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 <smmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_sigmoid_ukernel__sse41_p5_div_x16(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__sse41_p5_div_x16(
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 >= 16 * sizeof(float); n -= 16 * 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
45 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
46 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
47 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
48 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
49
50 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
51 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
52 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
53 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
54
55 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
56 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
57 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
58 const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
59
60 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
61 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
62 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
63 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
64
65 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
66 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
67 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
68 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
69
70 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
71 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
72 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
73 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
74
75 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4);
76 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4);
77 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4);
78 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc5, vtCDEF), vc4);
79
80 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
81 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
82 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
83 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
84
85 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
86 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
87 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
88 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
89
90 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1);
91 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1);
92 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1);
93 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc1);
94
95 vt0123 = _mm_mul_ps(vt0123, vs0123);
96 vt4567 = _mm_mul_ps(vt4567, vs4567);
97 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
98 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
99
100 __m128 ve0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123);
101 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567);
102 __m128 ve89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB);
103 __m128 veCDEF = _mm_add_ps(_mm_mul_ps(vtCDEF, vpCDEF), vsCDEF);
104
105 __m128 vd0123 = _mm_add_ps(ve0123, vone);
106 __m128 vd4567 = _mm_add_ps(ve4567, vone);
107 __m128 vd89AB = _mm_add_ps(ve89AB, vone);
108 __m128 vdCDEF = _mm_add_ps(veCDEF, vone);
109
110 __m128 vf0123 = _mm_div_ps(ve0123, vd0123);
111 __m128 vf4567 = _mm_div_ps(ve4567, vd4567);
112 __m128 vf89AB = _mm_div_ps(ve89AB, vd89AB);
113 __m128 vfCDEF = _mm_div_ps(veCDEF, vdCDEF);
114
115 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
116 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
117 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
118 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
119
120 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
121 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
122 vf89AB = _mm_blendv_ps(_mm_sub_ps(vone, vf89AB), vf89AB, vx89AB);
123 vfCDEF = _mm_blendv_ps(_mm_sub_ps(vone, vfCDEF), vfCDEF, vxCDEF);
124
125 _mm_storeu_ps(y, vf0123);
126 _mm_storeu_ps(y + 4, vf4567);
127 _mm_storeu_ps(y + 8, vf89AB);
128 _mm_storeu_ps(y + 12, vfCDEF);
129
130 x += 16;
131 y += 16;
132 }
133 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
134 const __m128 vx = _mm_loadu_ps(x);
135
136 const __m128 vz = _mm_or_ps(vx, vsign_mask);
137
138 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
139 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
140 vn = _mm_sub_ps(vn, vmagic_bias);
141
142 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
143 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
144
145 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
146 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
147 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
148 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
149
150 vt = _mm_mul_ps(vt, vs);
151 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
152
153 __m128 vd = _mm_add_ps(ve, vone);
154 __m128 vf = _mm_div_ps(ve, vd);
155
156 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
157 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
158
159 _mm_storeu_ps(y, vf);
160
161 x += 4;
162 y += 4;
163 }
164 if XNN_UNLIKELY(n != 0) {
165 const __m128 vx = _mm_loadu_ps(x);
166
167 const __m128 vz = _mm_or_ps(vx, vsign_mask);
168
169 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
170 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
171 vn = _mm_sub_ps(vn, vmagic_bias);
172
173 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
174 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
175
176 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
177 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
178 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
179 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
180
181 vt = _mm_mul_ps(vt, vs);
182 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
183
184 __m128 vd = _mm_add_ps(ve, vone);
185 __m128 vf = _mm_div_ps(ve, vd);
186
187 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
188 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
189
190 if (n & (2 * sizeof(float))) {
191 _mm_storel_pi((__m64*) y, vf);
192 vf = _mm_movehl_ps(vf, vf);
193 y += 2;
194 }
195 if (n & (1 * sizeof(float))) {
196 _mm_store_ss(y, vf);
197 }
198 }
199 }
200