 123 | + ********************/ |
| 124 | + |
| 125 | + namespace detail |
| 126 | + { |
| 127 | +#if XSIMD_WITH_LINUX_GETAUXVAL |
| 128 | + inline linux_getauxval_t linux_getauxval(linux_getauxval_t type) noexcept |
| 129 | + { |
| 130 | + return getauxval(type); |
| 131 | + } |
| 132 | +#else |
| 133 | + inline linux_getauxval_t linux_getauxval(linux_getauxval_t type) noexcept |
| 134 | + { |
| 135 | + return {}; // All bits set to 0 |
| 136 | + } |
| 137 | +#endif |
| 138 | + } |
| 139 | +} |
| 140 | + |
| 141 | +#endif |
0 commit comments