1 /*
2 * Copyright (c) 2017-2021 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_HELPERS_ASYMM_H
25 #define ARM_COMPUTE_HELPERS_ASYMM_H
26
27 #include "helpers.h"
28
29 /** Convert the given vector with round to nearest even rounding mode
30 *
31 * @param[in] x The target to be converted
32 * @param[in] type The target type
33 *
34 * @return The converted vector
35 */
36 #define CONVERT_DOWN_RTE_STR(x, type) (convert_##type##_rte((x)))
37 #define CONVERT_DOWN_RTE(x, type) CONVERT_DOWN_RTE_STR(x, type)
38
39 /** Quantize a floating-point scalar value to 8-bit asymmetric
40 *
41 * @param[in] input Input value to quantize
42 * @param[in] offset Quantization offset
43 * @param[in] scale Quantization scale
44 *
45 * @return quantized value
46 */
quantize_qasymm8(float input,float offset,float scale)47 inline uchar quantize_qasymm8(float input, float offset, float scale)
48 {
49 float out_f32 = input / scale + offset;
50 uchar res_u8 = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, int), uchar);
51 return res_u8;
52 }
53
54 /** Dequantize a scalar value from 8-bit asymmetric to floating-point
55 *
56 * @param[in] input Input value to quantize
57 * @param[in] offset Quantization offset
58 * @param[in] scale Quantization scale
59 *
60 * @return quantized value
61 */
dequantize_qasymm8(uchar input,float offset,float scale)62 inline float dequantize_qasymm8(uchar input, float offset, float scale)
63 {
64 return ((float)input - offset) * scale;
65 }
66
67 /** Dequantize a scalar value from signed 8-bit asymmetric to floating-point
68 *
69 * @param[in] input Input value to quantize
70 * @param[in] offset Quantization offset
71 * @param[in] scale Quantization scale
72 *
73 * @return quantized value
74 */
dequantize_qasymm8_signed(char input,float offset,float scale)75 inline float dequantize_qasymm8_signed(char input, float offset, float scale)
76 {
77 return ((float)input - offset) * scale;
78 }
79
80 /** Quantize a vector of values from floating-point
81 *
82 * @param[in] type Output data type.
83 * @param[in] size Size of vector.
84 *
85 * @return quantized values
86 */
87 #define QUANTIZE_IMPL(type, size) \
88 inline VEC_DATA_TYPE(type, size) quantize_##type##size(VEC_DATA_TYPE(float, size) input, float offset, float scale) \
89 { \
90 VEC_DATA_TYPE(float, size) \
91 out_f32 = input / (VEC_DATA_TYPE(float, size))(scale) + (VEC_DATA_TYPE(float, size))(offset); \
92 VEC_DATA_TYPE(type, size) \
93 res = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, VEC_DATA_TYPE(int, size)), VEC_DATA_TYPE(type, size)); \
94 return res; \
95 }
96
97 /** Dequantize a vector of values to floating-point
98 *
99 * @param[in] type Input data type.
100 * @param[in] size Size of vector.
101 *
102 * @return dequantized values in floating point
103 */
104 #define DEQUANTIZE_IMPL(type, size) \
105 inline VEC_DATA_TYPE(float, size) dequantize_##type##size(VEC_DATA_TYPE(type, size) input, float offset, float scale) \
106 { \
107 return (CONVERT(input, VEC_DATA_TYPE(float, size)) - offset) * scale; \
108 }
109
110 /** Correctly-rounded-to-nearest division by a power-of-two.
111 *
112 * @param[in] size Size of vector.
113 *
114 * @return Correctly-rounded-to-nearest division by a power-of-two.
115 */
116 #define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size) \
117 inline VEC_DATA_TYPE(int, size) asymm_rounding_divide_by_POW2_##size(VEC_DATA_TYPE(int, size) x, VEC_DATA_TYPE(int, size) exponent) \
118 { \
119 const VEC_DATA_TYPE(int, size) \
120 zero = (VEC_DATA_TYPE(int, size))0; \
121 const VEC_DATA_TYPE(int, size) \
122 one = (VEC_DATA_TYPE(int, size))1; \
123 VEC_DATA_TYPE(int, size) \
124 mask = (one << exponent) - one; \
125 VEC_DATA_TYPE(int, size) \
126 threshold = (mask >> 1) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))(x < 0)); \
127 return (x >> exponent) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))((x & mask) > threshold)); \
128 }
129
130 /** Product of two numbers, interpreting them as fixed-point values in the interval [-1, 1),
131 * rounding to the nearest value, and saturating -1 * -1 to the maximum value.
132 *
133 * @param[in] size Size of vector.
134 *
135 * @return Product of two fixed-point numbers.
136 */
137 #define ASYMM_MULT_IMPL(size) \
138 inline VEC_DATA_TYPE(int, size) asymm_mult##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
139 { \
140 VEC_DATA_TYPE(int, size) \
141 overflow = a == b && a == INT_MIN; \
142 VEC_DATA_TYPE(long, size) \
143 a_64 = convert_long##size(a); \
144 VEC_DATA_TYPE(long, size) \
145 b_64 = convert_long##size(b); \
146 VEC_DATA_TYPE(long, size) \
147 ab_64 = a_64 * b_64; \
148 /* Revert COMPMID-907 */ \
149 VEC_DATA_TYPE(long, size) \
150 mask1 = 1 << 30; \
151 VEC_DATA_TYPE(long, size) \
152 mask2 = 1 - (1 << 30); \
153 VEC_DATA_TYPE(long, size) \
154 is_positive_or_zero = ab_64 >= 0; \
155 VEC_DATA_TYPE(long, size) \
156 nudge = select(mask2, mask1, (SELECT_VEC_DATA_TYPE(long, size))(is_positive_or_zero)); \
157 VEC_DATA_TYPE(long, size) \
158 mask = 1ll << 31; \
159 VEC_DATA_TYPE(int, size) \
160 ab_x2_high32 = convert_int##size((ab_64 + nudge) / mask); \
161 return select(ab_x2_high32, INT_MAX, (SELECT_VEC_DATA_TYPE(int, size))(overflow)); \
162 }
163
164 /** Calculates \f$ exp(x) \f$ for x in [-1/4, 0).
165 *
166 * @param[in] size Size of vector.
167 *
168 * @return Result in fixed-point format Q0.
169 */
170 #define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size) \
171 inline VEC_DATA_TYPE(int, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(VEC_DATA_TYPE(int, size) a) \
172 { \
173 const VEC_DATA_TYPE(int, size) constant_term = 1895147668; \
174 const VEC_DATA_TYPE(int, size) constant_1_over_3 = 715827883; \
175 const int k_fractional_bits = 31; \
176 VEC_DATA_TYPE(int, size) \
177 x = a + (1 << (k_fractional_bits - 3)); \
178 VEC_DATA_TYPE(int, size) \
179 x2 = ASYMM_MULT(x, x, size); \
180 VEC_DATA_TYPE(int, size) \
181 x3 = ASYMM_MULT(x2, x, size); \
182 VEC_DATA_TYPE(int, size) \
183 x4 = ASYMM_MULT(x2, x2, size); \
184 VEC_DATA_TYPE(int, size) \
185 x4_over_4 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4, 2, size); \
186 VEC_DATA_TYPE(int, size) \
187 x4_over_24_plus_x3_over_6_plus_x2 = ASYMM_MULT((x4_over_4 + x3), constant_1_over_3, size) + x2; \
188 VEC_DATA_TYPE(int, size) \
189 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); \
190 return constant_term + ASYMM_MULT(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2, size); \
191 }
192
193 /** Each bit of the result is set to the corresponding bit of either then_val or
194 * else_val depending on whether the corresponding bit of if_mask is set.
195 * Equivalent to the VBSL instruction in Arm® Neon™.
196 *
197 * @param[in] size Size of vector.
198 *
199 * @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.
200 */
201 #define ASYMM_SELECT_USING_MASK_IMPL(size) \
202 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) \
203 { \
204 return (if_mask & then_val) ^ (~if_mask & else_val); \
205 }
206
207 /** For each element of input vector, the corresponding bits of the result item are set
208 * if the input item is zero.
209 *
210 * @param[in] size Size of vector.
211 *
212 * @returns Output vector with bits set when corresponding bit in @p a is zero.
213 */
214 #define ASYMM_MASK_IF_ZERO_IMPL(size) \
215 inline VEC_DATA_TYPE(int, size) asymm_mask_if_zero##size(VEC_DATA_TYPE(int, size) a) \
216 { \
217 const VEC_DATA_TYPE(int, size) all_zeros = 0; \
218 const VEC_DATA_TYPE(int, size) all_ones = ~0; \
219 return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a == 0)); \
220 }
221
222 /** For each element of input vector, the corresponding bits of the result item are set
223 * if the input item is non-zero.
224 *
225 * @param[in] size Size of vector.
226 *
227 * @returns Output vector with bits set when corresponding bit in @p a is non zero.
228 */
229 #define ASYMM_MASK_IF_NON_ZERO_IMPL(size) \
230 inline VEC_DATA_TYPE(int, size) asymm_mask_if_non_zero##size(VEC_DATA_TYPE(int, size) a) \
231 { \
232 const VEC_DATA_TYPE(int, size) all_zeros = 0; \
233 const VEC_DATA_TYPE(int, size) all_ones = ~0; \
234 return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a != 0)); \
235 }
236
237 #define EXP_BARREL_SHIFTER_IMPL(size) \
238 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) \
239 { \
240 if(k_integer_bits > exponent) \
241 { \
242 const int k_shift_amount = k_integer_bits > exponent ? k_fractional_bits + exponent : 0; \
243 return ASYMM_SELECT_USING_MASK( \
244 ASYMM_MASK_IF_NON_ZERO(remainder & (1 << k_shift_amount), size), \
245 ASYMM_MULT(result, fp_multiplier, size), result, size); \
246 } \
247 \
248 return result; \
249 }
250
251 /** Calculates \f$ exp(x) \f$ for x < 0.
252 *
253 * @param[in] size Size of vector.
254 *
255 * @return Result in fixed-point format Q0.
256 */
257 #define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size) \
258 inline VEC_DATA_TYPE(int, size) asymm_exp_on_negative_values##size(VEC_DATA_TYPE(int, size) a, int k_integer_bits) \
259 { \
260 const int k_fractional_bits = 31 - k_integer_bits; \
261 VEC_DATA_TYPE(int, size) \
262 k_one_quarter = 1 << (k_fractional_bits - 2); \
263 VEC_DATA_TYPE(int, size) \
264 mask = k_one_quarter - 1; \
265 VEC_DATA_TYPE(int, size) \
266 a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter; \
267 VEC_DATA_TYPE(int, size) \
268 a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits; \
269 VEC_DATA_TYPE(int, size) \
270 result = ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a_mod_quarter_minus_one_quarter_scaled, size); \
271 VEC_DATA_TYPE(int, size) \
272 remainder = a_mod_quarter_minus_one_quarter - a; \
273 \
274 result = EXP_BARREL_SHIFTER(result, -2, 1672461947, k_integer_bits, k_fractional_bits, remainder, size); \
275 result = EXP_BARREL_SHIFTER(result, -1, 1302514674, k_integer_bits, k_fractional_bits, remainder, size); \
276 result = EXP_BARREL_SHIFTER(result, +0, 790015084, k_integer_bits, k_fractional_bits, remainder, size); \
277 result = EXP_BARREL_SHIFTER(result, +1, 290630308, k_integer_bits, k_fractional_bits, remainder, size); \
278 result = EXP_BARREL_SHIFTER(result, +2, 39332535, k_integer_bits, k_fractional_bits, remainder, size); \
279 result = EXP_BARREL_SHIFTER(result, +3, 720401, k_integer_bits, k_fractional_bits, remainder, size); \
280 result = EXP_BARREL_SHIFTER(result, +4, 242, k_integer_bits, k_fractional_bits, remainder, size); \
281 \
282 if(k_integer_bits > 5) \
283 { \
284 const VEC_DATA_TYPE(int, size) clamp = -(1 << (k_fractional_bits + 5)); \
285 result = ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(a < clamp, size), 0, result, size); \
286 } \
287 \
288 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
289 return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_ZERO(a, size), Q0_one, result, size); \
290 }
291
292 /** Calculates the product of a integer value by a power of two, with either a positive exponent
293 * (equivalent to an arithmetic left shift, saturating) or a negative exponent
294 * (equivalent to an arithmetic right shift, rounding to nearest).
295 *
296 * @param[in] size Size of vector.
297 *
298 * @return Arithmetic left or right shift.
299 */
300 #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size) \
301 inline VEC_DATA_TYPE(int, size) asymm_saturating_rounding_mult_by_pow2##size(VEC_DATA_TYPE(int, size) x, int exponent) \
302 { \
303 if(exponent < 0) \
304 { \
305 return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size); \
306 } \
307 \
308 const VEC_DATA_TYPE(int, size) min = INT_MIN; \
309 const VEC_DATA_TYPE(int, size) max = INT_MAX; \
310 int threshold = ((1 << (31 - exponent)) - 1); \
311 VEC_DATA_TYPE(int, size) \
312 positive_mask = ASYMM_MASK_IF_NON_ZERO(x > threshold, size); \
313 VEC_DATA_TYPE(int, size) \
314 negative_mask = ASYMM_MASK_IF_NON_ZERO(x < -threshold, size); \
315 VEC_DATA_TYPE(int, size) \
316 result = x << exponent; \
317 result = ASYMM_SELECT_USING_MASK(positive_mask, max, result, size); \
318 result = ASYMM_SELECT_USING_MASK(negative_mask, min, result, size); \
319 return result; \
320 }
321
322 /** Calculates (a+b)/2, rounded to the nearest integer.
323 * Equivalent to VRHADD in the Arm Arm® Neon™ instruction set.
324 *
325 * @param[in] size Size of vector.
326 *
327 * @return (a+b)/2, rounded to the nearest integer.
328 */
329 #define ASYMM_ROUNDING_HALF_SUM_IMPL(size) \
330 inline VEC_DATA_TYPE(int, size) asymm_rounding_half_sum##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
331 { \
332 VEC_DATA_TYPE(long, size) \
333 a64 = convert_long##size(a); \
334 VEC_DATA_TYPE(long, size) \
335 b64 = convert_long##size(b); \
336 VEC_DATA_TYPE(long, size) \
337 sum = a64 + b64; \
338 const VEC_DATA_TYPE(long, size) one = 1; \
339 const VEC_DATA_TYPE(long, size) minus_one = -1; \
340 VEC_DATA_TYPE(long, size) \
341 sign = select(minus_one, one, (SELECT_VEC_DATA_TYPE(long, size))(sum >= 0)); \
342 return convert_int##size((sum + sign) / 2); \
343 }
344
345 /** Calculates \f$ 1 / (1 + x) \f$ for x in (0, 1).
346 *
347 * @param[in] size Size of vector.
348 *
349 * @return Result in fixed-point format Q0.
350 */
351 #define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size) \
352 inline VEC_DATA_TYPE(int, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(VEC_DATA_TYPE(int, size) a) \
353 { \
354 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
355 const VEC_DATA_TYPE(int, size) Q2_one = 1 << (31 - 2); \
356 VEC_DATA_TYPE(int, size) \
357 half_denominator = ASYMM_ROUNDING_HALF_SUM(a, Q0_one, size); \
358 const VEC_DATA_TYPE(int, size) Q2_48_over_17 = 1515870810; \
359 const VEC_DATA_TYPE(int, size) Q2_neg_32_over_17 = -1010580540; \
360 VEC_DATA_TYPE(int, size) \
361 x = Q2_48_over_17 + ASYMM_MULT(half_denominator, Q2_neg_32_over_17, size); \
362 for(int i = 0; i < 3; i++) \
363 { \
364 VEC_DATA_TYPE(int, size) \
365 half_denominator_times_x = ASYMM_MULT(half_denominator, x, size); \
366 VEC_DATA_TYPE(int, size) \
367 one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x; \
368 VEC_DATA_TYPE(int, size) \
369 tmp = ASYMM_MULT(x, one_minus_half_denominator_times_x, size); \
370 x = x + ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(tmp, 2, size); \
371 } \
372 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size); \
373 }
374
375 /** Considering the integer value as fixed-point, change the number of integer bits and update value accordingly.
376 *
377 * @param[in] size Size of vector.
378 *
379 * @return Rescaled value.
380 */
381 #define ASYMM_RESCALE_IMPL(size) \
382 inline VEC_DATA_TYPE(int, size) asymm_rescale##size(VEC_DATA_TYPE(int, size) value, int src_integer_bits, int dst_integer_bits) \
383 { \
384 int exponent = src_integer_bits - dst_integer_bits; \
385 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size); \
386 }
387
388 #define QUANTIZE_STR(input, offset, scale, type, size) quantize_##type##size(input, offset, scale)
389 #define QUANTIZE(input, offset, scale, type, size) QUANTIZE_STR(input, offset, scale, type, size)
390 #define DEQUANTIZE_STR(input, offset, scale, type, size) dequantize_##type##size(input, offset, scale)
391 #define DEQUANTIZE(input, offset, scale, type, size) DEQUANTIZE_STR(input, offset, scale, type, size)
392
393 #define ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size) asymm_rounding_divide_by_POW2_##size(x, exponent)
394 #define ASYMM_ROUNDING_DIVIDE_BY_POW2(x, exponent, size) ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size)
395 #define ASYMM_MULT_STR(a, b, size) asymm_mult##size(a, b)
396 #define ASYMM_MULT(a, b, size) ASYMM_MULT_STR(a, b, size)
397 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(x, quantized_multiplier, left_shift, size) \
398 ASYMM_MULT(x *((VEC_DATA_TYPE(int, size))(1) << (-left_shift)), quantized_multiplier, size)
399 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, quantized_multiplier, right_shift, size) \
400 ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(x, quantized_multiplier, size), right_shift, size)
401 #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)
402 #define ASYMM_SELECT_USING_MASK(if_mask, then_val, else_val, size) asymm_select_using_mask##size(if_mask, then_val, else_val)
403 #define ASYMM_MASK_IF_ZERO(a, size) asymm_mask_if_zero##size(a)
404 #define ASYMM_MASK_IF_NON_ZERO(a, size) asymm_mask_if_non_zero##size(a)
405 #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)
406 #define ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size) asymm_exp_on_negative_values##size(a, k_integer_bits)
407 #define ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, size) ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size)
408 #define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_STR(a, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(a)
409 #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_STR(a, size)
410 #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, exponent, size) asymm_saturating_rounding_mult_by_pow2##size(x, exponent)
411 #define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b)
412 #define ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size) asymm_rescale##size(value, src_integer_bits, dst_integer_bits)
413 #define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size)
414
415 #define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size) \
416 inline VEC_DATA_TYPE(int, size) multiply_by_quantized_multiplier##size(VEC_DATA_TYPE(int, size) input, int qmul, int shift) \
417 { \
418 const int left_shift = shift > 0 ? shift : 0; \
419 const int right_shift = shift > 0 ? 0 : -shift; \
420 return ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(input * (1 << left_shift), qmul, size), right_shift, size); \
421 }
422 #define MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, size) multiply_by_quantized_multiplier##size(input, qmul, shift)
423
424 QUANTIZE_IMPL(uchar, 1)
425 QUANTIZE_IMPL(char, 1)
426 QUANTIZE_IMPL(uint, 1)
427 QUANTIZE_IMPL(int, 1)
428 QUANTIZE_IMPL(uchar, 2)
429 QUANTIZE_IMPL(char, 2)
430 QUANTIZE_IMPL(uint, 2)
431 QUANTIZE_IMPL(int, 2)
432 QUANTIZE_IMPL(uchar, 3)
433 QUANTIZE_IMPL(char, 3)
434 QUANTIZE_IMPL(uint, 3)
435 QUANTIZE_IMPL(int, 3)
436 QUANTIZE_IMPL(uchar, 4)
437 QUANTIZE_IMPL(ushort, 4)
438 QUANTIZE_IMPL(short, 4)
439 QUANTIZE_IMPL(int, 4)
440 QUANTIZE_IMPL(uchar, 8)
441 QUANTIZE_IMPL(char, 8)
442 QUANTIZE_IMPL(uint, 8)
443 QUANTIZE_IMPL(int, 8)
444 QUANTIZE_IMPL(uchar, 16)
445 QUANTIZE_IMPL(char, 16)
446 QUANTIZE_IMPL(ushort, 16)
447 QUANTIZE_IMPL(short, 16)
448 QUANTIZE_IMPL(uint, 16)
449 QUANTIZE_IMPL(int, 16)
450
451 DEQUANTIZE_IMPL(uchar, 1)
452 DEQUANTIZE_IMPL(char, 1)
453 DEQUANTIZE_IMPL(uint, 1)
454 DEQUANTIZE_IMPL(int, 1)
455 DEQUANTIZE_IMPL(uchar, 2)
456 DEQUANTIZE_IMPL(char, 2)
457 DEQUANTIZE_IMPL(uint, 2)
458 DEQUANTIZE_IMPL(int, 2)
459 DEQUANTIZE_IMPL(uchar, 3)
460 DEQUANTIZE_IMPL(char, 3)
461 DEQUANTIZE_IMPL(uint, 3)
462 DEQUANTIZE_IMPL(int, 3)
463 DEQUANTIZE_IMPL(uchar, 4)
464 DEQUANTIZE_IMPL(ushort, 4)
465 DEQUANTIZE_IMPL(short, 4)
466 DEQUANTIZE_IMPL(int, 4)
467 DEQUANTIZE_IMPL(uchar, 8)
468 DEQUANTIZE_IMPL(char, 8)
469 DEQUANTIZE_IMPL(uint, 8)
470 DEQUANTIZE_IMPL(int, 8)
471 DEQUANTIZE_IMPL(uchar, 16)
472 DEQUANTIZE_IMPL(char, 16)
473 DEQUANTIZE_IMPL(ushort, 16)
474 DEQUANTIZE_IMPL(short, 16)
475 DEQUANTIZE_IMPL(uint, 16)
476 DEQUANTIZE_IMPL(int, 16)
477
478 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1)
479 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
480 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(3)
481 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4)
482 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8)
483 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
484
485 ASYMM_MULT_IMPL(1)
486 ASYMM_MULT_IMPL(2)
487 ASYMM_MULT_IMPL(3)
488 ASYMM_MULT_IMPL(4)
489 ASYMM_MULT_IMPL(8)
490 ASYMM_MULT_IMPL(16)
491
492 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(1)
493 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(2)
494 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(3)
495 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(4)
496 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(8)
497 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(16)
498
499 ASYMM_SELECT_USING_MASK_IMPL(1)
500 ASYMM_SELECT_USING_MASK_IMPL(2)
501 ASYMM_SELECT_USING_MASK_IMPL(3)
502 ASYMM_SELECT_USING_MASK_IMPL(4)
503 ASYMM_SELECT_USING_MASK_IMPL(8)
504 ASYMM_SELECT_USING_MASK_IMPL(16)
505
506 ASYMM_MASK_IF_ZERO_IMPL(1)
507 ASYMM_MASK_IF_ZERO_IMPL(2)
508 ASYMM_MASK_IF_ZERO_IMPL(3)
509 ASYMM_MASK_IF_ZERO_IMPL(4)
510 ASYMM_MASK_IF_ZERO_IMPL(8)
511 ASYMM_MASK_IF_ZERO_IMPL(16)
512
513 ASYMM_MASK_IF_NON_ZERO_IMPL(1)
514 ASYMM_MASK_IF_NON_ZERO_IMPL(2)
515 ASYMM_MASK_IF_NON_ZERO_IMPL(3)
516 ASYMM_MASK_IF_NON_ZERO_IMPL(4)
517 ASYMM_MASK_IF_NON_ZERO_IMPL(8)
518 ASYMM_MASK_IF_NON_ZERO_IMPL(16)
519
520 EXP_BARREL_SHIFTER_IMPL(1)
521 EXP_BARREL_SHIFTER_IMPL(2)
522 EXP_BARREL_SHIFTER_IMPL(3)
523 EXP_BARREL_SHIFTER_IMPL(4)
524 EXP_BARREL_SHIFTER_IMPL(8)
525 EXP_BARREL_SHIFTER_IMPL(16)
526
527 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(1)
528 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(2)
529 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(3)
530 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(4)
531 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(8)
532 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(16)
533
534 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(1)
535 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(2)
536 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(3)
537 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(4)
538 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(8)
539 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(16)
540
541 ASYMM_ROUNDING_HALF_SUM_IMPL(1)
542 ASYMM_ROUNDING_HALF_SUM_IMPL(2)
543 ASYMM_ROUNDING_HALF_SUM_IMPL(3)
544 ASYMM_ROUNDING_HALF_SUM_IMPL(4)
545 ASYMM_ROUNDING_HALF_SUM_IMPL(8)
546 ASYMM_ROUNDING_HALF_SUM_IMPL(16)
547
548 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(1)
549 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(2)
550 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(3)
551 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(4)
552 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(8)
553 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(16)
554
555 ASYMM_RESCALE_IMPL(1)
556 ASYMM_RESCALE_IMPL(2)
557 ASYMM_RESCALE_IMPL(3)
558 ASYMM_RESCALE_IMPL(4)
559 ASYMM_RESCALE_IMPL(8)
560 ASYMM_RESCALE_IMPL(16)
561
562 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(1)
563 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(2)
564 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(3)
565 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(4)
566 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(8)
567 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(16)
568
569 #endif // ARM_COMPUTE_HELPERS_ASYMM_H
570