1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
|
// This code is in the public domain -- castano@gmail.com
#pragma once
#ifndef NV_MATH_FTOI_H
#define NV_MATH_FTOI_H
#include "nvmath/nvmath.h"
#include <math.h>
namespace nv
{
// Optimized float to int conversions. See:
// http://cbloomrants.blogspot.com/2009/01/01-17-09-float-to-int.html
// http://www.stereopsis.com/sree/fpu2006.html
// http://assemblyrequired.crashworks.org/2009/01/12/why-you-should-never-cast-floats-to-ints/
// http://chrishecker.com/Miscellaneous_Technical_Articles#Floating_Point
union DoubleAnd64 {
uint64 i;
double d;
};
static const double floatutil_xs_doublemagic = (6755399441055744.0); // 2^52 * 1.5
static const double floatutil_xs_doublemagicdelta = (1.5e-8); // almost .5f = .5f + 1e^(number of exp bit)
static const double floatutil_xs_doublemagicroundeps = (0.5f - floatutil_xs_doublemagicdelta); // almost .5f = .5f - 1e^(number of exp bit)
NV_FORCEINLINE int ftoi_round_xs(double val, double magic) {
#if 1
DoubleAnd64 dunion;
dunion.d = val + magic;
return (int32) dunion.i; // just cast to grab the bottom bits
#else
val += magic;
return ((int*)&val)[0]; // @@ Assumes little endian.
#endif
}
NV_FORCEINLINE int ftoi_round_xs(float val) {
return ftoi_round_xs(val, floatutil_xs_doublemagic);
}
NV_FORCEINLINE int ftoi_floor_xs(float val) {
return ftoi_round_xs(val - floatutil_xs_doublemagicroundeps, floatutil_xs_doublemagic);
}
NV_FORCEINLINE int ftoi_ceil_xs(float val) {
return ftoi_round_xs(val + floatutil_xs_doublemagicroundeps, floatutil_xs_doublemagic);
}
NV_FORCEINLINE int ftoi_trunc_xs(float val) {
return (val<0) ? ftoi_ceil_xs(val) : ftoi_floor_xs(val);
}
#if NV_CPU_X86 || NV_CPU_X86_64
NV_FORCEINLINE int ftoi_round_sse(float f) {
return _mm_cvt_ss2si(_mm_set_ss(f));
}
NV_FORCEINLINE int ftoi_trunc_sse(float f) {
return _mm_cvtt_ss2si(_mm_set_ss(f));
}
#endif
#if NV_USE_SSE
NV_FORCEINLINE int ftoi_round(float val) {
return ftoi_round_sse(val);
}
NV_FORCEINLINE int ftoi_trunc(float f) {
return ftoi_trunc_sse(f);
}
// We can probably do better than this. See for example:
// http://dss.stephanierct.com/DevBlog/?p=8
NV_FORCEINLINE int ftoi_floor(float val) {
return ftoi_round(floorf(val));
}
NV_FORCEINLINE int ftoi_ceil(float val) {
return ftoi_round(ceilf(val));
}
#else
// In theory this should work with any double floating point math implementation, but it appears that MSVC produces incorrect code
// when SSE2 is targeted and fast math is enabled (/arch:SSE2 & /fp:fast). These problems go away with /fp:precise, which is the default mode.
NV_FORCEINLINE int ftoi_round(float val) {
return ftoi_round_xs(val);
}
NV_FORCEINLINE int ftoi_floor(float val) {
return ftoi_floor_xs(val);
}
NV_FORCEINLINE int ftoi_ceil(float val) {
return ftoi_ceil_xs(val);
}
NV_FORCEINLINE int ftoi_trunc(float f) {
return ftoi_trunc_xs(f);
}
#endif
inline void test_ftoi() {
// Round to nearest integer.
nvCheck(ftoi_round(0.1f) == 0);
nvCheck(ftoi_round(0.6f) == 1);
nvCheck(ftoi_round(-0.2f) == 0);
nvCheck(ftoi_round(-0.7f) == -1);
nvCheck(ftoi_round(10.1f) == 10);
nvCheck(ftoi_round(10.6f) == 11);
nvCheck(ftoi_round(-90.1f) == -90);
nvCheck(ftoi_round(-90.6f) == -91);
nvCheck(ftoi_round(0) == 0);
nvCheck(ftoi_round(1) == 1);
nvCheck(ftoi_round(-1) == -1);
nvCheck(ftoi_round(0.5f) == 0); // How are midpoints rounded? Bankers rounding.
nvCheck(ftoi_round(1.5f) == 2);
nvCheck(ftoi_round(2.5f) == 2);
nvCheck(ftoi_round(3.5f) == 4);
nvCheck(ftoi_round(4.5f) == 4);
nvCheck(ftoi_round(-0.5f) == 0);
nvCheck(ftoi_round(-1.5f) == -2);
// Truncation (round down if > 0, round up if < 0).
nvCheck(ftoi_trunc(0.1f) == 0);
nvCheck(ftoi_trunc(0.6f) == 0);
nvCheck(ftoi_trunc(-0.2f) == 0);
nvCheck(ftoi_trunc(-0.7f) == 0); // @@ When using /arch:SSE2 in Win32, msvc produce wrong code for this one. It is skipping the addition.
nvCheck(ftoi_trunc(1.99f) == 1);
nvCheck(ftoi_trunc(-1.2f) == -1);
// Floor (round down).
nvCheck(ftoi_floor(0.1f) == 0);
nvCheck(ftoi_floor(0.6f) == 0);
nvCheck(ftoi_floor(-0.2f) == -1);
nvCheck(ftoi_floor(-0.7f) == -1);
nvCheck(ftoi_floor(1.99f) == 1);
nvCheck(ftoi_floor(-1.2f) == -2);
nvCheck(ftoi_floor(0) == 0);
nvCheck(ftoi_floor(1) == 1);
nvCheck(ftoi_floor(-1) == -1);
nvCheck(ftoi_floor(2) == 2);
nvCheck(ftoi_floor(-2) == -2);
// Ceil (round up).
nvCheck(ftoi_ceil(0.1f) == 1);
nvCheck(ftoi_ceil(0.6f) == 1);
nvCheck(ftoi_ceil(-0.2f) == 0);
nvCheck(ftoi_ceil(-0.7f) == 0);
nvCheck(ftoi_ceil(1.99f) == 2);
nvCheck(ftoi_ceil(-1.2f) == -1);
nvCheck(ftoi_ceil(0) == 0);
nvCheck(ftoi_ceil(1) == 1);
nvCheck(ftoi_ceil(-1) == -1);
nvCheck(ftoi_ceil(2) == 2);
nvCheck(ftoi_ceil(-2) == -2);
}
// Safe versions using standard casts.
inline int iround(float f)
{
return ftoi_round(f);
//return int(floorf(f + 0.5f));
}
inline int iround(double f)
{
return int(::floor(f + 0.5));
}
inline int ifloor(float f)
{
return ftoi_floor(f);
//return int(floorf(f));
}
inline int iceil(float f)
{
return int(ceilf(f));
}
// I'm always confused about which quantizer to use. I think we should choose a quantizer based on how the values are expanded later and this is generally using the 'exact endpoints' rule.
// Some notes from cbloom: http://cbloomrants.blogspot.com/2011/07/07-26-11-pixel-int-to-float-options.html
// Quantize a float in the [0,1] range, using exact end points or uniform bins.
inline float quantizeFloat(float x, uint bits, bool exactEndPoints = true) {
nvDebugCheck(bits <= 16);
float range = float(1 << bits);
if (exactEndPoints) {
return floorf(x * (range-1) + 0.5f) / (range-1);
}
else {
return (floorf(x * range) + 0.5f) / range;
}
}
// This is the most common rounding mode:
//
// 0 1 2 3
// |___|_______|_______|___|
// 0 1
//
// You get that if you take the unit floating point number multiply by 'N-1' and round to nearest. That is, `i = round(f * (N-1))`.
// You reconstruct the original float dividing by 'N-1': `f = i / (N-1)`
// 0 1 2 3
// |_____|_____|_____|_____|
// 0 1
/*enum BinningMode {
RoundMode_ExactEndPoints,
RoundMode_UniformBins,
};*/
template <int N>
inline uint unitFloatToFixed(float f) {
return ftoi_round(f * ((1<<N)-1));
}
inline uint8 unitFloatToFixed8(float f) {
return (uint8)unitFloatToFixed<8>(f);
}
inline uint16 unitFloatToFixed16(float f) {
return (uint16)unitFloatToFixed<16>(f);
}
} // nv
#endif // NV_MATH_FTOI_H
|