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_x12(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__sse41_p5_div_x12(
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 >= 12 * sizeof(float); n -= 12 * 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
44 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
45 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
46 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
47
48 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
49 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
50 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
51
52 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
53 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
54 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
55
56 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
57 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
58 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
59
60 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
61 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
62 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
63
64 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
65 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
66 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
67
68 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4);
69 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4);
70 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4);
71
72 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
73 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
74 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
75
76 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
77 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
78 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
79
80 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1);
81 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1);
82 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1);
83
84 vt0123 = _mm_mul_ps(vt0123, vs0123);
85 vt4567 = _mm_mul_ps(vt4567, vs4567);
86 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
87
88 __m128 ve0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123);
89 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567);
90 __m128 ve89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB);
91
92 __m128 vd0123 = _mm_add_ps(ve0123, vone);
93 __m128 vd4567 = _mm_add_ps(ve4567, vone);
94 __m128 vd89AB = _mm_add_ps(ve89AB, vone);
95
96 __m128 vf0123 = _mm_div_ps(ve0123, vd0123);
97 __m128 vf4567 = _mm_div_ps(ve4567, vd4567);
98 __m128 vf89AB = _mm_div_ps(ve89AB, vd89AB);
99
100 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
101 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
102 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
103
104 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
105 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
106 vf89AB = _mm_blendv_ps(_mm_sub_ps(vone, vf89AB), vf89AB, vx89AB);
107
108 _mm_storeu_ps(y, vf0123);
109 _mm_storeu_ps(y + 4, vf4567);
110 _mm_storeu_ps(y + 8, vf89AB);
111
112 x += 12;
113 y += 12;
114 }
115 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
116 const __m128 vx = _mm_loadu_ps(x);
117
118 const __m128 vz = _mm_or_ps(vx, vsign_mask);
119
120 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
121 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
122 vn = _mm_sub_ps(vn, vmagic_bias);
123
124 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
125 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
126
127 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
128 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
129 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
130 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
131
132 vt = _mm_mul_ps(vt, vs);
133 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
134
135 __m128 vd = _mm_add_ps(ve, vone);
136 __m128 vf = _mm_div_ps(ve, vd);
137
138 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
139 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
140
141 _mm_storeu_ps(y, vf);
142
143 x += 4;
144 y += 4;
145 }
146 if XNN_UNLIKELY(n != 0) {
147 const __m128 vx = _mm_loadu_ps(x);
148
149 const __m128 vz = _mm_or_ps(vx, vsign_mask);
150
151 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
152 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
153 vn = _mm_sub_ps(vn, vmagic_bias);
154
155 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
156 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
157
158 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
159 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
160 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
161 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
162
163 vt = _mm_mul_ps(vt, vs);
164 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
165
166 __m128 vd = _mm_add_ps(ve, vone);
167 __m128 vf = _mm_div_ps(ve, vd);
168
169 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
170 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
171
172 if (n & (2 * sizeof(float))) {
173 _mm_storel_pi((__m64*) y, vf);
174 vf = _mm_movehl_ps(vf, vf);
175 y += 2;
176 }
177 if (n & (1 * sizeof(float))) {
178 _mm_store_ss(y, vf);
179 }
180 }
181 }
182