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_x20(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_x20(
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 >= 20 * sizeof(float); n -= 20 * 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
46 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
47 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
48 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
49 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
50 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask);
51
52 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
53 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
54 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
55 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
56 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
57
58 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
59 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
60 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
61 const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
62 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23));
63
64 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
65 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
66 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
67 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
68 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
69
70 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
71 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
72 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
73 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
74 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ);
75
76 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
77 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
78 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
79 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
80 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ);
81
82 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4);
83 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4);
84 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4);
85 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc5, vtCDEF), vc4);
86 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4);
87
88 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
89 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
90 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
91 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
92 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3);
93
94 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
95 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
96 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
97 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
98 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2);
99
100 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1);
101 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1);
102 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1);
103 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc1);
104 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1);
105
106 vt0123 = _mm_mul_ps(vt0123, vs0123);
107 vt4567 = _mm_mul_ps(vt4567, vs4567);
108 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
109 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
110 vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ);
111
112 __m128 ve0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123);
113 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567);
114 __m128 ve89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB);
115 __m128 veCDEF = _mm_add_ps(_mm_mul_ps(vtCDEF, vpCDEF), vsCDEF);
116 __m128 veGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ);
117
118 __m128 vd0123 = _mm_add_ps(ve0123, vone);
119 __m128 vd4567 = _mm_add_ps(ve4567, vone);
120 __m128 vd89AB = _mm_add_ps(ve89AB, vone);
121 __m128 vdCDEF = _mm_add_ps(veCDEF, vone);
122 __m128 vdGHIJ = _mm_add_ps(veGHIJ, vone);
123
124 __m128 vf0123 = _mm_div_ps(ve0123, vd0123);
125 __m128 vf4567 = _mm_div_ps(ve4567, vd4567);
126 __m128 vf89AB = _mm_div_ps(ve89AB, vd89AB);
127 __m128 vfCDEF = _mm_div_ps(veCDEF, vdCDEF);
128 __m128 vfGHIJ = _mm_div_ps(veGHIJ, vdGHIJ);
129
130 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
131 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
132 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
133 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
134 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ);
135
136 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
137 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
138 vf89AB = _mm_blendv_ps(_mm_sub_ps(vone, vf89AB), vf89AB, vx89AB);
139 vfCDEF = _mm_blendv_ps(_mm_sub_ps(vone, vfCDEF), vfCDEF, vxCDEF);
140 vfGHIJ = _mm_blendv_ps(_mm_sub_ps(vone, vfGHIJ), vfGHIJ, vxGHIJ);
141
142 _mm_storeu_ps(y, vf0123);
143 _mm_storeu_ps(y + 4, vf4567);
144 _mm_storeu_ps(y + 8, vf89AB);
145 _mm_storeu_ps(y + 12, vfCDEF);
146 _mm_storeu_ps(y + 16, vfGHIJ);
147
148 x += 20;
149 y += 20;
150 }
151 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
152 const __m128 vx = _mm_loadu_ps(x);
153
154 const __m128 vz = _mm_or_ps(vx, vsign_mask);
155
156 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
157 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
158 vn = _mm_sub_ps(vn, vmagic_bias);
159
160 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
161 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
162
163 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
164 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
165 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
166 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
167
168 vt = _mm_mul_ps(vt, vs);
169 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
170
171 __m128 vd = _mm_add_ps(ve, vone);
172 __m128 vf = _mm_div_ps(ve, vd);
173
174 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
175 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
176
177 _mm_storeu_ps(y, vf);
178
179 x += 4;
180 y += 4;
181 }
182 if XNN_UNLIKELY(n != 0) {
183 const __m128 vx = _mm_loadu_ps(x);
184
185 const __m128 vz = _mm_or_ps(vx, vsign_mask);
186
187 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
188 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
189 vn = _mm_sub_ps(vn, vmagic_bias);
190
191 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
192 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
193
194 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
195 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
196 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
197 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
198
199 vt = _mm_mul_ps(vt, vs);
200 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
201
202 __m128 vd = _mm_add_ps(ve, vone);
203 __m128 vf = _mm_div_ps(ve, vd);
204
205 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
206 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
207
208 if (n & (2 * sizeof(float))) {
209 _mm_storel_pi((__m64*) y, vf);
210 vf = _mm_movehl_ps(vf, vf);
211 y += 2;
212 }
213 if (n & (1 * sizeof(float))) {
214 _mm_store_ss(y, vf);
215 }
216 }
217 }
218