• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-f32-vcvt/sse-int32.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 <emmintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16 
17 
xnn_f16_f32_vcvt_ukernel__sse2_int32_x16(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__sse2_int32_x16(
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)]) XNN_OOB_READS
23 {
24   assert(n != 0);
25   assert(n % sizeof(uint16_t) == 0);
26   assert(input != NULL);
27   assert(output != NULL);
28 
29   const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse_int32.sign_mask);
30   const __m128i vexp_offset = _mm_load_si128((const __m128i*) params->sse_int32.exp_offset);
31   const __m128 vexp_scale = _mm_load_ps(params->sse_int32.exp_scale);
32   const __m128i vmagic_bias = _mm_load_si128((const __m128i*) params->sse_int32.magic_bias);
33   const __m128i vdenorm_cutoff = _mm_load_si128((const __m128i*) params->sse_int32.denorm_cutoff);
34 
35   const uint16_t* i = (const uint16_t*) input;
36   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
37     const __m128i vh0 = _mm_loadu_si128((const __m128i*) i);
38     const __m128i vh1 = _mm_loadu_si128((const __m128i*) (i + 8));
39     i += 16;
40 
41     const __m128i vw0 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh0);
42     const __m128i vw1 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh0);
43     const __m128i vw2 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh1);
44     const __m128i vw3 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh1);
45 
46     const __m128i vsign0 = _mm_and_si128(vw0, vsign_mask);
47     const __m128i vsign1 = _mm_and_si128(vw1, vsign_mask);
48     const __m128i vsign2 = _mm_and_si128(vw2, vsign_mask);
49     const __m128i vsign3 = _mm_and_si128(vw3, vsign_mask);
50 
51     const __m128i vnonsign0 = _mm_xor_si128(vw0, vsign0);
52     const __m128i vnonsign1 = _mm_xor_si128(vw1, vsign1);
53     const __m128i vnonsign2 = _mm_xor_si128(vw2, vsign2);
54     const __m128i vnonsign3 = _mm_xor_si128(vw3, vsign3);
55 
56     const __m128i vnorm0 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign0, 3), vexp_offset)), vexp_scale));
57     const __m128i vnorm1 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign1, 3), vexp_offset)), vexp_scale));
58     const __m128i vnorm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign2, 3), vexp_offset)), vexp_scale));
59     const __m128i vnorm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign3, 3), vexp_offset)), vexp_scale));
60 
61     const __m128i vdenorm0 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign0, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
62     const __m128i vdenorm1 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign1, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
63     const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign2, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
64     const __m128i vdenorm3 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign3, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
65 
66     const __m128i vmask0 = _mm_cmpgt_epi32(vnonsign0, vdenorm_cutoff);
67     const __m128i vmask1 = _mm_cmpgt_epi32(vnonsign1, vdenorm_cutoff);
68     const __m128i vmask2 = _mm_cmpgt_epi32(vnonsign2, vdenorm_cutoff);
69     const __m128i vmask3 = _mm_cmpgt_epi32(vnonsign3, vdenorm_cutoff);
70 
71     const __m128i vf0 = _mm_or_si128(vsign0,
72       _mm_or_si128(_mm_and_si128(vmask0, vnorm0), _mm_andnot_si128(vmask0, vdenorm0)));
73     const __m128i vf1 = _mm_or_si128(vsign1,
74       _mm_or_si128(_mm_and_si128(vmask1, vnorm1), _mm_andnot_si128(vmask1, vdenorm1)));
75     const __m128i vf2 = _mm_or_si128(vsign2,
76       _mm_or_si128(_mm_and_si128(vmask2, vnorm2), _mm_andnot_si128(vmask2, vdenorm2)));
77     const __m128i vf3 = _mm_or_si128(vsign3,
78       _mm_or_si128(_mm_and_si128(vmask3, vnorm3), _mm_andnot_si128(vmask3, vdenorm3)));
79 
80     _mm_storeu_ps(output, _mm_castsi128_ps(vf0));
81     _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf1));
82     _mm_storeu_ps(output + 8, _mm_castsi128_ps(vf2));
83     _mm_storeu_ps(output + 12, _mm_castsi128_ps(vf3));
84     output += 16;
85   }
86   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
87     const __m128i vh = _mm_loadu_si128((const __m128i*) i);
88     i += 8;
89 
90     const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
91     const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);
92 
93     const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
94     const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);
95 
96     const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
97     const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
98 
99     const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_lo, 3), vexp_offset)), vexp_scale));
100     const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offset)), vexp_scale));
101 
102     const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_lo, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
103     const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
104 
105     const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
106     const __m128i vf_lo = _mm_or_si128(vsign_lo,
107       _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));
108 
109     const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
110     const __m128i vf_hi = _mm_or_si128(vsign_hi,
111       _mm_or_si128(_mm_and_si128(vmask_hi, vnorm_hi), _mm_andnot_si128(vmask_hi, vdenorm_hi)));
112 
113     _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo));
114     _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi));
115     output += 8;
116   }
117   if XNN_UNPREDICTABLE(n != 0) {
118     const __m128i vh = _mm_loadu_si128((const __m128i*) i);
119 
120     const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
121     const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);
122 
123     const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
124     const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);
125 
126     const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
127     const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
128 
129     const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_lo, 3), vexp_offset)), vexp_scale));
130     const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offset)), vexp_scale));
131 
132     const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_lo, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
133     const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
134 
135     const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
136     __m128i vf = _mm_or_si128(vsign_lo,
137       _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));
138 
139     if (n & (4 * sizeof(uint16_t))) {
140       _mm_storeu_ps(output, _mm_castsi128_ps(vf));
141       output += 4;
142 
143       const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
144       vf = _mm_or_si128(vsign_hi,
145         _mm_or_si128(_mm_and_si128(vmask_hi, vnorm_hi), _mm_andnot_si128(vmask_hi, vdenorm_hi)));
146     }
147     if (n & (2 * sizeof(uint16_t))) {
148       _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf));
149       output += 2;
150 
151       vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf)));
152     }
153     if (n & (1 * sizeof(uint16_t))) {
154       _mm_store_ss(output, _mm_castsi128_ps(vf));
155     }
156   }
157 }
158