xref: /aosp_15_r20/external/XNNPACK/src/f32-f16-vcvt/gen/vcvt-scalar-fabsf-x1.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 
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)])18 void xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x1(
19     size_t n,
20     const float* input,
21     void* output,
22     const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26   assert(input != NULL);
27   assert(output != NULL);
28 
29   const float vscale_to_inf = params->scalar_fabsf.scale_to_inf;
30   const uint32_t vexp_bias = params->scalar_fabsf.exp_bias;
31   const float vscale_to_zero = params->scalar_fabsf.scale_to_zero;
32   const uint32_t vexpw_max = params->scalar_fabsf.expw_max;
33   const uint32_t vbias_min = params->scalar_fabsf.bias_min;
34   const uint16_t vexph_mask = params->scalar_fabsf.exph_mask;
35   const uint16_t vmanth_mask = params->scalar_fabsf.manth_mask;
36   const uint16_t vnanh = params->scalar_fabsf.nanh;
37 
38   uint16_t* o = (uint16_t*) output;
39   do {
40     const float vx = *input++;
41 
42     const float vabsx = fabsf(vx);
43     uint32_t vsignw = float_as_uint32(vx);
44 
45     const uint32_t vnonsignw = float_as_uint32(vabsx);
46     float vf = vabsx * vscale_to_inf;
47 
48     uint32_t vbias = vnonsignw + vexp_bias;
49     vsignw ^= vnonsignw;
50 
51     vf *= vscale_to_zero;
52     vbias &= vexpw_max;
53 
54     vbias = math_max_u32(vbias, vbias_min);
55 
56     vf += uint32_as_float(vbias);
57 
58     const uint32_t vbits = float_as_uint32(vf);
59 
60     const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
61     const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
62     const uint16_t vsignh = (uint16_t) (vsignw >> 16);
63 
64     uint16_t vh = vexph + vmanth;
65     if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
66       vh = vnanh;
67     }
68     vh |= vsignh;
69 
70     *o++ = vh;
71 
72     n -= sizeof(float);
73   } while (n != 0);
74 }
75