1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/sse-p5-div.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/vunary.h>
16
17
xnn_f32_sigmoid_ukernel__sse2_p5_div_x24(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__sse2_p5_div_x24(
19 size_t n,
20 const float* x,
21 float* y,
22 const void* params) XNN_DISABLE_TSAN
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m128 vsign_mask = _mm_set1_ps(-0.0f);
27 const __m128 vmagic_bias = _mm_set1_ps(0x1.8000FEp23f);
28 const __m128 vlog2e = _mm_set1_ps(0x1.715476p0f);
29 const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.62E400p-1f);
30 const __m128 vminus_ln2_lo = _mm_set1_ps(-0x1.7F7D1Cp-20f);
31 const __m128 vc5 = _mm_set1_ps(0x1.0F9F9Cp-7f);
32 const __m128 vc4 = _mm_set1_ps(0x1.573A1Ap-5f);
33 const __m128 vc3 = _mm_set1_ps(0x1.555A80p-3f);
34 const __m128 vc2 = _mm_set1_ps(0x1.FFFDC6p-2f);
35 const __m128 vc1 = _mm_set1_ps(0x1.FFFFF6p-1f);
36 const __m128 vone = _mm_set1_ps(1.0f);
37 const __m128 vdenorm_cutoff = _mm_set1_ps(-0x1.5D589Ep+6f);
38
39 for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
40 const __m128 vx0123 = _mm_loadu_ps(x);
41 const __m128 vx4567 = _mm_loadu_ps(x + 4);
42 const __m128 vx89AB = _mm_loadu_ps(x + 8);
43 const __m128 vxCDEF = _mm_loadu_ps(x + 12);
44 const __m128 vxGHIJ = _mm_loadu_ps(x + 16);
45 const __m128 vxKLMN = _mm_loadu_ps(x + 20);
46
47 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
48 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
49 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
50 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
51 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask);
52 const __m128 vzKLMN = _mm_or_ps(vxKLMN, vsign_mask);
53
54 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
55 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
56 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
57 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
58 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
59 __m128 vnKLMN = _mm_add_ps(_mm_mul_ps(vzKLMN, vlog2e), vmagic_bias);
60
61 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
62 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
63 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
64 const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
65 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23));
66 const __m128 vsKLMN = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnKLMN), 23));
67
68 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
69 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
70 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
71 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
72 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
73 vnKLMN = _mm_sub_ps(vnKLMN, vmagic_bias);
74
75 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
76 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
77 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
78 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
79 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ);
80 __m128 vtKLMN = _mm_add_ps(_mm_mul_ps(vnKLMN, vminus_ln2_hi), vzKLMN);
81
82 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
83 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
84 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
85 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
86 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ);
87 vtKLMN = _mm_add_ps(_mm_mul_ps(vnKLMN, vminus_ln2_lo), vtKLMN);
88
89 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4);
90 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4);
91 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4);
92 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc5, vtCDEF), vc4);
93 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4);
94 __m128 vpKLMN = _mm_add_ps(_mm_mul_ps(vc5, vtKLMN), vc4);
95
96 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
97 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
98 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
99 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
100 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3);
101 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc3);
102
103 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
104 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
105 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
106 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
107 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2);
108 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc2);
109
110 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1);
111 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1);
112 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1);
113 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc1);
114 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1);
115 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc1);
116
117 vt0123 = _mm_mul_ps(vt0123, vs0123);
118 vt4567 = _mm_mul_ps(vt4567, vs4567);
119 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
120 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
121 vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ);
122 vtKLMN = _mm_mul_ps(vtKLMN, vsKLMN);
123
124 __m128 ve0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123);
125 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567);
126 __m128 ve89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB);
127 __m128 veCDEF = _mm_add_ps(_mm_mul_ps(vtCDEF, vpCDEF), vsCDEF);
128 __m128 veGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ);
129 __m128 veKLMN = _mm_add_ps(_mm_mul_ps(vtKLMN, vpKLMN), vsKLMN);
130
131 __m128 vd0123 = _mm_add_ps(ve0123, vone);
132 __m128 vd4567 = _mm_add_ps(ve4567, vone);
133 __m128 vd89AB = _mm_add_ps(ve89AB, vone);
134 __m128 vdCDEF = _mm_add_ps(veCDEF, vone);
135 __m128 vdGHIJ = _mm_add_ps(veGHIJ, vone);
136 __m128 vdKLMN = _mm_add_ps(veKLMN, vone);
137
138 __m128 vf0123 = _mm_div_ps(ve0123, vd0123);
139 __m128 vf4567 = _mm_div_ps(ve4567, vd4567);
140 __m128 vf89AB = _mm_div_ps(ve89AB, vd89AB);
141 __m128 vfCDEF = _mm_div_ps(veCDEF, vdCDEF);
142 __m128 vfGHIJ = _mm_div_ps(veGHIJ, vdGHIJ);
143 __m128 vfKLMN = _mm_div_ps(veKLMN, vdKLMN);
144
145 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
146 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
147 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
148 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
149 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ);
150 vfKLMN = _mm_andnot_ps(_mm_cmplt_ps(vzKLMN, vdenorm_cutoff), vfKLMN);
151
152 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
153 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
154 const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
155 const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
156 const __m128 vmGHIJ = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxGHIJ)));
157 const __m128 vmKLMN = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxKLMN)));
158
159 vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123)));
160 vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567)));
161 vf89AB = _mm_or_ps(_mm_and_ps(vf89AB, vm89AB), _mm_andnot_ps(vm89AB, _mm_sub_ps(vone, vf89AB)));
162 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF)));
163 vfGHIJ = _mm_or_ps(_mm_and_ps(vfGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, _mm_sub_ps(vone, vfGHIJ)));
164 vfKLMN = _mm_or_ps(_mm_and_ps(vfKLMN, vmKLMN), _mm_andnot_ps(vmKLMN, _mm_sub_ps(vone, vfKLMN)));
165
166 _mm_storeu_ps(y, vf0123);
167 _mm_storeu_ps(y + 4, vf4567);
168 _mm_storeu_ps(y + 8, vf89AB);
169 _mm_storeu_ps(y + 12, vfCDEF);
170 _mm_storeu_ps(y + 16, vfGHIJ);
171 _mm_storeu_ps(y + 20, vfKLMN);
172
173 x += 24;
174 y += 24;
175 }
176 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
177 const __m128 vx = _mm_loadu_ps(x);
178
179 const __m128 vz = _mm_or_ps(vx, vsign_mask);
180
181 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
182 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
183 vn = _mm_sub_ps(vn, vmagic_bias);
184
185 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
186 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
187
188 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
189 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
190 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
191 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
192
193 vt = _mm_mul_ps(vt, vs);
194 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
195
196 __m128 vd = _mm_add_ps(ve, vone);
197 __m128 vf = _mm_div_ps(ve, vd);
198
199 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
200 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
201 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
202
203 _mm_storeu_ps(y, vf);
204
205 x += 4;
206 y += 4;
207 }
208 if XNN_UNLIKELY(n != 0) {
209 const __m128 vx = _mm_loadu_ps(x);
210
211 const __m128 vz = _mm_or_ps(vx, vsign_mask);
212
213 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
214 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
215 vn = _mm_sub_ps(vn, vmagic_bias);
216
217 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
218 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
219
220 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
221 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
222 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
223 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
224
225 vt = _mm_mul_ps(vt, vs);
226 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
227
228 __m128 vd = _mm_add_ps(ve, vone);
229 __m128 vf = _mm_div_ps(ve, vd);
230
231 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
232 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
233 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
234
235 if (n & (2 * sizeof(float))) {
236 _mm_storel_pi((__m64*) y, vf);
237 vf = _mm_movehl_ps(vf, vf);
238 y += 2;
239 }
240 if (n & (1 * sizeof(float))) {
241 _mm_store_ss(y, vf);
242 }
243 }
244 }
245