24.02.1
|
Go to the documentation of this file.
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) \
89 quantize_##type##size(VEC_DATA_TYPE(float, size) input, float offset, float scale) \
91 VEC_DATA_TYPE(float, size) \
92 out_f32 = input / (VEC_DATA_TYPE(float, size))(scale) + (VEC_DATA_TYPE(float, size))(offset); \
93 VEC_DATA_TYPE(type, size) \
94 res = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, VEC_DATA_TYPE(int, size)), VEC_DATA_TYPE(type, size)); \
105 #define DEQUANTIZE_IMPL(type, size) \
106 inline VEC_DATA_TYPE(float, size) \
107 dequantize_##type##size(VEC_DATA_TYPE(type, size) input, float offset, float scale) \
109 return (CONVERT(input, VEC_DATA_TYPE(float, size)) - offset) * scale; \
118 #define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size) \
119 inline VEC_DATA_TYPE(int, size) \
120 asymm_rounding_divide_by_POW2_##size(VEC_DATA_TYPE(int, size) x, VEC_DATA_TYPE(int, size) exponent) \
122 const VEC_DATA_TYPE(int, size) zero = (VEC_DATA_TYPE(int, size))0; \
123 const VEC_DATA_TYPE(int, size) one = (VEC_DATA_TYPE(int, size))1; \
124 VEC_DATA_TYPE(int, size) \
125 mask = (one << exponent) - one; \
126 VEC_DATA_TYPE(int, size) \
127 threshold = (mask >> 1) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))(x < 0)); \
128 return (x >> exponent) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))((x & mask) > threshold)); \
138 #define ASYMM_MULT_IMPL(size) \
139 inline VEC_DATA_TYPE(int, size) asymm_mult##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
141 VEC_DATA_TYPE(int, size) \
142 overflow = a == b && a == INT_MIN; \
143 VEC_DATA_TYPE(long, size) \
144 a_64 = convert_long##size(a); \
145 VEC_DATA_TYPE(long, size) \
146 b_64 = convert_long##size(b); \
147 VEC_DATA_TYPE(long, size) \
148 ab_64 = a_64 * b_64; \
150 VEC_DATA_TYPE(long, size) \
152 VEC_DATA_TYPE(long, size) \
153 mask2 = 1 - (1 << 30); \
154 VEC_DATA_TYPE(long, size) \
155 is_positive_or_zero = ab_64 >= 0; \
156 VEC_DATA_TYPE(long, size) \
157 nudge = select(mask2, mask1, (SELECT_VEC_DATA_TYPE(long, size))(is_positive_or_zero)); \
158 VEC_DATA_TYPE(long, size) \
160 VEC_DATA_TYPE(int, size) \
161 ab_x2_high32 = convert_int##size((ab_64 + nudge) / mask); \
162 return select(ab_x2_high32, INT_MAX, (SELECT_VEC_DATA_TYPE(int, size))(overflow)); \
171 #define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size) \
172 inline VEC_DATA_TYPE(int, size) \
173 asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(VEC_DATA_TYPE(int, size) a) \
175 const VEC_DATA_TYPE(int, size) constant_term = 1895147668; \
176 const VEC_DATA_TYPE(int, size) constant_1_over_3 = 715827883; \
177 const int k_fractional_bits = 31; \
178 VEC_DATA_TYPE(int, size) \
179 x = a + (1 << (k_fractional_bits - 3)); \
180 VEC_DATA_TYPE(int, size) \
181 x2 = ASYMM_MULT(x, x, size); \
182 VEC_DATA_TYPE(int, size) \
183 x3 = ASYMM_MULT(x2, x, size); \
184 VEC_DATA_TYPE(int, size) \
185 x4 = ASYMM_MULT(x2, x2, size); \
186 VEC_DATA_TYPE(int, size) \
187 x4_over_4 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4, 2, size); \
188 VEC_DATA_TYPE(int, size) \
189 x4_over_24_plus_x3_over_6_plus_x2 = ASYMM_MULT((x4_over_4 + x3), constant_1_over_3, size) + x2; \
190 VEC_DATA_TYPE(int, size) \
191 x4_over_24_plus_x3_over_6_plus_x2_over_2 = \
192 ASYMM_ROUNDING_DIVIDE_BY_POW2(x4_over_24_plus_x3_over_6_plus_x2, 1, size); \
193 return constant_term + ASYMM_MULT(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2, size); \
204 #define ASYMM_SELECT_USING_MASK_IMPL(size) \
205 inline VEC_DATA_TYPE(int, size) asymm_select_using_mask##size( \
206 VEC_DATA_TYPE(int, size) if_mask, VEC_DATA_TYPE(int, size) then_val, VEC_DATA_TYPE(int, size) else_val) \
208 return (if_mask & then_val) ^ (~if_mask & else_val); \
218 #define ASYMM_MASK_IF_ZERO_IMPL(size) \
219 inline VEC_DATA_TYPE(int, size) asymm_mask_if_zero##size(VEC_DATA_TYPE(int, size) a) \
221 const VEC_DATA_TYPE(int, size) all_zeros = 0; \
222 const VEC_DATA_TYPE(int, size) all_ones = ~0; \
223 return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a == 0)); \
233 #define ASYMM_MASK_IF_NON_ZERO_IMPL(size) \
234 inline VEC_DATA_TYPE(int, size) asymm_mask_if_non_zero##size(VEC_DATA_TYPE(int, size) a) \
236 const VEC_DATA_TYPE(int, size) all_zeros = 0; \
237 const VEC_DATA_TYPE(int, size) all_ones = ~0; \
238 return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a != 0)); \
241 #define EXP_BARREL_SHIFTER_IMPL(size) \
242 inline VEC_DATA_TYPE(int, size) \
243 exp_barrel_shifter##size(VEC_DATA_TYPE(int, size) result, int exponent, int fp_multiplier, int k_integer_bits, \
244 int k_fractional_bits, VEC_DATA_TYPE(int, size) remainder) \
246 if (k_integer_bits > exponent) \
248 const int k_shift_amount = k_integer_bits > exponent ? k_fractional_bits + exponent : 0; \
249 return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(remainder & (1 << k_shift_amount), size), \
250 ASYMM_MULT(result, fp_multiplier, size), result, size); \
262 #define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size) \
263 inline VEC_DATA_TYPE(int, size) asymm_exp_on_negative_values##size(VEC_DATA_TYPE(int, size) a, int k_integer_bits) \
265 const int k_fractional_bits = 31 - k_integer_bits; \
266 VEC_DATA_TYPE(int, size) \
267 k_one_quarter = 1 << (k_fractional_bits - 2); \
268 VEC_DATA_TYPE(int, size) \
269 mask = k_one_quarter - 1; \
270 VEC_DATA_TYPE(int, size) \
271 a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter; \
272 VEC_DATA_TYPE(int, size) \
273 a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits; \
274 VEC_DATA_TYPE(int, size) \
275 result = ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a_mod_quarter_minus_one_quarter_scaled, \
277 VEC_DATA_TYPE(int, size) \
278 remainder = a_mod_quarter_minus_one_quarter - a; \
280 result = EXP_BARREL_SHIFTER(result, -2, 1672461947, k_integer_bits, k_fractional_bits, remainder, size); \
281 result = EXP_BARREL_SHIFTER(result, -1, 1302514674, k_integer_bits, k_fractional_bits, remainder, size); \
282 result = EXP_BARREL_SHIFTER(result, +0, 790015084, k_integer_bits, k_fractional_bits, remainder, size); \
283 result = EXP_BARREL_SHIFTER(result, +1, 290630308, k_integer_bits, k_fractional_bits, remainder, size); \
284 result = EXP_BARREL_SHIFTER(result, +2, 39332535, k_integer_bits, k_fractional_bits, remainder, size); \
285 result = EXP_BARREL_SHIFTER(result, +3, 720401, k_integer_bits, k_fractional_bits, remainder, size); \
286 result = EXP_BARREL_SHIFTER(result, +4, 242, k_integer_bits, k_fractional_bits, remainder, size); \
288 if (k_integer_bits > 5) \
290 const VEC_DATA_TYPE(int, size) clamp = -(1 << (k_fractional_bits + 5)); \
291 result = ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(a < clamp, size), 0, result, size); \
294 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
295 return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_ZERO(a, size), Q0_one, result, size); \
306 #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size) \
307 inline VEC_DATA_TYPE(int, size) \
308 asymm_saturating_rounding_mult_by_pow2##size(VEC_DATA_TYPE(int, size) x, int exponent) \
312 return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size); \
315 const VEC_DATA_TYPE(int, size) min = INT_MIN; \
316 const VEC_DATA_TYPE(int, size) max = INT_MAX; \
317 int threshold = ((1 << (31 - exponent)) - 1); \
318 VEC_DATA_TYPE(int, size) \
319 positive_mask = ASYMM_MASK_IF_NON_ZERO(x > threshold, size); \
320 VEC_DATA_TYPE(int, size) \
321 negative_mask = ASYMM_MASK_IF_NON_ZERO(x < -threshold, size); \
322 VEC_DATA_TYPE(int, size) \
323 result = x << exponent; \
324 result = ASYMM_SELECT_USING_MASK(positive_mask, max, result, size); \
325 result = ASYMM_SELECT_USING_MASK(negative_mask, min, result, size); \
336 #define ASYMM_ROUNDING_HALF_SUM_IMPL(size) \
337 inline VEC_DATA_TYPE(int, size) \
338 asymm_rounding_half_sum##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
340 VEC_DATA_TYPE(long, size) \
341 a64 = convert_long##size(a); \
342 VEC_DATA_TYPE(long, size) \
343 b64 = convert_long##size(b); \
344 VEC_DATA_TYPE(long, size) \
346 const VEC_DATA_TYPE(long, size) one = 1; \
347 const VEC_DATA_TYPE(long, size) minus_one = -1; \
348 VEC_DATA_TYPE(long, size) \
349 sign = select(minus_one, one, (SELECT_VEC_DATA_TYPE(long, size))(sum >= 0)); \
350 return convert_int##size((sum + sign) / 2); \
359 #define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size) \
360 inline VEC_DATA_TYPE(int, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(VEC_DATA_TYPE(int, size) a) \
362 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
363 const VEC_DATA_TYPE(int, size) Q2_one = 1 << (31 - 2); \
364 VEC_DATA_TYPE(int, size) \
365 half_denominator = ASYMM_ROUNDING_HALF_SUM(a, Q0_one, size); \
366 const VEC_DATA_TYPE(int, size) Q2_48_over_17 = 1515870810; \
367 const VEC_DATA_TYPE(int, size) Q2_neg_32_over_17 = -1010580540; \
368 VEC_DATA_TYPE(int, size) \
369 x = Q2_48_over_17 + ASYMM_MULT(half_denominator, Q2_neg_32_over_17, size); \
370 for (int i = 0; i < 3; i++) \
372 VEC_DATA_TYPE(int, size) \
373 half_denominator_times_x = ASYMM_MULT(half_denominator, x, size); \
374 VEC_DATA_TYPE(int, size) \
375 one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x; \
376 VEC_DATA_TYPE(int, size) \
377 tmp = ASYMM_MULT(x, one_minus_half_denominator_times_x, size); \
378 x = x + ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(tmp, 2, size); \
380 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size); \
389 #define ASYMM_RESCALE_IMPL(size) \
390 inline VEC_DATA_TYPE(int, size) \
391 asymm_rescale##size(VEC_DATA_TYPE(int, size) value, int src_integer_bits, int dst_integer_bits) \
393 int exponent = src_integer_bits - dst_integer_bits; \
394 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size); \
397 #define QUANTIZE_STR(input, offset, scale, type, size) quantize_##type##size(input, offset, scale)
398 #define QUANTIZE(input, offset, scale, type, size) QUANTIZE_STR(input, offset, scale, type, size)
399 #define DEQUANTIZE_STR(input, offset, scale, type, size) dequantize_##type##size(input, offset, scale)
400 #define DEQUANTIZE(input, offset, scale, type, size) DEQUANTIZE_STR(input, offset, scale, type, size)
402 #define ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size) asymm_rounding_divide_by_POW2_##size(x, exponent)
403 #define ASYMM_ROUNDING_DIVIDE_BY_POW2(x, exponent, size) ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size)
404 #define ASYMM_MULT_STR(a, b, size) asymm_mult##size(a, b)
405 #define ASYMM_MULT(a, b, size) ASYMM_MULT_STR(a, b, size)
406 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(x, quantized_multiplier, left_shift, size) \
407 ASYMM_MULT(x *((VEC_DATA_TYPE(int, size))(1) << (-left_shift)), quantized_multiplier, size)
408 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, quantized_multiplier, right_shift, size) \
409 ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(x, quantized_multiplier, size), right_shift, size)
410 #define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a, size) \
411 asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(a)
412 #define ASYMM_SELECT_USING_MASK(if_mask, then_val, else_val, size) \
413 asymm_select_using_mask##size(if_mask, then_val, else_val)
414 #define ASYMM_MASK_IF_ZERO(a, size) asymm_mask_if_zero##size(a)
415 #define ASYMM_MASK_IF_NON_ZERO(a, size) asymm_mask_if_non_zero##size(a)
416 #define EXP_BARREL_SHIFTER(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder, size) \
417 exp_barrel_shifter##size(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder)
418 #define ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size) asymm_exp_on_negative_values##size(a, k_integer_bits)
419 #define ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, size) ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size)
420 #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)
421 #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)
422 #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, exponent, size) \
423 asymm_saturating_rounding_mult_by_pow2##size(x, exponent)
424 #define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b)
425 #define ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size) \
426 asymm_rescale##size(value, src_integer_bits, dst_integer_bits)
427 #define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) \
428 ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size)
430 #define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size) \
431 inline VEC_DATA_TYPE(int, size) \
432 multiply_by_quantized_multiplier##size(VEC_DATA_TYPE(int, size) input, int qmul, int shift) \
434 const int left_shift = shift > 0 ? shift : 0; \
435 const int right_shift = shift > 0 ? 0 : -shift; \
436 return ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(input * (1 << left_shift), qmul, size), right_shift, size); \
438 #define MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, size) \
439 multiply_by_quantized_multiplier##size(input, qmul, shift)
586 #endif // ARM_COMPUTE_HELPERS_ASYMM_H
#define EXP_BARREL_SHIFTER_IMPL(size)
#define CONVERT_SAT(x, type)
#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 DEQUANTIZE_IMPL(type, size)
Dequantize a vector of values to floating-point.
#define QUANTIZE_IMPL(type, size)
Quantize a vector of values from floating-point.
#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...
uchar quantize_qasymm8(float input, float offset, float scale)
Quantize a floating-point scalar value to 8-bit asymmetric.
#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_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size)
Calculates for x in [-1/4, 0).
__global uchar * offset(const Image *img, int x, int y)
Get the pointer position of a Image.
#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 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_RESCALE_IMPL(size)
Considering the integer value as fixed-point, change the number of integer bits and update value acco...
#define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size)
Calculates for x < 0.
#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size)
Calculates for x in (0, 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 ASYMM_ROUNDING_HALF_SUM_IMPL(size)
Calculates (a+b)/2, rounded to the nearest integer.
#define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size)