• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-f16-vcvt/scalar-fabsf.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 #include <math.h>
12 
13 #include <xnnpack/common.h>
14 #include <xnnpack/math.h>
15 #include <xnnpack/vcvt.h>
16 
17 #include <fp16.h>
18 
19 
xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x1(size_t n,const float * input,void * output,const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x1(
21     size_t n,
22     const float* input,
23     void* output,
24     const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26   assert(n != 0);
27   assert(n % sizeof(float) == 0);
28   assert(input != NULL);
29   assert(output != NULL);
30 
31   const float vscale_to_inf = params->scalar_fabsf.scale_to_inf;
32   const uint32_t vexp_bias = params->scalar_fabsf.exp_bias;
33   const float vscale_to_zero = params->scalar_fabsf.scale_to_zero;
34   const uint32_t vexpw_max = params->scalar_fabsf.expw_max;
35   const uint32_t vbias_min = params->scalar_fabsf.bias_min;
36   const uint16_t vexph_mask = params->scalar_fabsf.exph_mask;
37   const uint16_t vmanth_mask = params->scalar_fabsf.manth_mask;
38   const uint16_t vnanh = params->scalar_fabsf.nanh;
39 
40   uint16_t* o = (uint16_t*) output;
41   do {
42     const float vx = *input++;
43 
44     const float vabsx = fabsf(vx);
45     uint32_t vsignw = fp32_to_bits(vx);
46 
47     const uint32_t vnonsignw = fp32_to_bits(vabsx);
48     float vf = vabsx * vscale_to_inf;
49 
50     uint32_t vbias = vnonsignw + vexp_bias;
51     vsignw ^= vnonsignw;
52 
53     vf *= vscale_to_zero;
54     vbias &= vexpw_max;
55 
56     vbias = math_max_u32(vbias, vbias_min);
57 
58     vf += fp32_from_bits(vbias);
59 
60     const uint32_t vbits = fp32_to_bits(vf);
61 
62     const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
63     const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
64     const uint16_t vsignh = (uint16_t) (vsignw >> 16);
65 
66     uint16_t vh = vexph + vmanth;
67     if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
68       vh = vnanh;
69     }
70     vh |= vsignh;
71 
72     *o++ = vh;
73 
74     n -= sizeof(float);
75   } while (n != 0);
76 }
77