xref: /aosp_15_r20/external/XNNPACK/src/math/cvt-f32-f16-scalar-fabsf.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #include <assert.h>
7 #include <stddef.h>
8 #include <stdint.h>
9 #include <math.h>
10 
11 #include <xnnpack/math.h>
12 #include <xnnpack/math-stubs.h>
13 
14 
xnn_math_f32_f16_cvt__scalar_fabsf(size_t n,const float * input,void * output)15 void xnn_math_f32_f16_cvt__scalar_fabsf(
16     size_t n,
17     const float* input,
18     void* output)
19 {
20   assert(n % (sizeof(uint16_t)) == 0);
21 
22   const float vscale_to_inf = 0x1.0p+112f;
23   const uint32_t vexp_bias = UINT32_C(0x07800000);
24   const float vscale_to_zero = 0x1.0p-110f;
25   const uint32_t vexpw_max = UINT32_C(0x7F800000);
26   const uint32_t vbias_min = UINT32_C(0x40000000);
27   const uint16_t vexph_mask = UINT16_C(0x7C00);
28   const uint16_t vmanth_mask = UINT16_C(0x0FFF);
29   const uint16_t vnanh = UINT16_C(0x7E00);
30 
31   uint16_t* o = (uint16_t*) output;
32   for (; n != 0; n -= sizeof(uint16_t)) {
33     const float vx = *input++;
34 
35     const float vabsx = fabsf(vx);
36     uint32_t vsignw = float_as_uint32(vx);
37 
38     const uint32_t vnonsignw = float_as_uint32(vabsx);
39     float vf = vabsx * vscale_to_inf;
40 
41     uint32_t vbias = vnonsignw + vexp_bias;
42     vsignw ^= vnonsignw;
43 
44     vf *= vscale_to_zero;
45     vbias &= vexpw_max;
46 
47     vbias = math_max_u32(vbias, vbias_min);
48 
49     vf += uint32_as_float(vbias);
50 
51     const uint32_t vbits = float_as_uint32(vf);
52 
53     const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
54     const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
55     const uint16_t vsignh = (uint16_t) (vsignw >> 16);
56 
57     uint16_t vh = vexph + vmanth;
58     if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
59       vh = vnanh;
60     }
61     vh |= vsignh;
62 
63     *o++ = vh;
64   }
65 }
66