1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/scalar-imagic.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/common.h>
13 #include <xnnpack/math.h>
14 #include <xnnpack/vcvt.h>
15
16 #include <fp16.h>
17
18
xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x4(size_t n,const float * x,uint8_t * y,const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x4(
20 size_t n,
21 const float* x,
22 uint8_t* y,
23 const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
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;
35
36 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
37 float vx0 = x[0];
38 float vx1 = x[1];
39 float vx2 = x[2];
40 float vx3 = x[3];
41 x += 4;
42
43 vx0 *= vscale;
44 vx1 *= vscale;
45 vx2 *= vscale;
46 vx3 *= vscale;
47
48 vx0 += vmagic_bias;
49 vx1 += vmagic_bias;
50 vx2 += vmagic_bias;
51 vx3 += vmagic_bias;
52
53 int32_t vy0 = (int32_t) fp32_to_bits(vx0);
54 int32_t vy1 = (int32_t) fp32_to_bits(vx1);
55 int32_t vy2 = (int32_t) fp32_to_bits(vx2);
56 int32_t vy3 = (int32_t) fp32_to_bits(vx3);
57
58 vy0 = math_max_s32(vy0, vmagic_min);
59 vy1 = math_max_s32(vy1, vmagic_min);
60 vy2 = math_max_s32(vy2, vmagic_min);
61 vy3 = math_max_s32(vy3, vmagic_min);
62
63 vy0 = math_min_s32(vy0, vmagic_max);
64 vy1 = math_min_s32(vy1, vmagic_max);
65 vy2 = math_min_s32(vy2, vmagic_max);
66 vy3 = math_min_s32(vy3, vmagic_max);
67
68 vy0 -= vmagic_bias_less_zero_point;
69 vy1 -= vmagic_bias_less_zero_point;
70 vy2 -= vmagic_bias_less_zero_point;
71 vy3 -= vmagic_bias_less_zero_point;
72
73 y[0] = (uint8_t) vy0;
74 y[1] = (uint8_t) vy1;
75 y[2] = (uint8_t) vy2;
76 y[3] = (uint8_t) vy3;
77 y += 4;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 do {
81 float vx = *x++;
82 vx *= vscale;
83 vx += vmagic_bias;
84
85 int32_t vy = (int32_t) fp32_to_bits(vx);
86 vy = math_max_s32(vy, vmagic_min);
87 vy = math_min_s32(vy, vmagic_max);
88 vy -= vmagic_bias_less_zero_point;
89
90 *y++ = (uint8_t) vy;
91
92 n -= sizeof(float);
93 } while (n != 0);
94 }
95 }
96