• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-qs8-vcvt/scalar-fmagic.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_qu8_vcvt_ukernel__wasm_fmagic_x4(size_t n,const float * x,uint8_t * y,const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_qu8_vcvt_ukernel__wasm_fmagic_x4(
18     size_t n,
19     const float* x,
20     uint8_t* y,
21     const union xnn_f32_qu8_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_fmagic.scale;
29   const float voutput_min_less_zero_point = params->scalar_fmagic.output_min_less_zero_point;
30   const float voutput_max_less_zero_point = params->scalar_fmagic.output_max_less_zero_point;
31   const float vmagic_bias = params->scalar_fmagic.magic_bias;
32   const int32_t vmagic_bias_less_zero_point = params->scalar_fmagic.magic_bias_less_zero_point;
33 
34   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
35     float vx0 = x[0];
36     float vx1 = x[1];
37     float vx2 = x[2];
38     float vx3 = x[3];
39     x += 4;
40 
41     vx0 *= vscale;
42     vx1 *= vscale;
43     vx2 *= vscale;
44     vx3 *= vscale;
45 
46     vx0 = __builtin_wasm_max_f32(vx0, voutput_min_less_zero_point);
47     vx1 = __builtin_wasm_max_f32(vx1, voutput_min_less_zero_point);
48     vx2 = __builtin_wasm_max_f32(vx2, voutput_min_less_zero_point);
49     vx3 = __builtin_wasm_max_f32(vx3, voutput_min_less_zero_point);
50 
51     vx0 = __builtin_wasm_min_f32(vx0, voutput_max_less_zero_point);
52     vx1 = __builtin_wasm_min_f32(vx1, voutput_max_less_zero_point);
53     vx2 = __builtin_wasm_min_f32(vx2, voutput_max_less_zero_point);
54     vx3 = __builtin_wasm_min_f32(vx3, voutput_max_less_zero_point);
55 
56     vx0 += vmagic_bias;
57     vx1 += vmagic_bias;
58     vx2 += vmagic_bias;
59     vx3 += vmagic_bias;
60 
61     int32_t vy0 = (int32_t) float_as_uint32(vx0);
62     int32_t vy1 = (int32_t) float_as_uint32(vx1);
63     int32_t vy2 = (int32_t) float_as_uint32(vx2);
64     int32_t vy3 = (int32_t) float_as_uint32(vx3);
65 
66     vy0 -= vmagic_bias_less_zero_point;
67     vy1 -= vmagic_bias_less_zero_point;
68     vy2 -= vmagic_bias_less_zero_point;
69     vy3 -= vmagic_bias_less_zero_point;
70 
71     y[0] = (uint8_t) vy0;
72     y[1] = (uint8_t) vy1;
73     y[2] = (uint8_t) vy2;
74     y[3] = (uint8_t) vy3;
75     y += 4;
76   }
77   if XNN_UNLIKELY(n != 0) {
78     do {
79       float vx = *x++;
80       vx *= vscale;
81       vx = __builtin_wasm_max_f32(vx, voutput_min_less_zero_point);
82       vx = __builtin_wasm_min_f32(vx, voutput_max_less_zero_point);
83       vx += vmagic_bias;
84 
85       int32_t vy = (int32_t) float_as_uint32(vx);
86       vy -= vmagic_bias_less_zero_point;
87 
88       *y++ = (uint8_t) vy;
89 
90       n -= sizeof(float);
91     } while (n != 0);
92   }
93 }
94