Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 6b612f5fa1fee9528f2f87491fe7edb3887d9817 / . / src / core / CL / cl_kernels / helpers_asymm.h
blob: 53e6719cd7c82d9ef0cfe41d64fe94f0e5059a36 [file] [log] [blame]
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]1/*
Michele Di Giorgio4aff98f2019-08-28 16:27:26 +0100[diff] [blame]2 * Copyright (c) 2017-2019 ARM Limited.
Chunosovd621bca2017-11-03 17:33:15 +0700[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_HELPERS_ASYMM_H
25#define ARM_COMPUTE_HELPERS_ASYMM_H
26
27#include "helpers.h"
28
Michele Di Giorgio4aff98f2019-08-28 16:27:26 +0100[diff] [blame]29#define CONVERT_DOWN_RTE_STR(x, type) (convert_##type##_rte((x)))
30#define CONVERT_DOWN_RTE(x, type) CONVERT_DOWN_RTE_STR(x, type)
31
32/** Quantize a floating-point scalar value to 8-bit asymmetric
33 *
34 * @param[in] input Input value to quantize
35 * @param[in] offset Quantization offset
36 * @param[in] scale Quantization scale
37 *
38 * @return quantized value
39 */
40inline uchar quantize_qasymm8(float input, float offset, float scale)
41{
42 float out_f32 = input / scale + offset;
43 uchar res_u8 = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, int), uchar);
44 return res_u8;
45}
46
47/** Dequantize a scalar value from 8-bit asymmetric to floating-point
48 *
49 * @param[in] input Input value to quantize
50 * @param[in] offset Quantization offset
51 * @param[in] scale Quantization scale
52 *
53 * @return quantized value
54 */
55inline float dequantize_qasymm8(uchar input, float offset, float scale)
56{
57 return ((float)input - offset) * scale;
58}
59
60/** Quantize a vector of values from floating-point
61 *
62 * @param[in] type Output data type.
63 * @param[in] size Size of vector.
64 *
65 * @return quantized values
66 */
67#define QUANTIZE_IMPL(type, size) \
68 inline VEC_DATA_TYPE(type, size) quantize_##type##size(VEC_DATA_TYPE(float, size) input, float offset, float scale) \
69 { \
70 VEC_DATA_TYPE(float, size) \
71 out_f32 = input / (VEC_DATA_TYPE(float, size))(scale) + (VEC_DATA_TYPE(float, size))(offset); \
72 VEC_DATA_TYPE(type, size) \
73 res = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, VEC_DATA_TYPE(int, size)), VEC_DATA_TYPE(type, size)); \
74 return res; \
75 }
76
77/** Dequantize a vector of values to floating-point
78 *
79 * @param[in] type Input data type.
80 * @param[in] size Size of vector.
81 *
82 * @return dequantized values in floating point
83 */
84#define DEQUANTIZE_IMPL(type, size) \
85 inline VEC_DATA_TYPE(float, size) dequantize_##type##size(VEC_DATA_TYPE(type, size) input, float offset, float scale) \
86 { \
87 return (CONVERT(input, VEC_DATA_TYPE(float, 4)) - offset) * scale; \
88 }
89
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]90/** Correctly-rounded-to-nearest division by a power-of-two.
91 *
92 * @param[in] size Size of vector.
93 *
94 * @return Correctly-rounded-to-nearest division by a power-of-two.
95 */
96#define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size) \
97 inline VEC_DATA_TYPE(int, size) asymm_rounding_divide_by_POW2_##size(VEC_DATA_TYPE(int, size) x, int exponent) \
98 { \
99 VEC_DATA_TYPE(int, size) \
100 mask = (1 << exponent) - 1; \
101 const VEC_DATA_TYPE(int, size) zero = 0; \
102 const VEC_DATA_TYPE(int, size) one = 1; \
103 VEC_DATA_TYPE(int, size) \
104 threshold = (mask >> 1) + select(zero, one, x < 0); \
105 return (x >> exponent) + select(zero, one, (x & mask) > threshold); \
106 }
107
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]108/** Product of two numbers, interpreting them as fixed-point values in the interval [-1, 1),
109 * rounding to the nearest value, and saturating -1 * -1 to the maximum value.
110 *
111 * @param[in] size Size of vector.
112 *
113 * @return Product of two fixed-point numbers.
114 */
Giorgio Arenab99f00d2018-02-14 10:40:16 +0000[diff] [blame]115#define ASYMM_MULT_IMPL(size) \
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]116 inline VEC_DATA_TYPE(int, size) asymm_mult##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
117 { \
118 VEC_DATA_TYPE(int, size) \
119 overflow = a == b && a == INT_MIN; \
120 VEC_DATA_TYPE(long, size) \
121 a_64 = convert_long##size(a); \
122 VEC_DATA_TYPE(long, size) \
123 b_64 = convert_long##size(b); \
124 VEC_DATA_TYPE(long, size) \
125 ab_64 = a_64 * b_64; \
Giorgio Arena6232d042018-02-12 14:46:00 +0000[diff] [blame]126 /* COMPMID-907 */ \
Giorgio Arenab99f00d2018-02-14 10:40:16 +0000[diff] [blame]127 VEC_DATA_TYPE(int, size) \
Giorgio Arena6232d042018-02-12 14:46:00 +0000[diff] [blame]128 ab_x2_high32 = convert_int##size(((ab_64 + (1 << 30)) >> 31)); \
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]129 return select(ab_x2_high32, INT_MAX, overflow); \
130 }
131
Giorgio Arenab99f00d2018-02-14 10:40:16 +0000[diff] [blame]132/** Calculates \f$ exp(x) \f$ for x in [-1/4, 0).
133 *
134 * @param[in] size Size of vector.
135 *
136 * @return Result in fixed-point format Q0.
137 */
138#define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size) \
139 inline VEC_DATA_TYPE(int, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(VEC_DATA_TYPE(int, size) a) \
140 { \
141 const VEC_DATA_TYPE(int, size) constant_term = 1895147668; \
142 const VEC_DATA_TYPE(int, size) constant_1_over_3 = 715827883; \
143 const int k_fractional_bits = 31; \
144 VEC_DATA_TYPE(int, size) \
145 x = a + (1 << (k_fractional_bits - 3)); \
146 VEC_DATA_TYPE(int, size) \
147 x2 = ASYMM_MULT(x, x, size); \
148 VEC_DATA_TYPE(int, size) \
149 x3 = ASYMM_MULT(x2, x, size); \
150 VEC_DATA_TYPE(int, size) \
151 x4 = ASYMM_MULT(x2, x2, size); \
152 VEC_DATA_TYPE(int, size) \
153 x4_over_4 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4, 2, size); \
154 VEC_DATA_TYPE(int, size) \
155 x4_over_24_plus_x3_over_6_plus_x2 = ASYMM_MULT((x4_over_4 + x3), constant_1_over_3, size) + x2; \
156 VEC_DATA_TYPE(int, size) \
157 x4_over_24_plus_x3_over_6_plus_x2_over_2 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4_over_24_plus_x3_over_6_plus_x2, 1, size); \
158 return constant_term + ASYMM_MULT(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2, size); \
159 }
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]160
Giorgio Arenab99f00d2018-02-14 10:40:16 +0000[diff] [blame]161/** Each bit of the result is set to the corresponding bit of either then_val or
162 * else_val depending on whether the corresponding bit of if_mask is set.
163 * Equivalent to the VBSL instruction in ARM NEON.
164 *
165 * @param[in] size Size of vector.
166 *
167 * @returns Result contaning bits from @p then_val or from @p else_val depending on corresponding bit in @p if_mask is set or not.
168 */
169#define ASYMM_SELECT_USING_MASK_IMPL(size) \
170 inline VEC_DATA_TYPE(int, size) asymm_select_using_mask##size(VEC_DATA_TYPE(int, size) if_mask, VEC_DATA_TYPE(int, size) then_val, VEC_DATA_TYPE(int, size) else_val) \
171 { \
172 return (if_mask & then_val) ^ (~if_mask & else_val); \
173 }
174
175/** For each element of input vector, the corresponding bits of the result item are set
176 * if the input item is zero.
177 *
178 * @param[in] size Size of vector.
179 *
180 * @returns Output vector with bits set when corresponding bit in @p a is zero.
181 */
182#define ASYMM_MASK_IF_ZERO_IMPL(size) \
183 inline VEC_DATA_TYPE(int, size) asymm_mask_if_zero##size(VEC_DATA_TYPE(int, size) a) \
184 { \
185 const VEC_DATA_TYPE(int, size) all_zeros = 0; \
186 const VEC_DATA_TYPE(int, size) all_ones = ~0; \
187 return select(all_zeros, all_ones, a == 0); \
188 }
189
190/** For each element of input vector, the corresponding bits of the result item are set
191 * if the input item is non-zero.
192 *
193 * @param[in] size Size of vector.
194 *
195 * @returns Output vector with bits set when corresponding bit in @p a is non zero.
196 */
197#define ASYMM_MASK_IF_NON_ZERO_IMPL(size) \
198 inline VEC_DATA_TYPE(int, size) asymm_mask_if_non_zero##size(VEC_DATA_TYPE(int, size) a) \
199 { \
200 const VEC_DATA_TYPE(int, size) all_zeros = 0; \
201 const VEC_DATA_TYPE(int, size) all_ones = ~0; \
202 return select(all_zeros, all_ones, a != 0); \
203 }
204
205#define EXP_BARREL_SHIFTER_IMPL(size) \
206 inline VEC_DATA_TYPE(int, size) exp_barrel_shifter##size(VEC_DATA_TYPE(int, size) result, int exponent, int fp_multiplier, int k_integer_bits, int k_fractional_bits, VEC_DATA_TYPE(int, size) remainder) \
207 { \
208 if(k_integer_bits > exponent) \
209 { \
210 const int k_shift_amount = k_integer_bits > exponent ? k_fractional_bits + exponent : 0; \
211 return ASYMM_SELECT_USING_MASK( \
212 ASYMM_MASK_IF_NON_ZERO(remainder & (1 << k_shift_amount), size), \
213 ASYMM_MULT(result, fp_multiplier, size), result, size); \
214 } \
215 \
216 return result; \
217 }
218
219/** Calculates \f$ exp(x) \f$ for x < 0.
220 *
221 * @param[in] size Size of vector.
222 *
223 * @return Result in fixed-point format Q0.
224 */
225#define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size) \
226 inline VEC_DATA_TYPE(int, size) asymm_exp_on_negative_values##size(VEC_DATA_TYPE(int, size) a, int k_integer_bits) \
227 { \
228 const int k_fractional_bits = 31 - k_integer_bits; \
229 VEC_DATA_TYPE(int, size) \
230 k_one_quarter = 1 << (k_fractional_bits - 2); \
231 VEC_DATA_TYPE(int, size) \
232 mask = k_one_quarter - 1; \
233 VEC_DATA_TYPE(int, size) \
234 a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter; \
235 VEC_DATA_TYPE(int, size) \
236 a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits; \
237 VEC_DATA_TYPE(int, size) \
238 result = ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a_mod_quarter_minus_one_quarter_scaled, size); \
239 VEC_DATA_TYPE(int, size) \
240 remainder = a_mod_quarter_minus_one_quarter - a; \
241 \
242 result = EXP_BARREL_SHIFTER(result, -2, 1672461947, k_integer_bits, k_fractional_bits, remainder, size); \
243 result = EXP_BARREL_SHIFTER(result, -1, 1302514674, k_integer_bits, k_fractional_bits, remainder, size); \
244 result = EXP_BARREL_SHIFTER(result, +0, 790015084, k_integer_bits, k_fractional_bits, remainder, size); \
245 result = EXP_BARREL_SHIFTER(result, +1, 290630308, k_integer_bits, k_fractional_bits, remainder, size); \
246 result = EXP_BARREL_SHIFTER(result, +2, 39332535, k_integer_bits, k_fractional_bits, remainder, size); \
247 result = EXP_BARREL_SHIFTER(result, +3, 720401, k_integer_bits, k_fractional_bits, remainder, size); \
248 result = EXP_BARREL_SHIFTER(result, +4, 242, k_integer_bits, k_fractional_bits, remainder, size); \
249 \
250 if(k_integer_bits > 5) \
251 { \
252 const VEC_DATA_TYPE(int, size) clamp = -(1 << (k_fractional_bits + 5)); \
253 result = ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(a < clamp, size), 0, result, size); \
254 } \
255 \
256 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
257 return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_ZERO(a, size), Q0_one, result, size); \
258 }
259
260/** Calculates the product of a integer value by a power of two, with either a positive exponent
261 * (equivalent to an arithmetic left shift, saturating) or a negative exponent
262 * (equivalent to an arithmetic right shift, rounding to nearest).
263 *
264 * @param[in] size Size of vector.
265 *
266 * @return Arithmetic left or right shift.
267 */
268#define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size) \
269 inline VEC_DATA_TYPE(int, size) asymm_saturating_rounding_mult_by_pow2##size(VEC_DATA_TYPE(int, size) x, int exponent) \
270 { \
271 if(exponent < 0) \
272 { \
273 return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size); \
274 } \
275 \
276 const VEC_DATA_TYPE(int, size) min = INT_MIN; \
277 const VEC_DATA_TYPE(int, size) max = INT_MAX; \
278 int threshold = ((1 << (31 - exponent)) - 1); \
279 VEC_DATA_TYPE(int, size) \
280 positive_mask = ASYMM_MASK_IF_NON_ZERO(x > threshold, size); \
281 VEC_DATA_TYPE(int, size) \
282 negative_mask = ASYMM_MASK_IF_NON_ZERO(x < -threshold, size); \
283 VEC_DATA_TYPE(int, size) \
284 result = x << exponent; \
285 result = ASYMM_SELECT_USING_MASK(positive_mask, max, result, size); \
286 result = ASYMM_SELECT_USING_MASK(negative_mask, min, result, size); \
287 return result; \
288 }
289
290/** Calculates (a+b)/2, rounded to the nearest integer.
291 * Equivalent to VRHADD in the ARM NEON instruction set.
292 *
293 * @param[in] size Size of vector.
294 *
295 * @return (a+b)/2, rounded to the nearest integer.
296 */
297#define ASYMM_ROUNDING_HALF_SUM_IMPL(size) \
298 inline VEC_DATA_TYPE(int, size) asymm_rounding_half_sum##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
299 { \
300 VEC_DATA_TYPE(long, size) \
301 a64 = convert_long##size(a); \
302 VEC_DATA_TYPE(long, size) \
303 b64 = convert_long##size(b); \
304 VEC_DATA_TYPE(long, size) \
305 sum = a64 + b64; \
306 const VEC_DATA_TYPE(long, size) one = 1; \
307 const VEC_DATA_TYPE(long, size) minus_one = -1; \
308 VEC_DATA_TYPE(long, size) \
309 sign = select(minus_one, one, sum >= 0); \
310 return convert_int##size((sum + sign) / 2); \
311 }
312
313/** Calculates \f$ 1 / (1 + x) \f$ for x in (0, 1).
314 *
315 * @param[in] size Size of vector.
316 *
317 * @return Result in fixed-point format Q0.
318 */
319#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size) \
320 inline VEC_DATA_TYPE(int, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(VEC_DATA_TYPE(int, size) a) \
321 { \
322 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
323 const VEC_DATA_TYPE(int, size) Q2_one = 1 << (31 - 2); \
324 VEC_DATA_TYPE(int, size) \
325 half_denominator = ASYMM_ROUNDING_HALF_SUM(a, Q0_one, size); \
326 const VEC_DATA_TYPE(int, size) Q2_48_over_17 = 1515870810; \
327 const VEC_DATA_TYPE(int, size) Q2_neg_32_over_17 = -1010580540; \
328 VEC_DATA_TYPE(int, size) \
329 x = Q2_48_over_17 + ASYMM_MULT(half_denominator, Q2_neg_32_over_17, size); \
330 for(int i = 0; i < 3; i++) \
331 { \
332 VEC_DATA_TYPE(int, size) \
333 half_denominator_times_x = ASYMM_MULT(half_denominator, x, size); \
334 VEC_DATA_TYPE(int, size) \
335 one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x; \
336 VEC_DATA_TYPE(int, size) \
337 tmp = ASYMM_MULT(x, one_minus_half_denominator_times_x, size); \
338 x = x + ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(tmp, 2, size); \
339 } \
340 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size); \
341 }
342
343/** Considering the integer value as fixed-point, change the number of integer bits and update value accordingly.
344 *
345 * @param[in] size Size of vector.
346 *
347 * @return Rescaled value.
348 */
349#define ASYMM_RESCALE_IMPL(size) \
350 inline VEC_DATA_TYPE(int, size) asymm_rescale##size(VEC_DATA_TYPE(int, size) value, int src_integer_bits, int dst_integer_bits) \
351 { \
352 int exponent = src_integer_bits - dst_integer_bits; \
353 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size); \
354 }
355
Michele Di Giorgio4aff98f2019-08-28 16:27:26 +0100[diff] [blame]356#define QUANTIZE_STR(input, offset, scale, type, size) quantize_##type##size(input, offset, scale)
357#define QUANTIZE(input, offset, scale, type, size) QUANTIZE_STR(input, offset, scale, type, size)
358#define DEQUANTIZE_STR(input, offset, scale, type, size) dequantize_##type##size(input, offset, scale)
359#define DEQUANTIZE(input, offset, scale, type, size) DEQUANTIZE_STR(input, offset, scale, type, size)
360
Giorgio Arenab99f00d2018-02-14 10:40:16 +0000[diff] [blame]361#define ASYMM_ROUNDING_DIVIDE_BY_POW2(x, exponent, size) asymm_rounding_divide_by_POW2_##size(x, exponent)
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]362#define ASYMM_MULT(a, b, size) asymm_mult##size(a, b)
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]363#define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, quantized_multiplier, right_shift, size) \
364 ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(x, quantized_multiplier, size), right_shift, size)
Giorgio Arenab99f00d2018-02-14 10:40:16 +0000[diff] [blame]365#define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(a)
366#define ASYMM_SELECT_USING_MASK(if_mask, then_val, else_val, size) asymm_select_using_mask##size(if_mask, then_val, else_val)
367#define ASYMM_MASK_IF_ZERO(a, size) asymm_mask_if_zero##size(a)
368#define ASYMM_MASK_IF_NON_ZERO(a, size) asymm_mask_if_non_zero##size(a)
369#define EXP_BARREL_SHIFTER(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder, size) exp_barrel_shifter##size(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder)
370#define ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, size) asymm_exp_on_negative_values##size(a, k_integer_bits)
371#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1(a, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(a)
372#define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, exponent, size) asymm_saturating_rounding_mult_by_pow2##size(x, exponent)
373#define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b)
374#define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) asymm_rescale##size(value, src_integer_bits, dst_integer_bits)
Chunosovd621bca2017-11-03 17:33:15 +0700[diff] [blame]375
Michele Di Giorgio4aff98f2019-08-28 16:27:26 +0100[diff] [blame]376QUANTIZE_IMPL(uchar, 4)
377QUANTIZE_IMPL(ushort, 4)
Michele Di Giorgio6b612f52019-09-05 12:30:22 +0100[diff] [blame^]378QUANTIZE_IMPL(short, 4)
Michele Di Giorgio4aff98f2019-08-28 16:27:26 +0100[diff] [blame]379
380DEQUANTIZE_IMPL(uchar, 4)
381DEQUANTIZE_IMPL(ushort, 4)
Michele Di Giorgio6b612f52019-09-05 12:30:22 +0100[diff] [blame^]382DEQUANTIZE_IMPL(short, 4)
Michele Di Giorgio4aff98f2019-08-28 16:27:26 +0100[diff] [blame]383
Giorgio Arenab99f00d2018-02-14 10:40:16 +0000[diff] [blame]384ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
385ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4)
386ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8)
387ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
388
389ASYMM_MULT_IMPL(2)
390ASYMM_MULT_IMPL(4)
391ASYMM_MULT_IMPL(8)
392ASYMM_MULT_IMPL(16)
393
394ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(2)
395ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(4)
396ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(8)
397ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(16)
398
399ASYMM_SELECT_USING_MASK_IMPL(2)
400ASYMM_SELECT_USING_MASK_IMPL(4)
401ASYMM_SELECT_USING_MASK_IMPL(8)
402ASYMM_SELECT_USING_MASK_IMPL(16)
403
404ASYMM_MASK_IF_ZERO_IMPL(2)
405ASYMM_MASK_IF_ZERO_IMPL(4)
406ASYMM_MASK_IF_ZERO_IMPL(8)
407ASYMM_MASK_IF_ZERO_IMPL(16)
408
409ASYMM_MASK_IF_NON_ZERO_IMPL(2)
410ASYMM_MASK_IF_NON_ZERO_IMPL(4)
411ASYMM_MASK_IF_NON_ZERO_IMPL(8)
412ASYMM_MASK_IF_NON_ZERO_IMPL(16)
413
414EXP_BARREL_SHIFTER_IMPL(2)
415EXP_BARREL_SHIFTER_IMPL(4)
416EXP_BARREL_SHIFTER_IMPL(8)
417EXP_BARREL_SHIFTER_IMPL(16)
418
419ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(2)
420ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(4)
421ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(8)
422ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(16)
423
424ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(2)
425ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(4)
426ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(8)
427ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(16)
428
429ASYMM_ROUNDING_HALF_SUM_IMPL(2)
430ASYMM_ROUNDING_HALF_SUM_IMPL(4)
431ASYMM_ROUNDING_HALF_SUM_IMPL(8)
432ASYMM_ROUNDING_HALF_SUM_IMPL(16)
433
434ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(2)
435ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(4)
436ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(8)
437ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(16)
438
439ASYMM_RESCALE_IMPL(2)
440ASYMM_RESCALE_IMPL(4)
441ASYMM_RESCALE_IMPL(8)
442ASYMM_RESCALE_IMPL(16)
443
Michele Di Giorgio4aff98f2019-08-28 16:27:26 +0100[diff] [blame]444#endif // ARM_COMPUTE_HELPERS_ASYMM_H