| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* -*- c++ -*- */ | ||
| 2 | /* | ||
| 3 | * Copyright 2010, 2011, 2015-2017, 2019, 2020 Free Software Foundation, Inc. | ||
| 4 | * Copyright 2023 Magnus Lundmark <magnuslundmark@gmail.com> | ||
| 5 | * | ||
| 6 | * This file is part of VOLK | ||
| 7 | * | ||
| 8 | * SPDX-License-Identifier: LGPL-3.0-or-later | ||
| 9 | */ | ||
| 10 | |||
| 11 | #ifndef INCLUDED_LIBVOLK_COMMON_H | ||
| 12 | #define INCLUDED_LIBVOLK_COMMON_H | ||
| 13 | |||
| 14 | //////////////////////////////////////////////////////////////////////// | ||
| 15 | // Cross-platform attribute macros | ||
| 16 | //////////////////////////////////////////////////////////////////////// | ||
| 17 | #if _MSC_VER | ||
| 18 | #define __VOLK_ATTR_ALIGNED(x) __declspec(align(x)) | ||
| 19 | #define __VOLK_ATTR_UNUSED | ||
| 20 | #define __VOLK_ATTR_INLINE __forceinline | ||
| 21 | #define __VOLK_ATTR_DEPRECATED __declspec(deprecated) | ||
| 22 | #define __VOLK_ATTR_EXPORT __declspec(dllexport) | ||
| 23 | #define __VOLK_ATTR_IMPORT __declspec(dllimport) | ||
| 24 | #define __VOLK_PREFETCH(addr) | ||
| 25 | #define __VOLK_ASM __asm | ||
| 26 | #define __VOLK_VOLATILE | ||
| 27 | #elif defined(__clang__) | ||
| 28 | // AppleClang also defines __GNUC__, so do this check first. These | ||
| 29 | // will probably be the same as for __GNUC__, but let's keep them | ||
| 30 | // separate just to be safe. | ||
| 31 | #define __VOLK_ATTR_ALIGNED(x) __attribute__((aligned(x))) | ||
| 32 | #define __VOLK_ATTR_UNUSED __attribute__((unused)) | ||
| 33 | #define __VOLK_ATTR_INLINE __attribute__((always_inline)) | ||
| 34 | #define __VOLK_ATTR_DEPRECATED __attribute__((deprecated)) | ||
| 35 | #define __VOLK_ASM __asm__ | ||
| 36 | #define __VOLK_VOLATILE __volatile__ | ||
| 37 | #define __VOLK_ATTR_EXPORT __attribute__((visibility("default"))) | ||
| 38 | #define __VOLK_ATTR_IMPORT __attribute__((visibility("default"))) | ||
| 39 | #define __VOLK_PREFETCH(addr) __builtin_prefetch(addr) | ||
| 40 | #elif defined __GNUC__ | ||
| 41 | #define __VOLK_ATTR_ALIGNED(x) __attribute__((aligned(x))) | ||
| 42 | #define __VOLK_ATTR_UNUSED __attribute__((unused)) | ||
| 43 | #define __VOLK_ATTR_INLINE __attribute__((always_inline)) | ||
| 44 | #define __VOLK_ATTR_DEPRECATED __attribute__((deprecated)) | ||
| 45 | #define __VOLK_ASM __asm__ | ||
| 46 | #define __VOLK_VOLATILE __volatile__ | ||
| 47 | #if __GNUC__ >= 4 | ||
| 48 | #define __VOLK_ATTR_EXPORT __attribute__((visibility("default"))) | ||
| 49 | #define __VOLK_ATTR_IMPORT __attribute__((visibility("default"))) | ||
| 50 | #else | ||
| 51 | #define __VOLK_ATTR_EXPORT | ||
| 52 | #define __VOLK_ATTR_IMPORT | ||
| 53 | #endif | ||
| 54 | #define __VOLK_PREFETCH(addr) __builtin_prefetch(addr) | ||
| 55 | #elif _MSC_VER | ||
| 56 | #define __VOLK_ATTR_ALIGNED(x) __declspec(align(x)) | ||
| 57 | #define __VOLK_ATTR_UNUSED | ||
| 58 | #define __VOLK_ATTR_INLINE __forceinline | ||
| 59 | #define __VOLK_ATTR_DEPRECATED __declspec(deprecated) | ||
| 60 | #define __VOLK_ATTR_EXPORT __declspec(dllexport) | ||
| 61 | #define __VOLK_ATTR_IMPORT __declspec(dllimport) | ||
| 62 | #define __VOLK_PREFETCH(addr) | ||
| 63 | #define __VOLK_ASM __asm | ||
| 64 | #define __VOLK_VOLATILE | ||
| 65 | #else | ||
| 66 | #define __VOLK_ATTR_ALIGNED(x) | ||
| 67 | #define __VOLK_ATTR_UNUSED | ||
| 68 | #define __VOLK_ATTR_INLINE | ||
| 69 | #define __VOLK_ATTR_DEPRECATED | ||
| 70 | #define __VOLK_ATTR_EXPORT | ||
| 71 | #define __VOLK_ATTR_IMPORT | ||
| 72 | #define __VOLK_PREFETCH(addr) | ||
| 73 | #define __VOLK_ASM __asm__ | ||
| 74 | #define __VOLK_VOLATILE __volatile__ | ||
| 75 | #endif | ||
| 76 | |||
| 77 | //////////////////////////////////////////////////////////////////////// | ||
| 78 | // Ignore annoying warnings in MSVC | ||
| 79 | //////////////////////////////////////////////////////////////////////// | ||
| 80 | #if defined(_MSC_VER) | ||
| 81 | #pragma warning(disable : 4244) //'conversion' conversion from 'type1' to 'type2', | ||
| 82 | // possible loss of data | ||
| 83 | #pragma warning(disable : 4305) //'identifier' : truncation from 'type1' to 'type2' | ||
| 84 | #endif | ||
| 85 | |||
| 86 | //////////////////////////////////////////////////////////////////////// | ||
| 87 | // C-linkage declaration macros | ||
| 88 | // FIXME: due to the usage of complex.h, require gcc for c-linkage | ||
| 89 | //////////////////////////////////////////////////////////////////////// | ||
| 90 | #if defined(__cplusplus) && (__GNUC__) | ||
| 91 | #define __VOLK_DECL_BEGIN extern "C" { | ||
| 92 | #define __VOLK_DECL_END } | ||
| 93 | #else | ||
| 94 | #define __VOLK_DECL_BEGIN | ||
| 95 | #define __VOLK_DECL_END | ||
| 96 | #endif | ||
| 97 | |||
| 98 | //////////////////////////////////////////////////////////////////////// | ||
| 99 | // Define VOLK_API for library symbols | ||
| 100 | // http://gcc.gnu.org/wiki/Visibility | ||
| 101 | //////////////////////////////////////////////////////////////////////// | ||
| 102 | #ifdef volk_EXPORTS | ||
| 103 | #define VOLK_API __VOLK_ATTR_EXPORT | ||
| 104 | #else | ||
| 105 | #define VOLK_API __VOLK_ATTR_IMPORT | ||
| 106 | #endif | ||
| 107 | |||
| 108 | //////////////////////////////////////////////////////////////////////// | ||
| 109 | // The bit128 union used by some | ||
| 110 | //////////////////////////////////////////////////////////////////////// | ||
| 111 | #include <stdint.h> | ||
| 112 | |||
| 113 | #ifdef LV_HAVE_SSE | ||
| 114 | #ifdef _WIN32 | ||
| 115 | #include <intrin.h> | ||
| 116 | #else | ||
| 117 | #include <x86intrin.h> | ||
| 118 | #endif | ||
| 119 | #endif | ||
| 120 | |||
| 121 | union bit128 { | ||
| 122 | uint8_t i8[16]; | ||
| 123 | uint16_t i16[8]; | ||
| 124 | uint32_t i[4]; | ||
| 125 | float f[4]; | ||
| 126 | double d[2]; | ||
| 127 | |||
| 128 | #ifdef LV_HAVE_SSE | ||
| 129 | __m128 float_vec; | ||
| 130 | #endif | ||
| 131 | |||
| 132 | #ifdef LV_HAVE_SSE2 | ||
| 133 | __m128i int_vec; | ||
| 134 | __m128d double_vec; | ||
| 135 | #endif | ||
| 136 | }; | ||
| 137 | |||
| 138 | union bit256 { | ||
| 139 | uint8_t i8[32]; | ||
| 140 | uint16_t i16[16]; | ||
| 141 | uint32_t i[8]; | ||
| 142 | float f[8]; | ||
| 143 | double d[4]; | ||
| 144 | |||
| 145 | #ifdef LV_HAVE_AVX | ||
| 146 | __m256 float_vec; | ||
| 147 | __m256i int_vec; | ||
| 148 | __m256d double_vec; | ||
| 149 | #endif | ||
| 150 | }; | ||
| 151 | |||
| 152 | #define bit128_p(x) ((union bit128*)(x)) | ||
| 153 | #define bit256_p(x) ((union bit256*)(x)) | ||
| 154 | |||
| 155 | //////////////////////////////////////////////////////////////////////// | ||
| 156 | // log2f | ||
| 157 | //////////////////////////////////////////////////////////////////////// | ||
| 158 | #include <math.h> | ||
| 159 | // +-Inf -> +-127.0f in order to match the behaviour of the SIMD kernels | ||
| 160 | 1310758 | static inline float log2f_non_ieee(float f) | |
| 161 | { | ||
| 162 | 1310758 | float const result = log2f(f); | |
| 163 |
1/2✓ Branch 0 taken 1310758 times.
✗ Branch 1 not taken.
|
1310758 | return isinf(result) ? copysignf(127.0f, result) : result; |
| 164 | } | ||
| 165 | |||
| 166 | //////////////////////////////////////////////////////////////////////// | ||
| 167 | // Constant used to do log10 calculations as faster log2 | ||
| 168 | //////////////////////////////////////////////////////////////////////// | ||
| 169 | // precalculated 10.0 / log2f_non_ieee(10.0) to allow for constexpr | ||
| 170 | #define volk_log2to10factor (0x1.815182p1) // 3.01029995663981209120 | ||
| 171 | |||
| 172 | //////////////////////////////////////////////////////////////////////// | ||
| 173 | // arctan(x) | ||
| 174 | //////////////////////////////////////////////////////////////////////// | ||
| 175 | 262210 | static inline float volk_arctan_poly(const float x) | |
| 176 | { | ||
| 177 | /* | ||
| 178 | * arctan(x) polynomial expansion on the interval [-1, 1] | ||
| 179 | * Maximum relative error < 6.6e-7 | ||
| 180 | */ | ||
| 181 | 262210 | const float a1 = +0x1.ffffeap-1f; | |
| 182 | 262210 | const float a3 = -0x1.55437p-2f; | |
| 183 | 262210 | const float a5 = +0x1.972be6p-3f; | |
| 184 | 262210 | const float a7 = -0x1.1436ap-3f; | |
| 185 | 262210 | const float a9 = +0x1.5785aap-4f; | |
| 186 | 262210 | const float a11 = -0x1.2f3004p-5f; | |
| 187 | 262210 | const float a13 = +0x1.01a37cp-7f; | |
| 188 | |||
| 189 | 262210 | const float x_times_x = x * x; | |
| 190 | 262210 | float arctan = a13; | |
| 191 | 262210 | arctan = fmaf(x_times_x, arctan, a11); | |
| 192 | 262210 | arctan = fmaf(x_times_x, arctan, a9); | |
| 193 | 262210 | arctan = fmaf(x_times_x, arctan, a7); | |
| 194 | 262210 | arctan = fmaf(x_times_x, arctan, a5); | |
| 195 | 262210 | arctan = fmaf(x_times_x, arctan, a3); | |
| 196 | 262210 | arctan = fmaf(x_times_x, arctan, a1); | |
| 197 | 262210 | arctan *= x; | |
| 198 | |||
| 199 | 262210 | return arctan; | |
| 200 | } | ||
| 201 | |||
| 202 | 262210 | static inline float volk_arctan(const float x) | |
| 203 | { | ||
| 204 | /* | ||
| 205 | * arctan(x) + arctan(1 / x) == sign(x) * pi / 2 | ||
| 206 | */ | ||
| 207 | 262210 | const float pi_over_2 = 0x1.921fb6p0f; | |
| 208 | |||
| 209 |
1/2✓ Branch 0 taken 262210 times.
✗ Branch 1 not taken.
|
262210 | if (fabs(x) < 1.f) { |
| 210 | 262210 | return volk_arctan_poly(x); | |
| 211 | } else { | ||
| 212 | ✗ | return copysignf(pi_over_2, x) - volk_arctan_poly(1.f / x); | |
| 213 | } | ||
| 214 | } | ||
| 215 | |||
| 216 | #endif /*INCLUDED_LIBVOLK_COMMON_H*/ | ||
| 217 |