Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 6c928343b0fa2bf60ffdfe21aea28b598d742ed4 / . / src / core / CL / cl_kernels / fixed_point.h
blob: 7038d40e16d6cadd39c8763f0685c9b8c2880625 [file] [log] [blame]
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]1/*
2 * Copyright (c) 2017 ARM Limited.
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.
106 *
107 * @param[in] type the actual data type.
108 *
109 * @return The result of the fixed point absolute value.
110 */
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.
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.
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)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]171
172#define ADD_SAT_OP_EXPAND_STR(a, b, type, size) add_sat_##type##x##size((a), (b))
173#define ADD_SAT_OP_EXPAND(a, b, type, size) ADD_SAT_OP_EXPAND_STR(a, b, type, size)
174
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]175/** Computes saturated subtraction of fixed point types.
176 *
177 * @param[in] type the actual data type.
178 *
179 * @return The result of the fixed point subtraction. The result is saturated in case of overflow
180 */
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]181#define SUBQ_SAT_IMPL(type) \
182 inline type sub_sat_##type(type VopA, type VopB) \
183 { \
184 return sub_sat(VopA, VopB); \
185 }
186
187SUBQ_SAT_IMPL(qs8x1)
188SUBQ_SAT_IMPL(qs8x2)
189SUBQ_SAT_IMPL(qs8x4)
190SUBQ_SAT_IMPL(qs8x8)
191SUBQ_SAT_IMPL(qs8x16)
Georgios Pinitas09796752017-07-10 16:05:21 +0100[diff] [blame]192SUBQ_SAT_IMPL(qs16x1)
193SUBQ_SAT_IMPL(qs16x2)
194SUBQ_SAT_IMPL(qs16x4)
195SUBQ_SAT_IMPL(qs16x8)
196SUBQ_SAT_IMPL(qs16x16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]197
198#define SUB_SAT_OP_EXPAND_STR(a, b, type, size) sub_sat_##type##x##size((a), (b))
199#define SUB_SAT_OP_EXPAND(a, b, type, size) SUB_SAT_OP_EXPAND_STR(a, b, type, size)
200
Michele Di Giorgioab0a77e2017-06-21 15:36:24 +0100[diff] [blame]201/* Multiply of two fixed point numbers
202 *
203 * @param[in] type the actual data type.
204 * @param[in] itype the intermediate data type.
205 *
206 * @return The result of the fixed point multiplication.
207 */
208#define MULQ_IMPL(type, itype) \
209 inline type mul_##type(type VopA, type VopB, int fixed_point_position) \
210 { \
211 itype round_val = (itype)(1 << (fixed_point_position - 1)); \
212 itype res = CONVERT((VopA), itype) * CONVERT((VopB), itype) + round_val; \
213 return CONVERT((res >> (itype)fixed_point_position), type); \
214 }
215
216MULQ_IMPL(qs8x16, qs16x16)
217MULQ_IMPL(qs16x16, qs32x16)
218
219#define MUL_OP_EXPAND_STR(a, b, type, size, position) mul_##type##x##size((a), (b), (position))
220#define MUL_OP_EXPAND(a, b, type, size, position) MUL_OP_EXPAND_STR(a, b, type, size, position)
221
222/* Saturate multiply of two fixed point numbers
223 *
224 * @param[in] type the actual data type.
225 * @param[in] itype the intermediate data type.
226 *
227 * @return The result of the fixed point multiplication. The result is saturated in case of overflow
228 */
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]229#define MULQ_SAT_IMPL(type, itype) \
230 inline type mul_sat_##type(type VopA, type VopB, int fixed_point_position) \
231 { \
232 itype round_val = (itype)(1 << (fixed_point_position - 1)); \
233 itype res = mad_sat(CONVERT((VopA), itype), CONVERT((VopB), itype), round_val); \
234 return CONVERT_SAT((res >> (itype)fixed_point_position), type); \
235 }
236
237MULQ_SAT_IMPL(qs8x16, qs16x16)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]238MULQ_SAT_IMPL(qs16x8, qs32x8)
Michele Di Giorgioab0a77e2017-06-21 15:36:24 +0100[diff] [blame]239MULQ_SAT_IMPL(qs16x16, qs32x16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]240
241#define MUL_SAT_OP_EXPAND_STR(a, b, type, size, position) mul_sat_##type##x##size((a), (b), (position))
242#define MUL_SAT_OP_EXPAND(a, b, type, size, position) MUL_SAT_OP_EXPAND_STR(a, b, type, size, position)
243
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]244/** Saturate multiply-accumulate
245 *
246 * @param[in] type the actual data type.
247 * @param[in] itype the intermediate data type.
248 *
249 * @return The result of the fixed point multiply-accumulate. The result is saturated in case of overflow
250 */
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]251#define MLAQ_SAT_IMPL(type, itype) \
252 type mla_sat_##type(type VopA, type VopB, type VopC, int fixed_point_position) \
253 { \
254 itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \
255 return add_sat(VopA, CONVERT_SAT(res >> (itype)fixed_point_position, type)); \
256 }
257
258MLAQ_SAT_IMPL(qs8x8, qs16x8)
259MLAQ_SAT_IMPL(qs8x16, qs16x16)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]260MLAQ_SAT_IMPL(qs16x8, qs32x8)
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]261
262#define MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mla_sat_##type##x##size((a), (b), (c), (position))
263#define MLA_SAT_OP_EXPAND(a, b, c, type, size, position) MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position)
264
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]265/** Saturate multiply-accumulate long
266 *
267 * @param[in] type the actual data type.
268 * @param[in] itype the intermediate data type.
269 *
270 * @return The result of the fixed point multiply-accumulate long. The result is saturated in case of overflow
271 */
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]272#define MLALQ_SAT_IMPL(type, itype) \
273 itype mlal_sat_##type(itype VopA, type VopB, type VopC, int fixed_point_position) \
274 { \
275 itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \
276 return add_sat(VopA, res >> (itype)fixed_point_position); \
277 }
278
279MLALQ_SAT_IMPL(qs8x8, qs16x8)
Gian Marco Iodice8a383692017-07-03 17:41:47 +0100[diff] [blame]280MLALQ_SAT_IMPL(qs16x8, qs32x8)
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]281
282#define MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mlal_sat_##type##x##size((a), (b), (c), (position))
283#define MLAL_SAT_OP_EXPAND(a, b, c, type, size, position) MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position)
284
285/** Saturate division of two fixed point numbers
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]286 *
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]287 * @param[in] stype the actual scalar data type.
288 * @param[in] type the actual data type.
289 * @param[in] itype the intermediate data type.
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]290 *
291 * @return The result of the fixed point division. The result is saturated in case of overflow
292 */
293#define DIVQ_SAT_IMPL(stype, type, itype) \
294 inline type div_sat_##type(type VopA, type VopB, int fixed_point_position) \
295 { \
296 itype conv_a = CONVERT((VopA), itype); \
297 itype denominator = CONVERT((VopB), itype); \
298 itype numerator = conv_a << (itype)(fixed_point_position); \
299 itype res = select(numerator / denominator, select((itype)stype##_MAX, (itype)stype##_MIN, conv_a < (itype)0), denominator == (itype)0); \
300 return CONVERT_SAT((res), type); \
301 }
302
303DIVQ_SAT_IMPL(qs8, qs8x16, qs16x16)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]304DIVQ_SAT_IMPL(qs16, qs16x8, qs32x8)
Georgios Pinitas09796752017-07-10 16:05:21 +0100[diff] [blame]305DIVQ_SAT_IMPL(qs16, qs16x16, qs32x16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]306
307#define DIV_SAT_OP_EXPAND_STR(a, b, type, size, position) div_sat_##type##x##size((a), (b), (position))
308#define DIV_SAT_OP_EXPAND(a, b, type, size, position) DIV_SAT_OP_EXPAND_STR(a, b, type, size, position)
309
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]310/** Saturate exponential of a fixed point vector
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]311 *
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]312 * @note Implemented approach uses taylor polynomial to approximate the exponential function.
313 *
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]314 * @param[in] stype the actual scalar data type.
315 * @param[in] type the actual data type.
316 * @param[in] size the number of the calculated elements.
317 *
318 * @return The result of the fixed point exponential. The result is saturated in case of overflow
319 */
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]320#define EXPQ_IMPL(stype, type, size) \
321 inline type exp_sat_##type(type VopA, int fixed_point_position) \
322 { \
323 type const_one = (type)(1 << (fixed_point_position)); \
324 type ln2 = (type)((((0x58B9 >> (14 - fixed_point_position))) + 1) >> 1); \
325 type inv_ln2 = (type)((((0x38AA >> (14 - fixed_point_position)) + 1) >> 1)) | const_one; \
326 type A = (type)(((0x7FBA >> (14 - fixed_point_position)) + 1) >> 1); \
327 type B = (type)(((0x3FE9 >> (14 - fixed_point_position)) + 1) >> 1); \
328 type C = (type)(((0x1693 >> (14 - fixed_point_position)) + 1) >> 1); \
329 type D = (type)(((0x0592 >> (14 - fixed_point_position)) + 1) >> 1); \
330 type m = MUL_SAT_OP_EXPAND(VopA, inv_ln2, stype, size, fixed_point_position); \
331 type dec_m = m >> (type)fixed_point_position; \
332 type alpha = MUL_SAT_OP_EXPAND(dec_m << (type)fixed_point_position, ln2, stype, size, fixed_point_position); \
333 alpha = CONVERT(abs_diff(VopA, alpha), type); \
334 type sum = add_sat(MUL_SAT_OP_EXPAND(alpha, D, stype, size, fixed_point_position), C); \
335 sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), B); \
336 sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), A); \
337 sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), const_one); \
338 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]339 }
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]340
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]341EXPQ_IMPL(qs8, qs8x16, 16)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]342EXPQ_IMPL(qs16, qs16x8, 8)
Georgios Pinitas09796752017-07-10 16:05:21 +0100[diff] [blame]343EXPQ_IMPL(qs16, qs16x16, 16)
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]344
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]345#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]346#define EXP_OP_EXPAND(a, type, size, position) EXP_OP_EXPAND_STR(a, type, size, position)
347
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]348/** Saturate logarithm of a fixed point vector
349 *
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]350 * @note Implemented approach uses taylor polynomial to approximate the logarithm function.
351 *
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]352 * @param[in] stype the actual scalar data type.
353 * @param[in] type the actual data type.
354 * @param[in] size the number of the calculated elements.
355 *
356 * @return The result of the fixed point logarithm. The result is saturated in case of overflow
357 */
358#define LOGQ_IMPL(stype, type, size) \
359 inline type log_sat_##type(type VopA, int fixed_point_position) \
360 { \
361 type const_one = (type)(1 << (fixed_point_position)); \
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]362 type ln2 = (type)(0x58B9 >> (15 - fixed_point_position)); /* 1.4384189 */ \
363 type A = (type)(0x5C0F >> (14 - fixed_point_position)); /* 1.4384189 */ \
364 type B = -(type)(0x56AE >> (15 - fixed_point_position)); /* -0.6771900 */ \
365 type C = (type)(0x2933 >> (15 - fixed_point_position)); /* 0.3218538 */ \
366 type D = -(type)(0x0AA7 >> (15 - fixed_point_position)); /* -0.0832229 */ \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]367 type inter_a = select(VopA, DIV_SAT_OP_EXPAND(const_one, VopA, stype, size, fixed_point_position), VopA < const_one); \
368 type shift_val = (type)(15 - stype##_SHIFT) - clz(inter_a >> (type)fixed_point_position); \
369 inter_a = inter_a >> shift_val; \
370 inter_a = sub_sat(inter_a, const_one); \
371 type sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, D, stype, size, fixed_point_position), C); \
372 sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position), B); \
373 sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position), A); \
374 sum = MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position); \
375 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]376 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]377 }
378
379LOGQ_IMPL(qs8, qs8x16, 16)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]380LOGQ_IMPL(qs16, qs16x8, 8)
Michele Di Giorgio6c928342017-06-22 16:55:57 +0100[diff] [blame^]381LOGQ_IMPL(qs16, qs16x16, 16)
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]382
383#define LOG_OP_EXPAND_STR(a, type, size, position) log_sat_##type##x##size((a), (position))
384#define LOG_OP_EXPAND(a, type, size, position) LOG_OP_EXPAND_STR(a, type, size, position)
385
386/** Saturate inverse square root of a fixed point vector
387 *
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]388 * @note Implemented approach uses Newton's method to approximate the inverse square root function.
389 *
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]390 * @param[in] stype the actual scalar data type.
391 * @param[in] type the actual data type.
392 * @param[in] size the number of the calculated elements.
393 *
394 * @return The result of the fixed point inverse square root. The result is saturated in case of overflow
395 */
396#define INVSQRTQ_IMPL(stype, type, size) \
397 inline type invsqrt_sat_##type(type VopA, int fixed_point_position) \
398 { \
399 type const_three = (type)(3 << (fixed_point_position)); \
400 type shift_value = (type)(16 - stype##_SHIFT) - (clz(VopA) + (type)fixed_point_position); \
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]401 type temp = select(VopA >> shift_value, select((type)stype##_MAX, VopA << (-shift_value), clz(VopA) > (-shift_value)), shift_value < (type)0); \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]402 type x = temp; \
403 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; \
404 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; \
405 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]406 if(sizeof((stype)(1)) > 1) /* Perform more iterations if datatype is QS16 */ \
407 { \
408 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; \
409 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; \
410 } \
411 type shift_value2 = select(shift_value >> 1, (-shift_value) >> 1, shift_value < (type)0); \
412 return select(x >> shift_value2, select((type)stype##_MAX, x << shift_value2, clz(x) > shift_value2), shift_value < (type)0); /* Saturate result if needed */ \
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]413 }
414
415INVSQRTQ_IMPL(qs8, qs8x16, 16)
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]416INVSQRTQ_IMPL(qs16, qs16x8, 8)
Michalis Spyroud7e82812017-06-20 15:00:14 +0100[diff] [blame]417
418#define INVSQRT_OP_EXPAND_STR(a, type, size, position) invsqrt_sat_##type##x##size((a), (position))
419#define INVSQRT_OP_EXPAND(a, type, size, position) INVSQRT_OP_EXPAND_STR(a, type, size, position)
420
Georgios Pinitas00394ae2017-06-22 18:13:55 +0100[diff] [blame]421/** Saturate hyperbolic tangent of a fixed point vector
422 *
423 * tanh(x) = (e^2x - 1)/(e^2x + 1)
424 *
425 * @param[in] stype the actual scalar data type.
426 * @param[in] type the actual data type.
427 * @param[in] size the number of the calculated elements.
428 *
429 * @return The result of the fixed point hyperbolic tangent. The result is saturated in case of overflow
430 */
431#define TANHQ_IMPL(stype, type, size) \
432 inline type tanh_sat_##type(type VopA, int fixed_point_position) \
433 { \
434 type const_one = (type)(1 << (fixed_point_position)); \
435 type const_two = (type)(2 << (fixed_point_position)); \
436 type exp2x = EXP_OP_EXPAND(MUL_SAT_OP_EXPAND(const_two, VopA, stype, size, fixed_point_position), stype, size, fixed_point_position); \
437 type num = SUB_SAT_OP_EXPAND(exp2x, const_one, stype, size); \
438 type den = ADD_SAT_OP_EXPAND(exp2x, const_one, stype, size); \
439 return DIV_SAT_OP_EXPAND(num, den, stype, size, fixed_point_position); \
440 }
441
442TANHQ_IMPL(qs8, qs8x16, 16)
443TANHQ_IMPL(qs16, qs16x8, 8)
444
445#define TANH_OP_EXPAND_STR(a, type, size, position) tanh_sat_##type##x##size((a), (position))
446#define TANH_OP_EXPAND(a, type, size, position) TANH_OP_EXPAND_STR(a, type, size, position)
447
steniu01da37e2f2017-06-29 10:14:58 +0100[diff] [blame]448#define floatx16 float16
449#define float16_TYPE float16
450
451#define CONVERTQ_DOWN_IMPL(in_type, out_type) \
452 inline out_type convert_##out_type##_##in_type(in_type a, int fixed_point_position) \
453 { \
454 return CONVERT(a * (1 << fixed_point_position) + select((in_type)-0.5, (in_type)0.5, isgreater(a, (in_type)0)), out_type); \
455 }
456
457CONVERTQ_DOWN_IMPL(float16, qs8x16)
458CONVERTQ_DOWN_IMPL(float16, qs16x16)
459
460#define CONVERTQ_DOWN_SAT_IMPL(in_type, out_type) \
461 inline out_type convert_##out_type##_##in_type##_sat(in_type a, int fixed_point_position) \
462 { \
463 return CONVERT_SAT(a * (1 << fixed_point_position) + select((in_type)-0.5, (in_type)0.5, isgreater(a, (in_type)0)), out_type); \
464 }
465
466CONVERTQ_DOWN_SAT_IMPL(float16, qs8x16)
467CONVERTQ_DOWN_SAT_IMPL(float16, qs16x16)
468
469#define CONVERTQ_UP_IMPL(in_type, out_type) \
470 inline out_type convert_##out_type##_##in_type(in_type a, int fixed_point_position) \
471 { \
472 return CONVERT(a, out_type) / (1 << fixed_point_position); \
473 }
474
475CONVERTQ_UP_IMPL(qs8x16, float16)
476CONVERTQ_UP_IMPL(qs16x16, float16)
477
Michalis Spyrou172e5702017-06-26 14:18:47 +0100[diff] [blame]478#define SQCVT_SAT_IMPL(type) \
479 inline type sqcvt_##type##_sat(float a, int fixed_point_position) \
480 { \
481 return CONVERT_SAT((a * (1 << fixed_point_position) + ((a < 0) ? -0.5f : 0.5f)), type); \
482 }
483
484SQCVT_SAT_IMPL(qs8)
485SQCVT_SAT_IMPL(qs16)
486
487#define SQCVT_SAT_OP_EXPAND_STR(a, type, position) sqcvt_##type##_sat((a), (position))
488#define SQCVT_SAT_OP_EXPAND(a, type, position) SQCVT_SAT_OP_EXPAND_STR((a), type, position)
489
Georgios Pinitase5f8fd62017-06-23 18:03:44 +0100[diff] [blame]490#endif // ARM_COMPUTE_FIXED_POINT_H