1 /* 2 * Mesa 3-D graphics library 3 * 4 * Copyright (C) 1999-2007 Brian Paul All Rights Reserved. 5 * Copyright 2015 Philip Taylor <philip@zaynar.co.uk> 6 * Copyright 2018 Advanced Micro Devices, Inc. 7 * Copyright (C) 2018-2019 Intel Corporation 8 * 9 * Permission is hereby granted, free of charge, to any person obtaining a 10 * copy of this software and associated documentation files (the "Software"), 11 * to deal in the Software without restriction, including without limitation 12 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 13 * and/or sell copies of the Software, and to permit persons to whom the 14 * Software is furnished to do so, subject to the following conditions: 15 * 16 * The above copyright notice and this permission notice shall be included 17 * in all copies or substantial portions of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 23 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 24 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 25 * OTHER DEALINGS IN THE SOFTWARE. 26 */ 27 28 #include <math.h> 29 #include <assert.h> 30 #include "half_float.h" 31 #include "rounding.h" 32 #include "softfloat.h" 33 #include "macros.h" 34 #include "u_math.h" 35 36 typedef union { float f; int32_t i; uint32_t u; } fi_type; 37 38 /** 39 * Convert a 4-byte float to a 2-byte half float. 40 * 41 * Not all float32 values can be represented exactly as a float16 value. We 42 * round such intermediate float32 values to the nearest float16. When the 43 * float32 lies exactly between to float16 values, we round to the one with 44 * an even mantissa. 45 * 46 * This rounding behavior has several benefits: 47 * - It has no sign bias. 48 * 49 * - It reproduces the behavior of real hardware: opcode F32TO16 in Intel's 50 * GPU ISA. 51 * 52 * - By reproducing the behavior of the GPU (at least on Intel hardware), 53 * compile-time evaluation of constant packHalf2x16 GLSL expressions will 54 * result in the same value as if the expression were executed on the GPU. 55 */ 56 uint16_t _mesa_float_to_half_slow(float val)57 _mesa_float_to_half_slow(float val) 58 { 59 const fi_type fi = {val}; 60 const int flt_m = fi.i & 0x7fffff; 61 const int flt_e = (fi.i >> 23) & 0xff; 62 const int flt_s = (fi.i >> 31) & 0x1; 63 int s, e, m = 0; 64 uint16_t result; 65 66 /* sign bit */ 67 s = flt_s; 68 69 /* handle special cases */ 70 if ((flt_e == 0) && (flt_m == 0)) { 71 /* zero */ 72 /* m = 0; - already set */ 73 e = 0; 74 } 75 else if ((flt_e == 0) && (flt_m != 0)) { 76 /* denorm -- denorm float maps to 0 half */ 77 /* m = 0; - already set */ 78 e = 0; 79 } 80 else if ((flt_e == 0xff) && (flt_m == 0)) { 81 /* infinity */ 82 /* m = 0; - already set */ 83 e = 31; 84 } 85 else if ((flt_e == 0xff) && (flt_m != 0)) { 86 /* Retain the top bits of a NaN to make sure that the quiet/signaling 87 * status stays the same. 88 */ 89 m = flt_m >> 13; 90 if (!m) 91 m = 1; 92 e = 31; 93 } 94 else { 95 /* regular number */ 96 const int new_exp = flt_e - 127; 97 if (new_exp < -14) { 98 /* The float32 lies in the range (0.0, min_normal16) and is rounded 99 * to a nearby float16 value. The result will be either zero, subnormal, 100 * or normal. 101 */ 102 e = 0; 103 m = _mesa_lroundevenf((1 << 24) * fabsf(fi.f)); 104 } 105 else if (new_exp > 15) { 106 /* map this value to infinity */ 107 /* m = 0; - already set */ 108 e = 31; 109 } 110 else { 111 /* The float32 lies in the range 112 * [min_normal16, max_normal16 + max_step16) 113 * and is rounded to a nearby float16 value. The result will be 114 * either normal or infinite. 115 */ 116 e = new_exp + 15; 117 m = _mesa_lroundevenf(flt_m / (float) (1 << 13)); 118 } 119 } 120 121 assert(0 <= m && m <= 1024); 122 if (m == 1024) { 123 /* The float32 was rounded upwards into the range of the next exponent, 124 * so bump the exponent. This correctly handles the case where f32 125 * should be rounded up to float16 infinity. 126 */ 127 ++e; 128 m = 0; 129 } 130 131 result = (s << 15) | (e << 10) | m; 132 return result; 133 } 134 135 uint16_t _mesa_float_to_float16_rtz_slow(float val)136 _mesa_float_to_float16_rtz_slow(float val) 137 { 138 return _mesa_float_to_half_rtz_slow(val); 139 } 140 141 /** 142 * Convert a 2-byte half float to a 4-byte float. 143 * Based on code from: 144 * http://www.opengl.org/discussion_boards/ubb/Forum3/HTML/008786.html 145 */ 146 float _mesa_half_to_float_slow(uint16_t val)147 _mesa_half_to_float_slow(uint16_t val) 148 { 149 union fi infnan; 150 union fi magic; 151 union fi f32; 152 153 infnan.ui = 0x8f << 23; 154 infnan.f = 65536.0f; 155 magic.ui = 0xef << 23; 156 157 /* Exponent / Mantissa */ 158 f32.ui = (val & 0x7fff) << 13; 159 160 /* Adjust */ 161 f32.f *= magic.f; 162 /* XXX: The magic mul relies on denorms being available */ 163 164 /* Inf / NaN */ 165 if (f32.f >= infnan.f) 166 f32.ui |= 0xff << 23; 167 168 /* Sign */ 169 f32.ui |= (uint32_t)(val & 0x8000) << 16; 170 171 return f32.f; 172 } 173 174 /** 175 * Convert 0.0 to 0x00, 1.0 to 0xff. 176 * Values outside the range [0.0, 1.0] will give undefined results. 177 */ _mesa_half_to_unorm8(uint16_t val)178 uint8_t _mesa_half_to_unorm8(uint16_t val) 179 { 180 const int m = val & 0x3ff; 181 const int e = (val >> 10) & 0x1f; 182 ASSERTED const int s = (val >> 15) & 0x1; 183 184 /* v = round_to_nearest(1.mmmmmmmmmm * 2^(e-15) * 255) 185 * = round_to_nearest((1.mmmmmmmmmm * 255) * 2^(e-15)) 186 * = round_to_nearest((1mmmmmmmmmm * 255) * 2^(e-25)) 187 * = round_to_zero((1mmmmmmmmmm * 255) * 2^(e-25) + 0.5) 188 * = round_to_zero(((1mmmmmmmmmm * 255) * 2^(e-24) + 1) / 2) 189 * 190 * This happens to give the correct answer for zero/subnormals too 191 */ 192 assert(s == 0 && val <= FP16_ONE); /* check 0 <= this <= 1 */ 193 /* (implies e <= 15, which means the bit-shifts below are safe) */ 194 195 uint32_t v = ((1 << 10) | m) * 255; 196 v = ((v >> (24 - e)) + 1) >> 1; 197 return v; 198 } 199 200 /** 201 * Takes a uint16_t, divides by 65536, converts the infinite-precision 202 * result to fp16 with round-to-zero. Used by the ASTC decoder. 203 */ _mesa_uint16_div_64k_to_half(uint16_t v)204 uint16_t _mesa_uint16_div_64k_to_half(uint16_t v) 205 { 206 /* Zero or subnormal. Set the mantissa to (v << 8) and return. */ 207 if (v < 4) 208 return v << 8; 209 210 /* Count the leading 0s in the uint16_t */ 211 #ifdef HAVE___BUILTIN_CLZ 212 int n = __builtin_clz(v) - 16; 213 #else 214 int n = 16; 215 for (int i = 15; i >= 0; i--) { 216 if (v & (1 << i)) { 217 n = 15 - i; 218 break; 219 } 220 } 221 #endif 222 223 /* Shift the mantissa up so bit 16 is the hidden 1 bit, 224 * mask it off, then shift back down to 10 bits 225 */ 226 int m = ( ((uint32_t)v << (n + 1)) & 0xffff ) >> 6; 227 228 /* (0{n} 1 X{15-n}) * 2^-16 229 * = 1.X * 2^(15-n-16) 230 * = 1.X * 2^(14-n - 15) 231 * which is the FP16 form with e = 14 - n 232 */ 233 int e = 14 - n; 234 235 assert(e >= 1 && e <= 30); 236 assert(m >= 0 && m < 0x400); 237 238 return (e << 10) | m; 239 } 240