xref: /aosp_15_r20/external/XNNPACK/src/f32-f16-vcvt/gen/vcvt-scalar-fabsf-x2.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_x2(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_x2(
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   for (; n >= 2 * sizeof(float); n -= 2 * sizeof(float)) {
40     const float vx0 = input[0];
41     const float vx1 = input[1];
42     input += 2;
43 
44     const float vabsx0 = fabsf(vx0);
45     const float vabsx1 = fabsf(vx1);
46     uint32_t vsignw0 = float_as_uint32(vx0);
47     uint32_t vsignw1 = float_as_uint32(vx1);
48 
49     const uint32_t vnonsignw0 = float_as_uint32(vabsx0);
50     const uint32_t vnonsignw1 = float_as_uint32(vabsx1);
51     float vf0 = vabsx0 * vscale_to_inf;
52     float vf1 = vabsx1 * vscale_to_inf;
53 
54     uint32_t vbias0 = vnonsignw0 + vexp_bias;
55     uint32_t vbias1 = vnonsignw1 + vexp_bias;
56     vsignw0 ^= vnonsignw0;
57     vsignw1 ^= vnonsignw1;
58 
59     vf0 *= vscale_to_zero;
60     vf1 *= vscale_to_zero;
61     vbias0 &= vexpw_max;
62     vbias1 &= vexpw_max;
63 
64     vbias0 = math_max_u32(vbias0, vbias_min);
65     vbias1 = math_max_u32(vbias1, vbias_min);
66 
67     vf0 += uint32_as_float(vbias0);
68     vf1 += uint32_as_float(vbias1);
69 
70     const uint32_t vbits0 = float_as_uint32(vf0);
71     const uint32_t vbits1 = float_as_uint32(vf1);
72 
73     const uint16_t vexph0 = (uint16_t) (vbits0 >> 13) & vexph_mask;
74     const uint16_t vexph1 = (uint16_t) (vbits1 >> 13) & vexph_mask;
75     const uint16_t vmanth0 = (uint16_t) vbits0 & vmanth_mask;
76     const uint16_t vmanth1 = (uint16_t) vbits1 & vmanth_mask;
77     const uint16_t vsignh0 = (uint16_t) (vsignw0 >> 16);
78     const uint16_t vsignh1 = (uint16_t) (vsignw1 >> 16);
79 
80     uint16_t vh0 = vexph0 + vmanth0;
81     uint16_t vh1 = vexph1 + vmanth1;
82     if XNN_UNPREDICTABLE(vnonsignw0 > vexpw_max) {
83       vh0 = vnanh;
84     }
85     if XNN_UNPREDICTABLE(vnonsignw1 > vexpw_max) {
86       vh1 = vnanh;
87     }
88     vh0 |= vsignh0;
89     vh1 |= vsignh1;
90 
91     o[0] = vh0;
92     o[1] = vh1;
93     o += 2;
94   }
95   if XNN_UNLIKELY(n != 0) {
96     const float vx = *input;
97 
98     const float vabsx = fabsf(vx);
99     uint32_t vsignw = float_as_uint32(vx);
100 
101     const uint32_t vnonsignw = float_as_uint32(vabsx);
102     float vf = vabsx * vscale_to_inf;
103 
104     uint32_t vbias = vnonsignw + vexp_bias;
105     vsignw ^= vnonsignw;
106 
107     vf *= vscale_to_zero;
108     vbias &= vexpw_max;
109 
110     vbias = math_max_u32(vbias, vbias_min);
111 
112     vf += uint32_as_float(vbias);
113 
114     const uint32_t vbits = float_as_uint32(vf);
115 
116     const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
117     const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
118     const uint16_t vsignh = (uint16_t) (vsignw >> 16);
119 
120     uint16_t vh = vexph + vmanth;
121     if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
122       vh = vnanh;
123     }
124     vh |= vsignh;
125 
126     *o = vh;
127   }
128 }
129