1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx-rr2-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x32(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__avx_rr2_p5_nr2_x32(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vsign_mask = _mm256_load_ps(params->avx_rr2_p5.sign_mask);
27 const __m256 vmagic_bias = _mm256_load_ps(params->avx_rr2_p5.magic_bias);
28 const __m256 vlog2e = _mm256_load_ps(params->avx_rr2_p5.log2e);
29 const __m256 vminus_ln2_hi = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_hi);
30 const __m256 vminus_ln2_lo = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_lo);
31 const __m256 vc5 = _mm256_load_ps(params->avx_rr2_p5.c5);
32 const __m256 vc4 = _mm256_load_ps(params->avx_rr2_p5.c4);
33 const __m256 vc3 = _mm256_load_ps(params->avx_rr2_p5.c3);
34 const __m256 vc2 = _mm256_load_ps(params->avx_rr2_p5.c2);
35 const __m256 vc1 = _mm256_load_ps(params->avx_rr2_p5.c1);
36 const __m256 vone = _mm256_load_ps(params->avx_rr2_p5.one);
37 const __m256 vtwo = _mm256_load_ps(params->avx_rr2_p5.two);
38 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx_rr2_p5.denorm_cutoff);
39
40 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
41 const __m256 vx0 = _mm256_loadu_ps(x);
42 const __m256 vx1 = _mm256_loadu_ps(x + 8);
43 const __m256 vx2 = _mm256_loadu_ps(x + 16);
44 const __m256 vx3 = _mm256_loadu_ps(x + 24);
45 x += 32;
46
47 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
48 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
49 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
50 const __m256 vz3 = _mm256_or_ps(vx3, vsign_mask);
51
52 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
53 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
54 __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
55 __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
56
57 const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
58 const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
59 const __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo0), vs_hi0, 1);
60 const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
61 const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
62 const __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo1), vs_hi1, 1);
63 const __m128 vs_lo2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23));
64 const __m128 vs_hi2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 23));
65 const __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo2), vs_hi2, 1);
66 const __m128 vs_lo3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)), 23));
67 const __m128 vs_hi3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, 1)), 23));
68 const __m256 vs3 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo3), vs_hi3, 1);
69
70 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
71 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
72 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
73 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
74
75 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
76 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
77 __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
78 __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_hi), vz3);
79
80 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
81 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
82 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
83 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_lo), vt3);
84
85 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc5, vt0), vc4);
86 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc5, vt1), vc4);
87 __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc5, vt2), vc4);
88 __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc5, vt3), vc4);
89
90 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
91 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
92 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc3);
93 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc3);
94
95 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
96 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
97 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
98 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc2);
99
100 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc1);
101 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc1);
102 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc1);
103 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc1);
104
105 vt0 = _mm256_mul_ps(vt0, vs0);
106 vt1 = _mm256_mul_ps(vt1, vs1);
107 vt2 = _mm256_mul_ps(vt2, vs2);
108 vt3 = _mm256_mul_ps(vt3, vs3);
109
110 const __m256 ve0 = _mm256_add_ps(_mm256_mul_ps(vt0, vp0), vs0);
111 const __m256 ve1 = _mm256_add_ps(_mm256_mul_ps(vt1, vp1), vs1);
112 const __m256 ve2 = _mm256_add_ps(_mm256_mul_ps(vt2, vp2), vs2);
113 const __m256 ve3 = _mm256_add_ps(_mm256_mul_ps(vt3, vp3), vs3);
114
115 const __m256 vd0 = _mm256_add_ps(ve0, vone);
116 const __m256 vd1 = _mm256_add_ps(ve1, vone);
117 const __m256 vd2 = _mm256_add_ps(ve2, vone);
118 const __m256 vd3 = _mm256_add_ps(ve3, vone);
119
120 __m256 vr0 = _mm256_rcp_ps(vd0);
121 __m256 vr1 = _mm256_rcp_ps(vd1);
122 __m256 vr2 = _mm256_rcp_ps(vd2);
123 __m256 vr3 = _mm256_rcp_ps(vd3);
124
125 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
126 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
127 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
128 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
129 vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));
130 vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));
131 vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));
132 vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));
133
134 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
135 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
136 __m256 vf2 = _mm256_mul_ps(ve2, vr2);
137 __m256 vf3 = _mm256_mul_ps(ve3, vr3);
138
139 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
140 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
141 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
142 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vz3, vdenorm_cutoff, _CMP_LT_OS), vf3);
143
144 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
145 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
146 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
147 vf3 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf3), vf3, vx3);
148
149 _mm256_storeu_ps(y, vf0);
150 _mm256_storeu_ps(y + 8, vf1);
151 _mm256_storeu_ps(y + 16, vf2);
152 _mm256_storeu_ps(y + 24, vf3);
153 y += 32;
154 }
155 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
156 const __m256 vx = _mm256_loadu_ps(x);
157 x += 8;
158
159 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
160
161 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
162
163 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
164 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
165 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
166
167 vn = _mm256_sub_ps(vn, vmagic_bias);
168
169 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
170 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
171
172 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
173 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
174 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
175 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
176
177 vt = _mm256_mul_ps(vt, vs);
178 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
179
180 const __m256 vd = _mm256_add_ps(ve, vone);
181 __m256 vr = _mm256_rcp_ps(vd);
182 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
183 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
184 __m256 vf = _mm256_mul_ps(ve, vr);
185
186 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
187 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
188
189 _mm256_storeu_ps(y, vf);
190 y += 8;
191 }
192 if XNN_UNLIKELY(n != 0) {
193 assert(n >= 1 * sizeof(float));
194 assert(n <= 7 * sizeof(float));
195 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p5.mask_table[7] - n));
196
197 const __m256 vx = _mm256_maskload_ps(x, vmask);
198
199 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
200
201 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
202 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
203 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
204 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
205
206 vn = _mm256_sub_ps(vn, vmagic_bias);
207
208 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
209 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
210
211 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
212 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
213 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
214 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
215
216 vt = _mm256_mul_ps(vt, vs);
217 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
218
219 const __m256 vd = _mm256_add_ps(ve, vone);
220 __m256 vr = _mm256_rcp_ps(vd);
221 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
222 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
223 __m256 vf = _mm256_mul_ps(ve, vr);
224
225 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
226 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
227
228 __m128 vf_lo = _mm256_castps256_ps128(vf);
229 if (n & (4 * sizeof(float))) {
230 _mm_storeu_ps(y, vf_lo);
231 vf_lo = _mm256_extractf128_ps(vf, 1);
232 y += 4;
233 }
234 if (n & (2 * sizeof(float))) {
235 _mm_storel_pi((__m64*) y, vf_lo);
236 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
237 y += 2;
238 }
239 if (n & (1 * sizeof(float))) {
240 _mm_store_ss(y, vf_lo);
241 }
242 }
243 }
244