Marat Dukhan | 00a1085 | 2021-12-02 13:00:40 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
Marat Dukhan | bdf1099 | 2022-01-04 09:20:14 -0800 | [diff] [blame^] | 2 | // Template: src/f32-qs8-vcvt/scalar-imagic.c.in |
Marat Dukhan | 00a1085 | 2021-12-02 13:00:40 -0800 | [diff] [blame] | 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/common.h> |
| 13 | #include <xnnpack/math.h> |
| 14 | #include <xnnpack/vcvt.h> |
| 15 | |
| 16 | #include <fp16.h> |
| 17 | |
| 18 | |
Marat Dukhan | bdf1099 | 2022-01-04 09:20:14 -0800 | [diff] [blame^] | 19 | void xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x1( |
Marat Dukhan | 00a1085 | 2021-12-02 13:00:40 -0800 | [diff] [blame] | 20 | size_t n, |
| 21 | const float* x, |
| 22 | uint8_t* y, |
Marat Dukhan | d541fc0 | 2021-12-14 00:14:27 -0800 | [diff] [blame] | 23 | const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) |
Marat Dukhan | 00a1085 | 2021-12-02 13:00:40 -0800 | [diff] [blame] | 24 | { |
| 25 | assert(n != 0); |
Marat Dukhan | 4bdc9f5 | 2021-12-09 15:36:06 -0800 | [diff] [blame] | 26 | assert(n % sizeof(float) == 0); |
Marat Dukhan | 00a1085 | 2021-12-02 13:00:40 -0800 | [diff] [blame] | 27 | assert(x != NULL); |
| 28 | assert(y != NULL); |
| 29 | |
Marat Dukhan | bdf1099 | 2022-01-04 09:20:14 -0800 | [diff] [blame^] | 30 | const float vscale = params->scalar_imagic.scale; |
| 31 | const float vmagic_bias = params->scalar_imagic.magic_bias; |
| 32 | const int32_t vmagic_min = params->scalar_imagic.magic_min; |
| 33 | const int32_t vmagic_max = params->scalar_imagic.magic_max; |
| 34 | const int32_t vmagic_bias_less_zero_point = params->scalar_imagic.magic_bias_less_zero_point; |
Marat Dukhan | 00a1085 | 2021-12-02 13:00:40 -0800 | [diff] [blame] | 35 | |
| 36 | do { |
| 37 | float vx = *x++; |
| 38 | vx *= vscale; |
| 39 | vx += vmagic_bias; |
| 40 | |
| 41 | int32_t vy = (int32_t) fp32_to_bits(vx); |
| 42 | vy = math_max_s32(vy, vmagic_min); |
| 43 | vy = math_min_s32(vy, vmagic_max); |
| 44 | vy -= vmagic_bias_less_zero_point; |
| 45 | |
| 46 | *y++ = (uint8_t) vy; |
| 47 | |
Marat Dukhan | 4bdc9f5 | 2021-12-09 15:36:06 -0800 | [diff] [blame] | 48 | n -= sizeof(float); |
Marat Dukhan | 00a1085 | 2021-12-02 13:00:40 -0800 | [diff] [blame] | 49 | } while (n != 0); |
| 50 | } |