• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/vcvt.h>
14 
15 #include <fp16.h>
16 
17 
xnn_f16_f32_vcvt_ukernel__scalar_x1(size_t n,const void * input,float * output,const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_f32_vcvt_ukernel__scalar_x1(
19     size_t n,
20     const void* input,
21     float* output,
22     const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n != 0);
25   assert(n % sizeof(uint16_t) == 0);
26   assert(input != NULL);
27   assert(output != NULL);
28 
29   const uint32_t vsign_mask = params->scalar.sign_mask;
30   const uint32_t vexp_offset = params->scalar.exp_offset;
31   const float vexp_scale = params->scalar.exp_scale;
32   const uint32_t vmagic_mask = params->scalar.magic_mask;
33   const float vmagic_bias = params->scalar.magic_bias;
34   const uint32_t vdenorm_cutoff = params->scalar.denorm_cutoff;
35 
36   const uint16_t* i = (const uint16_t*) input;
37   uint32_t* o = (uint32_t*) output;
38   do {
39     const uint16_t vh = *i++;
40 
41     const uint32_t vw = (uint32_t) vh << 16;
42     const uint32_t vsign = vw & vsign_mask;
43     const uint32_t v2w = vw + vw;
44     const uint32_t vnorm = fp32_to_bits(fp32_from_bits((v2w >> 4) + vexp_offset) * vexp_scale);
45     const uint32_t vdenorm = fp32_to_bits(fp32_from_bits((v2w >> 17) | vmagic_mask) - vmagic_bias);
46     const uint32_t vf = vsign | (XNN_UNPREDICTABLE(v2w < vdenorm_cutoff) ? vdenorm : vnorm);
47 
48     *o++ = vf;
49 
50     n -= sizeof(uint16_t);
51   } while (n != 0);
52 }
53