1 /*
2 * Copyright (C) 2011 Marek Olšák <[email protected]>
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
22 */
23
24 /* Based on code from The OpenGL Programming Guide / 7th Edition, Appendix J.
25 * Available here: http://www.opengl-redbook.com/appendices/
26 * The algorithm in the book contains a bug though, which is fixed in the code
27 * below.
28 */
29
30 #ifndef FORMAT_R11G11B10F_H
31 #define FORMAT_R11G11B10F_H
32
33 #include <stdint.h>
34
35 #include "rounding.h"
36
37 #define UF11(e, m) ((e << 6) | (m))
38 #define UF11_EXPONENT_BIAS 15
39 #define UF11_EXPONENT_BITS 0x1F
40 #define UF11_EXPONENT_SHIFT 6
41 #define UF11_MANTISSA_BITS 0x3F
42 #define UF11_MANTISSA_SHIFT (23 - UF11_EXPONENT_SHIFT)
43 #define UF11_MAX_EXPONENT (UF11_EXPONENT_BITS << UF11_EXPONENT_SHIFT)
44
45 #define UF10(e, m) ((e << 5) | (m))
46 #define UF10_EXPONENT_BIAS 15
47 #define UF10_EXPONENT_BITS 0x1F
48 #define UF10_EXPONENT_SHIFT 5
49 #define UF10_MANTISSA_BITS 0x1F
50 #define UF10_MANTISSA_SHIFT (23 - UF10_EXPONENT_SHIFT)
51 #define UF10_MAX_EXPONENT (UF10_EXPONENT_BITS << UF10_EXPONENT_SHIFT)
52
53 #define F32_INFINITY 0x7f800000
54
f32_to_uf11(float val)55 static inline uint32_t f32_to_uf11(float val)
56 {
57 union {
58 float f;
59 uint32_t ui;
60 } f32 = {val};
61
62 uint16_t uf11 = 0;
63
64 /* Decode little-endian 32-bit floating-point value */
65 int sign = (f32.ui >> 16) & 0x8000;
66 /* Map exponent to the range [-127,128] */
67 int exponent = ((f32.ui >> 23) & 0xff) - 127;
68 int mantissa = f32.ui & 0x007fffff;
69
70 if (exponent == 128) { /* Infinity or NaN */
71 /* From the GL_EXT_packed_float spec:
72 *
73 * "Additionally: negative infinity is converted to zero; positive
74 * infinity is converted to positive infinity; and both positive and
75 * negative NaN are converted to positive NaN."
76 */
77 uf11 = UF11_MAX_EXPONENT;
78 if (mantissa) {
79 uf11 |= 1; /* NaN */
80 } else {
81 if (sign)
82 uf11 = 0; /* 0.0 */
83 }
84 } else if (sign) {
85 return 0;
86 } else if (val > 65024.0f) {
87 /* From the GL_EXT_packed_float spec:
88 *
89 * "Likewise, finite positive values greater than 65024 (the maximum
90 * finite representable unsigned 11-bit floating-point value) are
91 * converted to 65024."
92 */
93 uf11 = UF11(30, 63);
94 } else if (exponent > -15) { /* Normal value */
95 /* Dividing by 2^exponent gives us a number in the range [1, 2).
96 * Multiplying by 2^6=64 gives us our mantissa, plus an extra 1 which
97 * we'll mask off.
98 */
99 mantissa = _mesa_lroundevenf(ldexp(val, 6 - exponent));
100 if (mantissa >= 2 << UF11_EXPONENT_SHIFT) {
101 /* The float32 was rounded upwards into the range of the next
102 * exponent, so bump the exponent.
103 */
104 assert(mantissa == 2 << UF11_EXPONENT_SHIFT);
105 mantissa >>= 1;
106 exponent++;
107 }
108 assert((mantissa >> UF11_EXPONENT_SHIFT) == 1);
109 mantissa &= UF11_MANTISSA_BITS;
110 exponent += UF11_EXPONENT_BIAS;
111 uf11 = UF11(exponent, mantissa);
112 } else { /* Zero or denormal */
113 /* Since exponent <= -15, Multiplying by 2^14 gives us a number in the
114 * range [0, 1). Multiplying by 2^6=64 gives us our mantissa.
115 */
116 mantissa = _mesa_lroundevenf(ldexp(val, 6 + 14));
117
118 /* It's possible that we get a normal after rounding */
119 if ((mantissa >> UF11_EXPONENT_SHIFT) != 0) {
120 assert(mantissa == (1 << UF11_EXPONENT_SHIFT));
121 uf11 = UF11(1, 0);
122 } else {
123 uf11 = UF11(0, mantissa);
124 }
125 }
126
127 return uf11;
128 }
129
uf11_to_f32(uint16_t val)130 static inline float uf11_to_f32(uint16_t val)
131 {
132 union {
133 float f;
134 uint32_t ui;
135 } f32;
136
137 int exponent = (val & 0x07c0) >> UF11_EXPONENT_SHIFT;
138 int mantissa = (val & 0x003f);
139
140 f32.f = 0.0;
141
142 if (exponent == 0) {
143 if (mantissa != 0) {
144 const float scale = 1.0 / (1 << 20);
145 f32.f = scale * mantissa;
146 }
147 } else if (exponent == 31) {
148 f32.ui = F32_INFINITY | mantissa;
149 } else {
150 float scale, decimal;
151 exponent -= 15;
152 if (exponent < 0) {
153 scale = 1.0f / (1 << -exponent);
154 } else {
155 scale = (float) (1 << exponent);
156 }
157 decimal = 1.0f + (float) mantissa / 64;
158 f32.f = scale * decimal;
159 }
160
161 return f32.f;
162 }
163
f32_to_uf10(float val)164 static inline uint32_t f32_to_uf10(float val)
165 {
166 union {
167 float f;
168 uint32_t ui;
169 } f32 = {val};
170
171 uint16_t uf10 = 0;
172
173 /* Decode little-endian 32-bit floating-point value */
174 int sign = (f32.ui >> 16) & 0x8000;
175 /* Map exponent to the range [-127,128] */
176 int exponent = ((f32.ui >> 23) & 0xff) - 127;
177 int mantissa = f32.ui & 0x007fffff;
178
179 if (exponent == 128) {
180 /* From the GL_EXT_packed_float spec:
181 *
182 * "Additionally: negative infinity is converted to zero; positive
183 * infinity is converted to positive infinity; and both positive and
184 * negative NaN are converted to positive NaN."
185 */
186 uf10 = UF10_MAX_EXPONENT;
187 if (mantissa) {
188 uf10 |= 1; /* NaN */
189 } else {
190 if (sign)
191 uf10 = 0; /* 0.0 */
192 }
193 } else if (sign) {
194 return 0;
195 } else if (val > 64512.0f) {
196 /* From the GL_EXT_packed_float spec:
197 *
198 * "Likewise, finite positive values greater than 64512 (the maximum
199 * finite representable unsigned 10-bit floating-point value) are
200 * converted to 64512."
201 */
202 uf10 = UF10(30, 31);
203 } else if (exponent > -15) { /* Normal value */
204 /* Dividing by 2^exponent gives us a number in the range [1, 2).
205 * Multiplying by 2^5=32 gives us our mantissa, plus an extra 1 which
206 * we'll mask off.
207 */
208 mantissa = _mesa_lroundevenf(ldexp(val, 5 - exponent));
209 if (mantissa >= 2 << UF10_EXPONENT_SHIFT) {
210 /* The float32 was rounded upwards into the range of the next
211 * exponent, so bump the exponent.
212 */
213 assert(mantissa == 2 << UF10_EXPONENT_SHIFT);
214 mantissa >>= 1;
215 exponent++;
216 }
217 assert((mantissa >> UF10_EXPONENT_SHIFT) == 1);
218 mantissa &= UF10_MANTISSA_BITS;
219 exponent += UF10_EXPONENT_BIAS;
220 uf10 = UF10(exponent, mantissa);
221 } else { /* Zero or denormal */
222 /* Since exponent <= -15, Multiplying by 2^14 gives us a number in the
223 * range [0, 1). Multiplying by 2^5=32 gives us our mantissa.
224 */
225 mantissa = _mesa_lroundevenf(ldexp(val, 5 + 14));
226
227 /* It's possible that we get a normal after rounding */
228 if ((mantissa >> UF10_EXPONENT_SHIFT) != 0) {
229 assert(mantissa == (1 << UF10_EXPONENT_SHIFT));
230 uf10 = UF10(1, 0);
231 } else {
232 uf10 = UF10(0, mantissa);
233 }
234 }
235
236 return uf10;
237 }
238
uf10_to_f32(uint16_t val)239 static inline float uf10_to_f32(uint16_t val)
240 {
241 union {
242 float f;
243 uint32_t ui;
244 } f32;
245
246 int exponent = (val & 0x03e0) >> UF10_EXPONENT_SHIFT;
247 int mantissa = (val & 0x001f);
248
249 f32.f = 0.0;
250
251 if (exponent == 0) {
252 if (mantissa != 0) {
253 const float scale = 1.0 / (1 << 19);
254 f32.f = scale * mantissa;
255 }
256 } else if (exponent == 31) {
257 f32.ui = F32_INFINITY | mantissa;
258 } else {
259 float scale, decimal;
260 exponent -= 15;
261 if (exponent < 0) {
262 scale = 1.0f / (1 << -exponent);
263 }
264 else {
265 scale = (float) (1 << exponent);
266 }
267 decimal = 1.0f + (float) mantissa / 32;
268 f32.f = scale * decimal;
269 }
270
271 return f32.f;
272 }
273
float3_to_r11g11b10f(const float rgb[3])274 static inline uint32_t float3_to_r11g11b10f(const float rgb[3])
275 {
276 return ( f32_to_uf11(rgb[0]) & 0x7ff) |
277 ((f32_to_uf11(rgb[1]) & 0x7ff) << 11) |
278 ((f32_to_uf10(rgb[2]) & 0x3ff) << 22);
279 }
280
r11g11b10f_to_float3(uint32_t rgb,float retval[3])281 static inline void r11g11b10f_to_float3(uint32_t rgb, float retval[3])
282 {
283 retval[0] = uf11_to_f32( rgb & 0x7ff);
284 retval[1] = uf11_to_f32((rgb >> 11) & 0x7ff);
285 retval[2] = uf10_to_f32((rgb >> 22) & 0x3ff);
286 }
287
288 #endif /* FORMAT_R11G11B10F_H */
289