Marat Dukhan | 76e78c8 | 2021-07-20 21:11:23 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-vadd/scalar.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2021 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <xnnpack/math.h> |
| 13 | #include <xnnpack/vadd.h> |
| 14 | |
| 15 | |
| 16 | void xnn_qu8_vadd_minmax_ukernel__scalar_x1( |
| 17 | size_t n, |
| 18 | const uint8_t* input_a, |
| 19 | const uint8_t* input_b, |
| 20 | uint8_t* output, |
| 21 | const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
| 22 | { |
| 23 | const int32_t vbias = params->scalar.bias; |
| 24 | const int32_t va_multiplier = params->scalar.a_multiplier; |
| 25 | const int32_t vb_multiplier = params->scalar.b_multiplier; |
| 26 | const int32_t vrounding = params->scalar.rounding; |
| 27 | const uint32_t vshift = params->scalar.shift; |
| 28 | const int32_t voutput_min_less_zero_point = params->scalar.output_min_less_zero_point; |
| 29 | const int32_t voutput_max_less_zero_point = params->scalar.output_max_less_zero_point; |
| 30 | const int32_t voutput_zero_point = params->scalar.output_zero_point; |
| 31 | |
| 32 | do { |
| 33 | const int32_t va = *input_a++; |
| 34 | const int32_t vb = *input_b++; |
| 35 | const int32_t vacc = vbias + va * va_multiplier + vb * vb_multiplier; |
| 36 | |
| 37 | int32_t vout = asr_s32(vacc + vrounding, vshift); |
| 38 | vout = math_max_s32(vout, voutput_min_less_zero_point); |
| 39 | vout = math_min_s32(vout, voutput_max_less_zero_point); |
| 40 | *output++ = (uint8_t) (vout + voutput_zero_point); |
| 41 | |
| 42 | n -= sizeof(uint8_t); |
| 43 | } while (n != 0); |
| 44 | } |