compute/ARMComputeEx/src/core/CL/cl_kernels/helpers_asymm.h

   1 /*
   2  * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at
   7  *
   8  *      http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 /*
  18  * Copyright (c) 2017-2020 ARM Limited.
  19  *
  20  * SPDX-License-Identifier: MIT
  21  *
  22  * Permission is hereby granted, free of charge, to any person obtaining a copy
  23  * of this software and associated documentation files (the "Software"), to
  24  * deal in the Software without restriction, including without limitation the
  25  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  26  * sell copies of the Software, and to permit persons to whom the Software is
  27  * furnished to do so, subject to the following conditions:
  28  *
  29  * The above copyright notice and this permission notice shall be included in all
  30  * copies or substantial portions of the Software.
  31  *
  32  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  33  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  34  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  35  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  36  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  37  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  38  * SOFTWARE.
  39  */
  40 #ifndef ARM_COMPUTE_HELPERS_ASYMM_H
  41 #define ARM_COMPUTE_HELPERS_ASYMM_H
  42
  43 #include "helpers.h"
  44
  45 /** Convert the given vector with round to nearest even rounding mode
  46  *
  47  * @param[in] x    The target to be converted
  48  * @param[in] type The target type
  49  *
  50  * @return The converted vector
  51  */
  52 #define CONVERT_DOWN_RTE_STR(x, type) (convert_##type##_rte((x)))
  53 #define CONVERT_DOWN_RTE(x, type) CONVERT_DOWN_RTE_STR(x, type)
  54
  55 /** Quantize a floating-point scalar value to 8-bit asymmetric
  56  *
  57  * @param[in] input  Input value to quantize
  58  * @param[in] offset Quantization offset
  59  * @param[in] scale  Quantization scale
  60  *
  61  * @return quantized value
  62  */
  63 inline uchar quantize_qasymm8(float input, float offset, float scale)
  64 {
  65   float out_f32 = input / scale + offset;
  66   uchar res_u8 = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, int), uchar);
  67   return res_u8;
  68 }
  69
  70 /** Dequantize a scalar value from 8-bit asymmetric to floating-point
  71  *
  72  * @param[in] input  Input value to quantize
  73  * @param[in] offset Quantization offset
  74  * @param[in] scale  Quantization scale
  75  *
  76  * @return quantized value
  77  */
  78 inline float dequantize_qasymm8(uchar input, float offset, float scale)
  79 {
  80   return ((float)input - offset) * scale;
  81 }
  82
  83 /** Dequantize a scalar value from signed 8-bit asymmetric to floating-point
  84  *
  85  * @param[in] input  Input value to quantize
  86  * @param[in] offset Quantization offset
  87  * @param[in] scale  Quantization scale
  88  *
  89  * @return quantized value
  90  */
  91 inline float dequantize_qasymm8_signed(char input, float offset, float scale)
  92 {
  93   return ((float)input - offset) * scale;
  94 }
  95
  96 /** Quantize a vector of values from floating-point
  97  *
  98  * @param[in] type Output data type.
  99  * @param[in] size Size of vector.
 100  *
 101  * @return quantized values
 102  */
 103 #define QUANTIZE_IMPL(type, size)                                                                 \
 104   inline VEC_DATA_TYPE(type, size)                                                                \
 105       quantize_##type##size(VEC_DATA_TYPE(float, size) input, float offset, float scale)          \
 106   {                                                                                               \
 107     VEC_DATA_TYPE(float, size)                                                                    \
 108     out_f32 = input / (VEC_DATA_TYPE(float, size))(scale) + (VEC_DATA_TYPE(float, size))(offset); \
 109     VEC_DATA_TYPE(type, size)                                                                     \
 110     res = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, VEC_DATA_TYPE(int, size)),                        \
 111                       VEC_DATA_TYPE(type, size));                                                 \
 112     return res;                                                                                   \
 113   }
 114
 115 /** Dequantize a vector of values to floating-point
 116  *
 117  * @param[in] type Input data type.
 118  * @param[in] size Size of vector.
 119  *
 120  * @return dequantized values in floating point
 121  */
 122 #define DEQUANTIZE_IMPL(type, size)                                                       \
 123   inline VEC_DATA_TYPE(float, size)                                                       \
 124       dequantize_##type##size(VEC_DATA_TYPE(type, size) input, float offset, float scale) \
 125   {                                                                                       \
 126     return (CONVERT(input, VEC_DATA_TYPE(float, size)) - offset) * scale;                 \
 127   }
 128
 129 /** Correctly-rounded-to-nearest division by a power-of-two.
 130  *
 131  * @param[in] size Size of vector.
 132  *
 133  * @return Correctly-rounded-to-nearest division by a power-of-two.
 134  */
 135 #define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size)                        \
 136   inline VEC_DATA_TYPE(int, size) asymm_rounding_divide_by_POW2_##size( \
 137       VEC_DATA_TYPE(int, size) x, VEC_DATA_TYPE(int, size) exponent)    \
 138   {                                                                     \
 139     const VEC_DATA_TYPE(int, size) zero = (VEC_DATA_TYPE(int, size))0;  \
 140     const VEC_DATA_TYPE(int, size) one = (VEC_DATA_TYPE(int, size))1;   \
 141     VEC_DATA_TYPE(int, size)                                            \
 142     mask = (one << exponent) - one;                                     \
 143     VEC_DATA_TYPE(int, size)                                            \
 144     threshold = (mask >> 1) + select(zero, one, x < 0);                 \
 145     return (x >> exponent) + select(zero, one, (x & mask) > threshold); \
 146   }
 147
 148 /** Product of two numbers, interpreting them as fixed-point values in the interval [-1, 1),
 149  * rounding to the nearest value, and saturating -1 * -1 to the maximum value.
 150  *
 151  * @param[in] size Size of vector.
 152  *
 153  * @return Product of two fixed-point numbers.
 154  */
 155 #define ASYMM_MULT_IMPL(size)                                                  \
 156   inline VEC_DATA_TYPE(int, size)                                              \
 157       asymm_mult##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
 158   {                                                                            \
 159     VEC_DATA_TYPE(int, size)                                                   \
 160     overflow = a == b && a == INT_MIN;                                         \
 161     VEC_DATA_TYPE(long, size)                                                  \
 162     a_64 = convert_long##size(a);                                              \
 163     VEC_DATA_TYPE(long, size)                                                  \
 164     b_64 = convert_long##size(b);                                              \
 165     VEC_DATA_TYPE(long, size)                                                  \
 166     ab_64 = a_64 * b_64;                                                       \
 167     /* Revert COMPMID-907 */                                                   \
 168     VEC_DATA_TYPE(long, size)                                                  \
 169     mask1 = 1 << 30;                                                           \
 170     VEC_DATA_TYPE(long, size)                                                  \
 171     mask2 = 1 - (1 << 30);                                                     \
 172     VEC_DATA_TYPE(long, size)                                                  \
 173     is_positive_or_zero = ab_64 >= 0;                                          \
 174     VEC_DATA_TYPE(long, size)                                                  \
 175     nudge = select(mask2, mask1, is_positive_or_zero);                         \
 176     VEC_DATA_TYPE(long, size)                                                  \
 177     mask = 1ll << 31;                                                          \
 178     VEC_DATA_TYPE(int, size)                                                   \
 179     ab_x2_high32 = convert_int##size((ab_64 + nudge) / mask);                  \
 180     return select(ab_x2_high32, INT_MAX, overflow);                            \
 181   }
 182
 183 /** Calculates \f$ exp(x) \f$ for x in [-1/4, 0).
 184  *
 185  * @param[in] size Size of vector.
 186  *
 187  * @return Result in fixed-point format Q0.
 188  */
 189 #define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size)                   \
 190   inline VEC_DATA_TYPE(int, size)                                                                  \
 191       asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(VEC_DATA_TYPE(int, size) \
 192                                                                               a)                   \
 193   {                                                                                                \
 194     const VEC_DATA_TYPE(int, size) constant_term = 1895147668;                                     \
 195     const VEC_DATA_TYPE(int, size) constant_1_over_3 = 715827883;                                  \
 196     const int k_fractional_bits = 31;                                                              \
 197     VEC_DATA_TYPE(int, size)                                                                       \
 198     x = a + (1 << (k_fractional_bits - 3));                                                        \
 199     VEC_DATA_TYPE(int, size)                                                                       \
 200     x2 = ASYMM_MULT(x, x, size);                                                                   \
 201     VEC_DATA_TYPE(int, size)                                                                       \
 202     x3 = ASYMM_MULT(x2, x, size);                                                                  \
 203     VEC_DATA_TYPE(int, size)                                                                       \
 204     x4 = ASYMM_MULT(x2, x2, size);                                                                 \
 205     VEC_DATA_TYPE(int, size)                                                                       \
 206     x4_over_4 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4, 2, size);                                        \
 207     VEC_DATA_TYPE(int, size)                                                                       \
 208     x4_over_24_plus_x3_over_6_plus_x2 =                                                            \
 209         ASYMM_MULT((x4_over_4 + x3), constant_1_over_3, size) + x2;                                \
 210     VEC_DATA_TYPE(int, size)                                                                       \
 211     x4_over_24_plus_x3_over_6_plus_x2_over_2 =                                                     \
 212         ASYMM_ROUNDING_DIVIDE_BY_POW2(x4_over_24_plus_x3_over_6_plus_x2, 1, size);                 \
 213     return constant_term +                                                                         \
 214            ASYMM_MULT(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2, size);          \
 215   }
 216
 217 /** Each bit of the result is set to the corresponding bit of either then_val or
 218  * else_val depending on whether the corresponding bit of if_mask is set.
 219  * Equivalent to the VBSL instruction in ARM NEON.
 220  *
 221  * @param[in] size Size of vector.
 222  *
 223  * @returns Result contaning bits from @p then_val or from @p else_val depending on corresponding
 224  * bit in @p if_mask is set or not.
 225  */
 226 #define ASYMM_SELECT_USING_MASK_IMPL(size)                                                         \
 227   inline VEC_DATA_TYPE(int, size) asymm_select_using_mask##size(VEC_DATA_TYPE(int, size) if_mask,  \
 228                                                                 VEC_DATA_TYPE(int, size) then_val, \
 229                                                                 VEC_DATA_TYPE(int, size) else_val) \
 230   {                                                                                                \
 231     return (if_mask & then_val) ^ (~if_mask & else_val);                                           \
 232   }
 233
 234 /** For each element of input vector, the corresponding bits of the result item are set
 235  * if the input item is zero.
 236  *
 237  * @param[in] size Size of vector.
 238  *
 239  * @returns Output vector with bits set when corresponding bit in @p a is zero.
 240  */
 241 #define ASYMM_MASK_IF_ZERO_IMPL(size)                                                  \
 242   inline VEC_DATA_TYPE(int, size) asymm_mask_if_zero##size(VEC_DATA_TYPE(int, size) a) \
 243   {                                                                                    \
 244     const VEC_DATA_TYPE(int, size) all_zeros = 0;                                      \
 245     const VEC_DATA_TYPE(int, size) all_ones = ~0;                                      \
 246     return select(all_zeros, all_ones, a == 0);                                        \
 247   }
 248
 249 /** For each element of input vector, the corresponding bits of the result item are set
 250  * if the input item is non-zero.
 251  *
 252  * @param[in] size Size of vector.
 253  *
 254  * @returns Output vector with bits set when corresponding bit in @p a is non zero.
 255  */
 256 #define ASYMM_MASK_IF_NON_ZERO_IMPL(size)                                                  \
 257   inline VEC_DATA_TYPE(int, size) asymm_mask_if_non_zero##size(VEC_DATA_TYPE(int, size) a) \
 258   {                                                                                        \
 259     const VEC_DATA_TYPE(int, size) all_zeros = 0;                                          \
 260     const VEC_DATA_TYPE(int, size) all_ones = ~0;                                          \
 261     return select(all_zeros, all_ones, a != 0);                                            \
 262   }
 263
 264 #define EXP_BARREL_SHIFTER_IMPL(size)                                                          \
 265   inline VEC_DATA_TYPE(int, size) exp_barrel_shifter##size(                                    \
 266       VEC_DATA_TYPE(int, size) result, int exponent, int fp_multiplier, int k_integer_bits,    \
 267       int k_fractional_bits, VEC_DATA_TYPE(int, size) remainder)                               \
 268   {                                                                                            \
 269     if (k_integer_bits > exponent)                                                             \
 270     {                                                                                          \
 271       const int k_shift_amount = k_integer_bits > exponent ? k_fractional_bits + exponent : 0; \
 272       return ASYMM_SELECT_USING_MASK(                                                          \
 273           ASYMM_MASK_IF_NON_ZERO(remainder & (1 << k_shift_amount), size),                     \
 274           ASYMM_MULT(result, fp_multiplier, size), result, size);                              \
 275     }                                                                                          \
 276                                                                                                \
 277     return result;                                                                             \
 278   }
 279
 280 /** Calculates \f$ exp(x) \f$ for x < 0.
 281  *
 282  * @param[in] size Size of vector.
 283  *
 284  * @return Result in fixed-point format Q0.
 285  */
 286 #define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size)                                                   \
 287   inline VEC_DATA_TYPE(int, size)                                                                 \
 288       asymm_exp_on_negative_values##size(VEC_DATA_TYPE(int, size) a, int k_integer_bits)          \
 289   {                                                                                               \
 290     const int k_fractional_bits = 31 - k_integer_bits;                                            \
 291     VEC_DATA_TYPE(int, size)                                                                      \
 292     k_one_quarter = 1 << (k_fractional_bits - 2);                                                 \
 293     VEC_DATA_TYPE(int, size)                                                                      \
 294     mask = k_one_quarter - 1;                                                                     \
 295     VEC_DATA_TYPE(int, size)                                                                      \
 296     a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter;                                 \
 297     VEC_DATA_TYPE(int, size)                                                                      \
 298     a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits;   \
 299     VEC_DATA_TYPE(int, size)                                                                      \
 300     result = ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(                       \
 301         a_mod_quarter_minus_one_quarter_scaled, size);                                            \
 302     VEC_DATA_TYPE(int, size)                                                                      \
 303     remainder = a_mod_quarter_minus_one_quarter - a;                                              \
 304                                                                                                   \
 305     result = EXP_BARREL_SHIFTER(result, -2, 1672461947, k_integer_bits, k_fractional_bits,        \
 306                                 remainder, size);                                                 \
 307     result = EXP_BARREL_SHIFTER(result, -1, 1302514674, k_integer_bits, k_fractional_bits,        \
 308                                 remainder, size);                                                 \
 309     result = EXP_BARREL_SHIFTER(result, +0, 790015084, k_integer_bits, k_fractional_bits,         \
 310                                 remainder, size);                                                 \
 311     result = EXP_BARREL_SHIFTER(result, +1, 290630308, k_integer_bits, k_fractional_bits,         \
 312                                 remainder, size);                                                 \
 313     result = EXP_BARREL_SHIFTER(result, +2, 39332535, k_integer_bits, k_fractional_bits,          \
 314                                 remainder, size);                                                 \
 315     result = EXP_BARREL_SHIFTER(result, +3, 720401, k_integer_bits, k_fractional_bits, remainder, \
 316                                 size);                                                            \
 317     result =                                                                                      \
 318         EXP_BARREL_SHIFTER(result, +4, 242, k_integer_bits, k_fractional_bits, remainder, size);  \
 319                                                                                                   \
 320     if (k_integer_bits > 5)                                                                       \
 321     {                                                                                             \
 322       const VEC_DATA_TYPE(int, size) clamp = -(1 << (k_fractional_bits + 5));                     \
 323       result = ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(a < clamp, size), 0, result, size); \
 324     }                                                                                             \
 325                                                                                                   \
 326     const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX;                                              \
 327     return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_ZERO(a, size), Q0_one, result, size);            \
 328   }
 329
 330 /** Calculates the product of a integer value by a power of two, with either a positive exponent
 331  * (equivalent to an arithmetic left shift, saturating) or a negative exponent
 332  * (equivalent to an arithmetic right shift, rounding to nearest).
 333  *
 334  * @param[in] size Size of vector.
 335  *
 336  * @return Arithmetic left or right shift.
 337  */
 338 #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size)                                    \
 339   inline VEC_DATA_TYPE(int, size)                                                            \
 340       asymm_saturating_rounding_mult_by_pow2##size(VEC_DATA_TYPE(int, size) x, int exponent) \
 341   {                                                                                          \
 342     if (exponent < 0)                                                                        \
 343     {                                                                                        \
 344       return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size);                              \
 345     }                                                                                        \
 346                                                                                              \
 347     const VEC_DATA_TYPE(int, size) min = INT_MIN;                                            \
 348     const VEC_DATA_TYPE(int, size) max = INT_MAX;                                            \
 349     int threshold = ((1 << (31 - exponent)) - 1);                                            \
 350     VEC_DATA_TYPE(int, size)                                                                 \
 351     positive_mask = ASYMM_MASK_IF_NON_ZERO(x > threshold, size);                             \
 352     VEC_DATA_TYPE(int, size)                                                                 \
 353     negative_mask = ASYMM_MASK_IF_NON_ZERO(x < -threshold, size);                            \
 354     VEC_DATA_TYPE(int, size)                                                                 \
 355     result = x << exponent;                                                                  \
 356     result = ASYMM_SELECT_USING_MASK(positive_mask, max, result, size);                      \
 357     result = ASYMM_SELECT_USING_MASK(negative_mask, min, result, size);                      \
 358     return result;                                                                           \
 359   }
 360
 361 /** Calculates (a+b)/2, rounded to the nearest integer.
 362  * Equivalent to VRHADD in the ARM NEON instruction set.
 363  *
 364  * @param[in] size Size of vector.
 365  *
 366  * @return (a+b)/2, rounded to the nearest integer.
 367  */
 368 #define ASYMM_ROUNDING_HALF_SUM_IMPL(size)                                                  \
 369   inline VEC_DATA_TYPE(int, size)                                                           \
 370       asymm_rounding_half_sum##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
 371   {                                                                                         \
 372     VEC_DATA_TYPE(long, size)                                                               \
 373     a64 = convert_long##size(a);                                                            \
 374     VEC_DATA_TYPE(long, size)                                                               \
 375     b64 = convert_long##size(b);                                                            \
 376     VEC_DATA_TYPE(long, size)                                                               \
 377     sum = a64 + b64;                                                                        \
 378     const VEC_DATA_TYPE(long, size) one = 1;                                                \
 379     const VEC_DATA_TYPE(long, size) minus_one = -1;                                         \
 380     VEC_DATA_TYPE(long, size)                                                               \
 381     sign = select(minus_one, one, sum >= 0);                                                \
 382     return convert_int##size((sum + sign) / 2);                                             \
 383   }
 384
 385 /** Calculates \f$ 1 / (1 + x) \f$ for x in (0, 1).
 386  *
 387  * @param[in] size Size of vector.
 388  *
 389  * @return Result in fixed-point format Q0.
 390  */
 391 #define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size)                      \
 392   inline VEC_DATA_TYPE(int, size)                                              \
 393       asymm_one_over_one_plus_x_for_x_in_0_1##size(VEC_DATA_TYPE(int, size) a) \
 394   {                                                                            \
 395     const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX;                           \
 396     const VEC_DATA_TYPE(int, size) Q2_one = 1 << (31 - 2);                     \
 397     VEC_DATA_TYPE(int, size)                                                   \
 398     half_denominator = ASYMM_ROUNDING_HALF_SUM(a, Q0_one, size);               \
 399     const VEC_DATA_TYPE(int, size) Q2_48_over_17 = 1515870810;                 \
 400     const VEC_DATA_TYPE(int, size) Q2_neg_32_over_17 = -1010580540;            \
 401     VEC_DATA_TYPE(int, size)                                                   \
 402     x = Q2_48_over_17 + ASYMM_MULT(half_denominator, Q2_neg_32_over_17, size); \
 403     for (int i = 0; i < 3; i++)                                                \
 404     {                                                                          \
 405       VEC_DATA_TYPE(int, size)                                                 \
 406       half_denominator_times_x = ASYMM_MULT(half_denominator, x, size);        \
 407       VEC_DATA_TYPE(int, size)                                                 \
 408       one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x;  \
 409       VEC_DATA_TYPE(int, size)                                                 \
 410       tmp = ASYMM_MULT(x, one_minus_half_denominator_times_x, size);           \
 411       x = x + ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(tmp, 2, size);            \
 412     }                                                                          \
 413     return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size);                 \
 414   }
 415
 416 /** Considering the integer value as fixed-point, change the number of integer bits and update value
 417  * accordingly.
 418  *
 419  * @param[in] size Size of vector.
 420  *
 421  * @return Rescaled value.
 422  */
 423 #define ASYMM_RESCALE_IMPL(size)                                                                  \
 424   inline VEC_DATA_TYPE(int, size) asymm_rescale##size(VEC_DATA_TYPE(int, size) value,             \
 425                                                       int src_integer_bits, int dst_integer_bits) \
 426   {                                                                                               \
 427     int exponent = src_integer_bits - dst_integer_bits;                                           \
 428     return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size);                         \
 429   }
 430
 431 #define QUANTIZE_STR(input, offset, scale, type, size) quantize_##type##size(input, offset, scale)
 432 #define QUANTIZE(input, offset, scale, type, size) QUANTIZE_STR(input, offset, scale, type, size)
 433 #define DEQUANTIZE_STR(input, offset, scale, type, size) \
 434   dequantize_##type##size(input, offset, scale)
 435 #define DEQUANTIZE(input, offset, scale, type, size) \
 436   DEQUANTIZE_STR(input, offset, scale, type, size)
 437
 438 #define ASYMM_ROUNDING_DIVIDE_BY_POW2(x, exponent, size) \
 439   asymm_rounding_divide_by_POW2_##size(x, exponent)
 440 #define ASYMM_MULT(a, b, size) asymm_mult##size(a, b)
 441 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(x, quantized_multiplier, left_shift, size) \
 442   ASYMM_MULT(x *((VEC_DATA_TYPE(int, size))(1) << (-left_shift)), quantized_multiplier, size)
 443 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, quantized_multiplier, right_shift, size) \
 444   ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(x, quantized_multiplier, size), right_shift, size)
 445 #define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a, size) \
 446   asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(a)
 447 #define ASYMM_SELECT_USING_MASK(if_mask, then_val, else_val, size) \
 448   asymm_select_using_mask##size(if_mask, then_val, else_val)
 449 #define ASYMM_MASK_IF_ZERO(a, size) asymm_mask_if_zero##size(a)
 450 #define ASYMM_MASK_IF_NON_ZERO(a, size) asymm_mask_if_non_zero##size(a)
 451 #define EXP_BARREL_SHIFTER(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, \
 452                            remainder, size)                                                    \
 453   exp_barrel_shifter##size(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, \
 454                            remainder)
 455 #define ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, size) \
 456   asymm_exp_on_negative_values##size(a, k_integer_bits)
 457 #define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1(a, size) \
 458   asymm_one_over_one_plus_x_for_x_in_0_1##size(a)
 459 #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, exponent, size) \
 460   asymm_saturating_rounding_mult_by_pow2##size(x, exponent)
 461 #define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b)
 462 #define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) \
 463   asymm_rescale##size(value, src_integer_bits, dst_integer_bits)
 464
 465 #define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size)                                               \
 466   inline VEC_DATA_TYPE(int, size)                                                                 \
 467       multiply_by_quantized_multiplier##size(VEC_DATA_TYPE(int, size) input, int qmul, int shift) \
 468   {                                                                                               \
 469     const int left_shift = shift > 0 ? shift : 0;                                                 \
 470     const int right_shift = shift > 0 ? 0 : -shift;                                               \
 471     return ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(input * (1 << left_shift), qmul, size),       \
 472                                          right_shift, size);                                      \
 473   }
 474 #define MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, size) \
 475   multiply_by_quantized_multiplier##size(input, qmul, shift)
 476
 477 QUANTIZE_IMPL(uchar, 1)
 478 QUANTIZE_IMPL(char, 1)
 479 QUANTIZE_IMPL(uint, 1)
 480 QUANTIZE_IMPL(int, 1)
 481 QUANTIZE_IMPL(uchar, 4)
 482 QUANTIZE_IMPL(ushort, 4)
 483 QUANTIZE_IMPL(short, 4)
 484 QUANTIZE_IMPL(uchar, 16)
 485 QUANTIZE_IMPL(char, 16)
 486 QUANTIZE_IMPL(ushort, 16)
 487 QUANTIZE_IMPL(short, 16)
 488 QUANTIZE_IMPL(uint, 16)
 489 QUANTIZE_IMPL(int, 16)
 490
 491 DEQUANTIZE_IMPL(uchar, 1)
 492 DEQUANTIZE_IMPL(char, 1)
 493 DEQUANTIZE_IMPL(uint, 1)
 494 DEQUANTIZE_IMPL(int, 1)
 495 DEQUANTIZE_IMPL(uchar, 4)
 496 DEQUANTIZE_IMPL(ushort, 4)
 497 DEQUANTIZE_IMPL(short, 4)
 498 DEQUANTIZE_IMPL(uchar, 16)
 499 DEQUANTIZE_IMPL(char, 16)
 500 DEQUANTIZE_IMPL(ushort, 16)
 501 DEQUANTIZE_IMPL(short, 16)
 502 DEQUANTIZE_IMPL(uint, 16)
 503 DEQUANTIZE_IMPL(int, 16)
 504
 505 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1)
 506 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
 507 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4)
 508 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8)
 509 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
 510
 511 ASYMM_MULT_IMPL(1)
 512 ASYMM_MULT_IMPL(2)
 513 ASYMM_MULT_IMPL(4)
 514 ASYMM_MULT_IMPL(8)
 515 ASYMM_MULT_IMPL(16)
 516
 517 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(2)
 518 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(4)
 519 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(8)
 520 ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(16)
 521
 522 ASYMM_SELECT_USING_MASK_IMPL(1)
 523 ASYMM_SELECT_USING_MASK_IMPL(2)
 524 ASYMM_SELECT_USING_MASK_IMPL(4)
 525 ASYMM_SELECT_USING_MASK_IMPL(8)
 526 ASYMM_SELECT_USING_MASK_IMPL(16)
 527
 528 ASYMM_MASK_IF_ZERO_IMPL(1)
 529 ASYMM_MASK_IF_ZERO_IMPL(2)
 530 ASYMM_MASK_IF_ZERO_IMPL(4)
 531 ASYMM_MASK_IF_ZERO_IMPL(8)
 532 ASYMM_MASK_IF_ZERO_IMPL(16)
 533
 534 ASYMM_MASK_IF_NON_ZERO_IMPL(1)
 535 ASYMM_MASK_IF_NON_ZERO_IMPL(2)
 536 ASYMM_MASK_IF_NON_ZERO_IMPL(4)
 537 ASYMM_MASK_IF_NON_ZERO_IMPL(8)
 538 ASYMM_MASK_IF_NON_ZERO_IMPL(16)
 539
 540 EXP_BARREL_SHIFTER_IMPL(2)
 541 EXP_BARREL_SHIFTER_IMPL(4)
 542 EXP_BARREL_SHIFTER_IMPL(8)
 543 EXP_BARREL_SHIFTER_IMPL(16)
 544
 545 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(2)
 546 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(4)
 547 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(8)
 548 ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(16)
 549
 550 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(1)
 551 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(2)
 552 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(4)
 553 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(8)
 554 ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(16)
 555
 556 ASYMM_ROUNDING_HALF_SUM_IMPL(2)
 557 ASYMM_ROUNDING_HALF_SUM_IMPL(4)
 558 ASYMM_ROUNDING_HALF_SUM_IMPL(8)
 559 ASYMM_ROUNDING_HALF_SUM_IMPL(16)
 560
 561 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(2)
 562 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(4)
 563 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(8)
 564 ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(16)
 565
 566 ASYMM_RESCALE_IMPL(1)
 567 ASYMM_RESCALE_IMPL(2)
 568 ASYMM_RESCALE_IMPL(4)
 569 ASYMM_RESCALE_IMPL(8)
 570 ASYMM_RESCALE_IMPL(16)
 571
 572 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(1)
 573 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(2)
 574 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(4)
 575 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(8)
 576 MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(16)
 577
 578 #endif // ARM_COMPUTE_HELPERS_ASYMM_H