1 /*
2  * Copyright (C) 1999-2007  Brian Paul   All Rights Reserved.
3  * Copyright 2015 Philip Taylor <philip@zaynar.co.uk>
4  * Copyright 2018 Advanced Micro Devices, Inc.
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 #include <assert.h>
26 #include <math.h>
27 
28 #include "igt_halffloat.h"
29 #include "igt_x86.h"
30 
31 typedef union { float f; int32_t i; uint32_t u; } fi_type;
32 
33 /**
34  * Convert a 4-byte float to a 2-byte half float.
35  *
36  * Not all float32 values can be represented exactly as a float16 value. We
37  * round such intermediate float32 values to the nearest float16. When the
38  * float32 lies exactly between to float16 values, we round to the one with
39  * an even mantissa.
40  *
41  * This rounding behavior has several benefits:
42  *   - It has no sign bias.
43  *
44  *   - It reproduces the behavior of real hardware: opcode F32TO16 in Intel's
45  *     GPU ISA.
46  *
47  *   - By reproducing the behavior of the GPU (at least on Intel hardware),
48  *     compile-time evaluation of constant packHalf2x16 GLSL expressions will
49  *     result in the same value as if the expression were executed on the GPU.
50  */
_float_to_half(float val)51 static inline uint16_t _float_to_half(float val)
52 {
53 	const fi_type fi = {val};
54 	const int flt_m = fi.i & 0x7fffff;
55 	const int flt_e = (fi.i >> 23) & 0xff;
56 	const int flt_s = (fi.i >> 31) & 0x1;
57 	int s, e, m = 0;
58 	uint16_t result;
59 
60 	/* sign bit */
61 	s = flt_s;
62 
63 	/* handle special cases */
64 	if ((flt_e == 0) && (flt_m == 0)) {
65 		/* zero */
66 		/* m = 0; - already set */
67 		e = 0;
68 	} else if ((flt_e == 0) && (flt_m != 0)) {
69 		/* denorm -- denorm float maps to 0 half */
70 		/* m = 0; - already set */
71 		e = 0;
72 	} else if ((flt_e == 0xff) && (flt_m == 0)) {
73 		/* infinity */
74 		/* m = 0; - already set */
75 		e = 31;
76 	} else if ((flt_e == 0xff) && (flt_m != 0)) {
77 		/* NaN */
78 		m = 1;
79 		e = 31;
80 	} else {
81 		/* regular number */
82 		const int new_exp = flt_e - 127;
83 		if (new_exp < -14) {
84 			/* The float32 lies in the range (0.0, min_normal16) and
85 			 * is rounded to a nearby float16 value. The result will
86 			 * be either zero, subnormal, or normal.
87 			 */
88 			e = 0;
89 			m = lrintf((1 << 24) * fabsf(fi.f));
90 		} else if (new_exp > 15) {
91 			/* map this value to infinity */
92 			/* m = 0; - already set */
93 			e = 31;
94 		} else {
95 			/* The float32 lies in the range
96 			 *   [min_normal16, max_normal16 + max_step16)
97 			 * and is rounded to a nearby float16 value. The result
98 			 * will be either normal or infinite.
99 			 */
100 			e = new_exp + 15;
101 			m = lrintf(flt_m / (float)(1 << 13));
102 		}
103 	}
104 
105 	assert(0 <= m && m <= 1024);
106 	if (m == 1024) {
107 		/* The float32 was rounded upwards into the range of the next
108 		 * exponent, so bump the exponent. This correctly handles the
109 		 * case where f32 should be rounded up to float16 infinity.
110 		 */
111 		++e;
112 		m = 0;
113 	}
114 
115 	result = (s << 15) | (e << 10) | m;
116 	return result;
117 }
118 
119 /**
120  * Convert a 2-byte half float to a 4-byte float.
121  * Based on code from:
122  * http://www.opengl.org/discussion_boards/ubb/Forum3/HTML/008786.html
123  */
_half_to_float(uint16_t val)124 static inline float _half_to_float(uint16_t val)
125 {
126 	/* XXX could also use a 64K-entry lookup table */
127 	const int m = val & 0x3ff;
128 	const int e = (val >> 10) & 0x1f;
129 	const int s = (val >> 15) & 0x1;
130 	int flt_m, flt_e, flt_s;
131 	fi_type fi;
132 
133 	/* sign bit */
134 	flt_s = s;
135 
136 	/* handle special cases */
137 	if ((e == 0) && (m == 0)) {
138 		/* zero */
139 		flt_m = 0;
140 		flt_e = 0;
141 	} else if ((e == 0) && (m != 0)) {
142 		/* denorm -- denorm half will fit in non-denorm single */
143 		const float half_denorm = 1.0f / 16384.0f; /* 2^-14 */
144 		float mantissa = ((float) (m)) / 1024.0f;
145 		float sign = s ? -1.0f : 1.0f;
146 		return sign * mantissa * half_denorm;
147 	} else if ((e == 31) && (m == 0)) {
148 		/* infinity */
149 		flt_e = 0xff;
150 		flt_m = 0;
151 	} else if ((e == 31) && (m != 0)) {
152 		/* NaN */
153 		flt_e = 0xff;
154 		flt_m = 1;
155 	} else {
156 		/* regular */
157 		flt_e = e + 112;
158 		flt_m = m << 13;
159 	}
160 
161 	fi.i = (flt_s << 31) | (flt_e << 23) | flt_m;
162 	return fi.f;
163 }
164 
165 #if defined(__x86_64__) && !defined(__clang__)
166 #pragma GCC push_options
167 #pragma GCC target("f16c")
168 
169 #include <immintrin.h>
170 
float_to_half_f16c(const float * f,uint16_t * h,unsigned int num)171 static void float_to_half_f16c(const float *f, uint16_t *h, unsigned int num)
172 {
173 	for (int i = 0; i < num; i++)
174 		h[i] = _cvtss_sh(f[i], 0);
175 }
176 
half_to_float_f16c(const uint16_t * h,float * f,unsigned int num)177 static void half_to_float_f16c(const uint16_t *h, float *f, unsigned int num)
178 {
179 	for (int i = 0; i < num; i++)
180 		f[i] = _cvtsh_ss(h[i]);
181 }
182 
183 #pragma GCC pop_options
184 
float_to_half(const float * f,uint16_t * h,unsigned int num)185 static void float_to_half(const float *f, uint16_t *h, unsigned int num)
186 {
187 	for (int i = 0; i < num; i++)
188 		h[i] = _float_to_half(f[i]);
189 }
190 
half_to_float(const uint16_t * h,float * f,unsigned int num)191 static void half_to_float(const uint16_t *h, float *f, unsigned int num)
192 {
193 	for (int i = 0; i < num; i++)
194 		f[i] = _half_to_float(h[i]);
195 }
196 
resolve_float_to_half(void)197 static void (*resolve_float_to_half(void))(const float *f, uint16_t *h, unsigned int num)
198 {
199 	if (igt_x86_features() & F16C)
200 		return float_to_half_f16c;
201 
202 	return float_to_half;
203 }
204 
205 void igt_float_to_half(const float *f, uint16_t *h, unsigned int num)
206 	__attribute__((ifunc("resolve_float_to_half")));
207 
resolve_half_to_float(void)208 static void (*resolve_half_to_float(void))(const uint16_t *h, float *f, unsigned int num)
209 {
210 	if (igt_x86_features() & F16C)
211 		return half_to_float_f16c;
212 
213 	return half_to_float;
214 }
215 
216 void igt_half_to_float(const uint16_t *h, float *f, unsigned int num)
217 	__attribute__((ifunc("resolve_half_to_float")));
218 
219 #else
220 
igt_float_to_half(const float * f,uint16_t * h,unsigned int num)221 void igt_float_to_half(const float *f, uint16_t *h, unsigned int num)
222 {
223 	for (int i = 0; i < num; i++)
224 		h[i] = _float_to_half(f[i]);
225 }
226 
igt_half_to_float(const uint16_t * h,float * f,unsigned int num)227 void igt_half_to_float(const uint16_t *h, float *f, unsigned int num)
228 {
229 	for (int i = 0; i < num; i++)
230 		f[i] = _half_to_float(h[i]);
231 }
232 
233 #endif
234 
235