24 #ifndef ARM_COMPUTE_HELPERS_ASYMM_H 25 #define ARM_COMPUTE_HELPERS_ASYMM_H 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) 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) \ 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)); \ 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) \ 107 return (CONVERT(input, VEC_DATA_TYPE(float, size)) - offset) * scale; \ 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) \ 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)); \ 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) \ 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; \ 149 VEC_DATA_TYPE(long, size) \ 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) \ 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)); \ 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) \ 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); \ 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) \ 204 return (if_mask & then_val) ^ (~if_mask & else_val); \ 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) \ 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)); \ 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) \ 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)); \ 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) \ 240 if(k_integer_bits > exponent) \ 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); \ 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) \ 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; \ 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); \ 282 if(k_integer_bits > 5) \ 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); \ 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); \ 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) \ 305 return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size); \ 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); \ 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) \ 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) \ 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); \ 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) \ 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++) \ 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); \ 372 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size); \ 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) \ 384 int exponent = src_integer_bits - dst_integer_bits; \ 385 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size); \ 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) 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) 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) \ 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); \ 422 #define MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, size) multiply_by_quantized_multiplier##size(input, qmul, shift) 569 #endif // ARM_COMPUTE_HELPERS_ASYMM_H __global uchar * offset(const Image *img, int x, int y)
Get the pointer position of a Image.
#define ASYMM_ROUNDING_HALF_SUM_IMPL(size)
Calculates (a+b)/2, rounded to the nearest integer.
#define DEQUANTIZE_IMPL(type, size)
Dequantize a vector of values to floating-point.
#define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size)
Calculates for x < 0.
#define EXP_BARREL_SHIFTER_IMPL(size)
#define ASYMM_SELECT_USING_MASK_IMPL(size)
Each bit of the result is set to the corresponding bit of either then_val or else_val depending on wh...
#define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size)
Calculates for x in [-1/4, 0).
#define CONVERT_SAT(a, b)
#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size)
Calculates for x in (0, 1).
#define ASYMM_MASK_IF_ZERO_IMPL(size)
For each element of input vector, the corresponding bits of the result item are set if the input item...
#define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size)
Correctly-rounded-to-nearest division by a power-of-two.
float dequantize_qasymm8(uchar input, float offset, float scale)
Dequantize a scalar value from 8-bit asymmetric to floating-point.
#define ASYMM_RESCALE_IMPL(size)
Considering the integer value as fixed-point, change the number of integer bits and update value acco...
#define QUANTIZE_IMPL(type, size)
Quantize a vector of values from floating-point.
uchar quantize_qasymm8(float input, float offset, float scale)
Quantize a floating-point scalar value to 8-bit asymmetric.
#define CONVERT_DOWN_RTE(x, type)
float dequantize_qasymm8_signed(char input, float offset, float scale)
Dequantize a scalar value from signed 8-bit asymmetric to floating-point.
#define ASYMM_MASK_IF_NON_ZERO_IMPL(size)
For each element of input vector, the corresponding bits of the result item are set if the input item...
#define ASYMM_MULT_IMPL(size)
Product of two numbers, interpreting them as fixed-point values in the interval [-1,...
#define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size)
Calculates the product of a integer value by a power of two, with either a positive exponent (equival...
#define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size)