xref: /aosp_15_r20/external/XNNPACK/src/math/cvt-f32-f16-sse2.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 
10 #include <emmintrin.h>
11 
12 #include <xnnpack/math-stubs.h>
13 
14 
xnn_math_f32_f16_cvt__sse2(size_t n,const float * input,void * output)15 void xnn_math_f32_f16_cvt__sse2(
16     size_t n,
17     const float* input,
18     void* output)
19 {
20   assert(n % (8 * sizeof(uint16_t)) == 0);
21 
22   const __m128 vnonsign_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF));
23   const __m128i vexp_bias = _mm_set1_epi32(0x07800000);
24   const __m128 vscale_to_inf = _mm_set1_ps(0x1.0p+112f);
25   const __m128i vexpw_max = _mm_set1_epi32(0x7F800000);
26   const __m128 vscale_to_zero = _mm_set1_ps(0x1.0p-110f);
27   const __m128i vbias_min = _mm_set1_epi32(0x40008000);
28   const __m128i vmanth_mask = _mm_set1_epi32(0x0FFF);
29   const __m128i vexph_mask = _mm_set1_epi32(0x7C00);
30   const __m128i vnanh = _mm_set1_epi16(0x7E00);
31 
32   uint16_t* o = (uint16_t*) output;
33   for (; n != 0; n -= 8 * sizeof(uint16_t)) {
34     const __m128 vx_lo = _mm_loadu_ps(input);
35     const __m128 vx_hi = _mm_loadu_ps(input + 4);
36     input += 8;
37 
38     const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask);
39     const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask);
40 
41     const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo);
42     const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi);
43     __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias);
44     __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias);
45     __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf);
46     __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf);
47     const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max);
48     const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max);
49 
50     vbias_lo = _mm_and_si128(vbias_lo, vexpw_max);
51     vbias_hi = _mm_and_si128(vbias_hi, vexpw_max);
52     vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero);
53     vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero);
54     const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi);
55     const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi));
56 
57     vbias_lo = _mm_max_epi16(vbias_lo, vbias_min);
58     vbias_hi = _mm_max_epi16(vbias_hi, vbias_min);
59     __m128i vh = _mm_and_si128(vnanh, vnanmaskh);
60 
61     vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo));
62     vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi));
63     vh = _mm_or_si128(vh, vsignh);
64 
65     __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13);
66     __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13);
67     const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask);
68     const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask);
69 
70     vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask);
71     vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask);
72 
73     const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo);
74     const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi);
75 
76     const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi);
77 
78     vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh));
79 
80     _mm_storeu_si128((__m128i*) o, vh);
81     o += 8;
82   }
83 }
84