xref: /aosp_15_r20/external/XNNPACK/src/f32-f16-vcvt/gen/vcvt-scalar-bitcast-x3.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-f16-vcvt/scalar-bitcast.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_f16_vcvt_ukernel__scalar_bitcast_x3(size_t n,const float * input,void * output,const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x3(
18     size_t n,
19     const float* input,
20     void* output,
21     const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23   assert(n != 0);
24   assert(n % sizeof(float) == 0);
25   assert(input != NULL);
26   assert(output != NULL);
27 
28   const uint32_t vnonsign_mask = params->scalar_bitcast.nonsign_mask;
29   const uint32_t vexp_bias = params->scalar_bitcast.exp_bias;
30   const float vscale_to_inf = params->scalar_bitcast.scale_to_inf;
31   const uint32_t vexpw_max = params->scalar_bitcast.expw_max;
32   const float vscale_to_zero = params->scalar_bitcast.scale_to_zero;
33   const uint32_t vbias_min = params->scalar_bitcast.bias_min;
34   const uint16_t vexph_mask = params->scalar_bitcast.exph_mask;
35   const uint16_t vmanth_mask = params->scalar_bitcast.manth_mask;
36   const uint16_t vnanh = params->scalar_bitcast.nanh;
37 
38   const uint32_t* i = (const uint32_t*) input;
39   uint16_t* o = (uint16_t*) output;
40   for (; n >= 3 * sizeof(float); n -= 3 * sizeof(float)) {
41     const uint32_t vw0 = i[0];
42     const uint32_t vw1 = i[1];
43     const uint32_t vw2 = i[2];
44     i += 3;
45 
46     const uint32_t vnonsignw0 = vw0 & vnonsign_mask;
47     const uint32_t vnonsignw1 = vw1 & vnonsign_mask;
48     const uint32_t vnonsignw2 = vw2 & vnonsign_mask;
49 
50     float vf0 = uint32_as_float(vnonsignw0);
51     float vf1 = uint32_as_float(vnonsignw1);
52     float vf2 = uint32_as_float(vnonsignw2);
53     const uint32_t vsignw0 = vw0 ^ vnonsignw0;
54     const uint32_t vsignw1 = vw1 ^ vnonsignw1;
55     const uint32_t vsignw2 = vw2 ^ vnonsignw2;
56     uint32_t vbias0 = vnonsignw0 + vexp_bias;
57     uint32_t vbias1 = vnonsignw1 + vexp_bias;
58     uint32_t vbias2 = vnonsignw2 + vexp_bias;
59 
60     vf0 *= vscale_to_inf;
61     vf1 *= vscale_to_inf;
62     vf2 *= vscale_to_inf;
63     vbias0 &= vexpw_max;
64     vbias1 &= vexpw_max;
65     vbias2 &= vexpw_max;
66 
67     vf0 *= vscale_to_zero;
68     vf1 *= vscale_to_zero;
69     vf2 *= vscale_to_zero;
70     vbias0 = math_max_u32(vbias0, vbias_min);
71     vbias1 = math_max_u32(vbias1, vbias_min);
72     vbias2 = math_max_u32(vbias2, vbias_min);
73 
74     vf0 += uint32_as_float(vbias0);
75     vf1 += uint32_as_float(vbias1);
76     vf2 += uint32_as_float(vbias2);
77 
78     const uint32_t vbits0 = float_as_uint32(vf0);
79     const uint32_t vbits1 = float_as_uint32(vf1);
80     const uint32_t vbits2 = float_as_uint32(vf2);
81 
82     const uint16_t vexph0 = (uint16_t) (vbits0 >> 13) & vexph_mask;
83     const uint16_t vexph1 = (uint16_t) (vbits1 >> 13) & vexph_mask;
84     const uint16_t vexph2 = (uint16_t) (vbits2 >> 13) & vexph_mask;
85     const uint16_t vmanth0 = (uint16_t) vbits0 & vmanth_mask;
86     const uint16_t vmanth1 = (uint16_t) vbits1 & vmanth_mask;
87     const uint16_t vmanth2 = (uint16_t) vbits2 & vmanth_mask;
88     const uint16_t vsignh0 = (uint16_t) (vsignw0 >> 16);
89     const uint16_t vsignh1 = (uint16_t) (vsignw1 >> 16);
90     const uint16_t vsignh2 = (uint16_t) (vsignw2 >> 16);
91 
92     uint16_t vh0 = vexph0 + vmanth0;
93     uint16_t vh1 = vexph1 + vmanth1;
94     uint16_t vh2 = vexph2 + vmanth2;
95     if XNN_UNPREDICTABLE(vnonsignw0 > vexpw_max) {
96       vh0 = vnanh;
97     }
98     if XNN_UNPREDICTABLE(vnonsignw1 > vexpw_max) {
99       vh1 = vnanh;
100     }
101     if XNN_UNPREDICTABLE(vnonsignw2 > vexpw_max) {
102       vh2 = vnanh;
103     }
104     vh0 |= vsignh0;
105     vh1 |= vsignh1;
106     vh2 |= vsignh2;
107 
108     o[0] = vh0;
109     o[1] = vh1;
110     o[2] = vh2;
111     o += 3;
112   }
113   if XNN_UNLIKELY(n != 0) {
114     do {
115       const uint32_t vw = *i++;
116 
117       const uint32_t vnonsignw = vw & vnonsign_mask;
118 
119       float vf = uint32_as_float(vnonsignw);
120       const uint32_t vsignw = vw ^ vnonsignw;
121       uint32_t vbias = vnonsignw + vexp_bias;
122 
123       vf *= vscale_to_inf;
124       vbias &= vexpw_max;
125 
126       vf *= vscale_to_zero;
127       vbias = math_max_u32(vbias, vbias_min);
128 
129       vf += uint32_as_float(vbias);
130 
131       const uint32_t vbits = float_as_uint32(vf);
132 
133       const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
134       const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
135       const uint16_t vsignh = (uint16_t) (vsignw >> 16);
136 
137       uint16_t vh = vexph + vmanth;
138       if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
139         vh = vnanh;
140       }
141       vh |= vsignh;
142 
143       *o++ = vh;
144 
145       n -= sizeof(float);
146     } while (n != 0);
147   }
148 }
149