common/alnumeric.h

   1 #ifndef AL_NUMERIC_H
   2 #define AL_NUMERIC_H
   3
   4 #include <algorithm>
   5 #include <cmath>
   6 #include <cstddef>
   7 #include <cstdint>
   8 #ifdef HAVE_INTRIN_H
   9 #include <intrin.h>
  10 #endif
  11 #ifdef HAVE_SSE_INTRINSICS
  12 #include <xmmintrin.h>
  13 #endif
  14
  15 #include "opthelpers.h"
  16
  17
  18 inline constexpr int64_t operator "" _i64(unsigned long long int n) noexcept { return static_cast<int64_t>(n); }
  19 inline constexpr uint64_t operator "" _u64(unsigned long long int n) noexcept { return static_cast<uint64_t>(n); }
  20
  21
  22 constexpr inline float minf(float a, float b) noexcept
  23 { return ((a > b) ? b : a); }
  24 constexpr inline float maxf(float a, float b) noexcept
  25 { return ((a > b) ? a : b); }
  26 constexpr inline float clampf(float val, float min, float max) noexcept
  27 { return minf(max, maxf(min, val)); }
  28
  29 constexpr inline double mind(double a, double b) noexcept
  30 { return ((a > b) ? b : a); }
  31 constexpr inline double maxd(double a, double b) noexcept
  32 { return ((a > b) ? a : b); }
  33 constexpr inline double clampd(double val, double min, double max) noexcept
  34 { return mind(max, maxd(min, val)); }
  35
  36 constexpr inline unsigned int minu(unsigned int a, unsigned int b) noexcept
  37 { return ((a > b) ? b : a); }
  38 constexpr inline unsigned int maxu(unsigned int a, unsigned int b) noexcept
  39 { return ((a > b) ? a : b); }
  40 constexpr inline unsigned int clampu(unsigned int val, unsigned int min, unsigned int max) noexcept
  41 { return minu(max, maxu(min, val)); }
  42
  43 constexpr inline int mini(int a, int b) noexcept
  44 { return ((a > b) ? b : a); }
  45 constexpr inline int maxi(int a, int b) noexcept
  46 { return ((a > b) ? a : b); }
  47 constexpr inline int clampi(int val, int min, int max) noexcept
  48 { return mini(max, maxi(min, val)); }
  49
  50 constexpr inline int64_t mini64(int64_t a, int64_t b) noexcept
  51 { return ((a > b) ? b : a); }
  52 constexpr inline int64_t maxi64(int64_t a, int64_t b) noexcept
  53 { return ((a > b) ? a : b); }
  54 constexpr inline int64_t clampi64(int64_t val, int64_t min, int64_t max) noexcept
  55 { return mini64(max, maxi64(min, val)); }
  56
  57 constexpr inline uint64_t minu64(uint64_t a, uint64_t b) noexcept
  58 { return ((a > b) ? b : a); }
  59 constexpr inline uint64_t maxu64(uint64_t a, uint64_t b) noexcept
  60 { return ((a > b) ? a : b); }
  61 constexpr inline uint64_t clampu64(uint64_t val, uint64_t min, uint64_t max) noexcept
  62 { return minu64(max, maxu64(min, val)); }
  63
  64 constexpr inline size_t minz(size_t a, size_t b) noexcept
  65 { return ((a > b) ? b : a); }
  66 constexpr inline size_t maxz(size_t a, size_t b) noexcept
  67 { return ((a > b) ? a : b); }
  68 constexpr inline size_t clampz(size_t val, size_t min, size_t max) noexcept
  69 { return minz(max, maxz(min, val)); }
  70
  71
  72 constexpr inline float lerpf(float val1, float val2, float mu) noexcept
  73 { return val1 + (val2-val1)*mu; }
  74 constexpr inline float cubic(float val1, float val2, float val3, float val4, float mu) noexcept
  75 {
  76     const float mu2{mu*mu}, mu3{mu2*mu};
  77     const float a0{-0.5f*mu3 +       mu2 + -0.5f*mu};
  78     const float a1{ 1.5f*mu3 + -2.5f*mu2            + 1.0f};
  79     const float a2{-1.5f*mu3 +  2.0f*mu2 +  0.5f*mu};
  80     const float a3{ 0.5f*mu3 + -0.5f*mu2};
  81     return val1*a0 + val2*a1 + val3*a2 + val4*a3;
  82 }
  83
  84
  85 /** Find the next power-of-2 for non-power-of-2 numbers. */
  86 inline uint32_t NextPowerOf2(uint32_t value) noexcept
  87 {
  88     if(value > 0)
  89     {
  90         value--;
  91         value |= value>>1;
  92         value |= value>>2;
  93         value |= value>>4;
  94         value |= value>>8;
  95         value |= value>>16;
  96     }
  97     return value+1;
  98 }
  99
 100 /** Round up a value to the next multiple. */
 101 inline size_t RoundUp(size_t value, size_t r) noexcept
 102 {
 103     value += r-1;
 104     return value - (value%r);
 105 }
 106
 107
 108 /**
 109  * Fast float-to-int conversion. No particular rounding mode is assumed; the
 110  * IEEE-754 default is round-to-nearest with ties-to-even, though an app could
 111  * change it on its own threads. On some systems, a truncating conversion may
 112  * always be the fastest method.
 113  */
 114 inline int fastf2i(float f) noexcept
 115 {
 116 #if defined(HAVE_SSE_INTRINSICS)
 117     return _mm_cvt_ss2si(_mm_set_ss(f));
 118
 119 #elif defined(_MSC_VER) && defined(_M_IX86_FP)
 120
 121     int i;
 122     __asm fld f
 123     __asm fistp i
 124     return i;
 125
 126 #elif (defined(__GNUC__) || defined(__clang__)) && (defined(__i386__) || defined(__x86_64__))
 127
 128     int i;
 129 #ifdef __SSE_MATH__
 130     __asm__("cvtss2si %1, %0" : "=r"(i) : "x"(f));
 131 #else
 132     __asm__ __volatile__("fistpl %0" : "=m"(i) : "t"(f) : "st");
 133 #endif
 134     return i;
 135
 136 #else
 137
 138     return static_cast<int>(f);
 139 #endif
 140 }
 141 inline unsigned int fastf2u(float f) noexcept
 142 { return static_cast<unsigned int>(fastf2i(f)); }
 143
 144 /** Converts float-to-int using standard behavior (truncation). */
 145 inline int float2int(float f) noexcept
 146 {
 147 #if defined(HAVE_SSE_INTRINSICS)
 148     return _mm_cvtt_ss2si(_mm_set_ss(f));
 149
 150 #elif (defined(_MSC_VER) && defined(_M_IX86_FP) && _M_IX86_FP == 0) \
 151     || ((defined(__GNUC__) || defined(__clang__)) && (defined(__i386__) || defined(__x86_64__)) \
 152         && !defined(__SSE_MATH__))
 153     int sign, shift, mant;
 154     union {
 155         float f;
 156         int i;
 157     } conv;
 158
 159     conv.f = f;
 160     sign = (conv.i>>31) | 1;
 161     shift = ((conv.i>>23)&0xff) - (127+23);
 162
 163     /* Over/underflow */
 164     if UNLIKELY(shift >= 31 || shift < -23)
 165         return 0;
 166
 167     mant = (conv.i&0x7fffff) | 0x800000;
 168     if LIKELY(shift < 0)
 169         return (mant >> -shift) * sign;
 170     return (mant << shift) * sign;
 171
 172 #else
 173
 174     return static_cast<int>(f);
 175 #endif
 176 }
 177 inline unsigned int float2uint(float f) noexcept
 178 { return static_cast<unsigned int>(float2int(f)); }
 179
 180 /** Converts double-to-int using standard behavior (truncation). */
 181 inline int double2int(double d) noexcept
 182 {
 183 #if defined(HAVE_SSE_INTRINSICS)
 184     return _mm_cvttsd_si32(_mm_set_sd(d));
 185
 186 #elif (defined(_MSC_VER) && defined(_M_IX86_FP) && _M_IX86_FP < 2) \
 187     || ((defined(__GNUC__) || defined(__clang__)) && (defined(__i386__) || defined(__x86_64__)) \
 188         && !defined(__SSE2_MATH__))
 189     int sign, shift;
 190     int64_t mant;
 191     union {
 192         double d;
 193         int64_t i64;
 194     } conv;
 195
 196     conv.d = d;
 197     sign = (conv.i64 >> 63) | 1;
 198     shift = ((conv.i64 >> 52) & 0x7ff) - (1023 + 52);
 199
 200     /* Over/underflow */
 201     if UNLIKELY(shift >= 63 || shift < -52)
 202         return 0;
 203
 204     mant = (conv.i64 & 0xfffffffffffff_i64) | 0x10000000000000_i64;
 205     if LIKELY(shift < 0)
 206         return (int)(mant >> -shift) * sign;
 207     return (int)(mant << shift) * sign;
 208
 209 #else
 210
 211     return static_cast<int>(d);
 212 #endif
 213 }
 214
 215 /**
 216  * Rounds a float to the nearest integral value, according to the current
 217  * rounding mode. This is essentially an inlined version of rintf, although
 218  * makes fewer promises (e.g. -0 or -0.25 rounded to 0 may result in +0).
 219  */
 220 inline float fast_roundf(float f) noexcept
 221 {
 222 #if (defined(__GNUC__) || defined(__clang__)) && (defined(__i386__) || defined(__x86_64__)) \
 223     && !defined(__SSE_MATH__)
 224
 225     float out;
 226     __asm__ __volatile__("frndint" : "=t"(out) : "0"(f));
 227     return out;
 228
 229 #elif (defined(__GNUC__) || defined(__clang__)) && defined(__aarch64__)
 230
 231     float out;
 232     __asm__ volatile("frintx %s0, %s1" : "=w"(out) : "w"(f));
 233     return out;
 234
 235 #else
 236
 237     /* Integral limit, where sub-integral precision is not available for
 238      * floats.
 239      */
 240     static const float ilim[2]{
 241          8388608.0f /*  0x1.0p+23 */,
 242         -8388608.0f /* -0x1.0p+23 */
 243     };
 244     unsigned int sign, expo;
 245     union {
 246         float f;
 247         unsigned int i;
 248     } conv;
 249
 250     conv.f = f;
 251     sign = (conv.i>>31)&0x01;
 252     expo = (conv.i>>23)&0xff;
 253
 254     if UNLIKELY(expo >= 150/*+23*/)
 255     {
 256         /* An exponent (base-2) of 23 or higher is incapable of sub-integral
 257          * precision, so it's already an integral value. We don't need to worry
 258          * about infinity or NaN here.
 259          */
 260         return f;
 261     }
 262     /* Adding the integral limit to the value (with a matching sign) forces a
 263      * result that has no sub-integral precision, and is consequently forced to
 264      * round to an integral value. Removing the integral limit then restores
 265      * the initial value rounded to the integral. The compiler should not
 266      * optimize this out because of non-associative rules on floating-point
 267      * math (as long as you don't use -fassociative-math,
 268      * -funsafe-math-optimizations, -ffast-math, or -Ofast, in which case this
 269      * may break).
 270      */
 271     f += ilim[sign];
 272     return f - ilim[sign];
 273 #endif
 274 }
 275
 276
 277 template<typename T>
 278 constexpr const T& clamp(const T& value, const T& min_value, const T& max_value) noexcept
 279 {
 280     return std::min(std::max(value, min_value), max_value);
 281 }
 282
 283 // Converts level (mB) to gain.
 284 inline float level_mb_to_gain(float x)
 285 {
 286     if(x <= -10'000.0f)
 287         return 0.0f;
 288     return std::pow(10.0f, x / 2'000.0f);
 289 }
 290
 291 // Converts gain to level (mB).
 292 inline float gain_to_level_mb(float x)
 293 {
 294     if (x <= 0.0f)
 295         return -10'000.0f;
 296     return maxf(std::log10(x) * 2'000.0f, -10'000.0f);
 297 }
 298
 299 #endif /* AL_NUMERIC_H */