• 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_x3(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_x3(
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   for (; n >= 3 * sizeof(float); n -= 3 * sizeof(float)) {
42     const float vx0 = input[0];
43     const float vx1 = input[1];
44     const float vx2 = input[2];
45     input += 3;
46 
47     const float vabsx0 = fabsf(vx0);
48     const float vabsx1 = fabsf(vx1);
49     const float vabsx2 = fabsf(vx2);
50     uint32_t vsignw0 = fp32_to_bits(vx0);
51     uint32_t vsignw1 = fp32_to_bits(vx1);
52     uint32_t vsignw2 = fp32_to_bits(vx2);
53 
54     const uint32_t vnonsignw0 = fp32_to_bits(vabsx0);
55     const uint32_t vnonsignw1 = fp32_to_bits(vabsx1);
56     const uint32_t vnonsignw2 = fp32_to_bits(vabsx2);
57     float vf0 = vabsx0 * vscale_to_inf;
58     float vf1 = vabsx1 * vscale_to_inf;
59     float vf2 = vabsx2 * vscale_to_inf;
60 
61     uint32_t vbias0 = vnonsignw0 + vexp_bias;
62     uint32_t vbias1 = vnonsignw1 + vexp_bias;
63     uint32_t vbias2 = vnonsignw2 + vexp_bias;
64     vsignw0 ^= vnonsignw0;
65     vsignw1 ^= vnonsignw1;
66     vsignw2 ^= vnonsignw2;
67 
68     vf0 *= vscale_to_zero;
69     vf1 *= vscale_to_zero;
70     vf2 *= vscale_to_zero;
71     vbias0 &= vexpw_max;
72     vbias1 &= vexpw_max;
73     vbias2 &= vexpw_max;
74 
75     vbias0 = math_max_u32(vbias0, vbias_min);
76     vbias1 = math_max_u32(vbias1, vbias_min);
77     vbias2 = math_max_u32(vbias2, vbias_min);
78 
79     vf0 += fp32_from_bits(vbias0);
80     vf1 += fp32_from_bits(vbias1);
81     vf2 += fp32_from_bits(vbias2);
82 
83     const uint32_t vbits0 = fp32_to_bits(vf0);
84     const uint32_t vbits1 = fp32_to_bits(vf1);
85     const uint32_t vbits2 = fp32_to_bits(vf2);
86 
87     const uint16_t vexph0 = (uint16_t) (vbits0 >> 13) & vexph_mask;
88     const uint16_t vexph1 = (uint16_t) (vbits1 >> 13) & vexph_mask;
89     const uint16_t vexph2 = (uint16_t) (vbits2 >> 13) & vexph_mask;
90     const uint16_t vmanth0 = (uint16_t) vbits0 & vmanth_mask;
91     const uint16_t vmanth1 = (uint16_t) vbits1 & vmanth_mask;
92     const uint16_t vmanth2 = (uint16_t) vbits2 & vmanth_mask;
93     const uint16_t vsignh0 = (uint16_t) (vsignw0 >> 16);
94     const uint16_t vsignh1 = (uint16_t) (vsignw1 >> 16);
95     const uint16_t vsignh2 = (uint16_t) (vsignw2 >> 16);
96 
97     uint16_t vh0 = vexph0 + vmanth0;
98     uint16_t vh1 = vexph1 + vmanth1;
99     uint16_t vh2 = vexph2 + vmanth2;
100     if XNN_UNPREDICTABLE(vnonsignw0 > vexpw_max) {
101       vh0 = vnanh;
102     }
103     if XNN_UNPREDICTABLE(vnonsignw1 > vexpw_max) {
104       vh1 = vnanh;
105     }
106     if XNN_UNPREDICTABLE(vnonsignw2 > vexpw_max) {
107       vh2 = vnanh;
108     }
109     vh0 |= vsignh0;
110     vh1 |= vsignh1;
111     vh2 |= vsignh2;
112 
113     o[0] = vh0;
114     o[1] = vh1;
115     o[2] = vh2;
116     o += 3;
117   }
118   if XNN_UNLIKELY(n != 0) {
119     do {
120       const float vx = *input++;
121 
122       const float vabsx = fabsf(vx);
123       uint32_t vsignw = fp32_to_bits(vx);
124 
125       const uint32_t vnonsignw = fp32_to_bits(vabsx);
126       float vf = vabsx * vscale_to_inf;
127 
128       uint32_t vbias = vnonsignw + vexp_bias;
129       vsignw ^= vnonsignw;
130 
131       vf *= vscale_to_zero;
132       vbias &= vexpw_max;
133 
134       vbias = math_max_u32(vbias, vbias_min);
135 
136       vf += fp32_from_bits(vbias);
137 
138       const uint32_t vbits = fp32_to_bits(vf);
139 
140       const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
141       const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
142       const uint16_t vsignh = (uint16_t) (vsignw >> 16);
143 
144       uint16_t vh = vexph + vmanth;
145       if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
146         vh = vnanh;
147       }
148       vh |= vsignh;
149 
150       *o++ = vh;
151 
152       n -= sizeof(float);
153     } while (n != 0);
154   }
155 }
156