2 * Copyright (c) 2017 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #ifndef ARM_COMPUTE_ASYMM_HELPER_H
25 #define ARM_COMPUTE_ASYMM_HELPER_H
27 // Algoriths for these functions were taken from
28 // https://github.com/google/gemmlowp/blob/master/fixedpoint/fixedpoint.h
29 // and adapted to operate on integer vectors.
31 /** For each element of input vector, the corresponding bits of the result item are set
32 * if the input item is zero.
34 * @param[in] a Input vector whose zero bits define which corresponding bits in result will be set.
36 * @returns Output vector with bits set when corresponding bit in @p a is zero.
38 inline int16 asymm_mask_if_zero(int16 a)
40 const int16 all_zeros = 0;
41 const int16 all_ones = ~0;
42 return select(all_zeros, all_ones, a == 0);
45 /** For each element of input vector, the corresponding bits of the result item are set
46 * if the input item is non-zero.
48 * @param[in] a Input vector whose non-zero bits define which corresponding bits in result will be set.
50 * @returns Output vector with bits set when corresponding bit in @p a is non zero.
52 inline int16 asymm_mask_if_non_zero(int16 a)
54 const int16 all_zeros = 0;
55 const int16 all_ones = ~0;
56 return select(all_zeros, all_ones, a != 0);
59 /** Each bit of the result is set to the corresponding bit of either then_val or
60 * else_val depending on whether the corresponding bit of if_mask is set.
61 * Equivalent to the VBSL instruction in ARM NEON.
63 * @param[in] if_mask Mask defines will bit be taken from @p then_val or @p else_val depending on corresponding bit in mask is set or not.
64 * @param[in] then_val Value whose bit will be used for result when corresponding bit in @p if_mask is set.
65 * @param[in] else_val Value whose bit will be used for result when corresponding bit in @p if_mask is not set.
67 * @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.
69 inline int16 asymm_select_using_mask(int16 if_mask, int16 then_val, int16 else_val)
71 return (if_mask & then_val) ^ (~if_mask & else_val);
74 /** Correctly rounded to nearest division by a power of two.
75 * Also known as a rounding arithmetic right shift.
77 * @param[in] x Value needed to be divided by power of two.
78 * @param[in] exponent Power of two, must be positive number.
80 * @return Arithmetic right shift.
82 inline int16 asymm_rounding_divide_by_pow2(int16 x, int exponent)
84 int16 mask = (1 << exponent) - 1;
87 int16 threshold = (mask >> 1) + select(zero, one, x < 0);
88 return (x >> exponent) + select(zero, one, (x & mask) > threshold);
91 /** Calculates the product of a integer value by a power of two, with either a positive exponent
92 * (equivalent to an arithmetic left shift, saturating) or a negative exponent
93 * (equivalent to an arithmetic right shift, rounding to nearest).
95 * @param[in] x Value needed to be multiplied or divided by power of two depending on sign of @p exponent.
96 * @param[in] exponent Power of two, can be positive or negative number.
98 * @return Arithmetic left or right shift.
100 inline int16 asymm_saturating_rounding_mult_by_pow2(int16 x, int exponent)
104 return asymm_rounding_divide_by_pow2(x, -exponent);
107 const int16 min = INT_MIN;
108 const int16 max = INT_MAX;
109 int threshold = ((1 << (31 - exponent)) - 1);
110 int16 positive_mask = asymm_mask_if_non_zero(x > threshold);
111 int16 negative_mask = asymm_mask_if_non_zero(x < -threshold);
112 int16 result = x << exponent;
113 result = asymm_select_using_mask(positive_mask, max, result);
114 result = asymm_select_using_mask(negative_mask, min, result);
118 /** Calculates (a+b)/2, rounded to the nearest integer.
119 * Equivalent to VRHADD in the ARM NEON instruction set.
121 * @param[in] a First term of half-sum.
122 * @param[in] b Second term of half-sum.
124 * @return (a+b)/2, rounded to the nearest integer.
126 inline int16 asymm_rounding_half_sum(int16 a, int16 b)
128 long16 a64 = convert_long16(a);
129 long16 b64 = convert_long16(b);
130 long16 sum = a64 + b64;
131 const long16 one = 1;
132 const long16 minus_one = -1;
133 long16 sign = select(minus_one, one, sum >= 0);
134 return convert_int16((sum + sign) / 2);
137 /** Product of two numbers, interpreting them as fixed-point values in the interval [-1, 1),
138 * rounding to the nearest value, and saturating -1 * -1 to the maximum value.
139 * This is equivalent to the VQRDMULH instruction in ARM NEON.
141 * @param[in] a First term of product.
142 * @param[in] b Second term of product.
144 * @return Product of two numbers.
146 inline int16 asymm_saturating_rounding_doubling_high_mul(int16 a, int16 b)
148 int16 overflow = (a == b) && (a == INT_MIN);
149 long16 a_64 = convert_long16(a);
150 long16 b_64 = convert_long16(b);
151 long16 ab_64 = a_64 * b_64;
152 long16 mask1 = 1 << 30;
153 long16 mask2 = 1 - (1 << 30);
154 long16 nudge = select(mask2, mask1, ab_64 >= 0);
155 long16 mask = 1ll << 31;
156 int16 ab_x2_high32 = convert_int16((ab_64 + nudge) / mask);
157 return select(ab_x2_high32, INT_MAX, overflow);
160 /** Fixed-point multiplication.
162 * @param[in] a Argument 1 in fixed-point format Q(a).
163 * @param[in] b Argument 2 in fixed-point format Q(b).
165 * @return Result in fixed-point format Q(a+b).
167 inline int16 asymm_mult(int16 a, int16 b)
169 return asymm_saturating_rounding_doubling_high_mul(a, b);
172 /** Calculates \f$ exp(x) \f$ for x in [-1/4, 0).
174 * @param[in] a Argument in fixed-point format Q0.
176 * @return Result in fixed-point format Q0.
178 inline int16 asymm_exp_on_interval_between_negative_one_quarter_and_0_excl(int16 a)
180 const int16 constant_term = 1895147668;
181 const int16 constant_1_over_3 = 715827883;
182 const int k_fractional_bits = 31;
183 int16 x = a + (1 << (k_fractional_bits - 3));
184 int16 x2 = asymm_mult(x, x);
185 int16 x3 = asymm_mult(x2, x);
186 int16 x4 = asymm_mult(x2, x2);
187 int16 x4_over_4 = asymm_rounding_divide_by_pow2(x4, 2);
188 int16 x4_over_24_plus_x3_over_6_plus_x2 = asymm_mult((x4_over_4 + x3), constant_1_over_3) + x2;
189 int16 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);
190 return constant_term + asymm_mult(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2);
193 /** Calculates \f$ exp(x) \f$ for x < 0.
195 * @param[in] a Argument in fixed-point format Q(k_integer_bits).
196 * @param[in] k_integer_bits Number of integer bit in argument.
198 * @return Result in fixed-point format Q0.
200 inline int16 asymm_exp_on_negative_values(int16 a, int k_integer_bits)
202 const int k_fractional_bits = 31 - k_integer_bits;
203 int16 k_one_quarter = 1 << (k_fractional_bits - 2);
204 int16 mask = k_one_quarter - 1;
205 int16 a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter;
206 int16 a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits;
207 int16 result = asymm_exp_on_interval_between_negative_one_quarter_and_0_excl(a_mod_quarter_minus_one_quarter_scaled);
208 int16 remainder = a_mod_quarter_minus_one_quarter - a;
210 #define EXP_BARREL_SHIFTER(Exponent, FixedPointMultiplier) \
211 if(k_integer_bits > Exponent) \
213 const int k_shift_amount = k_integer_bits > Exponent ? k_fractional_bits + Exponent : 0; \
214 result = asymm_select_using_mask( \
215 asymm_mask_if_non_zero(remainder & (1 << k_shift_amount)), \
216 asymm_mult(result, FixedPointMultiplier), result); \
218 EXP_BARREL_SHIFTER(-2, 1672461947);
219 EXP_BARREL_SHIFTER(-1, 1302514674);
220 EXP_BARREL_SHIFTER(+0, 790015084);
221 EXP_BARREL_SHIFTER(+1, 290630308);
222 EXP_BARREL_SHIFTER(+2, 39332535);
223 EXP_BARREL_SHIFTER(+3, 720401);
224 EXP_BARREL_SHIFTER(+4, 242);
225 #undef EXP_BARREL_SHIFTER
227 if(k_integer_bits > 5)
229 const int16 clamp = -(1 << (k_fractional_bits + 5));
230 result = asymm_select_using_mask(asymm_mask_if_non_zero(a < clamp), 0, result);
233 const int16 Q0_one = INT_MAX;
234 return asymm_select_using_mask(asymm_mask_if_zero(a), Q0_one, result);
237 /** Calculates \f$ 1 / (1 + x) \f$ for x in (0, 1).
239 * @param[in] a Argument in fixed-point format Q0.
241 * @return Result in fixed-point format Q0.
243 inline int16 asymm_one_over_one_plus_x_for_x_in_0_1(int16 a)
245 const int16 Q0_one = INT_MAX;
246 const int16 Q2_one = 1 << (31 - 2);
247 int16 half_denominator = asymm_rounding_half_sum(a, Q0_one);
248 const int16 Q2_48_over_17 = 1515870810;
249 const int16 Q2_neg_32_over_17 = -1010580540;
250 int16 x = Q2_48_over_17 + asymm_mult(half_denominator, Q2_neg_32_over_17);
251 for(int i = 0; i < 3; i++)
253 int16 half_denominator_times_x = asymm_mult(half_denominator, x);
254 int16 one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x;
255 int16 tmp = asymm_mult(x, one_minus_half_denominator_times_x);
256 x = x + asymm_saturating_rounding_mult_by_pow2(tmp, 2);
258 return asymm_saturating_rounding_mult_by_pow2(x, 1);
261 /** Considering the integer value as fixed-point, change the number of integer bits and update value accordingly.
263 * @param[in] value Value to be rescaled.
264 * @param[in] src_integer_bits Old number of integer bits.
265 * @param[in] dst_integer_bits New number of integer bits.
267 * @return Rescaled value.
269 inline int16 asymm_rescale(int16 value, int src_integer_bits, int dst_integer_bits)
271 int exponent = src_integer_bits - dst_integer_bits;
272 return asymm_saturating_rounding_mult_by_pow2(value, exponent);
275 #endif // ARM_COMPUTE_ASYMM_HELPER_H