xref: /aosp_15_r20/external/XNNPACK/src/f16-f32-vcvt/gen/vcvt-scalar-x2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-f32-vcvt/scalar.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_f16_f32_vcvt_ukernel__scalar_x2(size_t n,const void * input,float * output,const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_f32_vcvt_ukernel__scalar_x2(
18     size_t n,
19     const void* input,
20     float* output,
21     const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23   assert(n != 0);
24   assert(n % sizeof(uint16_t) == 0);
25   assert(input != NULL);
26   assert(output != NULL);
27 
28   const uint32_t vsign_mask = params->scalar.sign_mask;
29   const uint32_t vexp_offset = params->scalar.exp_offset;
30   const float vexp_scale = params->scalar.exp_scale;
31   const uint32_t vmagic_mask = params->scalar.magic_mask;
32   const float vmagic_bias = params->scalar.magic_bias;
33   const uint32_t vdenorm_cutoff = params->scalar.denorm_cutoff;
34 
35   const uint16_t* i = (const uint16_t*) input;
36   uint32_t* o = (uint32_t*) output;
37   for (; n >= 2 * sizeof(uint16_t); n -= 2 * sizeof(uint16_t)) {
38     const uint16_t vh0 = i[0];
39     const uint16_t vh1 = i[1];
40     i += 2;
41 
42     const uint32_t vw0 = (uint32_t) vh0 << 16;
43     const uint32_t vw1 = (uint32_t) vh1 << 16;
44 
45     const uint32_t vsign0 = vw0 & vsign_mask;
46     const uint32_t vsign1 = vw1 & vsign_mask;
47 
48     const uint32_t v2w0 = vw0 + vw0;
49     const uint32_t v2w1 = vw1 + vw1;
50 
51     const uint32_t vnorm0 = float_as_uint32(uint32_as_float((v2w0 >> 4) + vexp_offset) * vexp_scale);
52     const uint32_t vnorm1 = float_as_uint32(uint32_as_float((v2w1 >> 4) + vexp_offset) * vexp_scale);
53 
54     const uint32_t vdenorm0 = float_as_uint32(uint32_as_float((v2w0 >> 17) | vmagic_mask) - vmagic_bias);
55     const uint32_t vdenorm1 = float_as_uint32(uint32_as_float((v2w1 >> 17) | vmagic_mask) - vmagic_bias);
56 
57     const uint32_t vf0 = vsign0 | (XNN_UNPREDICTABLE(v2w0 < vdenorm_cutoff) ? vdenorm0 : vnorm0);
58     const uint32_t vf1 = vsign1 | (XNN_UNPREDICTABLE(v2w1 < vdenorm_cutoff) ? vdenorm1 : vnorm1);
59 
60     o[0] = vf0;
61     o[1] = vf1;
62     o += 2;
63   }
64   if XNN_UNLIKELY(n != 0) {
65     const uint16_t vh = *i;
66 
67     const uint32_t vw = (uint32_t) vh << 16;
68     const uint32_t vsign = vw & vsign_mask;
69     const uint32_t v2w = vw + vw;
70     const uint32_t vnorm = float_as_uint32(uint32_as_float((v2w >> 4) + vexp_offset) * vexp_scale);
71     const uint32_t vdenorm = float_as_uint32(uint32_as_float((v2w >> 17) | vmagic_mask) - vmagic_bias);
72     const uint32_t vf = vsign | (XNN_UNPREDICTABLE(v2w < vdenorm_cutoff) ? vdenorm : vnorm);
73 
74     *o = vf;
75   }
76 }
77