• 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_x32(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_x32(
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 >= 32 * sizeof(uint16_t); n -= 32 * 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     const __m128i vh2 = _mm_loadu_si128((const __m128i*) (i + 16));
40     const __m128i vh3 = _mm_loadu_si128((const __m128i*) (i + 24));
41     i += 32;
42 
43     const __m128i vw0 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh0);
44     const __m128i vw1 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh0);
45     const __m128i vw2 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh1);
46     const __m128i vw3 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh1);
47     const __m128i vw4 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh2);
48     const __m128i vw5 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh2);
49     const __m128i vw6 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh3);
50     const __m128i vw7 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh3);
51 
52     const __m128i vsign0 = _mm_and_si128(vw0, vsign_mask);
53     const __m128i vsign1 = _mm_and_si128(vw1, vsign_mask);
54     const __m128i vsign2 = _mm_and_si128(vw2, vsign_mask);
55     const __m128i vsign3 = _mm_and_si128(vw3, vsign_mask);
56     const __m128i vsign4 = _mm_and_si128(vw4, vsign_mask);
57     const __m128i vsign5 = _mm_and_si128(vw5, vsign_mask);
58     const __m128i vsign6 = _mm_and_si128(vw6, vsign_mask);
59     const __m128i vsign7 = _mm_and_si128(vw7, vsign_mask);
60 
61     const __m128i vnonsign0 = _mm_xor_si128(vw0, vsign0);
62     const __m128i vnonsign1 = _mm_xor_si128(vw1, vsign1);
63     const __m128i vnonsign2 = _mm_xor_si128(vw2, vsign2);
64     const __m128i vnonsign3 = _mm_xor_si128(vw3, vsign3);
65     const __m128i vnonsign4 = _mm_xor_si128(vw4, vsign4);
66     const __m128i vnonsign5 = _mm_xor_si128(vw5, vsign5);
67     const __m128i vnonsign6 = _mm_xor_si128(vw6, vsign6);
68     const __m128i vnonsign7 = _mm_xor_si128(vw7, vsign7);
69 
70     const __m128i vnorm0 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign0, 3), vexp_offset)), vexp_scale));
71     const __m128i vnorm1 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign1, 3), vexp_offset)), vexp_scale));
72     const __m128i vnorm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign2, 3), vexp_offset)), vexp_scale));
73     const __m128i vnorm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign3, 3), vexp_offset)), vexp_scale));
74     const __m128i vnorm4 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign4, 3), vexp_offset)), vexp_scale));
75     const __m128i vnorm5 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign5, 3), vexp_offset)), vexp_scale));
76     const __m128i vnorm6 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign6, 3), vexp_offset)), vexp_scale));
77     const __m128i vnorm7 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign7, 3), vexp_offset)), vexp_scale));
78 
79     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)));
80     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)));
81     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)));
82     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)));
83     const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign4, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
84     const __m128i vdenorm5 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign5, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
85     const __m128i vdenorm6 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign6, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
86     const __m128i vdenorm7 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign7, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
87 
88     const __m128i vmask0 = _mm_cmpgt_epi32(vnonsign0, vdenorm_cutoff);
89     const __m128i vmask1 = _mm_cmpgt_epi32(vnonsign1, vdenorm_cutoff);
90     const __m128i vmask2 = _mm_cmpgt_epi32(vnonsign2, vdenorm_cutoff);
91     const __m128i vmask3 = _mm_cmpgt_epi32(vnonsign3, vdenorm_cutoff);
92     const __m128i vmask4 = _mm_cmpgt_epi32(vnonsign4, vdenorm_cutoff);
93     const __m128i vmask5 = _mm_cmpgt_epi32(vnonsign5, vdenorm_cutoff);
94     const __m128i vmask6 = _mm_cmpgt_epi32(vnonsign6, vdenorm_cutoff);
95     const __m128i vmask7 = _mm_cmpgt_epi32(vnonsign7, vdenorm_cutoff);
96 
97     const __m128i vf0 = _mm_or_si128(vsign0,
98       _mm_or_si128(_mm_and_si128(vmask0, vnorm0), _mm_andnot_si128(vmask0, vdenorm0)));
99     const __m128i vf1 = _mm_or_si128(vsign1,
100       _mm_or_si128(_mm_and_si128(vmask1, vnorm1), _mm_andnot_si128(vmask1, vdenorm1)));
101     const __m128i vf2 = _mm_or_si128(vsign2,
102       _mm_or_si128(_mm_and_si128(vmask2, vnorm2), _mm_andnot_si128(vmask2, vdenorm2)));
103     const __m128i vf3 = _mm_or_si128(vsign3,
104       _mm_or_si128(_mm_and_si128(vmask3, vnorm3), _mm_andnot_si128(vmask3, vdenorm3)));
105     const __m128i vf4 = _mm_or_si128(vsign4,
106       _mm_or_si128(_mm_and_si128(vmask4, vnorm4), _mm_andnot_si128(vmask4, vdenorm4)));
107     const __m128i vf5 = _mm_or_si128(vsign5,
108       _mm_or_si128(_mm_and_si128(vmask5, vnorm5), _mm_andnot_si128(vmask5, vdenorm5)));
109     const __m128i vf6 = _mm_or_si128(vsign6,
110       _mm_or_si128(_mm_and_si128(vmask6, vnorm6), _mm_andnot_si128(vmask6, vdenorm6)));
111     const __m128i vf7 = _mm_or_si128(vsign7,
112       _mm_or_si128(_mm_and_si128(vmask7, vnorm7), _mm_andnot_si128(vmask7, vdenorm7)));
113 
114     _mm_storeu_ps(output, _mm_castsi128_ps(vf0));
115     _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf1));
116     _mm_storeu_ps(output + 8, _mm_castsi128_ps(vf2));
117     _mm_storeu_ps(output + 12, _mm_castsi128_ps(vf3));
118     _mm_storeu_ps(output + 16, _mm_castsi128_ps(vf4));
119     _mm_storeu_ps(output + 20, _mm_castsi128_ps(vf5));
120     _mm_storeu_ps(output + 24, _mm_castsi128_ps(vf6));
121     _mm_storeu_ps(output + 28, _mm_castsi128_ps(vf7));
122     output += 32;
123   }
124   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
125     const __m128i vh = _mm_loadu_si128((const __m128i*) i);
126     i += 8;
127 
128     const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
129     const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);
130 
131     const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
132     const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);
133 
134     const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
135     const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
136 
137     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));
138     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));
139 
140     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)));
141     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)));
142 
143     const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
144     const __m128i vf_lo = _mm_or_si128(vsign_lo,
145       _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));
146 
147     const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
148     const __m128i vf_hi = _mm_or_si128(vsign_hi,
149       _mm_or_si128(_mm_and_si128(vmask_hi, vnorm_hi), _mm_andnot_si128(vmask_hi, vdenorm_hi)));
150 
151     _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo));
152     _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi));
153     output += 8;
154   }
155   if XNN_UNPREDICTABLE(n != 0) {
156     const __m128i vh = _mm_loadu_si128((const __m128i*) i);
157 
158     const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
159     const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);
160 
161     const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
162     const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);
163 
164     const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
165     const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
166 
167     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));
168     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));
169 
170     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)));
171     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)));
172 
173     const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
174     __m128i vf = _mm_or_si128(vsign_lo,
175       _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));
176 
177     if (n & (4 * sizeof(uint16_t))) {
178       _mm_storeu_ps(output, _mm_castsi128_ps(vf));
179       output += 4;
180 
181       const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
182       vf = _mm_or_si128(vsign_hi,
183         _mm_or_si128(_mm_and_si128(vmask_hi, vnorm_hi), _mm_andnot_si128(vmask_hi, vdenorm_hi)));
184     }
185     if (n & (2 * sizeof(uint16_t))) {
186       _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf));
187       output += 2;
188 
189       vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf)));
190     }
191     if (n & (1 * sizeof(uint16_t))) {
192       _mm_store_ss(output, _mm_castsi128_ps(vf));
193     }
194   }
195 }
196