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
xnn_f32_qs8_vcvt_ukernel__scalar_imagic_x2(size_t n,const float * x,int8_t * y,const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_qs8_vcvt_ukernel__scalar_imagic_x2(
18 size_t n,
19 const float* x,
20 int8_t* y,
21 const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23 assert(n != 0);
24 assert(n % sizeof(float) == 0);
25 assert(x != NULL);
26 assert(y != NULL);
27
28 const float vscale = params->scalar_imagic.scale;
29 const float vmagic_bias = params->scalar_imagic.magic_bias;
30 const int32_t vmagic_min = params->scalar_imagic.magic_min;
31 const int32_t vmagic_max = params->scalar_imagic.magic_max;
32 const int32_t vmagic_bias_less_zero_point = params->scalar_imagic.magic_bias_less_zero_point;
33
34 for (; n >= 2 * sizeof(float); n -= 2 * sizeof(float)) {
35 float vx0 = x[0];
36 float vx1 = x[1];
37 x += 2;
38
39 vx0 *= vscale;
40 vx1 *= vscale;
41
42 vx0 += vmagic_bias;
43 vx1 += vmagic_bias;
44
45 int32_t vy0 = (int32_t) float_as_uint32(vx0);
46 int32_t vy1 = (int32_t) float_as_uint32(vx1);
47
48 vy0 = math_max_s32(vy0, vmagic_min);
49 vy1 = math_max_s32(vy1, vmagic_min);
50
51 vy0 = math_min_s32(vy0, vmagic_max);
52 vy1 = math_min_s32(vy1, vmagic_max);
53
54 vy0 -= vmagic_bias_less_zero_point;
55 vy1 -= vmagic_bias_less_zero_point;
56
57 y[0] = (int8_t) vy0;
58 y[1] = (int8_t) vy1;
59 y += 2;
60 }
61 if XNN_UNLIKELY(n != 0) {
62 float vx = *x;
63 vx *= vscale;
64 vx += vmagic_bias;
65
66 int32_t vy = (int32_t) float_as_uint32(vx);
67 vy = math_max_s32(vy, vmagic_min);
68 vy = math_min_s32(vy, vmagic_max);
69 vy -= vmagic_bias_less_zero_point;
70
71 *y = (int8_t) vy;
72 }
73 }
74