1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/sse-rr2-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_vsigmoid_ukernel__sse41_rr2_p5_div_x8(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m128 vsign_mask = _mm_load_ps(params->sse2_rr2_p5.sign_mask);
27 const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_p5.magic_bias);
28 const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_p5.log2e);
29 const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_p5.minus_ln2_hi);
30 const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_p5.minus_ln2_lo);
31 const __m128 vc5 = _mm_load_ps(params->sse2_rr2_p5.c5);
32 const __m128 vc4 = _mm_load_ps(params->sse2_rr2_p5.c4);
33 const __m128 vc3 = _mm_load_ps(params->sse2_rr2_p5.c3);
34 const __m128 vc2 = _mm_load_ps(params->sse2_rr2_p5.c2);
35 const __m128 vc1 = _mm_load_ps(params->sse2_rr2_p5.c1);
36 const __m128 vone = _mm_load_ps(params->sse2_rr2_p5.one);
37 const __m128 vdenorm_cutoff = _mm_load_ps(params->sse2_rr2_p5.denorm_cutoff);
38
39 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
40 const __m128 vx0123 = _mm_loadu_ps(x);
41 const __m128 vx4567 = _mm_loadu_ps(x + 4);
42
43 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
44 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
45
46 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
47 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
48
49 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
50 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
51
52 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
53 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
54
55 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
56 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
57
58 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
59 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
60
61 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4);
62 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4);
63
64 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
65 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
66
67 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
68 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
69
70 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1);
71 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1);
72
73 vt0123 = _mm_mul_ps(vt0123, vs0123);
74 vt4567 = _mm_mul_ps(vt4567, vs4567);
75
76 __m128 ve0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123);
77 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567);
78
79 __m128 vd0123 = _mm_add_ps(ve0123, vone);
80 __m128 vd4567 = _mm_add_ps(ve4567, vone);
81
82 __m128 vf0123 = _mm_div_ps(ve0123, vd0123);
83 __m128 vf4567 = _mm_div_ps(ve4567, vd4567);
84
85 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
86 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
87
88 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
89 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
90
91 _mm_storeu_ps(y, vf0123);
92 _mm_storeu_ps(y + 4, vf4567);
93
94 x += 8;
95 y += 8;
96 }
97 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
98 const __m128 vx = _mm_loadu_ps(x);
99
100 const __m128 vz = _mm_or_ps(vx, vsign_mask);
101
102 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
103 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
104 vn = _mm_sub_ps(vn, vmagic_bias);
105
106 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
107 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
108
109 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
110 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
111 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
112 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
113
114 vt = _mm_mul_ps(vt, vs);
115 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
116
117 __m128 vd = _mm_add_ps(ve, vone);
118 __m128 vf = _mm_div_ps(ve, vd);
119
120 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
121 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
122
123 _mm_storeu_ps(y, vf);
124
125 x += 4;
126 y += 4;
127 }
128 if XNN_UNLIKELY(n != 0) {
129 const __m128 vx = _mm_loadu_ps(x);
130
131 const __m128 vz = _mm_or_ps(vx, vsign_mask);
132
133 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
134 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
135 vn = _mm_sub_ps(vn, vmagic_bias);
136
137 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
138 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
139
140 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
141 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
142 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
143 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
144
145 vt = _mm_mul_ps(vt, vs);
146 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
147
148 __m128 vd = _mm_add_ps(ve, vone);
149 __m128 vf = _mm_div_ps(ve, vd);
150
151 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
152 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
153
154 if (n & (2 * sizeof(float))) {
155 _mm_storel_pi((__m64*) y, vf);
156 vf = _mm_movehl_ps(vf, vf);
157 y += 2;
158 }
159 if (n & (1 * sizeof(float))) {
160 _mm_store_ss(y, vf);
161 }
162 }
163 }
164