• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2021 Alyssa Rosenzweig <alyssa@rosenzweig.io>
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21  * SOFTWARE.
22  */
23 
24 #ifndef __AGX_MINIFLOAT_H_
25 #define __AGX_MINIFLOAT_H_
26 
27 #include <math.h>
28 #include "util/macros.h"
29 
30 /* AGX includes an 8-bit floating-point format for small dyadic immediates,
31  * consisting of 3 bits for the exponent, 4 bits for the mantissa, and 1-bit
32  * for sign, in the usual order. Zero exponent has special handling. */
33 
34 static inline float
agx_minifloat_decode(uint8_t imm)35 agx_minifloat_decode(uint8_t imm)
36 {
37    float sign = (imm & 0x80) ? -1.0 : 1.0;
38    signed exp = (imm & 0x70) >> 4;
39    unsigned mantissa = (imm & 0xF);
40 
41    if (exp)
42       return ldexpf(sign * (float) (mantissa | 0x10), exp - 7);
43    else
44       return ldexpf(sign * ((float) mantissa), -6);
45 }
46 
47 /* Encodes a float. Results are only valid if the float can be represented
48  * exactly, if not the result of this function is UNDEFINED. signbit() is used
49  * to ensure -0.0 is handled correctly. */
50 
51 static inline uint8_t
agx_minifloat_encode(float f)52 agx_minifloat_encode(float f)
53 {
54    unsigned sign = signbit(f) ? 0x80 : 0;
55    f = fabsf(f);
56 
57    /* frac is in [0.5, 1) and f = frac * 2^exp */
58    int exp = 0;
59    float frac = frexpf(f, &exp);
60 
61    if (f >= 0.25) {
62       unsigned mantissa = (frac * 32.0);
63       exp -= 5; /* 2^5 = 32 */
64       exp = CLAMP(exp + 7, 0, 7);
65 
66       assert(mantissa >= 0x10 && mantissa < 0x20);
67       assert(exp >= 1);
68 
69       return sign | (exp << 4) | (mantissa & 0xF);
70    } else {
71       unsigned mantissa = (f * 64.0f);
72       assert(mantissa < 0x10);
73 
74       return sign | mantissa;
75    }
76 }
77 
78 static inline bool
agx_minifloat_exact(float f)79 agx_minifloat_exact(float f)
80 {
81    float f_ = agx_minifloat_decode(agx_minifloat_encode(f));
82    return memcmp(&f, &f_, sizeof(float)) == 0;
83 }
84 
85 #endif
86