xref: /aosp_15_r20/external/mesa3d/src/util/double.h (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Mesa 3-D graphics library
3  *
4  * Copyright (C) 2018-2019 Intel Corporation
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included
14  * in all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
17  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  */
24 
25 #ifndef _DOUBLE_H_
26 #define _DOUBLE_H_
27 
28 #include "half_float.h"
29 #include "u_math.h"
30 
31 #ifdef __cplusplus
32 extern "C" {
33 #endif
34 
35 /*
36  * This API is no more than a wrapper to the counterpart softfloat.h
37  * calls. Still, softfloat.h conversion API is meant to be kept private. In
38  * other words, only use the API published here, instead of calling directly
39  * the softfloat.h one.
40  */
41 
42 float _mesa_double_to_float(double val);
43 float _mesa_double_to_float_rtz(double val);
44 
45 static inline float
_mesa_double_to_float_rtne(double val)46 _mesa_double_to_float_rtne(double val)
47 {
48    return _mesa_double_to_float(val);
49 }
50 
51 /*
52  * We round down from double to half float by going through float in between,
53  * but this can give us inaccurate results in some cases.
54  * One such case is 0x40ee6a0000000001, which should round to 0x7b9b, but
55  * going through float first turns into 0x7b9a instead. This is because the
56  * first non-fitting bit is set, so we get a tie, but with the least
57  * significant bit of the original number set, the tie should break rounding
58  * up.
59  * The cast to float, however, turns into 0x47735000, which when going to half
60  * still ties, but now we lost the tie-up bit, and instead we round to the
61  * nearest even, which in this case is down.
62  *
63  * To fix this, we check if the original would have tied, and if the tie would
64  * have rounded up, and if both are true, set the least significant bit of the
65  * intermediate float to 1, so that a tie on the next cast rounds up as well.
66  * If the rounding already got rid of the tie, that set bit will just be
67  * truncated anyway and the end result doesn't change.
68  *
69  * Another failing case is 0x40effdffffffffff. This one doesn't have the tie
70  * from double to half, so it just rounds down to 0x7bff (65504.0), but going
71  * through float first, it turns into 0x477ff000, which does have the tie bit
72  * for half set, and when that one gets rounded it turns into 0x7c00
73  * (Infinity).
74  * The fix for that one is to make sure the intermediate float does not have
75  * the tie bit set if the original didn't have it.
76  */
77 static inline uint16_t
_mesa_double_to_float16_rtne(double val)78 _mesa_double_to_float16_rtne(double val)
79 {
80    int significand_bits16 = 10;
81    int significand_bits32 = 23;
82    int significand_bits64 = 52;
83    int f64_to_16_tie_bit = significand_bits64 - significand_bits16 - 1;
84    int f32_to_16_tie_bit = significand_bits32 - significand_bits16 - 1;
85    uint64_t f64_rounds_up_mask = ((1ULL << f64_to_16_tie_bit) - 1);
86 
87    union di src;
88    union fi dst;
89 
90    src.d = val;
91    dst.f = val;
92 
93    bool f64_has_tie = (src.ui & (1ULL << f64_to_16_tie_bit)) != 0;
94    bool f64_rounds_up = (src.ui & f64_rounds_up_mask) != 0;
95 
96    dst.ui |= (f64_has_tie && f64_rounds_up);
97    if (!f64_has_tie)
98       dst.ui &= ~(1U << f32_to_16_tie_bit);
99 
100    return _mesa_float_to_float16_rtne(dst.f);
101 }
102 
103 /*
104  * double -> float -> half with RTZ doesn't have as many complications as
105  * RTNE, but we do need to ensure that the double -> float cast also uses RTZ.
106  */
107 static inline uint16_t
_mesa_double_to_float16_rtz(double val)108 _mesa_double_to_float16_rtz(double val)
109 {
110    return _mesa_float_to_float16_rtz(_mesa_double_to_float_rtz(val));
111 }
112 
113 #ifdef __cplusplus
114 } /* extern C */
115 #endif
116 
117 #endif /* _DOUBLE_H_ */
118