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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x16(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__sse2_rr2_p5_div_x16(
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 >= 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 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
121 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
122 const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
123 const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
124
125 vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123)));
126 vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567)));
127 vf89AB = _mm_or_ps(_mm_and_ps(vf89AB, vm89AB), _mm_andnot_ps(vm89AB, _mm_sub_ps(vone, vf89AB)));
128 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF)));
129
130 _mm_storeu_ps(y, vf0123);
131 _mm_storeu_ps(y + 4, vf4567);
132 _mm_storeu_ps(y + 8, vf89AB);
133 _mm_storeu_ps(y + 12, vfCDEF);
134
135 x += 16;
136 y += 16;
137 }
138 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
139 const __m128 vx = _mm_loadu_ps(x);
140
141 const __m128 vz = _mm_or_ps(vx, vsign_mask);
142
143 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
144 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
145 vn = _mm_sub_ps(vn, vmagic_bias);
146
147 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
148 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
149
150 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
151 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
152 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
153 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
154
155 vt = _mm_mul_ps(vt, vs);
156 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
157
158 __m128 vd = _mm_add_ps(ve, vone);
159 __m128 vf = _mm_div_ps(ve, vd);
160
161 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
162 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
163 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
164
165 _mm_storeu_ps(y, vf);
166
167 x += 4;
168 y += 4;
169 }
170 if XNN_UNLIKELY(n != 0) {
171 const __m128 vx = _mm_loadu_ps(x);
172
173 const __m128 vz = _mm_or_ps(vx, vsign_mask);
174
175 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
176 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
177 vn = _mm_sub_ps(vn, vmagic_bias);
178
179 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
180 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
181
182 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
183 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
184 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
185 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
186
187 vt = _mm_mul_ps(vt, vs);
188 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
189
190 __m128 vd = _mm_add_ps(ve, vone);
191 __m128 vf = _mm_div_ps(ve, vd);
192
193 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
194 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
195 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
196
197 if (n & (2 * sizeof(float))) {
198 _mm_storel_pi((__m64*) y, vf);
199 vf = _mm_movehl_ps(vf, vf);
200 y += 2;
201 }
202 if (n & (1 * sizeof(float))) {
203 _mm_store_ss(y, vf);
204 }
205 }
206 }
207