Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / d051e97e36b9981f411093904cc019c2c7f9ac75 / . / src / core / CL / cl_kernels / fixed_point.h
blob: 46fa645c2ba8caf10432976297c61997810b477e [file] [log] [blame]
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]1/*
Gian Marco76faef82018-01-29 12:15:32 +0000[diff] [blame]2 * Copyright (c) 2017-2018 ARM Limited.
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24#ifndef ARM_COMPUTE_FIXED_POINT_H
25#define ARM_COMPUTE_FIXED_POINT_H
26
27#define TYPE_ALIAS(type, alias) \
28 typedef type alias; \
29 typedef type alias##x##1; \
30 typedef type##2 alias##x##2; \
31 typedef type##3 alias##x##3; \
32 typedef type##4 alias##x##4; \
33 typedef type##8 alias##x##8; \
34 typedef type##16 alias##x##16;
35
36TYPE_ALIAS(char, qs8)
37TYPE_ALIAS(short, qs16)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]38TYPE_ALIAS(int, qs32)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]39
40#define qs8_MIN ((char)CHAR_MIN)
41#define qs8_MAX ((char)CHAR_MAX)
42#define qs16_MIN ((short)SHRT_MIN)
43#define qs16_MAX ((short)SHRT_MAX)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]44#define qs32_MIN ((int)INT_MIN)
45#define qs32_MAX ((int)INT_MAX)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]46
47#define qu8_MIN ((uchar)0)
48#define qu8_MAX ((uchar)UCHAR_MAX)
49#define qu16_MIN ((ushort)0)
50#define qu16_MAX ((ushort)USHRT_MAX)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]51#define qu32_MIN ((uint)0)
52#define qu32_MAX ((uint)UINT_MAX)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]53
54#define qs8_TYPE char
55#define qs8x1_TYPE char
56#define qs8x2_TYPE char2
Gian Marco Iodice3a623242017-07-25 10:25:53 +0100[diff] [blame]57#define qs8x3_TYPE char3
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]58#define qs8x4_TYPE char4
59#define qs8x8_TYPE char8
60#define qs8x16_TYPE char16
61
62#define qs16_TYPE short
63#define qs16x1_TYPE short
64#define qs16x2_TYPE short2
Gian Marco Iodice3a623242017-07-25 10:25:53 +0100[diff] [blame]65#define qs16x3_TYPE short3
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]66#define qs16x4_TYPE short4
67#define qs16x8_TYPE short8
68#define qs16x16_TYPE short16
69
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]70#define qs32_TYPE int
71#define qs32x1_TYPE int
72#define qs32x2_TYPE int2
Gian Marco Iodice3a623242017-07-25 10:25:53 +0100[diff] [blame]73#define qs32x3_TYPE int3
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]74#define qs32x4_TYPE int4
75#define qs32x8_TYPE int8
76#define qs32x16_TYPE int16
77
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]78/* All internal constants are represented in the maximum supported fixed point format (QS16),
79 * thus we define an additional shift parameter required to convert the constant
80 * from the maximum supported format to the require one.
81 */
82#define qs8_SHIFT 8
83#define qs16_SHIFT 0
84
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]85#undef VEC_DATA_TYPE_STR
86#undef VEC_DATA_TYPE
87#undef CONVERT_STR
88#undef CONVERT
89#undef CONVERT_SAT_STR
90#undef CONVERT_SAT
91
92#define VEC_DATA_TYPE_STR(type, size) type##x##size
93#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)
94
95#define CONVERT_STR3(x, type, rtype) (convert_##rtype((x)))
96#define CONVERT_STR2(x, type, rtype) CONVERT_STR3(x, type, rtype)
97#define CONVERT_STR(x, type) CONVERT_STR2(x, type, type##_TYPE)
98#define CONVERT(x, type) CONVERT_STR(x, type)
99
100#define CONVERT_SAT_STR3(x, type, rtype) (convert_##rtype##_sat((x)))
101#define CONVERT_SAT_STR2(x, type, rtype) CONVERT_SAT_STR3(x, type, rtype)
102#define CONVERT_SAT_STR(x, type) CONVERT_SAT_STR2(x, type, type##_TYPE)
103#define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)
104
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]105/** Computes saturating absolute value of fixed point vector.
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]106 *
107 * @param[in] type the actual data type.
108 *
109 * @return The result of the fixed point absolute value.
110 */
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]111#define ABSQ_SAT_IMPL(type) \
112 inline type abs_##type##_sat(type VopA) \
113 { \
114 return CONVERT_SAT(abs(VopA), type); \
115 }
116
117ABSQ_SAT_IMPL(qs8x16)
118ABSQ_SAT_IMPL(qs16x8)
119
120#define ABS_SAT_OP_EXPAND_STR(a, type, size) abs_##type##x##size##_sat((a))
121#define ABS_SAT_OP_EXPAND(a, type, size) ABS_SAT_OP_EXPAND_STR(a, type, size)
122
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]123/** Computes max of fixed point types.
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]124 *
125 * @param[in] type the actual data type.
126 *
127 * @return The result of the fixed point maximum.
128 */
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]129#define MAXQ_IMPL(type) \
130 inline type max_##type(type VopA, type VopB) \
131 { \
132 return max(VopA, VopB); \
133 }
134
135MAXQ_IMPL(qs8x1)
136MAXQ_IMPL(qs8x2)
137MAXQ_IMPL(qs8x4)
138MAXQ_IMPL(qs8x8)
139MAXQ_IMPL(qs8x16)
Georgios Pinitas09796752017-07-10 16:05:21 +0100[diff] [blame]140MAXQ_IMPL(qs16x1)
141MAXQ_IMPL(qs16x2)
142MAXQ_IMPL(qs16x4)
143MAXQ_IMPL(qs16x8)
144MAXQ_IMPL(qs16x16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]145
146#define MAX_OP_EXPAND_STR(a, b, type, size) max_##type##x##size((a), (b))
147#define MAX_OP_EXPAND(a, b, type, size) MAX_OP_EXPAND_STR(a, b, type, size)
148
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]149/** Computes saturated addition of fixed point types.
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]150 *
151 * @param[in] type the actual data type.
152 *
153 * @return The result of the fixed point addition. The result is saturated in case of overflow
154 */
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]155#define ADDQ_SAT_IMPL(type) \
156 inline type add_sat_##type(type VopA, type VopB) \
157 { \
158 return add_sat(VopA, VopB); \
159 }
160
161ADDQ_SAT_IMPL(qs8x1)
162ADDQ_SAT_IMPL(qs8x2)
163ADDQ_SAT_IMPL(qs8x4)
164ADDQ_SAT_IMPL(qs8x8)
165ADDQ_SAT_IMPL(qs8x16)
Gian Marco Iodice7d323a62017-07-05 20:05:23 +0100[diff] [blame]166ADDQ_SAT_IMPL(qs16x1)
167ADDQ_SAT_IMPL(qs16x2)
168ADDQ_SAT_IMPL(qs16x4)
169ADDQ_SAT_IMPL(qs16x8)
170ADDQ_SAT_IMPL(qs16x16)
Michalis Spyroudef665a2017-08-14 11:26:37 +0100[diff] [blame]171ADDQ_SAT_IMPL(qs32x1)
172ADDQ_SAT_IMPL(qs32x2)
173ADDQ_SAT_IMPL(qs32x4)
174ADDQ_SAT_IMPL(qs32x8)
175ADDQ_SAT_IMPL(qs32x16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]176
177#define ADD_SAT_OP_EXPAND_STR(a, b, type, size) add_sat_##type##x##size((a), (b))
178#define ADD_SAT_OP_EXPAND(a, b, type, size) ADD_SAT_OP_EXPAND_STR(a, b, type, size)
179
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]180/** Computes saturated subtraction of fixed point types.
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]181 *
182 * @param[in] type the actual data type.
183 *
184 * @return The result of the fixed point subtraction. The result is saturated in case of overflow
185 */
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]186#define SUBQ_SAT_IMPL(type) \
187 inline type sub_sat_##type(type VopA, type VopB) \
188 { \
189 return sub_sat(VopA, VopB); \
190 }
191
192SUBQ_SAT_IMPL(qs8x1)
193SUBQ_SAT_IMPL(qs8x2)
194SUBQ_SAT_IMPL(qs8x4)
195SUBQ_SAT_IMPL(qs8x8)
196SUBQ_SAT_IMPL(qs8x16)
Georgios Pinitas09796752017-07-10 16:05:21 +0100[diff] [blame]197SUBQ_SAT_IMPL(qs16x1)
198SUBQ_SAT_IMPL(qs16x2)
199SUBQ_SAT_IMPL(qs16x4)
200SUBQ_SAT_IMPL(qs16x8)
201SUBQ_SAT_IMPL(qs16x16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]202
203#define SUB_SAT_OP_EXPAND_STR(a, b, type, size) sub_sat_##type##x##size((a), (b))
204#define SUB_SAT_OP_EXPAND(a, b, type, size) SUB_SAT_OP_EXPAND_STR(a, b, type, size)
205
Michele Di Giorgioab0a77e2017-06-21 15:36:24 +0100[diff] [blame]206/* Multiply of two fixed point numbers
207 *
208 * @param[in] type the actual data type.
209 * @param[in] itype the intermediate data type.
210 *
211 * @return The result of the fixed point multiplication.
212 */
213#define MULQ_IMPL(type, itype) \
214 inline type mul_##type(type VopA, type VopB, int fixed_point_position) \
215 { \
216 itype round_val = (itype)(1 << (fixed_point_position - 1)); \
217 itype res = CONVERT((VopA), itype) * CONVERT((VopB), itype) + round_val; \
218 return CONVERT((res >> (itype)fixed_point_position), type); \
219 }
220
Michalis Spyroudef665a2017-08-14 11:26:37 +0100[diff] [blame]221MULQ_IMPL(qs8x8, qs16x8)
222MULQ_IMPL(qs16x8, qs32x8)
Michele Di Giorgioab0a77e2017-06-21 15:36:24 +0100[diff] [blame]223MULQ_IMPL(qs8x16, qs16x16)
224MULQ_IMPL(qs16x16, qs32x16)
225
226#define MUL_OP_EXPAND_STR(a, b, type, size, position) mul_##type##x##size((a), (b), (position))
227#define MUL_OP_EXPAND(a, b, type, size, position) MUL_OP_EXPAND_STR(a, b, type, size, position)
228
229/* Saturate multiply of two fixed point numbers
230 *
231 * @param[in] type the actual data type.
232 * @param[in] itype the intermediate data type.
233 *
234 * @return The result of the fixed point multiplication. The result is saturated in case of overflow
235 */
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]236#define MULQ_SAT_IMPL(type, itype) \
237 inline type mul_sat_##type(type VopA, type VopB, int fixed_point_position) \
238 { \
239 itype round_val = (itype)(1 << (fixed_point_position - 1)); \
240 itype res = mad_sat(CONVERT((VopA), itype), CONVERT((VopB), itype), round_val); \
241 return CONVERT_SAT((res >> (itype)fixed_point_position), type); \
242 }
243
Georgios Pinitascdf51452017-08-31 14:21:36 +0100[diff] [blame]244MULQ_SAT_IMPL(qs8x1, qs16x1)
245MULQ_SAT_IMPL(qs8x2, qs16x2)
246MULQ_SAT_IMPL(qs8x3, qs16x3)
247MULQ_SAT_IMPL(qs8x4, qs16x4)
Michalis Spyroudef665a2017-08-14 11:26:37 +0100[diff] [blame]248MULQ_SAT_IMPL(qs8x8, qs16x8)
Michalis Spyroudef665a2017-08-14 11:26:37 +0100[diff] [blame]249MULQ_SAT_IMPL(qs8x16, qs16x16)
Georgios Pinitascdf51452017-08-31 14:21:36 +0100[diff] [blame]250MULQ_SAT_IMPL(qs16x1, qs32x1)
251MULQ_SAT_IMPL(qs16x2, qs32x2)
252MULQ_SAT_IMPL(qs16x3, qs32x3)
253MULQ_SAT_IMPL(qs16x4, qs32x4)
254MULQ_SAT_IMPL(qs16x8, qs32x8)
Michele Di Giorgioab0a77e2017-06-21 15:36:24 +0100[diff] [blame]255MULQ_SAT_IMPL(qs16x16, qs32x16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]256
257#define MUL_SAT_OP_EXPAND_STR(a, b, type, size, position) mul_sat_##type##x##size((a), (b), (position))
258#define MUL_SAT_OP_EXPAND(a, b, type, size, position) MUL_SAT_OP_EXPAND_STR(a, b, type, size, position)
259
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]260/** Saturate multiply-accumulate
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]261 *
262 * @param[in] type the actual data type.
263 * @param[in] itype the intermediate data type.
264 *
265 * @return The result of the fixed point multiply-accumulate. The result is saturated in case of overflow
266 */
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]267#define MLAQ_SAT_IMPL(type, itype) \
268 type mla_sat_##type(type VopA, type VopB, type VopC, int fixed_point_position) \
269 { \
270 itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \
271 return add_sat(VopA, CONVERT_SAT(res >> (itype)fixed_point_position, type)); \
272 }
273
274MLAQ_SAT_IMPL(qs8x8, qs16x8)
275MLAQ_SAT_IMPL(qs8x16, qs16x16)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]276MLAQ_SAT_IMPL(qs16x8, qs32x8)
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]277
278#define MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mla_sat_##type##x##size((a), (b), (c), (position))
279#define MLA_SAT_OP_EXPAND(a, b, c, type, size, position) MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position)
280
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]281/** Saturate multiply-accumulate long
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]282 *
283 * @param[in] type the actual data type.
284 * @param[in] itype the intermediate data type.
285 *
286 * @return The result of the fixed point multiply-accumulate long. The result is saturated in case of overflow
287 */
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]288#define MLALQ_SAT_IMPL(type, itype) \
289 itype mlal_sat_##type(itype VopA, type VopB, type VopC, int fixed_point_position) \
290 { \
291 itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \
292 return add_sat(VopA, res >> (itype)fixed_point_position); \
293 }
294
295MLALQ_SAT_IMPL(qs8x8, qs16x8)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]296MLALQ_SAT_IMPL(qs16x8, qs32x8)
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]297
298#define MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mlal_sat_##type##x##size((a), (b), (c), (position))
299#define MLAL_SAT_OP_EXPAND(a, b, c, type, size, position) MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position)
300
steniu010c7614f2017-06-23 17:00:26 +0100[diff] [blame]301/** Saturate division of two fixed point vectors
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]302 *
303 * @param[in] stype the actual scalar data type.
304 * @param[in] type the actual data type.
305 * @param[in] itype the intermediate data type.
306 *
307 * @return The result of the fixed point division. The result is saturated in case of overflow
308 */
steniu010c7614f2017-06-23 17:00:26 +0100[diff] [blame]309#define DIVQ_SAT_IMPL(stype, type, itype) \
310 inline type div_sat_##type(type VopA, type VopB, int fixed_point_position) \
311 { \
312 itype conv_a = CONVERT((VopA), itype); \
313 itype denominator = CONVERT((VopB), itype); \
314 itype numerator = conv_a << (itype)(fixed_point_position); \
315 itype res = select((itype)(numerator / denominator), select((itype)stype##_MAX, (itype)stype##_MIN, (itype)(conv_a < (itype)0)), (itype)(denominator == (itype)0)); \
316 return CONVERT_SAT((res), type); \
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]317 }
318
319DIVQ_SAT_IMPL(qs8, qs8x16, qs16x16)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]320DIVQ_SAT_IMPL(qs16, qs16x8, qs32x8)
Georgios Pinitas09796752017-07-10 16:05:21 +0100[diff] [blame]321DIVQ_SAT_IMPL(qs16, qs16x16, qs32x16)
steniu010c7614f2017-06-23 17:00:26 +0100[diff] [blame]322DIVQ_SAT_IMPL(qs8, qs8, qs16)
323DIVQ_SAT_IMPL(qs16, qs16, qs32)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]324
steniu010c7614f2017-06-23 17:00:26 +0100[diff] [blame]325#define DIV_SAT_OP_EXPAND_STR(a, b, type, position) div_sat_##type((a), (b), (position))
326#define DIV_SAT_OP_EXPAND(a, b, type, position) DIV_SAT_OP_EXPAND_STR(a, b, type, position)
327
328#define DIV_SAT_OP_VEC_EXPAND_STR(a, b, type, size, position) div_sat_##type##x##size((a), (b), (position))
329#define DIV_SAT_OP_VEC_EXPAND(a, b, type, size, position) DIV_SAT_OP_VEC_EXPAND_STR(a, b, type, size, position)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]330
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]331/** Saturate exponential of a fixed point vector
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]332 *
333 * @note Implemented approach uses taylor polynomial to approximate the exponential function.
334 *
335 * @param[in] stype the actual scalar data type.
336 * @param[in] type the actual data type.
337 * @param[in] size the number of the calculated elements.
338 *
339 * @return The result of the fixed point exponential. The result is saturated in case of overflow
340 */
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]341#define EXPQ_IMPL(stype, type, size) \
342 inline type exp_sat_##type(type VopA, int fixed_point_position) \
343 { \
344 type const_one = (type)(1 << (fixed_point_position)); \
345 type ln2 = (type)((((0x58B9 >> (14 - fixed_point_position))) + 1) >> 1); \
346 type inv_ln2 = (type)((((0x38AA >> (14 - fixed_point_position)) + 1) >> 1)) | const_one; \
347 type A = (type)(((0x7FBA >> (14 - fixed_point_position)) + 1) >> 1); \
348 type B = (type)(((0x3FE9 >> (14 - fixed_point_position)) + 1) >> 1); \
349 type C = (type)(((0x1693 >> (14 - fixed_point_position)) + 1) >> 1); \
350 type D = (type)(((0x0592 >> (14 - fixed_point_position)) + 1) >> 1); \
351 type m = MUL_SAT_OP_EXPAND(VopA, inv_ln2, stype, size, fixed_point_position); \
352 type dec_m = m >> (type)fixed_point_position; \
353 type alpha = MUL_SAT_OP_EXPAND(dec_m << (type)fixed_point_position, ln2, stype, size, fixed_point_position); \
354 alpha = CONVERT(abs_diff(VopA, alpha), type); \
355 type sum = add_sat(MUL_SAT_OP_EXPAND(alpha, D, stype, size, fixed_point_position), C); \
356 sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), B); \
357 sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), A); \
358 sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), const_one); \
359 return select((type)stype##_MAX, select(sum << dec_m, sum >> -dec_m, dec_m < (type)0), clz(sum) > dec_m); /* Saturate result if needed */ \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]360 }
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]361
Chunosovd6afedc2017-11-06 22:09:45 +0700[diff] [blame]362EXPQ_IMPL(qs8, qs8x2, 2)
363EXPQ_IMPL(qs8, qs8x4, 4)
364EXPQ_IMPL(qs8, qs8x8, 8)
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]365EXPQ_IMPL(qs8, qs8x16, 16)
Chunosovd6afedc2017-11-06 22:09:45 +0700[diff] [blame]366EXPQ_IMPL(qs16, qs16x2, 2)
367EXPQ_IMPL(qs16, qs16x4, 4)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]368EXPQ_IMPL(qs16, qs16x8, 8)
Georgios Pinitas09796752017-07-10 16:05:21 +0100[diff] [blame]369EXPQ_IMPL(qs16, qs16x16, 16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]370
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]371#define EXP_OP_EXPAND_STR(a, type, size, position) exp_sat_##type##x##size((a), (position))
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]372#define EXP_OP_EXPAND(a, type, size, position) EXP_OP_EXPAND_STR(a, type, size, position)
373
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]374/** Saturate logarithm of a fixed point vector
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]375 *
376 * @note Implemented approach uses taylor polynomial to approximate the logarithm function.
377 *
378 * @param[in] stype the actual scalar data type.
379 * @param[in] type the actual data type.
380 * @param[in] size the number of the calculated elements.
381 *
382 * @return The result of the fixed point logarithm. The result is saturated in case of overflow
383 */
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]384#define LOGQ_IMPL(stype, type, size) \
385 inline type log_sat_##type(type VopA, int fixed_point_position) \
386 { \
387 type const_one = (type)(1 << (fixed_point_position)); \
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]388 type ln2 = (type)(0x58B9 >> (15 - fixed_point_position)); /* 1.4384189 */ \
389 type A = (type)(0x5C0F >> (14 - fixed_point_position)); /* 1.4384189 */ \
390 type B = -(type)(0x56AE >> (15 - fixed_point_position)); /* -0.6771900 */ \
391 type C = (type)(0x2933 >> (15 - fixed_point_position)); /* 0.3218538 */ \
392 type D = -(type)(0x0AA7 >> (15 - fixed_point_position)); /* -0.0832229 */ \
steniu010c7614f2017-06-23 17:00:26 +0100[diff] [blame]393 type inter_a = select(VopA, DIV_SAT_OP_VEC_EXPAND(const_one, VopA, stype, size, fixed_point_position), VopA < const_one); \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]394 type shift_val = (type)(15 - stype##_SHIFT) - clz(inter_a >> (type)fixed_point_position); \
395 inter_a = inter_a >> shift_val; \
396 inter_a = sub_sat(inter_a, const_one); \
397 type sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, D, stype, size, fixed_point_position), C); \
398 sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position), B); \
399 sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position), A); \
400 sum = MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position); \
401 sum = MUL_SAT_OP_EXPAND(add_sat(sum, shift_val << (type)fixed_point_position), ln2, stype, size, fixed_point_position); \
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]402 return select(select(sum, -sum, VopA < const_one), (type)0, VopA < (type)0); /* Saturate result if needed */ \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]403 }
404
405LOGQ_IMPL(qs8, qs8x16, 16)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]406LOGQ_IMPL(qs16, qs16x8, 8)
Michele Di Giorgio6c928342017-06-22 16:55:57 +0100[diff] [blame]407LOGQ_IMPL(qs16, qs16x16, 16)
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]408
409#define LOG_OP_EXPAND_STR(a, type, size, position) log_sat_##type##x##size((a), (position))
410#define LOG_OP_EXPAND(a, type, size, position) LOG_OP_EXPAND_STR(a, type, size, position)
411
412/** Saturate inverse square root of a fixed point vector
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]413 *
414 * @note Implemented approach uses Newton's method to approximate the inverse square root function.
415 *
416 * @param[in] stype the actual scalar data type.
417 * @param[in] type the actual data type.
418 * @param[in] size the number of the calculated elements.
419 *
420 * @return The result of the fixed point inverse square root. The result is saturated in case of overflow
421 */
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]422#define INVSQRTQ_IMPL(stype, type, size) \
423 inline type invsqrt_sat_##type(type VopA, int fixed_point_position) \
424 { \
425 type const_three = (type)(3 << (fixed_point_position)); \
426 type shift_value = (type)(16 - stype##_SHIFT) - (clz(VopA) + (type)fixed_point_position); \
Georgios Pinitascdf51452017-08-31 14:21:36 +0100[diff] [blame]427 type temp = select((type)(VopA >> shift_value), select((type)stype##_MAX, (type)(VopA << (-shift_value)), (type)(clz(VopA) > (-shift_value))), (type)(shift_value < (type)0)); \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]428 type x = temp; \
429 x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
430 x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
431 x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]432 if(sizeof((stype)(1)) > 1) /* Perform more iterations if datatype is QS16 */ \
433 { \
434 x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
435 x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
436 } \
437 type shift_value2 = select(shift_value >> 1, (-shift_value) >> 1, shift_value < (type)0); \
Georgios Pinitascdf51452017-08-31 14:21:36 +0100[diff] [blame]438 return select((type)(x >> shift_value2), select((type)stype##_MAX, (type)(x << shift_value2), (type)(clz(x) > shift_value2)), (type)(shift_value < (type)0)); /* Saturate result if needed */ \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]439 }
440
Georgios Pinitascdf51452017-08-31 14:21:36 +0100[diff] [blame]441INVSQRTQ_IMPL(qs8, qs8x1, 1)
442INVSQRTQ_IMPL(qs16, qs16x1, 1)
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]443INVSQRTQ_IMPL(qs8, qs8x16, 16)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]444INVSQRTQ_IMPL(qs16, qs16x8, 8)
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]445
446#define INVSQRT_OP_EXPAND_STR(a, type, size, position) invsqrt_sat_##type##x##size((a), (position))
447#define INVSQRT_OP_EXPAND(a, type, size, position) INVSQRT_OP_EXPAND_STR(a, type, size, position)
448
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]449/** Saturate hyperbolic tangent of a fixed point vector
Anthony Barbierf202e502017-11-23 18:02:04 +0000[diff] [blame]450 *
451 * tanh(x) = (e^2x - 1)/(e^2x + 1)
452 *
453 * @param[in] stype the actual scalar data type.
454 * @param[in] type the actual data type.
455 * @param[in] size the number of the calculated elements.
456 *
457 * @return The result of the fixed point hyperbolic tangent. The result is saturated in case of overflow
458 */
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]459#define TANHQ_IMPL(stype, type, size) \
460 inline type tanh_sat_##type(type VopA, int fixed_point_position) \
461 { \
462 type const_one = (type)(1 << (fixed_point_position)); \
463 type const_two = (type)(2 << (fixed_point_position)); \
464 type exp2x = EXP_OP_EXPAND(MUL_SAT_OP_EXPAND(const_two, VopA, stype, size, fixed_point_position), stype, size, fixed_point_position); \
465 type num = SUB_SAT_OP_EXPAND(exp2x, const_one, stype, size); \
466 type den = ADD_SAT_OP_EXPAND(exp2x, const_one, stype, size); \
steniu010c7614f2017-06-23 17:00:26 +0100[diff] [blame]467 return DIV_SAT_OP_VEC_EXPAND(num, den, stype, size, fixed_point_position); \
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]468 }
469
470TANHQ_IMPL(qs8, qs8x16, 16)
471TANHQ_IMPL(qs16, qs16x8, 8)
472
473#define TANH_OP_EXPAND_STR(a, type, size, position) tanh_sat_##type##x##size((a), (position))
474#define TANH_OP_EXPAND(a, type, size, position) TANH_OP_EXPAND_STR(a, type, size, position)
475
steniu01da37e2f2017-06-29 10:14:58 +0100[diff] [blame]476#define floatx16 float16
477#define float16_TYPE float16
478
Gian Marco76faef82018-01-29 12:15:32 +0000[diff] [blame]479#define CONVERTQ_DOWN_IMPL(in_type, out_type) \
480 inline out_type convert_##out_type##_##in_type(in_type a, int fixed_point_position) \
481 { \
482 return CONVERT(a * (1 << fixed_point_position) + select((in_type)-0.5f, (in_type)0.5f, isgreater(a, (in_type)0)), out_type); \
steniu01da37e2f2017-06-29 10:14:58 +0100[diff] [blame]483 }
484
485CONVERTQ_DOWN_IMPL(float16, qs8x16)
486CONVERTQ_DOWN_IMPL(float16, qs16x16)
487
Gian Marco76faef82018-01-29 12:15:32 +0000[diff] [blame]488#define CONVERTQ_DOWN_SAT_IMPL(in_type, out_type) \
489 inline out_type convert_##out_type##_##in_type##_sat(in_type a, int fixed_point_position) \
490 { \
491 return CONVERT_SAT(a * (1 << fixed_point_position) + select((in_type)-0.5f, (in_type)0.5f, isgreater(a, (in_type)0)), out_type); \
steniu01da37e2f2017-06-29 10:14:58 +0100[diff] [blame]492 }
493
494CONVERTQ_DOWN_SAT_IMPL(float16, qs8x16)
495CONVERTQ_DOWN_SAT_IMPL(float16, qs16x16)
496
497#define CONVERTQ_UP_IMPL(in_type, out_type) \
498 inline out_type convert_##out_type##_##in_type(in_type a, int fixed_point_position) \
499 { \
500 return CONVERT(a, out_type) / (1 << fixed_point_position); \
501 }
502
503CONVERTQ_UP_IMPL(qs8x16, float16)
504CONVERTQ_UP_IMPL(qs16x16, float16)
505
Michalis Spyrou172e5702017-06-26 14:18:47 +0100[diff] [blame]506#define SQCVT_SAT_IMPL(type) \
507 inline type sqcvt_##type##_sat(float a, int fixed_point_position) \
508 { \
509 return CONVERT_SAT((a * (1 << fixed_point_position) + ((a < 0) ? -0.5f : 0.5f)), type); \
510 }
511
512SQCVT_SAT_IMPL(qs8)
513SQCVT_SAT_IMPL(qs16)
514
515#define SQCVT_SAT_OP_EXPAND_STR(a, type, position) sqcvt_##type##_sat((a), (position))
516#define SQCVT_SAT_OP_EXPAND(a, type, position) SQCVT_SAT_OP_EXPAND_STR((a), type, position)
517
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]518#endif // ARM_COMPUTE_FIXED_POINT_H